diff --git a/.github/workflows/html2video-for-mcode-smoke.yml b/.github/workflows/html2video-for-mcode-smoke.yml
new file mode 100644
index 00000000..24dbc216
--- /dev/null
+++ b/.github/workflows/html2video-for-mcode-smoke.yml
@@ -0,0 +1,97 @@
+# html2video-for-mcode · scoped smoke CI
+# 官方主 CI(ci.yml)只跑 validate + node --test(无 ffmpeg/playwright, 渲染冒烟会 skip)。
+# 本 workflow 按 CONTRIBUTING/ci.yml 注释给出的"单插件 scoped workflow"模式,
+# 装齐依赖后真实执行全部测试 —— 这是 PR #41 评审要求的可执行测试证据。
+# 触发条件: 仅本插件目录(.github/workflows/ 本文件除外)或本文件自身变更。
+#
+# 第三轮 review 修正(2026-09-18, 1.7.1 才真正推上来 —— 1.7.0 时这份修正只改了本地没推送,
+# PR 分支上一直是旧枚举版, 漏跑 cover-transition / review-round2 / review-round3 / subtitles-invalidate):
+#   ① 测试文件不再逐个枚举, 改用 shell glob `tests/*.test.mjs` —— 枚举写法每加一个测试文件就要
+#      手工记得改这里, 历史上已经漏过 css-kit / tokens-fx / chart-kit / table-kit 与上述四个文件。
+#   ② 补 permissions: contents: read(与本仓库其余 scoped workflow 一致, 只读仓库)。
+#   ③ actions 按仓库主 ci.yml 的同一批版本做**完整 SHA pin**(附版本注释), 不用浮动 tag;
+#      SHA 由 gh api repos/<owner>/<repo>/commits/<tag> 解析得到, 40 位整(主 ci.yml 的 checkout
+#      那行多了一位字符, 那是宿主仓库自己的笔误, 不去动它 —— 见 PR comment)。
+#
+# 第四轮 CHANGES_REQUESTED 修正(2026-09-21): 评审要求 exact-head 证据覆盖 **Windows** ——
+#   新增 windows-latest 作业(同一条命令, 依赖换成 choco ffmpeg + chromium)。两个作业的命令与
+#   依赖组合已先在技能镜像仓的 CI 上双平台跑绿(248 pass / 0 fail / 0 skip, 两个平台各一次),
+#   Node 版本也从 22 提到 24 与那份证过的配置对齐(此 workflow 此前因审批门从未真正执行过)。
+#   ffmpeg 一律走系统包(PATH 发现): 测试会把脚本放进临时项目目录当 cwd 跑, node_modules 发现
+#   依赖 cwd/上两层锚点, 在插件这种深层布局里够不到仓库根的 node_modules。
+name: html2video-for-mcode smoke
+
+on:
+  pull_request:
+    paths:
+      - 'plugins/Wzdhehe/html2video-for-mcode/**'
+      - '.github/workflows/html2video-for-mcode-smoke.yml'
+  push:
+    branches: [main]
+    paths:
+      - 'plugins/Wzdhehe/html2video-for-mcode/**'
+      - '.github/workflows/html2video-for-mcode-smoke.yml'
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  smoke-linux:
+    name: smoke (ubuntu-latest · ffmpeg + chromium)
+    runs-on: ubuntu-latest
+    timeout-minutes: 20
+    steps:
+      - name: Checkout
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+
+      - name: Set up Node
+        uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
+        with:
+          node-version: 24
+
+      - name: Install ffmpeg
+        run: sudo apt-get update && sudo apt-get install -y --no-install-recommends ffmpeg
+
+      - name: Install playwright (no package.json / lockfile changes)
+        run: |
+          npm install --no-save --no-package-lock playwright
+          npx playwright install --with-deps chromium
+
+      # 一条命令跑完 tests/ 下的**全部** *.test.mjs(安全 / 策略 / 模板 / 受管块 / 渲染冒烟)。
+      # glob 交给 shell 展开: 以后新增测试文件自动进 CI, 不需要改这里(旧枚举写法漏过 8 个文件)。
+      # 单测超时 300s: 渲染冒烟里的 capture / build-video 在 CI 上要几十秒。
+      - name: All tests (security · policy · templates · css blocks · render smoke)
+        run: |
+          node --test --test-timeout=300000 \
+            plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/*.test.mjs
+
+  smoke-windows:
+    name: smoke (windows-latest · ffmpeg + chromium)
+    runs-on: windows-latest
+    timeout-minutes: 30
+    steps:
+      - name: Checkout
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+
+      - name: Set up Node
+        uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
+        with:
+          node-version: 24
+
+      # choco 的 ffmpeg 包含 ffmpeg.exe 与 ffprobe.exe, 落在 PATH 上
+      - name: Install ffmpeg
+        shell: bash
+        run: choco install ffmpeg -y --no-progress
+
+      - name: Install playwright (no package.json / lockfile changes)
+        shell: bash
+        run: |
+          npm install --no-save --no-package-lock playwright
+          npx playwright install chromium
+
+      - name: All tests (security · policy · templates · css blocks · render smoke)
+        shell: bash
+        run: |
+          node --test --test-timeout=300000 \
+            plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/*.test.mjs
diff --git a/plugins/Wzdhehe/html2video-for-mcode/.claude-plugin/plugin.json b/plugins/Wzdhehe/html2video-for-mcode/.claude-plugin/plugin.json
new file mode 100644
index 00000000..1e3e089f
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/.claude-plugin/plugin.json
@@ -0,0 +1,8 @@
+{
+  "name": "html2video-for-mcode",
+  "version": "1.9.19",
+  "description": "Turn a topic, outline, or script into a narrated MP4: HTML slides with staged entrance animations, TTS voiceover, ffmpeg assembly, and ASR verification.",
+  "skills": [
+    "./skills/html2video-for-mcode/SKILL.md"
+  ]
+}
diff --git a/plugins/Wzdhehe/html2video-for-mcode/.gitattributes b/plugins/Wzdhehe/html2video-for-mcode/.gitattributes
new file mode 100644
index 00000000..c304db3a
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/.gitattributes
@@ -0,0 +1,10 @@
+# Force LF for this plugin's text files. Added in 1.9.3: nine files carried CRLF
+# (authored on Windows), which made `git diff --check` report trailing-whitespace
+# errors on the PR under default whitespace rules. Same convention as
+# plugins/antianqi/mcode-island/.gitattributes.
+# Override at clone time: `git config core.autocrlf input`.
+*           text=auto eol=lf
+*.md        text eol=lf
+*.mjs       text eol=lf
+*.json      text eol=lf
+LICENSE     text eol=lf
diff --git a/plugins/Wzdhehe/html2video-for-mcode/CHANGELOG.md b/plugins/Wzdhehe/html2video-for-mcode/CHANGELOG.md
new file mode 100644
index 00000000..b00cb86d
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/CHANGELOG.md
@@ -0,0 +1,583 @@
+# Changelog
+
+## 1.9.19 — 2026-10-10
+
+**One regression, found by reviewing our own 1.9.18 fix — and a test that structurally could not have caught it.**
+
+- **`--open` stopped opening the browser on Windows.** The 1.9.18 fix for M1 added `windowsVerbatimArguments` so the target would be quoted explicitly instead of relying on Node's quote-only-when-it-contains-spaces rule. But verbatim mode also stops Node from quoting the **empty-string argument**, so `start`'s empty *title* vanished: `start "" "<path>"` became `start "<path>"`. START treats the first quoted token as the **window title**, so with no command left to run nothing opens — while the script still prints "已在浏览器打开(若没弹出…)". A security fix that broke the feature it was securing.
+  - The rule was measured, not assumed. With `mkdir` as a marker: `cmd /c start "T" cmd /c <cmd>` runs `<cmd>` (T became the title), while `cmd /c start T cmd /c <cmd>` does not (T became a command that does not exist). One quoted token and no command after it therefore opens nothing.
+  - Fix: the empty title is now a **literal** `'""'`, and the argv construction moved into an exported `openCmdArgv(target, platform)` so it can be asserted directly rather than inferred from a spawn.
+- **Why our own M1 test missed it, and what changed.** That test substituted `echo` for `start`. `echo` has no title semantics, so "the quoted path is consumed as a title" was outside anything the test could observe — the proxy proved quoting, not START's parser. The replacement asserts on `openCmdArgv`'s argv, **red-proves** that the pre-fix shape fails the same assertion, and then runs **`start` itself**: the quoted form must create the marker, the unquoted form must not.
+- **Housekeeping:** the new behavioural test runs `start /b`. Without `/b` every suite run flashes a console window on the machine running it (measured: 3639 ms and one new conhost with `/b` absent; 134 ms and none with it).
+
+**Tests +3 → 317 in sixteen files.** Suite re-run before pushing: **317 tests, 313 pass, 0 fail, 4 named capability skips** — all four the file-symlink canaries, skipped with the reason "当前环境建不了文件符号链接(Windows 需开发者模式; ubuntu CI 真跑)" because this Windows machine has Developer Mode off. The directory-junction equivalents of those guards (`safeRel` escape, `L1` image gate) run and pass here via the `mklink /J` fallback, and the render-smoke suite ran for real on this machine rather than skipping.
+
+## 1.9.18 — 2026-10-10
+
+**Independent security audit of the "parse and generate" dimension — the one the hosted review rounds never covered.** The auditor was barred from reading the earlier review records so as not to anchor; it reported 2 HIGH / 4 MEDIUM / 6 LOW. Every finding was reproduced on our own tree before acting, three of its claims did not survive that, and what it confirmed is fixed here.
+
+**HIGH — fixed**
+
+- **H1 · attribute injection in the play page (`preview-page.mjs`).** `setAttr` re-serialised a value it had just read out of a **single-quoted** attribute into double quotes **without escaping**, so an author's `<html class='x" onmouseover=window.FIRED=1 data-x='>` came back from `addNoFx` with a live event handler — and `X` (the animation toggle the docs push) is the trigger. Reproduced by calling the real function. Values are now HTML-escaped on the way out (`&` `"` `<` `>`); the same applies to `injectHtmlVars`'s `style` path. The regression test **parses attribute names** instead of searching for `on*=` text — after the fix the text legitimately lives inside the attribute value, so a text search would report a false red — and it carries a detector self-check (the unescaped shape must be recognised) so the pair cannot be green for free.
+- **H2 · SRT cue injection (`build-video.mjs`).** `clause` text was spliced into `out/subs.srt` verbatim, so a clause containing `\n\n99\n00:59:00,000 --> 01:00:00,000\n…` produced **more timecodes than cues** — forged cues in a file that gets uploaded to a platform. SRT generation moved to a single source, `tools.buildSrt`, which sanitises each line: bidi and C0/C1 control characters stripped, all whitespace runs folded to one space (so a cue is always one line), a residual inline `-->` turned into `→` because lenient parsers — ffmpeg's srt demuxer scans lines for timecodes — would otherwise still mis-read it. The invariant asserted is *line-anchored*: line-anchored timecode lines === cue count.
+
+**MEDIUM — fixed**
+
+- **M1 · `--open` could run a second command.** `spawn('cmd', ['/c','start','',target])` relied on Node quoting the target, and Node only quotes when the argument contains spaces or quotes — a Windows path may contain `&` without a space. Reproduced with a controlled origin: the second command really executed. The target is now quoted explicitly with `windowsVerbatimArguments`. The test runs both shapes and requires the unquoted one to still be exploitable, so it cannot pass for the wrong reason.
+- **M2 · `fx-*` classes were trusted by prefix.** `class="fx-notreal"` satisfied the data-stage rule while nothing animates, and a locally defined `.fx-mine` with a keyframe lacking `opacity` was never inspected (only `tokens.css` classes were). Both now check the slide's own `<style>` as well; a class with no animation declaration is named.
+- **M3 · HTML comments could stand in for a definition.** `<!-- --c-fake: 1 -->` satisfied the undefined-variable gate. Comments (closed and unterminated) are stripped before both the definition and the usage scan.
+- **M4 · external-resource gate had holes.** Only double-quoted `src|href="https://…"` was seen, so `<IMG SRC=https://…>`, protocol-relative `//host/x.png`, `@import url(…)` and `background:url(…)` all passed. The scan is now case-insensitive, covers unquoted values and `srcset`, and covers CSS regions — while deliberately **not** flagging links written as slide prose.
+
+**LOW — addressed**
+
+- **L1** the image gate claimed to prevent second-order escapes but compared lexically; it now also compares canonical paths when the file exists (junction under `slides/` no longer reads outside). **L2** Windows `cmd` expands `%VAR%` even inside quotes and cannot be escaped on a command line — the ASR path now refuses such paths with an actionable message instead of mangling them (the old comment understated this). **L3** `data-style`/`data-class` no longer shadow the real attributes, a `>` inside a quoted value no longer truncates the tag, and a decoy `<html>` inside a comment is no longer patched. **L4** control and bidi characters are stripped from transcript text before comparison, printing and `checklist.md`. **L5** `isBlockedHost` now catches the expanded IPv4-mapped form (`0:0:0:0:0:ffff:7f00:1`) that only `net.isIP()` recognised. **L6** ffmpeg's concat list cannot represent a quote inside a quoted path, so such paths now fail loudly via `tools.assertConcatPathSafe` instead of writing a silently broken list.
+
+**Reported, reproduced, refuted** — recorded so they are not re-litigated: the auditor's claim that `rmSync(recursive)` follows junctions into the project parent (it does not — the guards are load-bearing, as we had already established), its suggestion that the missing exact-head runs could be self-served (they cannot; base-repo admin approval returns 403), and the suspicion that our escape guards might not run at all (the junction matrices came back 10/10 blocked).
+
+**Tests +19 → 314 in sixteen files**, including a Windows-only behavioural test for M1 that requires the control case to still fail. Seven red-proofs: reverting each of the seven guards turns exactly its own case red, and every restore is byte-compared. One red-proof initially did **not** redden, which exposed dead code rather than a weak test — the sanitizer had a dedicated CRLF fold that the trailing whitespace collapse already subsumed; it was removed and the proof now pins the load-bearing step.
+
+## 1.9.17 — 2026-10-10
+
+**Self-review round on the round-7 head (two-axis `/code-review`): the security findings were re-derived and one residual hardening landed; two reported findings did not survive reproduction and are recorded as such**
+
+- **Both refusal layers are now separately load-bearing.** `policyGet` applies the host policy before connecting (`where:'policyGet'`), and the proxy's absolute-form HTTP branch pre-checks with `where:'代理 HTTP'`. Removing either one alone left the suite green, so neither was actually pinned; the new tests assert each layer's own label, which turns both into load-bearing checks (red-proofs A and B: each removal now turns exactly its own assertion red). The deliberate redundancy stays, and the reason is now stated where the code is: `policyGet` is the single entry point for every caller including the download paths, while the proxy pre-check keeps a refusal attributable without depending on `policyGet` internals.
+- **The CONNECT branch's refusals are attributable too.** It used to `destroy()` the client socket silently — the HTTP branch said "blocked-host", the CONNECT branch said nothing, and the comment in the HTTP branch claiming they were "the same shape" was simply wrong. CONNECT now answers `502 Bad Gateway` with the reason in the body (never a tunnel, never an upstream connect). Two existing tests asserted the old silent behaviour and were updated by contract; red-proof C (reverting to a bare `destroy()`) turns both red again.
+- **The browser-side invariant is pinned, because it was a library default doing the work.** Chromium has an implicit proxy bypass for loopback, so "every browser request goes through the veto proxy" was resting on Playwright's defaults. Measured with the real browser: loopback subresources do go through the proxy and come back `502`, with the origin receiving 0 requests — but that is now a test (`capture`-shaped launch + a slide whose image points at a counting origin) instead of an assumption. If a future Playwright/Chromium changes the default, CI reports it rather than the pipeline silently losing its choke point.
+- **Reported, reproduced, and refuted (kept here so the next reviewer does not re-litigate them):**
+  - *"Chromium bypasses the proxy for loopback, so `capture.mjs` has an egress hole"* — not reproducible in this configuration: with an explicit `proxy.server`, a slide's `http://127.0.0.1:…/` subresource goes through the veto proxy and is refused (`502`, origin hits 0), with or without `bypass:'<-loopback>'`. The new test above pins the observable outcome anyway.
+  - *"the new ancestor tolerance has no test, relaxing it leaves the suite green"* — refuted by mutation: relaxing that single gate makes `safe-paths` fail (the escape-shape test depends on it). The gate is pinned.
+  - Also checked and dismissed: the exotic IPv4 spellings (`127.1`, `0x7f.0.0.1`, `2130706433`, `0177.0.0.1`, `::ffff:127.0.0.1`) are all normalised by the URL parser to `127.0.0.1` and refused, while a public address still passes.
+- **Housekeeping from the same review:** the root-side canonicalisation was byte-identical in `safeRel` and `assertContained` (the reviewer's "duplicated code", and the reason the mutation anchors appeared twice). Both now call one exported `canonicalRoot(root, where)`, which is also where the rule and the fail-closed behaviour live. **Tests +4 → 295 in fifteen files.**
+
+## 1.9.16 — 2026-10-10
+
+**Maintainer review round 7 (CHANGES_REQUESTED on head 51443138): two of the three blockers closed, each reproduced on our tree before fixing**
+
+- **Blocker 1 (security) — the veto proxy let IP-literal targets through.** The absolute-form HTTP branch of `startVetoProxy` fed straight into `policyGet`, and `policyGet` only carried the veto on the request's `lookup` option. For an **IP literal** target Node never calls `lookup` (it connects to the address directly), so the veto was a no-op there. Reproduced with the production proxy against a controlled local origin: `http://127.0.0.1:<port>/secret` returned **200 and the origin received the request**. Note the other blocked literals only produced 502 because those addresses are *unreachable*, not because policy said no — the canary therefore asserts both **zero requests** and that the 502 body names the policy reason (`blocked-host`), so an unreachable target can never make the guard look green. Fix in two layers: `policyGet` now applies the URL/host policy before connecting (the single literal entry point for every caller), and the proxy's HTTP branch pre-checks with an attributable error, mirroring the CONNECT branch. `checkedLookup` stays on the request for the DNS/rebinding case; a positive-control test keeps the allowed-hostname path working so the pipe isn't welded shut.
+- **Blocker 2 (macOS) — a path alias made legitimate paths look like escapes.** `safeRel`/`assertContained` canonicalized the probe side with `realpath` but, when the root did not exist yet, compared it against the **uncanonicalized** `path.resolve(root)`. On macOS `/var/…` and `/private/var/…` are two spellings of one directory, so a project path was rejected and the command exited 1 (the reporter saw `preview-page` refuse valid thumbnail/audio paths). Both sides now use the same `canonicalPath` rule — deepest existing ancestor's `realpath` plus the not-yet-created tail — and `assertContained` gained the same "root doesn't exist yet, so an ancestor probe is legitimate" tolerance `safeRel` already had (still withheld whenever the root exists, which is the shape that lets a project-internal link point at the project's parent). Fail-closed is preserved: an existing root whose `realpath` throws is still rejected. macOS itself isn't reachable from here, so the regression reproduces the *mechanism* with a junction alias and runs on every platform; a negative case pins that a genuine escape through a link inside the project is still refused.
+- **Tests +6 → 291 in fifteen files:** four production-proxy canaries (loopback literal → zero requests and a policy-named 502; private / metadata / IPv6 / zero / dotless literals all policy-named; hostname resolving to loopback still vetoed at connect time by `checkedLookup`; allowed hostname still served) and two alias cases (the legitimate path is allowed; the genuine escape is not). Red-proofs: removing both vetoes turns the loopback canary red; removing only `policyGet`'s assertion keeps it green (the layering control); removing the `canonicalPath` root side turns the alias regression red. Every restore is byte-compared.
+- **Blocker 3 is unchanged and still ours-to-explain:** exact-head execution evidence needs a base-repo administrator to approve the queued runs (the fork author gets 403, documented upthread); the mirror repository's dual-platform CI remains the substitute evidence.
+
+## 1.9.15 — 2026-10-03
+
+**Marketplace submission fix: the SKILL.md frontmatter failed real YAML parsing** (submission PLUGIN-202610030176 rejected with `SKILL_FRONTMATTER_INVALID`; the first submission, PLUGIN-202610020174, had failed earlier on the missing `.minimax-plugin/plugin.json` — fixed by the repo's plugin-package root layout)
+
+- **Root cause, reproduced locally with PyYAML:** the frontmatter `description` carried one unquoted ASCII `": "` — "(regulated subjects): finance/…" — and an unquoted plain scalar containing colon-space is a YAML parse error (`ScannerError: mapping values are not allowed here`). Every strict parser rejects it; the community repo's validator reads the line with a regex, which is why this passed 28 rounds of hosted validation. The guide's canonical example and the one marketplace-passed precedent both avoid the shape.
+- **Fix (description only, 977 → 970 chars, every trigger phrase kept):** the colon becomes an em-dash, and the four `**bold**` markers are stripped (the marketplace renders manifest/SKILL text as plain text — markdown emphasis shows up literally). The frontmatter now parses clean under PyYAML and stays within the 1024-char cap.
+- **Version discipline:** package content changed, so the version increments in all three manifests (`plugin.json`, `.claude-plugin/plugin.json`, `.minimax-plugin/plugin.json`), per the submission guide's "any content change must bump the version" rule. No code or behavior changed; tests stay at 285 in fifteen files (verified green on the dev tree before pushing).
+
+## 1.9.14 — 2026-09-28
+
+**Field handoff (Mavis, billing-arena project): the subtitle pill's `max-width` was unreachable dead code — subtitles wrapped at half the designed width, and the band gate couldn't see the collision** (A/B reproduced on our tree with the reporter's probe before fixing: 22 real narration sentences, 8 wrapped wrong; the fix yields 0 wraps)
+
+- **P0 · `width:max-content` (root cause, one line):** the pill is absolutely positioned with `left:50%`, so its shrink-to-fit available width is the containing block minus the left offset = **half the canvas** (960px at 1920); `translateX(-50%)` only shifts, it doesn't widen. The `max-width: 1400px` rule could never bind — subtitles wrapped at ≈22 chars instead of the designed ≈33, double-height pills reached 230px from the bottom and covered captions/disclaimers on two slides. `left:50% + translateX(-50%)` centering and `max-width` are mutually exclusive; the pill CSS now lives once in `tools.subPillCss()` (single source with `SUB_GEOMETRY`, which also gains `padYPx/radiusPx/bottomFrac/sub2Em` so no literal survives outside tools) **with `width:max-content`**, and capture injects it verbatim. Consequence worth stating: `subLineCap`'s pinned numbers (24/33/35) were the *designed* values all along — the implementation had never matched them until now.
+- **P0 · the band gate was both too small and blind:** `check-slides` 5e hard-coded a 168px band (single line actually measures 170; a wrapped pill reaches 230 — a 62px collision zone never reported while two slides were really covered). The band now comes from `tools.subBandTop()` computed from the same geometry (single-line vs bilingual double-line tiers, reading each slide's `text2`), and a new heuristic warns on **flow-anchored bottoms** (`margin-top:auto` / flex-spacer shapes) that the absolute-only scan never saw — both point at `grab-frames` for the definitive check, because a still preview cannot show subtitles at all.
+- **P1 · `.disclaimer` no longer bakes in `max-width:1240px`** (previously reported, unfixed): that cap was designed for one-line source attributions and trapped full-width multi-line disclaimer blocks (an explicit `width:100%` couldn't override it). Projects that want it narrower override it themselves. This changes the generated tokens.css body, so old projects' managed region turns stale until `--upgrade-css` — the designed propagation.
+- **P2 · `--ids` partial matches no longer render silently less:** unknown ids in `--ids` are now named in a warning (with the effective id list printed) in both `capture` and `grab-frames`; the message also covers the PowerShell shape where a bare `03,05` is evaluated as a number array and loses its leading zeros. (The reporter's earlier "`--ids` only honours the last one" was **retracted** — quoting the argument fixes it; the real skill-side gap was this silence.)
+- **Tests +4 → 285 in fifteen files:** a browser-level geometry test (30-char clause must stay single-line and wider than 1000px — the old CSS caps at 960 and wraps, both asserts redden), band-edge cases (169 must warn, 185 silent, 185 warns on a bilingual slide), the flow-bottom heuristic, the `.disclaimer` pin, and mixed-`--ids` CLI cases (capture + grab-frames). The round-22 single-source pin was updated by contract (the consumer is now `subPillCss`, not raw `SUB_GEOMETRY` fields — the pin caught the rewiring, which is its job). Four red-proofs: removing `max-content`, reverting the band to 168, disabling the flow heuristic, and dropping the `--ids` warning each redden exactly their own case.
+
+## 1.9.13 — 2026-09-28
+
+**External play-page review: six findings (2×P1, 2×P2, 2×P3), all reproduced personally before fixing** (the reporter had scripted the play page and hit every one of these live)
+
+- **P1 · self-contained overview thumbnails:** the overview grid hardcoded `../<id>.png`, pointing into the parent `preview/` directory — copying `play/` alone (or sending `index.html`) killed every thumbnail, and since `preview/*.png` is capture's output, a project that hadn't run capture silently got all placeholders. Thumbnails are now copied into `play/thumbs/` at generation (the directory is rebuilt fresh, so removed slides leave no stale images), the grid references them relatively, and the terminal counts what it copied (`仅 N/M 张` warning when capture hasn't run for some slides). The honest boundary is documented too: the snapshot copies read the project's `slides/` via `<base href>`, so `play/` is the minimum shareable unit, not `index.html` alone.
+- **P1 · a programmatic interface:** page state used to live entirely in closures, forcing automation to simulate keys and scrape `innerText` for "级 1/2" and guess the visible iframe — the reporter's script got this wrong. The page now exposes **`window.playPage`** (`state` getter with index/total/step/steps/fx/narr/trans/overview, plus `go/next/prev/setFx/setNarr/setTrans/overview`), documented in SKILL.md and the eval expectations.
+- **P2 · key hints and boundary feedback:** the footer listed ←→XPOFT while the handler also supported Space/PageDown/PageUp/Home/End/Esc — all keys are now listed (bilingual). And next/prev at the first/last frame used to return silently; they now flash a "已是第一张 / 已是最后一张" toast (the no-op itself stays: the old behaviour of replaying the last slide was removed on purpose).
+- **P3 · xfade `.show` leak + shadowed parameter:** in dissolve mode the delayed 400ms un-show is generation-guarded, so a superseding action within that window left the old frame at `opacity:1` — invisible to the eye (z-order masks it) but breaking any external "current frame" detection via `offsetParent`. `show()` now normalises the back buffer to hidden at entry, closing every leak path at the source. And `addStepScript`'s parameter is renamed `stages` → `stageIds` — it holds stage **ids** while the same file's `stages` is the timings dictionary; same name, different meaning, a maintenance trap.
+- **Tests +6 → 281 in fifteen files** (`preview-page`): relative thumbs path (and no `'../' + s.id` back), the full `playPage` surface, footer key hints in both languages, boundary-toast wiring (next/prev edge branches), the show-entry normalisation pin, and a CLI-level case asserting `preview/01.png` lands in `play/thumbs/01.png`, an uncaptured slide gets no fake image, the terminal counts 1/2, and stale thumbs are cleaned on regeneration. Four guards red-proofed (thumbs reverted → red; `playPage` removed → red; edge feedback removed → red; entry normalisation removed → red). One pre-existing test updated by contract (the placeholder text now names the capture step).
+
+## 1.9.12 — 2026-09-28
+
+**Field report: TTS reads numbers by its own rules — display form and spoken form are now separated (`clauses[].say`)** (user-reported: `2026年` was read as "两千零二十六年"; wanted: subtitle shows `2026年`, narration says "二零二六年"; same for `82.3%` → "百分之八十二点三" and `SWE2.3` → "SWE二点三")
+
+- **New optional field `clauses[].say`** — the spoken form of that sentence. `text` remains the on-screen/subtitle form and keeps the natural written form (Arabic numerals, original orthography — spelled-out digits like `八十二点三` belong in `say`, never on screen). **TTS input and the ASR checklist expectation are always `say ?? text`**; the burned-in subtitles and `out/subs.srt` always come from `text`. `say` propagates through `plan-timings` into `timings.json` (which is what `build-video --asr` reads); `say` applies to the main line only (`text2` stays display-only).
+- **Year-reading gate (plan-timings, zh/yue):** a clause whose effective spoken text (`say ?? text`) contains a four-digit year followed by 年 gets a warning that names the digit-by-digit broadcast form and hands over the ready-to-paste `say` value (`加 say:"二零二六年…"`). The gate is deliberately narrow — whether a number reads digit-by-digit or as an integer is intent (`2026年` digit-by-digit, `2026点` as an integer), so only the unambiguous year shape is gated and the rest of the number discipline lives in the docs (SKILL.md schema section + tts-and-timing.md, with all three reported examples).
+- **Tests +1 → 274 in fifteen files:** the year gate (no-`say` named with the paste-ready form / with-`say` silent / `en` not over-warned / `say` passed through to timings verbatim and never invented for `say`-less clauses), and the smoke pipeline now runs `build-video --asr` on a `say`-bearing clause and asserts the checklist expects the spoken form while the srt keeps the display form (`say` must not leak into subtitles). Each guard red-proofed: disabling the year gate, dropping the `say` propagation, or reverting the checklist to `c.text` reddens exactly its own case.
+- **Round-24 self-review corrections (same release):** the two-axis review of this increment found four real items in the fix itself, all closed here — ① **clause-start allocation weighted by the display form**: the audio speaks `say` ("82.3%" shows 5 chars, "百分之八十二点三" speaks 8), so time weights, per-clause `chars` and the pacing check now use `say ?? text` (subtitle-wrap geometry still uses `text` — the pill is on screen); the year-gate test now pins the weighted start against both candidate formulas. ② the year gate's regex only matched `19xx/20xx` while the comments claimed "four-digit year" — now any `\d{4}年` (1897 warns too). ③ the zh-guard used the layout metric `emPer===1` as a language proxy — replaced by an explicit `yearGate` flag on `zh`/`yue` (unknown-language fallback stays ungated). ④ two `evals.json` cases still taught the old "write the spoken form into the narration text" move (the EN number-swap line and the ASR-mismatch fix) — both now teach the `say` mechanism. Each correction red-proofed (weight reverted → weighted-start assert red; regex reverted → 1897 red; gate off → warn red).
+- **Round-25 self-review corrections (same release, tests +1 → 275 in fifteen files):** five smaller items — ① **`check-timing`'s silence-to-clause alignment still used the display form**: `matchBoundaries` moved to `tools.mjs` as the single source and now counts pause punctuation on `say ?? text` (with `say`/`text` punctuation disagreeing, exact mode misaligned and `--calibrate` would write back wrong starts); unit-tested with a fixture where the two forms produce different methods/values. ② the year regex matched four digits **inside** longer digit runs ("12026年" sliced to "2026年" and suggested a digit-dropping `say`) — now digit-bounded on both sides, with a non-warn case pinned. ③ a self-referencing comment cited a line number that the same release had moved. ④ `LANG_CFG` now states `yearGate` explicitly on every branch and shares one `zh`/`yue` literal. ⑤ the `evals.json` ASR-mismatch case now says the `say` rewrite must actually change the pronunciation (九成二-style), not restate what TTS already says (百分之九十二 wouldn't fix the mishearing). Both behavioural corrections red-proofed (spoken-marks reverted → alignment test red; boundary reverted → the 12026 case red).
+
+## 1.9.11 — 2026-09-28
+
+**Round-6 review: the SSRF policy now fails closed and is bound to the actual connection** (maintainer finding on 1.9.10: `assertResolvedHost` treated a DNS lookup error as "pass", and the real fetch/navigate resolved the hostname again on its own — exactly the DNS-rebinding window)
+
+- **Fail closed:** `assertResolvedHost` no longer swallows lookup errors — a failed **or empty** answer is a rejection (`dns-error`), never a silent pass.
+- **Bound to the connected address:** new `checkedLookup` (`url-policy.mjs`) wraps Node's `lookup` and vets **every** address the socket could connect to; new `policyGet` is the tool's egress GET (`http(s).request` + that lookup, follows no redirects — the caller re-validates each hop — and caps the body **mid-stream**, not by trusting `content-length`). Both download routes (`--url` and `--get`) now egress through it; the old paths called global `fetch()` / `context.request.get()`, each resolving DNS a second time after the pre-check had passed.
+- **The browser cannot inject a lookup, so its traffic is tunnelled:** new `startVetoProxy` is a loopback-only proxy. **Both** Chromium launch sites now go through it — the fetch tool's browser (navigation, self-followed redirects, every subresource) **and the render browser in `capture.mjs`** (whose slide HTML is locally generated and externally gated by `check-slides`, but a gate is not a connection point; with the tunnel, a bypassed gate still cannot reach private/metadata addresses). Upstream connects run the same `checkedLookup` (CONNECT) or `policyGet` (absolute-form http), so the veto sits on the resolution the tunnel actually uses. The per-request route interceptor keeps the string policy and now re-resolves **uncached** — the old per-host cache widened the rebinding window it was meant to guard.
+- **Tests +16 → 273 in fifteen files** (new `connect-binding.test.mjs`, 15 cases, plus a geometry single-source pin in `render-smoke`): lookup-error/empty-answer rejections; `checkedLookup` public-passthrough/loopback-veto/resolver-error; `policyGet` rebinding with the target receiving **zero requests on the real `net` lookup path**, response plumbing, and the mid-stream cap; veto-proxy CONNECT tunnel/refusal/literal plus absolute-form relay/refusal (again zero upstream requests); a browser-mode wiring smoke (file:// page with the proxy in the chain); a structural pin that the render browser's launch carries the proxy; and the pin that no script but `tools.mjs` may carry the pill-width literal. Every guard red-proofed against its own defect: restoring fail-open, removing the connect-time veto, disabling the CONNECT pre-veto, dropping the render-browser proxy, or re-hard-coding the pill width reddens exactly its own cases (the red-proof also caught a vacuous-green test — assertions made inside a `lookup` callback get swallowed by the promise chain, so the new tests capture first and assert after).
+- **Round-22 residuals closed:** `evals/evals.json`'s English-narration case still called "42 characters per line" plan-timings' rule — now the same craft-guideline-vs-pill-capacity split as the other five surfaces (Latin capacity ≈55 chars @1080 / ≈75 @1920); and `review-round3`'s redirect-loop structure pin now also requires both download paths to egress via `policyGet`.
+- **Tool discovery walks the whole ancestor chain (pre-existing, surfaced by this round's three-tree run):** `candidatePaths` only looked two levels above the scripts dir for `node_modules`, so in the deeply nested plugin layout (`plugins/<owner>/<name>/skills/<name>/scripts`) a spawned `asr.mjs` with a temp-project cwd could not see a workspace-root install the test runner itself had found — one suite, two verdicts depending on layout. It now walks every ancestor (same shape as Node's own module resolution), which is also the layout the monorepo scoped workflow will run in.
+- **Round-23 self-review pass (two-axis, on this release):** the render-browser gap above and four judgement items were found by our own review round and closed in the same release — misleading indentation in the restructured browser block, inconsistent manual-close sites before `process.exit` (now also close the veto proxy), a dead parameter and missing timeout guards in the new tests.
+
+## 1.9.10 — 2026-09-23
+
+**Round-21 review: subtitle line-width geometry corrected, and the 18-char rule's drift closed** (two-axis review of 1.9.9 found one hard inconsistency and one wrong baseline — both reproduced before fixing)
+
+- **Rule drift (hard):** 1.9.9 rewrote the subtitle wrap rule in `SKILL.md` and `references/authoring.md`, but three surfaces kept the old ">18 chars" rule — `references/render.md` ("A single sentence >18 chars … gets a warning from plan-timings"), `references/tts-and-timing.md` ("the single-line subtitle limit is 18 chars for Chinese / 42 characters for English") and `evals/evals.json` ("单句建议 ≤18 字"). All three now carry the width-aware rule (and `render.md` is a file 1.9.9 itself had edited — so this was a missed spot, not an untouched file).
+- **The 1080 baseline published in 1.9.9 was wrong** — correction to that entry: its "≈18 CJK chars per line at 1080, ≈32 at 1920" figures were a restatement of the superseded vertical baseline. The pill's font **and** padding both scale with `--sub-scale = clamp(W/1920, 0.75, 1.25)`, so per-line capacity is **not** linear in width: true values are ≈**24** CJK chars at 1080, ≈**33** at 1920, ≈**35** at 2560. The `18 × W/1080` estimate over-warned at 1080 by ~27% and — the dangerous direction — missed real three-line wraps at 2560 (estimated 42 vs true 35.6). All five doc surfaces now state the corrected numbers and separate the **craft guideline** (≈18 CJK / ≈42 English for one-liners) from the **pill capacity** the gate uses.
+- **Single source:** the pill geometry (box width 0.729167, side padding 34px, font 40px, the scale clamp) now lives once in `tools.SUB_GEOMETRY` / `subScale` / `subLineCap`; `plan-timings` derives the line cap from it and `capture` builds the `.kit-sub` CSS from the same constants — the two copies had already drifted (the review's Duplicated Code finding).
+- **Entry guard (same review):** the same-basename fallback now **announces itself on stderr** when it fires — the trade is "a loud, self-annotated extra run" against "a silent no-op" (the silent kind cost a 10-minute mystery in the field); its stated justification no longer claims cases `realpath` already covers. Also from the review's small items: the unused `fileURLToPath` import in `preview-page.mjs` is gone, the `real` helper is renamed `realpathOf`, and the guard's `!entry` branch is now actually covered (explicit `null` — passing `undefined` hits the default parameter and never reached it).
+- **Tests +1 (257 in fourteen files):** the geometry numbers are pinned (`subLineCap` = 24/33/35 CJK, 55 Latin; reverting to the linear formula turns it red), and the plan-timings case now exercises 2/3/4-line samples so the "three or more" predicate is tested at its exact boundary (the review noted the 1.9.9 title claimed a 3-line case it didn't run).
+
+## 1.9.9 — 2026-09-22
+
+**Cloud-sandbox field batch: entry guard could silently no-op; subtitle wrap warning mis-fired** (report from a mavis-style web sandbox session, sorted by the reporter into skill issues vs environment quirks)
+
+- **🔴 Silent no-op entry guard (the core bug).** `preview-page.mjs`'s "run `main` only when executed directly" check compared `path.resolve(argv[1])` with `fileURLToPath(import.meta.url)` — for an entry reached through a **symlinked path** the module URL is the real path while `argv[1]` is the link path, so the comparison is always false: the script exits 0 with **no output and no work done** (the reporter burned ~10 minutes proving `main` hadn't run). The check is now `tools.isMainModule` (single source): realpath both sides, plus a same-basename fallback so the "silent skip" shape is impossible in practice. Clarified in the same comment: `node -e "import(…)"` does not run `main` by design (import semantics) — run CLIs as `node <script> …`. (The reporter attributed this to ESM module caching; caches do not span processes — the path-shape mismatch is the mechanism, and the symlink case is the one that reproduces.)
+- **🟡 Subtitle wrap warning was calibrated on portrait and mis-fired on landscape.** The `> 18 characters` limit is the single-line baseline at 1080 width (≈18 CJK chars per line); at 1920 the pill fits ≈32, so 20–32-char sentences were warned "will wrap" without wrapping at all — and two wrapped lines read fine anyway. `plan-timings` now estimates line width from the canvas width and warns at **three or more** lines (naming the estimated width and line count).
+- **🟡 Capture runtime expectations documented** (`references/render.md`): frame count scales with entrance-choreography windows (measured 5 × 15s ≈ 2250 frames ≈ 6–8 min); quick passes can drop `fps` to 24, trim choreography, or shoot in `--ids` batches (frame dirs are per-slide).
+- **🟢 Environment quirks recorded in `references/symptoms.md`** (not skill defects, but sandbox users hit them repeatedly): partial Playwright browser installs in cloud sandboxes (`chromium-*` vs `chromium_headless_shell-*` layout, copy+rename stopgap) and `git clone` SSL failures (use the `codeload.github.com` zip).
+- **Tests +2 (256 in fourteen files):** `isMainModule` (direct / imported / same-name fallback / symlinked entry — the symlink case runs wherever file symlinks exist, including Linux CI) and the plan-timings width-aware threshold (2 lines silent, 3+ named). Red-proofed: reverting the guard to raw-path comparison reddens the entry test.
+
+## 1.9.8 — 2026-09-22
+
+**Layout-overflow protection: safe-center scaffold + a capture-time geometry self-check** (field report: a dense table/data-viz slide had its kicker pushed to y≈17px — clipped at the canvas top — and the screenshot review passed it as fine)
+
+- Root-cause chain (field postmortem, numbers reproduced): content ≈925px inside an 800px content box (280px subtitle-band padding) with plain `justify-content: center` on the whole `.layout` — a centered flex line that overflows is pushed **above** the canvas and clipped there; the attempted `margin-top: 80px` patch made it worse (margins join the content height in flex centering).
+- **Template fix (root):** the scaffold `_template.html`'s `.layout` now ships `justify-content: safe center` + `padding: 120px 160px 190px` (top mirrors the brand bar, bottom is the subtitle safe area). Under `safe`, overflowing content falls back to top alignment — the kicker can never be clipped at the top edge again. `references/authoring.md` states the pitfall and the wrong "margin-top patch" explicitly.
+- **Check fix (rendered truth):** `capture.mjs` gained a layout-geometry self-check next to the broken-image check — it measures the layout's content extent vs its box and the topmost content element's y, and names the numbers ("内容总高 X > 可用 Y（溢出 ≈Z）… 顶部元素被推到 y≈N px"). Static text checks cannot see content height; this one measures, so the failure cannot pass another eyeball screenshot review as "just not rendered yet".
+- **Tests +2 (254 in fourteen files):** the geometry self-check warns (with the numbers) on the reproduced overflow shape and stays silent on healthy short content; the scaffold contract (`safe center` + the exact padding) is pinned. Each guard red-proofed individually — removing it reddens exactly its own case.
+
+## 1.9.7 — 2026-09-22
+
+**Chart-structure gate: percentage-height bars must not sit bare in a flex column** (field report: an agent's bar chart came out with the bars collapsed/missing, and the screenshot review passed it as "just not rendered yet")
+
+- The authoring docs already carried the warning (`references/authoring.md` chart disciplines: bars in a flex column with percentage heights get squeezed by `flex-shrink` — measured: 84% and 72% drawn at the same height; fix = wrap in `.chart-plot-cell`, the grid `1fr` plot row), and the symptoms table has the entry. The agent read the content-volume and entrance sections but not the chart-toolkit section — and `check-slides.mjs` had no structural check to catch the mistake, so a silent visual failure sailed to `build-video`.
+- New `check-slides` finding (hint-level, does not block): bar-semantic element with a `< 100%` inline height **in a slide that contains a `flex-direction: column` and no `.chart-plot-cell`** → "these bars may collapse; use `.chart-col > .chart-plot-cell.grid > .chart-bar-wrap{height:N%}`". Three conjuncts on purpose — percentage heights outside flex columns are legitimate and must not warn (covered by a false-positive regression case).
+- **Tests +3 (252 in fourteen files):** pitfall shape warns (and stays hint-level), the documented recipe shape stays silent, the no-flex-column case stays silent. Red-proofed: disabling the check turns the pitfall case red.
+
+## 1.9.6 — 2026-09-22
+
+**mcode-first tool choice made explicit** (field report: an agent running inside the mcode sandbox reached for mmx-cli and forgot the mcode tools were available)
+
+- The runtime table already marked the mcode column `(preferred)` and the TTS cell `(the main path)`, and mmx-cli setup was scoped to "(non-mcode environments)" — but the discipline paragraph named `asr.mjs` **unconditionally**, and since 1.9.0 `asr.mjs` auto-picks `mmx speech transcribe` whenever mmx-cli is installed. Following that sentence for voice verification in the mcode sandbox leads straight to mmx-cli.
+- Two documentation fixes, no code change: (1) a hard environment rule above the table — in the mcode sandbox use the left column only, do not install or call mmx-cli there (`mcode-tools` covers TTS, music and ASR); (2) the voice-verification sentence now names the per-environment tool (mcode sandbox: `mcode-tools upload_temp_url` + `connector__matrix__listen_audio`; other environments: `asr.mjs`).
+
+## 1.9.5 — 2026-09-22
+
+**Round-20 review: two regressions introduced by 1.9.4's own fix — found by fresh-eyes review, reproduced visually, fixed and guarded**
+
+- **The cover baked in the last subtitle.** The substill pin (stop the animation + inline styles) left the subtitle elements with no animation for `captureCover`'s hide step to seek (`a.currentTime = 0`), so the last subtitle survived into `preview/cover.png` — which ships as `attached_pic`, `out/cover.png` and the opening 0.25s dissolve. The cover and the still-mode base now hide subtitles explicitly via inline styles instead of relying on animation state.
+- **The still-mode base baked in the last subtitle.** The "last clause emits no fade-out tail" shape left the animation's terminal value at `opacity: 1`, so `finish()` no longer hid subtitles — and the still-mode terminal image is consumed by `build-video` as the subtitle-free base before the first clause. The keyframe shape now satisfies both constraints at once: the plateau holds to **99.999%** and the tail occupies **100%** alone (visible fade-out ≈ 0.001% — the grey-subtitle bug stays fixed) while the terminal value returns to 0 (the "finish() hides" contract the cover and base rely on). The unit contract test now asserts exactly this — first and last keyframe groups must declare `opacity: 0` — which fails against 1.9.4's shape and would have caught it.
+- **Guards + red-proofs:** one pixel assertion each for the cover and the base ("must differ from a sub-bearing still" — the leak measured as diff **0** in both regressions). Red-proofed individually: the cover guard goes red with only the cover hide removed; the base guard and the unit contract go red with the 1.9.4 shape restored.
+- **Corrections to 1.9.4's entry** (it stands as written; this sets the record): ① "all three trees 249 pass / 0 fail / 3 capability skips" is arithmetically impossible — the measurement is **249 tests: 246 pass / 0 fail / 3 skips**; ② "thresholds were recalibrated" — the threshold values were unchanged, the fixture text was (the same sentence's second half describes this correctly); ③ "red-proofed … both new regression tests fail" applies to the keyframe-shape test and the brightness assertions — the dark-floor guard targets the first-cut cross-still leak and is not expected to fail on the pre-fix shape; ④ "the platform was eaten" is a mistranslation — the word is **plateau**.
+- Judgement cleanups from the same review: the duplicated pair-diff helpers in `subtitles-invalidate.test.mjs` merged into one; comment wording fixed (per-screenshot page reuse; the dark-floor comparison's discrimination depends on the fixture's light theme and now says so).
+- Verified: three trees **249 tests / 246 pass / 0 fail / 3 skips**; the affected project re-rendered end-to-end — 12 subtitle stills at full brightness (235), the 1.05s/2.20s frames identical (Δ0), the **cover clean** (title fully revealed, no subtitle — visually checked), and the still-mode base subtitle-free (guarded in the suite).
+
+## 1.9.4 — 2026-09-22
+
+**Grey/flickering subtitles fixed ("subtitle appears and goes grey within a second, then stays grey") — found in a real production render (the `ai-daily-2026-09-21` project), diagnosed by handoff analysis and reproduced byte-for-byte before fixing**
+
+- **Root cause (keyframe collision).** The last clause's subtitle keyframes emitted a fade-out tail whose first offset collided with the plateau's end: for the last clause `b = 100`, so the generated CSS read `…22.471%,100.000%{opacity:1}100.000%,100%{opacity:0}` — the same offset declared twice, and per CSS the later declaration wins. The platform was eaten and the whole rest of the clause became a slow linear fade-out (back-calculated 0.441 at the splice point; measured 0.44). The spliced static subtitle still, sampled at the window midpoint, caught exactly that mid-fade value (band peak 138 instead of 235), so the video showed a bright subtitle for the frame-sequence half of the window and a grey one from the splice to the end of the slide. The last clause now emits **no fade-out tail** — the platform holds to the slide end as designed.
+- **Substills no longer sample a guessed time.** The still used to seek the subtitle animation to the window midpoint ("midpoint = full-opacity platform", true only while the keyframes are right). The target subtitle is now **pinned** (animation off, opacity 1) and every other subtitle is re-pinned hidden on each capture pass — structurally decoupled from keyframe shape. (First cut of this fix leaked pinned subtitles into later stills on the same page — a dark double-box patch; caught by inspecting the output image before release, and now guarded by an assertion comparing the dark floor between stills.)
+- **Self-check false positive removed.** The overlap warning added the fade-out length to `prevB`, which is already the fade-out end — a double count that fired on every clause pair ("overlapping X s" where X was always the fade length). Real overlap is only possible when a short window is pushed out by the 0.5% floor, and that is what the warning now reports.
+- **Window end clamped to 100** (a clause starting after 99.5% could emit offsets > 100%, which CSS keyframe selectors do not accept).
+- Keyframe generation moved into `tools.mjs` (`subtitleWindows` / `subtitleKeyframes`) as the single source shared by the page injection and the window self-check — the two copies of the arithmetic can no longer drift.
+- **Tests: 249 in fourteen files (+1).** New tool-level keyframe-shape test (no same-offset collision anywhere, last clause tail-less, non-last keeps its tail, offsets monotonic within [0,100]) plus full-brightness and cross-still-leak assertions on the spliced stills. **Two render-smoke thresholds were recalibrated** and this is disclosed on purpose: their old `> 4.0` cross-clause margins had been passing partly on the bug's own grey artifacts (the fixtures used equal-length `A…`/`B…` strings, so once both stills render at full brightness the pure glyph difference is only ~3.7). The second fixture clause is now a different-length string, making the discrimination geometric (box width) and font-independent; threshold semantics are unchanged (a frozen frame yields ~0 and still fails).
+- Verified: all three trees 249 pass / 0 fail / 3 capability skips; red-proofed by restoring the pre-fix shape (both new regression tests fail, unrelated tests unaffected); re-render of the affected project — all 12 subtitle stills at full brightness (band peak 235 vs the bug's 138) and the finished video's 1.05s/2.20s frames now identical in brightness (previously a 0.88 → 0.44 jump at the splice).
+
+## 1.9.3 — 2026-09-21
+
+**Diff hygiene: text files normalized to LF (maintainer review gate on the PR head)**
+
+- Nine published files carried CRLF line endings — 1,854 lines in total — and under default whitespace rules (no `cr-at-eol`) `git diff --check` flags every CR-terminated line, so on the PR head that printed 3,708 lines of report. All nine are now LF, matching the rest of the repository (root README, every other plugin, the maintainer's own plugin). One of the nine was mixed (`tests/fetch-policy.test.mjs`, later edits were already LF); all are now consistently LF.
+- `.gitattributes` (`* text=auto eol=lf`) added at the plugin root, byte-identical in both published trees — the convention `plugins/antianqi/mcode-island` already uses — so CRLF cannot re-enter through a Windows checkout.
+- No functional change, verified: the conversion was byte-checked (refuses to write unless the old and new versions are byte-identical after normalizing line endings only), `tokens-template.mjs`'s generated CSS was already LF (JS template literals normalize CRLF to LF, so generated artifacts and `TOKENS_REV` are unchanged), and the full suite was re-run on all three trees afterwards.
+- Internal `docs/` files (never published) were normalized in the same pass.
+
+## 1.9.2 — 2026-09-20
+
+**Seventeenth review round (two-axis, over the 1.9.1 increment): both axes caught the same two defects in 1.9.1's own fix — corrected here**
+
+- **The env scrub was at one call site, not the source.** 1.9.1 stripped `MINIMAX_API_KEY`/`MINIMAX_BASE_URL` only from the `speech transcribe` child; the `mmx --version` capability probe still inherited them, so 1.9.1's "the key never reaches the external CLI even by inheritance" was overstated. The scrub now lives inside `mmxRun` itself — the single spawn point — so every mmx child (probe included) gets it.
+- **The `ENVKEY=absent` assertion was vacuous.** It read the first `ENVKEY=` line in the shim log, which was written by the test's own setup probe (whose env never held the key) — it passed regardless of the transcribe child's env, and would have failed spuriously on machines with `MINIMAX_API_KEY` exported. The assertion now requires **every** `ENVKEY=` line in the log (probe and transcribe alike) to be `absent`; the test helper's own probe scrubs the same way for determinism. Red-proofed: removing the scrub from `mmxRun` turns this test red (verified locally before restoring).
+- **Disclosure correction:** the `a&b c` metacharacter-directory regression exercises the cmd.exe quoting path only on Windows, where the CLI is invoked through `cmd.exe /c`; on POSIX the child is spawned directly and the metacharacters are inert by construction (the Linux CI run therefore proves path handling there, not the quoting). 1.9.1's blanket "regression-covered" phrasing was Windows-only in effect.
+
+## 1.9.1 — 2026-09-20
+
+**Sixteenth review round (two-axis, over the 1.9.0 increment): one hard finding fixed plus four hardening/correction items**
+
+- **The plugin-root READMEs were missed by 1.9.0** (both review axes caught it): the 1.9.0 doc updates were applied to the skill-repo copies, but the copies inside `plugins/Wzdhehe/html2video-for-mcode/` — the tree this PR actually publishes — still described ASR as "Optional: `MINIMAX_API_KEY` … direct REST" and carried the old 244-test count. Both plugin-root READMEs are now byte-identical to the skill-repo versions (requirements/disclosure bullets + the 248-test count).
+- **Wording correction:** 1.9.0 claimed "the external CLI never touches user paths" — overstated. The input audio is passed to mmx read-only at its original path; only the **output** is staged through `os.tmpdir()` before landing in jailed paths. In-file comments now say exactly that.
+- **Windows cmd.exe quoting hardened:** args were quoted only on whitespace, so a legal path containing `&`/`|`/`<`/`>`/`(`/`)`/`^` would be re-interpreted by cmd as control characters. Any metacharacter now triggers quoting (documented residual: `%VAR%` inside quotes still expands; the worst case is mmx failing loudly on a mangled path). Regression-covered by running the shim test inside a project directory named `a&b c`.
+- **`MINIMAX_API_KEY` / `MINIMAX_BASE_URL` are stripped from the mmx child environment** — the key never reaches the external CLI even by inheritance; the shim test asserts `ENVKEY=absent`.
+- **The shim test now really asserts the full argv mapping**, including `--response-format verbose_json --timestamp-level word` (1.9.0's changelog listed `--timestamp-level`, which the test did not actually cover), and the duplicated two-test setup was extracted into one helper.
+- No test-count change: **248 tests / 14 files**.
+
+## 1.9.0 — 2026-09-20
+
+**ASR toolchain simplification: `mmx speech transcribe` is now the default provider (mmx-cli ≥ 1.0.26, merged upstream in MiniMax-AI/cli#262)**
+
+- `scripts/asr.mjs` now picks its provider automatically: with mmx-cli ≥ 1.0.26 on `PATH` it shells out to `mmx speech transcribe` — the same login as TTS, so the API key never passes through this script at all. Results are staged through a dedicated `os.tmpdir()` scratch directory before landing in jailed paths; the external CLI never touches user paths directly. `--provider api` forces the previous direct-REST path for environments without mmx-cli, and `--from` (offline compare) no longer trips the missing-key gate.
+- The capability probe reads `mmx --version` — `--help` exits 0 even for unknown subcommands, so it cannot gate. On Windows the CLI is invoked through `cmd.exe /c` because Node cannot spawn npm-global commands directly (bare name → ENOENT, `.cmd` → EINVAL).
+- The endpoint allowlist is unchanged and provider-independent: a non-official `--base-url` / `MINIMAX_BASE_URL` is still rejected before any provider resolution.
+- Docs corrected where they claimed "mmx-cli has no ASR subcommand" (false since 1.0.26): SKILL.md (scripts table, runtime comparison table, first-time setup, Phase 5), `references/tts-and-timing.md` (ASR paths B/C), and both READMEs' requirements/disclosure bullets — the mmx provider spends the same paid, quota-metered MiniMax account as REST.
+- Tests +4 (invalid `--provider` guidance; keyless `--provider api` guidance; an mmx-shim integration pair asserting the exact argv mapping — `speech transcribe --model asr-1.0 --response-format … --timestamp-level … --language …` — tmpdir staging/cleanup, and failure passthrough with the upgrade hint). REST-stub tests now pin `--provider api` so they stay deterministic on machines that do have mmx. **248 tests / 14 files**: verified locally with tools (245 pass / 0 fail / 3 capability skips) and end-to-end against real mmx 1.0.26 (keyless run, transcript exact including the number).
+
+## 1.8.1 — 2026-09-19
+
+**Thirteenth-review residue (found by the independent sweep over the published `e845cbd8` tree): `SKILL.md` was the one install surface 1.8.0 left behind**
+
+- **The 1.8.0 channel fix did not reach the file the agent actually reads.** That change collapsed both READMEs' install sections to "copy the reviewed skill directory out of the hosted plugin — not from any other source", but `SKILL.md`'s "Installing into other Agent environments" block still carried a bare `cp -r` whose source was unstated, plus a generic skills-CLI route — which re-admits exactly the "install it from a repository of your choosing" path 1.8.0 removed. It matters more than a README here: `.claude-plugin/plugin.json` points `skills` at this file. Both routes are gone; the block now sources the copy from the hosted plugin directory, worded like the READMEs.
+- No code or test changes (nothing asserts on this block); the version is bumped so published copies carry the corrected text. The removed routes now occur in the published trees only inside the changelog entries that describe their removal.
+
+## 1.8.0 — 2026-09-19
+
+**Thirteenth review (both axes at the exact head): two documented-contract violations fixed, three duplications closed, the rest disclosed**
+
+- **The standalone repository is no longer documented as an install channel.** The host's `docs/architecture.md` is explicit — "the hosted directory is the publication unit … contributors do not create a second repository" — yet this plugin's `homepage` pointed at a second repository and both READMEs offered `npx skills add Wzdhehe/…` / bare `cp -r` as install routes, which routes users around the review gate that guards `main`. `homepage` now points at the hosted plugin directory, and both install sections route through that hosted plugin only ("copy the reviewed skill directory out of the hosted plugin — not from any other source"). The standalone repository remains as a read-only history mirror, no longer a channel.
+- **The paid account is now disclosed as paid.** `CONTRIBUTING.md` requires disclosing "accounts, paid services"; the READMEs named `mmx-cli` / `MINIMAX_API_KEY` without saying they spend a **MiniMax account's paid, quota-metered balance** (billed per character/minute; each transcription call billed by audio duration). Both READMEs now say so, in both languages.
+- **The two redirect-following loops are one function.** `fetch-official-images` had the hop/policy loop implemented twice — once for the `--url` direct-download path, once inline for `--get` — and the second copy lacked the ping-pong cycle guard and the zero-byte body guard. A single `followRedirects(src, get)` core (client injected as an adapter) now serves both paths, so `--get` gained both guards; a structural regression test pins "exactly one hop loop, both call sites go through it, both zero-byte guards present" (a behavioural cycle test cannot run locally: the SSRF policy refuses loopback by design, and public URLs are unreliable in CI).
+- **ffprobe probing consolidated.** The verbatim `format=duration` probe existed three times and the `stream=width,height` probe twice (`asr`, `plan-timings`, `build-video`, `prep-image`); both now live once in `tools.mjs` (`probeDuration`, `probeSize`), behaviour-identical, with local thin wrappers at the call sites.
+- The preview-page file-symlink canary now covers a **per-slide leaf** (`01.html`) in addition to `index.html` — three write sites, previously two of them only code-covered.
+- Judgement calls from the same review, disclosed rather than changed: `review-round2`/`review-round3` are named by review provenance, not by topic (each `describe` block is topical; renaming would orphan the suite list documented in both READMEs); the subtitle fade constants in `capture.mjs` are deliberately re-derived in the Node self-check so drift between the page script and the check would be *caught* rather than hidden by a shared constant; the subtitle-band geometry (render-side bar vs check-side zone) uses intentionally different numbers whose unification would change gate behaviour mid-review; the css-kit region-range merging appears twice in different operations (locate vs remove); and `preview-page`'s `buildPlayPage` embeds the play page as one long template — a structural trade-off, not a defect.
+- The PR body's evidence transcript now matches a fresh validator run (the host has merged unrelated plugins since the old transcript: 26 → 27) and names `publish/validate-plugin.mjs` as an authoring-side tool that is deliberately not part of the shipped plugin tree.
+- Tests **243 → 244 in fourteen files**, all green in the development tree and both published trees.
+
+## 1.7.9 — 2026-09-19
+
+**Docs-only: corrected the 1.7.8 entry's site-class arithmetic** (the per-class numbers did not sum to the headline because the classes overlap — the order-protected site is also a jailed one). The conclusion stands on its own without the tally: an exhaustive primitive-level inventory of every write/delete site across all twenty scripts found no exploitable unjailed write. No code or test changes; versions bumped so the published copies carry the corrected entry.
+
+## 1.7.8 — 2026-09-19
+
+**Eleventh-round audit: the write-site inventory converges; one shipped-in-1.7.7 canary had an unsatisfiable assertion**
+
+- **The 1.7.7 transcode canary's "nothing appears in the input directory" assertion could never pass** on any platform where it actually runs: the test itself plants a symlink at the old temp path, and `existsSync` follows the link to its existing target — so the assertion compared against the test's own plant and would have failed on Linux CI (it never executed anywhere: file symlinks are uncreatable on the author's Windows, and CI runs sit at `action_required`). The assertions are rewritten to what the fix actually guarantees: a before/after diff of the input directory shows **no new entries**, the outside canary file is byte-unchanged, the planted path is still a symlink (not replaced by a regular file), and the `os.tmpdir()` transcode directory is removed by the time the run ends.
+- **PR-body counts re-phrased as capability invariants.** "243 pass, 0 fail, 0 skip" had gone stale the same way twice before — "0 skip" is only true where file symlinks are creatable. The body now states what is invariant (**0 fail in every environment measured**) and describes skips as capability-dependent and always named, instead of carrying per-environment counts.
+- The standards axis' primitive-exhaustive write-site inventory (method changed from file-intuition to primitive enumeration across all twenty scripts) found **no exploitable unjailed write**: jailed sites, CLI-given paths per the documented contract, internal temp files, and one order-protected site (the `build-video` `part-*` writes, disclosed in 1.7.7).
+- Test count unchanged (243 in fourteen files): the canary's assertions were corrected, not added to.
+
+## 1.7.7 — 2026-09-19
+
+**Tenth-round audit: the 1.7.6 "no other member of this class" claim was wrong — one more leaf existed**
+
+- **`asr.mjs` no longer writes its transcode temp next to the input file.** An independent sweep (the spec axis re-did the write-site inventory rather than trusting the changelog) found the one site the 1.7.6 sweep missed: oversized inputs are transcoded to `.asr-<name>.16k.mp3` **beside the input**, so for an input inside the project a pre-planted file symlink at that derived name would be written through by `ffmpeg -y`. The temp now goes to a per-invocation `mkdtemp` directory under `os.tmpdir()`, removed whole when the transcription attempt ends. New canary drives a real 501-second input (past the 500 s limit → real transcode, which runs **before** any network request) with the old temp name planted as a file symlink next to the input: nothing appears in the input directory, the canary is untouched. Runs on Linux CI; names its skip on stock Windows.
+- **Correction to 1.7.6's claim, stated plainly:** its changelog said a sweep found "no other member of this class". That was false — the `asr` transcode temp above was exactly such a member, missed by our own sweep and caught by an independent one. Lesson recorded: a sweep claim about absence is only as good as the inventory it ran on, and "we looked" is not evidence.
+- Disclosed without change: `build-video`'s `part-*.mp3` writes are protected **by order** (the preceding `safeOut`-jailed delete of every old `part-*` rejects a pre-planted leaf symlink before the ffmpeg write) — the invariant is now a code comment so the ordering is not accidentally reordered.
+- Tests **242 → 243 in fourteen files**, all green in the development tree and both published trees (the three file-symlink canaries name their skip where Windows lacks Developer Mode).
+
+## 1.7.6 — 2026-09-19
+
+**Ninth-round audit: the last uncontained leaf, and a body sentence fixed to stop going stale**
+
+- **`grab-frames` frame writes are now leaf-contained.** Its output *directory* was jailed, but `build/introspect/frame-<id>-<tag>.png` was written with plain `path.join` (the comment argued the `safeId` whitelist rules out path injection — true for traversal, blind to a pre-planted **file symlink**, which `ffmpeg -y` would have written through). The leaf goes through `safeOut` now, with a canary (needs ffmpeg + file symlinks → runs on Linux CI, names its skip on stock Windows). A sweep over all remaining raw write/delete sites in `scripts/` found no other member of this class.
+- **The PR body's tool-less-sandbox sentence no longer carries per-version numbers.** It named "23 / 28 skipped" — true when written (1.7.4), stale twice since, because every release adds tests and the skip count legitimately drifts with the environment. The body now states the invariant ("every test that needs a tool skips with its stated reason — nothing pretends to pass") and points at the CHANGELOG for per-version measurements.
+- Hygiene: the ancestor-link canary file in `safe-paths` is removed after the test instead of lingering in the shared temp root.
+- Tests **241 → 242 in fourteen files**, all green in the development tree and both published trees (the two file-symlink canaries name their skip where Windows lacks Developer Mode).
+
+## 1.7.5 — 2026-09-19
+
+**Third maintainer review, current head: all four code findings fixed**
+
+- **`safeRel` no longer accepts a project-local link to the project's own ancestor.** The "resolved path may sit above the root" fallback — meant for a not-yet-existing project root — also matched a link inside an existing project pointing at its parent, so a read path under the link resolved *outside* the project (reproduced live: a write through `safeRel(root, 'link/new.txt')` landed in the parent directory). The fallback now applies **only while the root itself does not exist**; once it exists, the resolved path must be the root or beneath it. Regression case in `safe-paths` covers both the new-target (the hole) and existing-target shapes.
+- **All three `preview-page` leaf writes are contained.** The output directory was jailed, but `01.html`, `01.nofx.html` and `index.html` inside it were written with plain `path.join` — a pre-planted file symlink redirected the write outside the project. All three go through `safeOut` now, with a file-symlink canary (file links need Windows Developer Mode, so the canary names its skip there and really runs on Linux CI).
+- **`asr.mjs --out` is contained and refuses to clobber.** It previously wrote any path raw and silently overwrote. It now accepts only a project-relative path inside the project (absolute paths rejected with a named error), refuses an existing output without `--force`, and — so the refusals need no network and are testable on bare CI — the checks run *before* the transcription request.
+- **`prep-image --crop` is now a declared exception, not an inconsistency.** Its input and output are given explicitly on the command line (the same trust level as running ffmpeg yourself), so both READMEs carve it out of the write-boundary promise explicitly (overwriting still needs `--force`); every *derived* path remains fully contained.
+- The review also confirmed the earlier rounds' fixes on this head, and noted the fork's CI runs sit at `action_required` — they execute only after a maintainer approves them; the scoped workflow is ready to install ffmpeg + Chromium and run the full suite on approval.
+- Tests **238 → 241 in fourteen files**, all green with 0 skips in the development tree and in both published trees.
+
+## 1.7.4 — 2026-09-18
+
+**Eighth-round audit: no serious findings — one stale evidence line corrected, three small hardenings**
+
+- **The PR body's sandbox sentence had gone stale**: it still said both tool-less sandbox shapes report "0 fail and 23 skipped", which the 1.7.3 evidence itself supersedes (the links-not-creatable shape reports 28 skipped). The body now carries both shape numbers.
+- **`capture` now validates `script.width`/`script.height`** the same way `preview-page` and `build-video` already do (integer, 16–16384) — previously a non-integer died with a raw Playwright stack deep in the browser launch, and an out-of-range value silently produced off-canvas frames. The gate sits **before** the Playwright load, so the new regression case runs on tool-less CI as well.
+- Test-infra hygiene: the two remaining string-concatenated `file://` imports in the test helpers now go through `pathToFileURL` (a checkout path containing `#` or `%` used to break them loudly), and the suite's temp directories are removed on process exit instead of accumulating in `%TEMP%`.
+- Disclosed, not changed: `preview-page`'s `?s=k` step-stepping silently disables itself when a hand-written slide has no `<head>` for the script to attach to (generated slides always have one).
+- Tests **237 → 238 in fourteen files**, all green with 0 skips in the development tree and in both published trees; both tool-less sandbox shapes 0 fail (23 and 28 skipped, reasons stated).
+
+## 1.7.3 — 2026-09-18
+
+**Fifth-round audit: five containment tests could pass vacuously where links cannot be created**
+
+- Five tests that build a symlink/junction to exercise the path jail (`safeRel` escape, `safeRel` dangling link, unit-level `safeOut`, `init-project --upgrade-css` containment, `init-project` skeleton containment) ended with a bare early `return` when the link could not be created — green, but testing nothing, which contradicts the README's "nothing pretends to pass" promise (the seventh-round review caught it; two of the five were pointed out, a repo-wide grep found all five). They now skip **with a named reason**, and the `safeRel` escape test gained the junction fallback so it really runs on stock Windows instead of skipping there.
+- Verified in both tool-less sandbox shapes: links creatable → 237 tests, 0 fail, 23 skipped; links not creatable → 209 pass, 0 fail, 28 skipped — the five formerly vacuous passes are now the five extra named skips, and no shape reports a pass it did not earn.
+- Small dedups from the same pass: `build-video` reads `--transition` through the shared `flagValue` instead of an inline reimplementation; the e2e image-listing fixture builds its `file://` URL with `pathToFileURL` (temp paths with spaces/non-ASCII). Disclosed, not changed: the `--get` download loop re-walks redirects without self-loop detection — it is bounded by `MAX_REDIRECTS`, so this is redundancy, not a hole; the test-side Playwright probe runs one synchronous `npm root -g` at discovery.
+
+## 1.7.2 — 2026-09-18
+
+**Final audit round: one silent feature regression, one environment-dependent test, and a corrected evidence claim**
+
+- **`--json` never printed JSON in its documented position.** The flag is boolean, but it was read through the value-taking reader, so in the documented trailing position (`fetch-official-images.mjs <url> --json`) it resolved to `undefined` and the listing silently degraded to the human-readable form. It is now read by presence, and a new end-to-end case drives a local `file://` fixture page through the real listing path with the flag in **both** positions, asserting actual JSON array output (the pre-existing case only asserted that the run reached URL validation, so it could not see this).
+- **The two symlink-canary cases could redden a tool-less CI.** They had no Playwright guard, and both `capture` and `preview-page` load Playwright *before* their containment check: on a runner where symlinks/junctions are creatable but Playwright is absent (an ubuntu CI), `capture` exits `未找到 playwright` and the case failed on its expected-message assertion — the same class of host-CI reddening fixed in 1.7.1, in a different environment shape. Both cases now skip with a stated reason when Playwright is missing.
+- **Evidence correction to 1.7.1.** Its "clean sandbox: 0 fail, 22 skipped" was measured in a Windows sandbox where junction creation itself failed (the shell was unreachable on the stripped PATH), which masked exactly the case above. Corrected measurements, after the fixes, in two isolated tool-less sandbox shapes — symlinks creatable and not creatable — agree: **237 tests, 0 fail, 23 skipped**, every skip naming its reason.
+- Tests **236 → 237 in fourteen files**, all green with 0 skips in the development tree and in both published trees.
+
+## 1.7.1 — 2026-09-18
+
+**Fourth pass over the *published* PR: three claims that were not true of the published tree, and the one test that could redden the host CI**
+
+- **The 1.7.0 CI rewrite had never been pushed.** The PR branch still carried the old scoped workflow — explicit file enumeration (so `cover-transition`, `review-round2`, `review-round3` and `subtitles-invalidate` never ran in any CI), no `permissions:` block, floating `@v4` action tags — while the 1.7.0 entry above already described the corrected form. The corrected workflow (shell glob `tests/*.test.mjs`, `permissions: contents: read`, full-SHA pins with version comments) is what the branch carries now.
+- **The host monorepo's own CI would have gone red.** `npm run check` = validate + a root-level `node --test`, which auto-discovers this plugin's tests — on runners that have neither ffmpeg nor Playwright. Exactly one case lacked the tool guard every other rendering case has (`asr.mjs` exits with `找不到 ffmpeg` at startup, before any request is made): the ASR-503 case now skips with its stated reason like the rest. Verified in a clean sandbox with neither tool resolvable: **236 tests, 0 fail, 22 skipped with stated reasons** — and the pipeline scripts themselves still fail loudly rather than pretend to run.
+- **The 1.7.0 README sentence "the rendering suites fail rather than skip" was itself wrong** — it rested on an observation made where `ffmpeg-static` was silently discoverable through a workspace-level `node_modules`. Both READMEs now state what actually happens: tests that need a tool skip with their stated reason; the scripts exit with `找不到 ffmpeg` instead of running half a pipeline.
+- **Plan B2's third manifest check is implemented.** `build-video` now also compares the subtitle manifest's `framesCover` with the current frame sequence's real duration (it previously checked only the slide duration and the clause indices) — a re-captured animation window no longer splices subtitle stills at stale offsets. New pixel-level case in `subtitles-invalidate`: a stale manifest is ignored with a named warning (measured luma difference 45 against the stale still — it is not in the picture), an aligned one splices it in (difference 0).
+- Smaller findings from the same pass: `VALUE_FLAGS` no longer lists `--speed` / `--voice` (nothing parses them, and a stray `--speed` silently swallowed the next positional), the six per-script `flag()` helpers collapsed into the one `flagValue` in `tools.mjs`, and the internal `KIT_PROJECT_DIR` variable is now documented in both READMEs.
+- Tests **235 → 236 in fourteen files**, all green with 0 skips in the development tree and in both published trees.
+
+## 1.7.0 — 2026-09-18
+
+**Third review round: the CSS-propagation rewrite, stale-CSS blocking, and six silent-failure paths closed**
+
+**CSS propagation is now a single managed region (`tokens`), and staleness blocks rendering.**
+The skill writes its generated CSS into the project's `slides/tokens.css`. The old mechanism delimited only the three toolkit blocks (no-fx / charts / tables) and decided "up to date" by looking at *the one block it recognised*, which left four paths that reported `ok` while the project still rendered old rules: ① the same toolkit present **twice** (the later copy wins the cascade and was never examined); ② a stale copy **outside** the managed blocks (written by an older, append-style upgrade, or pasted by hand); ③ a **mangled closing delimiter** (the regex no longer recognises the block, so the old existence probe reported `ok`, and the upgrade emitted a *second* opening marker); ④ drift in the **generated body itself** (theme tokens, `--fs-*` scale, fx keyframes/utilities, `.fx-stagger`) was not managed at all. Now the whole generated body is wrapped in one `tokens` region with the three toolkit blocks nested inside, and the status vocabulary is explicit: `ok` / `stale` / `duplicate` / `broken` / `legacy-outside` / `missing`. `--upgrade-css` replaces the region in place, merges duplicates, removes orphan markers, and cleans leftover full copies of a toolkit outside the region — **rule by rule**, so a single project override sitting next to a leftover copy survives (only the verbatim module rules are removed), with the original text in `tokens.css.bak`. A legacy file that is still the old "bare text" format but byte-identical to the current version reports `ok` and is wrapped in place on upgrade rather than duplicated. An adversarial re-review of the new mechanism found **two more silent paths inside it**: a current region **plus a stale copy of the generated body outside it** still reported `ok` (the outside scan only recognised toolkit selectors, not body-shaped rules — it now classifies those too and reports `legacy-outside`), and the `ok` note for a genuine project override did not say **where** the override sits, which is what decides whether it wins: before the region it loses the cascade, after it wins. The note now names the position and which of the two applies. **`check-slides`, `capture` and `build-video` now refuse to run while the region is stale** (exit 1 with the issue list and the `--upgrade-css` command). The gate at the two rendering entry points has an explicit escape — `--allow-stale-css` downgrades it to a warning — while `check-slides` is a static gate and stays blocking by design (it is the check you run *before* rendering; `SKILL.md` records that it has no such flag). The play page's fallback injection states plainly that a correct preview page does **not** mean the rendered video is current. Measured along the way: an ordering bug of my own — the leftover-copy cleanup ran *before* the region step — deleted the toolkit text out of a bare-text body, so the in-place match missed and the upgrade both duplicated the body and left the old rules **winning the cascade**; the regression test asserts the region lands *after* legacy content (later declaration wins) and that the render actually changes, pixel-level (a stale project's frame is measurably darker, and after the upgrade it returns to the skill's current picture within 0.05 luma).
+
+**A two-axis review of this very change set** (standards vs. spec, run against the published 1.6.0 tree) found a further batch, all closed here: `build-video` discovered its tools **before** validating the configuration, so on a machine without ffmpeg a bad `width`/`fps`/`bgm` was reported as "ffmpeg not found" (exit 2) and the real error was swallowed — tool discovery now happens after every config check; `requireFreshCss` was the one entry point that read `tokens.css` **without** the 2 MB scan cap the other three enforce (it now refuses an oversized file on the same rule); `init-project` contained only its `--upgrade-css` path — the **skeleton** writes (8 directories + 5 files) were still raw `path.join`, so a pre-existing `assets/` or `slides/` link pointing outside the project received the generated files (now contained, with the project root created first so the jail has a real root to compare against); `fetch-official-images` contained each downloaded **file** but created its output **directory** unchecked; `readTransition` now names the source it rejected (`--transition` vs `script.transition`); two `generateTokensCss()` re-assemblies collapsed into one shared `expectedTokens()` used by `check-slides` and the play page; and the dead `needsUpgrade` export was removed. Doc-vs-code drift fixed in the same pass: the READMEs still promised "three rev-stamped blocks … rules outside the blocks are untouched" (the pre-1.7.0 contract, and the opposite of what the leftover cleanup does), the README module list named six of eight modules and kept the old mechanism's name, the `init-project` and `check-slides` headers still described their pre-1.7.0 behaviour (`check-slides`' staleness item is error level, and it deliberately has no `--allow-stale-css`), `capture`'s usage line said `--dsf 1|2` and omitted two flags, and the README's "where those are missing they skip with a stated reason" was **false** for the rendering suites — without ffmpeg they fail, so the sentence now says so (only the cases with a declared reason skip).
+
+**The plugin validation entry point no longer validates a stale copy.** The host's `npm run check` scans `_official-plugins/plugins/**`, so validating while the staged copy was one release behind produced a **green result that proved nothing** — which is exactly how the 1024-character `description` limit (`validation.mjs:82`) was missed twice. New `publish/validate-plugin.mjs`: stages the plugin tree into the host checkout, **fingerprints every file** (sha256) and refuses to continue if the staged copy differs from the source tree, then runs the upstream validator and prints its exit code. The skill's `description` is now 977 characters, with the symptom list extracted to `references/symptoms.md` and a pointer to it from `SKILL.md`.
+
+**Second class of output escape closed (a directory inside the project can itself be a symlink).** `tools.mjs`'s path jail walks ancestors with `lstat` now (a **dangling** symlink is detected instead of being treated as "does not exist" and skipped) and **fails closed** when `realpath` cannot be resolved (it used to return the absolute path and let the write through). Output containment (`safeOut`) is applied to every remaining write site — `plan-timings`, `check-timing`, `build-video` (including leaf files `concat.txt`, `audio-timeline.wav`, `audio-mix.wav`, `final.mp4`, `asr/checklist.md`), `asr`, `capture`, `fetch-official-images`, `init-project` — with a unit test asserting "a junction parent pointing outside the project is refused".
+
+**Subtitle stills are invalidated together with the subtitles.** `capture` deletes `build/substills/<id>/` and its manifest before writing new ones, and `build-video` validates the manifest against the current slide (`duration`, `stills[].k < clauses.length`, `framesCover`) and ignores it with a named warning if it does not match — otherwise deleting clauses (or `--no-subs`) and rebuilding spliced the *old* subtitles back into the video.
+
+**SSRF policy completed.** Blocked ranges now cover `0/8`, `10/8`, `100.64/10`, `127/8`, `169.254/16`, `172.16/12`, `192.0.0/24`, `192.0.2/24`, `192.168/16`, `198.18/15`, `198.51.100/24`, `203.0.113/24`, `224/4`, `240/4`, plus `fec0::/10` and `2001:db8::/32` on the IPv6 side; a trailing-dot FQDN (`localhost.`) is normalised before the check, embedded IPv4 addresses inside `::ffff:` **NAT64 `64:ff9b::/96`** and **6to4 `2002::/16`** are decoded recursively, and a dead `localhost` branch was removed.
+
+**Seven functional defects (each one a documented way to get a wrong artefact or a misleading exit code).**
+- `check-timing --calibrate` could write back a **mixture** of exact-punctuation-aligned and nearest-neighbour boundaries while reporting `method: "nearest"` (the nearest pass reused the array the failed exact pass had already partly filled). The two passes now allocate separately, `nearest` is only reported when **every** boundary was measured, `sparse` is display-only, and a run that calibrated nothing no longer rewrites the file at all.
+- Positional-argument parsing: `init-project.mjs --topic 主题 <project>` treated the topic's **value** as the project directory and silently created a skeleton named after the topic in the current directory. Every script that takes a positional project directory now resolves it through `tools.positionalDir`, which skips the values of value-taking flags and shares one `VALUE_FLAGS` list with the rest of the module (`flagValue` is exported for callers that also need a flag's value, with the same list behind it).
+- `fetch-official-images`: `--json` was listed as a value-taking flag, so `--json <page URL>` swallowed the URL and reported a usage error (position-independent now); `--min` accepts `WxH` (`800x600`, `x600`) instead of comparing width only; and a `--get` run where **every** download failed exits 1 instead of 0.
+- Argument validation at the entry point: `capture --mode` accepts only `still|motion`, `--dsf` only the integers 1–4, `grab-frames --at` only values in `(0,1]` (a typo used to reach Playwright as `deviceScaleFactor: NaN` or be silently clamped to the first frame); `grab-frames` exits 1 when it captured zero frames, and a missing ffmpeg is exit 2 (the convention `tools.requireTool` already had).
+- The subtitle-band gate matched `padding-bottom`/`margin-bottom` (no word boundary) and flagged elements that are not positioned at all. It now requires `position: absolute|fixed` in the same rule, handles both `<style>` blocks and inline `style` attributes, and stays a warning level.
+- Slide transitions are parsed in exactly one place (`tools.readTransition`). Before, `check-slides` accepted the bare string `"xfade"` while the renderer only looked at `object.type`, so the gate could report a cross-dissolve that the video rendered as a hard cut.
+- `build-video --dry-run` exited 1: the self-check ran even though a dry run encodes nothing and writes no artefacts, so "show me the plan" — the way an agent verifies the cut/dissolve decision and the segment list before spending minutes on an encode — always looked like a failure. It now prints the plan and exits 0. (The test for the transition fix above is what hit this, since it inspects the plan rather than encoding.)
+
+**CI and packaging.** The scoped workflow runs `node --test tests/*.test.mjs` (a shell glob — it used to enumerate files, and had already missed eight of them), has `permissions: contents: read`, and pins `actions/checkout` / `actions/setup-node` to full SHAs with version comments (the pins were resolved through the API; note the host repository's own `checkout` pin in `ci.yml` is one hex character too long, which is why it cannot be copied verbatim). The plugin-root `README.md` / `README.zh-CN.md` / `THIRD-PARTY-NOTICES.md` were synchronised with the standalone repository (they still described the removed `R`-replay key and 146 tests in ten files), the test count is stated once (**235 tests in fourteen files**) wherever it appears, and the small doc drifts were fixed: `.fx-stagger`'s baseline is `var(--t2)`, `--dsf` defaults to 1, `check-theme` checks **five** contrast pairs, the `evals.json` play-page expectation no longer asserts the removed replay key, the 2 MB scan cap moved to `limits.mjs` (it applies to slide HTML as well as tokens.css) and the `inside()` doc comment now matches its implementation.
+
+- Tests **174 → 235 in fourteen files** (`review-round3` adds the D1–D6 regressions, the CLI value-flag guard, the junction-containment cases (`--upgrade-css` **and** the skeleton writes) and the "one transition parser on both ends" case; `subtitles-invalidate` is new and asserts at pixel level that deleting clauses — or `--no-subs` — and rebuilding removes the old subtitles from the video; `css-kit` grows from 16 to 29 cases covering duplicate blocks, leftover copies, broken delimiters, whole-region replacement and the two false-green paths above; `safe-paths` gains the dangling-symlink case **and a unit-level `safeOut` assertion** that calls the jail directly (an integration case can pass for the wrong reason if another check exits first); `render-smoke` gains a pixel-level "the upgrade really changes the picture" proof), all green with 0 skips in the development tree and in both published trees.
+
+## 1.6.0 — 2026-09-18
+
+**Cover / slide transitions / speech-rate alignment (three field reports from real deliveries)**
+
+- **Cover and a non-black first frame.** Every segment carried `fade=t=in`, so frame 0 of the final video was pure black — the thumbnail people saw when the file was forwarded. Fix: `capture.mjs` now also writes `preview/cover.png` for the first slide (base animations finished, subtitles hidden — the fully revealed design), the first segment dissolves from that cover over 0.25 s (**frame 0 *is* the cover**, no added duration), `build-video.mjs` embeds it as `attached_pic` (main video stream still `-c:v copy`; note `-shortest` must be dropped when a cover is embedded) and exports `out/cover.png` at canvas size for platform uploads.
+- **No more black frames between slides.** Each segment faded to black (0.3 s) and the next faded in from black (0.25 s) ≈ 0.55 s of black at every page change. Now: hard **cut** by default (only the last segment keeps its outro fade; the first gets the cover dissolve above), plus an optional **cross-dissolve** (`script.json` → `"transition": {"type":"xfade","duration":0.4}`, or `--transition xfade`). The dissolve never passes through black: each segment holds `duration` extra tail frames, which the xfade consumes, so the total duration still equals `timings.total`. The play page compares the two live (bottom-bar toggle / `T`).
+- **Speech rate: default 1.1, and the audition now asks.** Default was `speed 1.0` (≈4.8 Chinese chars/sec) — field feedback: "the default pace is a bit slow". Default is now **1.1** (≈5.3 chars/sec; first/last slides 1.05) across SKILL.md, `tts-and-timing.md`, `authoring.md`, the project template and evals. More importantly the **audition step asks two questions at once**: the three candidate voices *and* the same probe line in the chosen voice at three speeds (1.05 / 1.15 / 1.25); the answer is written to `script.json` (`speed`), and `plan-timings.mjs` now **reads that field** (it used to be declarative only) and warns when the measured rate deviates >20% from `baseline × speed`, or when speed is outside 0.8–1.4.
+- **Gate 4 now simulates the preview and asks two more questions.** After the HTML is done: ① slide transition (cut / dissolve / custom duration — comparable live in the play page) ② entrance layering (title + sub-info all at once, or staged as `data-stage` does today). `check-slides.mjs` prints the current transition value and each slide's level count so the agent asks with real numbers instead of impressions. To collapse staging, set the layers to the same stage and re-run `capture --mode motion` for that slide.
+- **Self-check gained three visual gates** (all fail the run non-zero): cover embedded, first-frame luma above black level, and no black frame around any cut point.
+- Tests **171 → 174** (new `cover-transition.test.mjs`: cover present + `attached_pic` stream + non-black first frame + no black at cuts in both cut and xfade modes + duration unchanged + a single-level slide still renders), all green with 0 skips.
+
+**Documentation is now English.** SKILL.md, all six references, CHANGELOG and THIRD-PARTY-NOTICES are written in English so the skill is usable outside Chinese-language teams; `README.zh-CN.md` stays Chinese. Source examples were internationalised at the same time: the announcement/media screenshot path and the source tiering now give **both** international (Reuters, AP, Bloomberg, FT, WSJ, BBC, CNBC) and Chinese (Xinhua, CCTV, The Paper, Caixin, Tencent News, Sina, NetEase) examples, ranked by primary-ness and editorial accountability rather than by country; the regulated-topic guide names jurisdictions (SEC/FTC/FDA, FCA/ESMA/EMA, CSRC/NMPA/SAMR) and states the up/down colour convention per market instead of assuming one.
+
+## 1.5.0 — 2026-09-18
+
+**Round-2 review fixes: output containment, the network boundary at the real request point, the subtitle timeline (3 P1 + 4 P2)**
+
+- **[P1] Subtitle timeline (a finished-video defect reproduced by measurement)**: subtitles are "full-duration percentage animations", while the frame sequence only covers the animation window and the rest was frozen on the last frame by `tpad` — so **every subtitle change after the animation window never made it into the picture** (measured: SRT clause 4 sits at 5.32s, yet the frame at 6.0s still shows clause 2); on the `no-fx`/static path it is worse — `finish()` jumps the subtitle animation to `opacity:0`, so the whole video carries not one subtitle frame. capture now emits, per clause, a "final-state base + that clause's subtitle" still (`build/substills/<id>/`, with a manifest `build/substills/<id>.json` recording `framesCover` and every clause's window), and build-video uses `concat` to stitch "frame-sequence segments + the subtitle segments" into the whole slide (the segment lengths must sum to the slide duration, backed by the existing self-check). The regression test uses a **pixel-level criterion**: inside clause 2's window the finished video must match clause 2's subtitle still (measured 0.93) and stay far from clause 1's (7.0); under the old behaviour those two numbers invert exactly (0.99 / 7.1) — the test goes red.
+- **[P1] Output containment (the second class of escape)**: only input paths used to be contained; derived output (`preview/`, `build/frames/<id>/`) is "a fixed position inside the project + a validated id", but **a directory segment inside the project can itself be a symlink pointing outside it**, and `rmSync(recursive)` deletes straight through it while the command reports success. New `tools.assertContained/safeOut` (the realpath of every existing ancestor segment must still sit inside the root's realpath), wired into every write/delete site in capture (frame-directory recursive delete, preview writes), build-video (out/build/asr writes and cleanup), preview-page, plan-timings and grab-frames. The canary regression test (on Windows the link is made with a junction, no administrator needed) asserts "the command must refuse + the canary is intact".
+- **[P1] The network boundary now sits at the real request point**: previously only "the one URL supplied at the start" was validated. Now `url-policy.assertResolvedHost` resolves DNS and then validates the IP for **every real request** (page navigation, the 30x redirects Chromium follows on its own, browser subresources, per-hop image downloads); on the browser side it hooks `context.route` with a per-hostname resolution cache. The end-to-end test asserts "a refused target must receive 0 requests", and an injectable fake resolver covers rebinding scenarios (no real DNS needed).
+- **[P2] ASR request errors count as failures**: authentication / rate-limit / network failures used to land in `results` without entering the tally, and `1 通过 / 0 不通过 / 退出 0` slipped the gate through. They are now counted separately, excluded from the pass count, and the whole run exits non-zero naming the affected segments.
+- **[P2] The play page's attribute parsing accepts three quoting styles**: `<html class='theme' style='--t2:800ms'>` (single quotes) and unquoted attributes were not recognised, so a second, duplicate set of attributes was injected — and the browser keeps **the group that appears first** → the measured delay, the canvas size and `no-fx` all failed. The attribute is now rewritten in place and no duplicates are produced.
+- **[P2] ffmpeg discovery gained cwd and the project directory**: the README states that installing `ffmpeg-static`/`ffprobe-static` inside the video project is supported, but `tools.mjs` never looked at `process.cwd()`, and `prep-image.mjs`/`asr.mjs` never passed the project directory → "it is installed in the project, yet it reports not found". Both are now passed, with a source-level assertion against regressions.
+- **[P2] `--out-dir` compares canonical paths**: on macOS `/var/...` and `/private/var/...` are the same directory, and a plain string comparison misread a legitimate in-project path as an escape (hit for real on the official CI's macOS environment); the check now takes the realpath of the deepest existing ancestor before comparing.
+
+**Audit fixes folded into the same batch (after 1.4.0, never released on their own)**
+
+- `preview-page`: a 2MB cap on tokens.css (the same gate as check-slides/init-project — a 6MB input once ran 129s in measurement), numeric validation of `script.width/height` (a string written by mistake kills the play page's script outright), no-ops at the first/last-slide boundaries (pressing → on the last slide no longer resets it to level 1 and replays it), falling back to whole-slide paging when the filename contains `?`/`#` (the stepping query string breaks), and slides in subdirectories are skipped with an explicit explanation.
+- `build-video`: `bgm.volume/fadeIn/fadeOut` and `fps` are now all numerically validated (the same injection surface as `width/height` — previously only W/H was covered); validation moved ahead of encoding (a bad config fails at once instead of after minutes of rendering).
+- `asr`: `String.replace` now uses a function replacement string (a `$&`/`$1` inside the transcript would corrupt the checklist).
+- **Root fix in the template**: all nine `.fx-stagger` rules gained `:not([data-stage])` — the `nth-child` delay's specificity (0,2,0) is higher than the fx class's (0,1,0), and it silently overrode the entrance moment of `data-stage` children inside the container (measured: they popped up 3 seconds early).
+- `check-slides` gained two gates: **5d** `.fx-stagger` and `data-stage` coexisting on one slide (a prompt to confirm the layout), and **5e** an absolutely positioned `bottom` falling inside the subtitle band (the most expensive class in measurement: invisible in the still preview, and in the finished video the subtitles sit on the figure caption).
+- New **`grab-frames.mjs`** (the 12th CLI): computes each slide's absolute start from `timings.json` and pulls frames from `out/final.mp4` to check them — the final reconciliation tool for the class of problem "a still cannot show it, only the finished video reveals it".
+- Tests **151 → 171** (new `review-round2.test.mjs` + two end-to-end subtitle-timeline cases + a grab-frames smoke test), all green with 0 skips.
+
+## 1.4.0 — 2026-09-18
+
+**Play page interaction model reworked: animations on = manual level-by-level entrance**
+
+- Measured feedback: with animations on, ←/→ could only turn whole slides, and second-level sub-headings or small charts had to be waited for as they appeared on schedule — that does not match the feel. Now **with animations on, → reveals the next level first** (the next level animates in on the spot, and only once it is out does the page turn; ← steps back level by level, and stepping back does not replay animation but goes straight to the final state); with animations off it still turns whole slides.
+- **How level stepping works**: a script injected into the snapshot copy's `<head>` runs before the CSS animations start and, given `?s=k`, resets each level's `--tN` delay — passed levels at `-60s` (final state), the current level at `0ms` (enters on the spot), future levels at `+60s` (stay hidden until the key is pressed); the parent page only swaps the iframe src, the same mechanism as the snapshot, with no cross-document access (under file:// the iframe is a separate origin). The top bar's page number shows the current level (e.g. `1 / 8 · 级 2/3` — page 1 of 8, level 2 of 3).
+- **Replay button and the `R` key removed** (under level stepping, replay has no meaning); the touch bar's forward/back arrows enlarged (76px wide, 24px type — measured feedback: too small).
+- **Frame switching no longer flashes white** (measured feedback: a big slide change flashed a white screen): now double-buffered — two stacked iframes, the new address is fed to the hidden one first, and only after `load` plus two more rAF frames (the new document's first frame has painted) do they swap for display, so the old picture is always on screen; level stepping benefits from the same.
+- **Bilingual UI** (measured feedback: an English project got a whole set of Chinese buttons): buttons / hints / panel titles / placeholder copy adapt between Chinese and English from `script.json`'s `lang` (`en*` → Motion on / Narration on / Overview…), while terminal output stays Chinese (the agent's view).
+- **Toggles moved to the bottom bar, labelled with their state** (measured feedback: "动效/口播" gives no clue whether it is on or off, and the top-bar toggle is redundant): the bottom-bar buttons read "动效开/动效关" "口播开/口播关" (motion on / motion off, narration on / narration off — the off state lit in a warning colour), shown on desktop too, and the top-bar animation toggle is removed.
+- Tests +5 (the toggle labels and the top-bar removal / the double-buffer structure and frame-wait order / **behaviour-level** cases: at `?s=1` level 2 must still be waiting, at `?s=2&anim=1` it enters on the spot, stepping back goes straight to the final state / an English page's copy contains no Chinese; a failed assertion also closes the browser in `finally` — otherwise the event loop keeps the runner alive forever, hit in measurement), **151 tests** in total. The play-page instructions in SKILL.md / render.md were kept in sync; also verified live through Playwright (in steady state the double buffer shows exactly one frame, the X/P toggles flip, the en page's copy).
+
+## 1.3.7 — 2026-09-18
+
+**Kickoff alignment checklist reordered + Gate 3 gained the "pick assets / add assets" questions**
+
+- The must-ask kickoff checklist was reordered into 11 rows per the user's measured feedback: **new "Audience"** (domain professionals / general tech-savvy public / management briefing / client demo / internal training — it decides term density and tone of voice, and previously hid inside "Topic and domain" without a row of its own); **new "Sensitive content and disclaimer"** (the disclaimer question expanded from "regulated domains only" to any hit of financial opinions / medical / legal / policy / negative-event reporting / personal information, default yes); "Asset boundaries" renamed to "**Images and asset boundaries**" (ask whether images are wanted at all first, and for negative-event subjects add the announcement/news-screenshot question); "Voice" annotated "a rough direction is enough"; row titles completed with the "theme color" / "aspect ratio" wording.
+- **Gate 3 became three questions** (once asset collection is done and both the facts and candidate images are in hand): ① pick "which ones to use" slide by slide; ② **"is there anything to add"** — links / screenshots / logos / pictures / data files / a reference voiceover script can all be handed to the agent, and user-provided assets have the highest priority (images pass `prep-image --check` and are registered in MANIFEST as "source: user-provided", links go to disk per paths A/B/D); ③ confirm sources and licensing. Previously Gate 3 had only a blanket confirmation, leaving the user no way in to pick slide by slide or to add assets.
+
+## 1.3.6 — 2026-09-18
+
+**Kickoff alignment's "asset boundaries" gained the negative-event question**
+
+- Hit in measurement: for a negative-news video the agent asked about topic / charts / audience but never "do you want news or media-platform screenshots as illustrations" — path D (1.3.4) had been written into the image SOP, but was never wired into the must-ask kickoff checklist, so without the question nothing was left but pure typography. The "asset boundaries" row now carries a note: **for negative-event subjects (collapse / penalty / lawsuit / controversy) one more question is mandatory — "do you want announcement or news-report screenshots as illustrations"** (default suggestion: yes, 1–2; source tiering and screenshot discipline in image-sources.md path D).
+
+## 1.3.5 — 2026-09-18
+
+**Iron rule 4 tightened: kickoff alignment must be "ask, then wait for the answer" — never a packaged announcement of defaults**
+
+- Hit in measurement: inside mcode the agent turned the must-ask checklist into "one message firing out a default block and then pressing on" — the message read 『默认项(不特别说的话就按这个):中文普通话 · 中文字幕 · 1920×1080 · 温润男声 · 10–12 张…』 ("Defaults (unless you say otherwise, these apply): Mandarin Chinese · Chinese subtitles · 1920×1080 · warm male voice · 10–12 slides…") — the user never got the chance to take a position on each item, which amounts to no alignment at all. The root cause was the old wording "(items the user does not answer may take the default, but you **must ask first**)", which left room to read it as "inform first; no objection means consent".
+- The rule now reads: the first response = **ask each question, one by one, in question form, then stop and wait for the answers**; each item may carry a recommended default, but it must be "ask, then wait"; only items the user explicitly answered "go with the defaults" or left unanswered take the default, and before moving to the next step you restate "which defaults were finally adopted". The verbatim counter-example from the field was written into SKILL.md as a prohibited case.
+
+## 1.3.4 — 2026-09-18
+
+**Image SOP gained "path D · announcement and authoritative-coverage screenshots" (negative-event subjects)**
+
+- For collapse / regulatory penalty / lawsuit / scandal subjects the company's official site can never supply the assets (paths A/B are guaranteed to fail) — and the SOP did not cover this real path. New path D: assets come from **the primary carrier of the event itself**, tiered as "exchange filings / regulatory documents > company official statements > media coverage"; **top-tier authorities (Xinhua / CCTV / The Paper) and mainstream portals (Tencent News / Sina / NetEase / Phoenix and the like) are all qualified sources**, and the only things excluded are self-media, aggregator accounts and content farms.
+- Screenshot discipline: the frame must keep the outlet name and date (a screenshot with the source cropped away is unusable), close the cookie banner first, capture the article body only; save to `assets/` + register in MANIFEST with "source URL / publication date / reasonable reporting-based citation"; bring it into the frame with `.img-frame.contain` + `.img-cap` naming the source; negative financial subjects are equally bound by compliance.md. A minimal Playwright screenshot command is included (wait for lazy loading, then capture the article/main body area — the built-in browser's screenshot works the same way).
+- SKILL.md's Phase 3 priority chain and compliance self-check were kept in sync; the subject index gained "collapse / regulatory penalty / lawsuit / scandal → path D".
+
+## 1.3.3 — 2026-09-18
+
+**Workflow: Gate 1's presentation format gained the "table and animation" column**
+
+- Measured feedback: the table the user reviewed at Gate 1 previously held only "layout / frame / narration", and slides using table primitives or carrying entrance animations were left unexplained — the user had to imagine them, saw them move for the first time only at Gate 4, and dissatisfaction meant rework at the most expensive step before rendering. Gate 1's presentation format is now "# / layout / frame / **table and animation** / verbatim voiceover script", and **the table-and-animation column is mandatory**: table slides must state their shape (how many rows and columns, which primitive, which row/column is highlighted); animation slides must name the effect and the order (staggered fade-in item by item / a rolling number / bars growing one after another).
+
+## 1.3.2 — 2026-09-18
+
+**Workflow: kickoff alignment gained the mandatory "data and charts" row**
+
+- Iron rule 4's must-ask kickoff checklist had no entry point for charts: the chart toolkit (merged during 1.1–1.2) hid only in the Phase 1/4 documents, so when the subject contained data the agent did not necessarily think to ask. A new row "**Data and charts**" was added: mandatory when the subject contains key numbers / comparisons / proportions / trends — ① is there any data worth a slide of its own as a visualisation; ② do you want chart slides, and roughly how many (default suggestion: hard data + a suitable layout → 1–2 slides); ③ style preference (horizontal bars / columns / donut / line / progress bar, or let the agent pick per the data shape); ④ numbers on the chart are equally bound by Gate 0 (≥2 independent sources, with scope and point in time). **Even if the user does not raise charts, data-dense subjects must be offered this option proactively**.
+- Phase 1's content-volume hard rules gained "**for data-dense subjects, proactively make charts**": when the voiceover contains ≥2 comparable numbers / proportions / trends, prefer the `data-viz` layout + chart toolkit over listing the numbers as bullets; bar heights / bar lengths are computed from the values (the nine disciplines).
+
+## 1.3.1 — 2026-09-18
+
+**Docs: which environment owns each research tool, and how open the toolset is**
+
+- The two-toolchain comparison table's **research row** was missing mcode's built-in browser (the mcode cell held only `web_search`/`web_fetch`, while the image row did list the built-in browser — the same table, two different standards): the mcode cell now adds "for SPA/JS-rendered pages, open them in the **built-in browser** to read the body text", and the other cell adds "for SPA pages, render the body text with the project's Playwright **headless browser**". The image row's "same as above" became an explicit statement too (this environment uses `fetch-official-images`, which ships its own Playwright headless rendering).
+- Phase 0's "how to get the body text" now marks each environment's owner: the **mcode built-in browser** (a host capability of the mcode environment) vs the **Playwright headless browser** (any environment); research.md's search-tool table and SPA row were kept in sync.
+- Made explicit that **the tools are open, the discipline is not**: if other search skills/plugins are installed on this machine, or any headless browser, use whichever one can pull the body text (try it once to confirm result quality before depending on it); source tiering, two-source cross-checking and scope labelling lose not one rule. research.md's tool table gained the row "any environment | other search skills/plugins installed on this machine".
+
+## 1.3.0 — 2026-09-18
+
+**Three-axis audit fixes (doc alignment / security / process): one high-severity SSRF + the vertical play page + a batch of process breaks**
+
+- **Security · the hexadecimal IPv4-mapped IPv6 form slipped past the egress block (high severity, penetrated in measurement)**: the WHATWG URL parser normalises `http://[::ffff:127.0.0.1]/`'s hostname to `::ffff:7f00:1` before it reaches the check, while the old mapped-address handling only stripped the prefix and the remaining hex groups matched no block rule — loopback / private networks / cloud metadata (`::ffff:a9fe:a9fe`) could all slip past `fetch-official-images`' SSRF line this way (the end-to-end measurement reached a local server). The address is now converted back to dotted IPv4 before the check; the public mapped form (`::ffff:808:808` = 8.8.8.8) is still allowed through. `fetch-policy` gained negative cases that go through the `new URL()` normalisation path.
+- **The play page's iframe forces relative resolution**: when a slide filename contains `:` (e.g. `javascript:alert(1).html`, legal on Linux/macOS), `iframe.src` was treated as a URL scheme and executed same-origin in the play page; a `./` prefix is now always added.
+- **Vertical (1080×1920) play page fixed**: the stage `#fit/#frame` dimensions, the scale divisor and the overview thumbnail ratio were hard-coded to 1920×1080, so the vertical Gate 4 picture was cut off by nearly half; they are now all parameterised from `script.json`'s canvas. authoring.md gained the vertical subtitle-safe-area numbers (≈150–215px from the bottom, ≈787px wide centred, starting from `padding-bottom: 240px`).
+- **A still re-capture silently swallowed the animations**: after capture invalidates that slide's frame directory on the still path (to prevent stale-frame contamination), build-video's fallback to a static image used to ship with no warning at all; it now names the slide with a ⚠ and gives the command to re-run (`capture --mode motion --ids <id>`). render.md's "three-piece visual verification" and SKILL.md's Phase 5 were updated to state "after re-checking in still mode you must re-run motion for the same ids".
+- **The managed-block mechanism hardened**: ① content comparison is normalised to LF uniformly (a CRLF file checked out by git autocrlf used to false-positive all three blocks as "hand-edited"); ② the upgrade runs in two passes (replace everything in place by rev first, then re-assess the result and fill the gaps) — under a pathological layout, "nofx bare text nested inside an old chart block" no longer produces the self-inconsistency of "reported success last time, reported stale again this time"; ③ an existing `.bak` now warns explicitly before it is overwritten; ④ tokens.css / slide HTML over 2MB is refused a scan (crafted input makes managed-block location quadratic — 4MB measured at 39 seconds); ⑤ `--upgrade-css` ends its output by naming the re-run list for "generated preview/frames/out still show the old CSS".
+- **Process gaps filled (SKILL.md)**: the workflow gained the step "to resume an old project, run `--check-css` first" (including the Phase-detection markers and the post-upgrade re-run matrix); Phase 4's check sequence gained the conditional check-theme item (when tokens.css was changed, contrast problems no longer leak through to the most expensive Gate 5); Phase 6's delivery list gained `out/subs.srt` and `build/timings.json`; init-project's closing note had its Gate number corrected (final-state preview = Gate 4); Gate 2's acceptance artefacts no longer hard-code "8 audio segments"; check-slides 5c's `--check-css` pointer is scoped to the toolkit scenario.
+- **Doc alignment**: the plugin package's top-level THIRD-PARTY-NOTICES.md was resynced with the copy inside the skill (it had been left behind as an old version); the README's verification section now gives commands that hold in both repository contexts (monorepo / standalone skill repo); authoring.md's chart-discipline heading changed from "six" to "nine" and the order of items 8/9 was corrected; evals had 4 fixes (id24's animation wording updated to `fx-sweep`/`fx-grow-w`, id0's Gate number, renumbering after a gap, the missing `files` field filled in); SKILL.md's script table gained the flags that actually exist (`--topic`/`--ids`/`--pacing`/`--min`/`--json`/`--dsf`).
+- Tests **132 → 146** (fetch-policy +6, css-kit +4, preview-page +3, render-smoke +1), all green locally.
+
+## 1.2.0 — 2026-09-18
+
+**CSS toolkit managed blocks: the root fix for "you changed the CSS but the project does not pick it up"**
+
+- The three toolkits (`nofx-css.mjs` / `chart-css.mjs` / `table-css.mjs`) become **managed blocks carrying a content rev** in the project's `tokens.css` (`/* >>> html2video:<id> rev=<sha256> >>> */ … /* <<< html2video:<id> <<< */`), with the rev hashing the block's CSS content — change the content and the rev changes.
+- `init-project --upgrade-css` was rewritten from "presence probe + append" to **replacement in place by rev**: an old / hand-edited block → swapped for the new one in its original position; undelimited but current content → wrapped in place (position unchanged); missing → appended at the end of the file with a note that the old rules will be overridden by what follows. Rules outside the managed blocks (including project-side overrides) are left untouched, and `tokens.css.bak` is backed up before writing. This removes two classes of defect from the old implementation: after one patch it always reported "no upgrade needed", and later source-module changes never propagated; and the append-style upgrade crushed the user's overrides while leaving a whole duplicate block behind.
+- New `init-project --check-css` (inspect only, never modify: outdated / missing items reported one by one, ending with exit code 1); `check-slides` gained two info-level checks: tokens.css toolkit staleness (naming which block), and a `class` that exists in neither `tokens.css` nor the slide's own `<style>` (silently unstyled — the `.grid g4` / `<p class="dim">` class of trap).
+- `preview-page`'s fallback injection expanded from no-fx alone to all three toolkits: the no-fx rules are still injected only into the "motion off" copy, while the chart / table toolkits are injected into every copy (otherwise the old rules would render the new drawing approach broken), and the page and the terminal report the staleness honestly.
+- Fixed `.matrix`'s highlight column not covering the header row: `.matrix td.hi` → now also covers `.matrix th.hi` (`<th class="hi">` silently got no background, so the 「我们」 ("us") column looked only half lit).
+- Doc corrections: `authoring.md`'s layout-snippet notes (the layout skeleton classes come from each slide's inline `<style>`, not tokens.css), and the comparison snippet's undefined `<p class="dim">` class changed to `style="color:var(--muted)"`.
+- New `tests/css-kit.test.mjs` (12 cases: a block that is present but stale / hand-edited must be detected and repaired in place, project-side overrides are not crushed, legacy files are wrapped in place with duplicates removed, `--check-css` exit-code semantics), **132 tests** in total, all green.
+
+## 1.1.0 — 2026-09-18
+
+**Security boundary (answering the five Request changes from the PR #41 review)**
+
+- **Path containment**: `script.json` is an agent-editable file, and its `slides[].id/html/audio` and `bgm.file` used to be spliced straight into file paths. `tools.mjs` now adds `safeId / safeRel / inside / validateScriptPaths / validateTimingsIds`, and the six consuming scripts (capture / build-video / plan-timings / check-timing / check-slides / asr) validate immediately after `JSON.parse`: ids must match the allowlist `^[A-Za-z0-9_-]{1,64}$`, paths reject absolute forms, must resolve inside the project directory, and are re-checked against symlinks via realpath. Previously `id="../../victim"` could trigger a **recursive delete outside the project directory** (`capture.mjs`'s `rmSync(build/frames/<id>, {recursive:true})`).
+- **Clobber refusal**: `init-project.mjs` refuses an existing, non-empty directory outright (listing the 5 generated files that would be overwritten) and needs an explicit `--force`; `fetch-official-images.mjs`'s `--out-dir` defaults to inside the working directory and never overwrites existing files; `prep-image.mjs --crop` needs `--force` when the output exists. Also fixed `--topic` being inserted into the template HTML without escaping.
+- **ASR endpoint allowlist**: the API key is sent only to the official domains (`api.minimaxi.com` / `api.minimax.io`); any other `--base-url` / `MINIMAX_BASE_URL` is rejected outright, and a self-hosted gateway needs an explicit `--allow-any-endpoint` (which prints a prominent warning). A swapped environment variable could previously send the key to any endpoint at all.
+- **Image fetching SSRF tightened**: new `scripts/url-policy.mjs` (pure functions). It blocks loopback / link-local (including cloud metadata 169.254.169.254) / private networks / CGNAT / dotless hostnames; only http(s) is allowed, `file://` needs an explicit `--allow-file`; URLs carrying userinfo are banned; `maxRedirects:0` follows redirects manually and re-uses the same policy at **every hop**; responses are size-capped at 30MB by default (`--max-mb`); and the on-disk filename sanitiser gained Windows reserved names.
+- **Executable tests**: new `tests/` (node:test, zero dependencies, auto-discovered by `node --test` from the repository root → really executed by `npm run check`) — `safe-paths` (malicious ids/paths + canary intactness + symlink escapes), `no-clobber` (clobber refusal), `endpoint-allowlist` (allowlist refusals + positive evidence that a local fake server receives a fake Bearer key), `fetch-policy` (host/URL/redirect/filename, 44 cases), `render-smoke` (init → silent audio → plan-timings → check-slides → capture → build-video, the whole chain through to a finished video). The first round was 72 cases (later `preview-page` and `tokens-fx` were added, see below — 120 cases across 9 files now), all green locally. Also included: the scoped workflow `.github/workflows/html2video-for-mcode-smoke.yml` (path-filtered so it only runs when this plugin changes; it installs ffmpeg + playwright and really executes the suite, including the render smoke test).
+
+**Charts and motion effects**
+
+- `references/authoring.md` gained a **plain CSS/SVG charts** section: five drawing methods — horizontal bars / columns / donut (`conic-gradient`) / line (inline SVG) / progress bar — plus a selection cheat sheet; three bottom lines = no external chart libraries (unreachable offline, and canvas animation cannot be seeked frame by frame), numbers must come from a verified scope, and AI-generated images are forbidden as charts.
+- New `fx-grow-x` / `fx-grow-y` (bars growing from the left, columns from the bottom), and the line's `fx-draw` stroke draw-in; **motion can be switched off in one move**: add `no-fx` to the root element or any container, and motion capture automatically degrades to static frames with durations and audio-visual sync unchanged (measured: the animation's final state vs a `no-fx` frame at PSNR 51.7dB — the picture matches).
+- The `timeline` layout became a whole-axis fade-in plus the time points entering one by one, staggered.
+
+**Silent failures (more gates added)**
+
+- **fx keyframes without opacity → an element that stays invisible forever**: `[data-stage]`'s base state is `opacity:0` and only the animation lifts it back to 1; animations that only do transform/stroke (`fx-grow-x/y`, the existing `fx-draw`) would never become visible unless they touched opacity. The keyframes now carry `opacity:1`, and `check-slides.mjs` gained **item 5b**, a static block (an fx class whose keyframes do not declare opacity → ✗).
+- **A stale frame directory contaminating the finished video**: after switching to `no-fx` or to still mode, the previous motion run's frame directory was still there, and `build-video` preferred the leftover frames, mixing stale animation into the finished video; capture now actively invalidates that slide's frame directory on every path that produces a static image.
+- **The column-chart template's collapsed height**: writing `align-items:flex-end` on the outer container collapsed the column wrapper to its content height, turning the columns' percentage heights into 0 (silently not drawn); the template now defaults to stretch + `justify-content:flex-end` inside the column, and the trap is documented.
+
+**Play page (you can screen it once before rendering)**
+
+- New `scripts/preview-page.mjs <项目> [--open] [--no-script]` → `preview/play/index.html`: a single-file, zero-dependency play page you double-click over `file://`. **It does exactly one thing — "play the HTML picture through once"**: `←` `→` (or a touch swipe) turn pages, `R` replays the entrance animations, `X` compares motion on / motion off, `P` toggles the voiceover text, `O` is the overview, `F` fullscreen. **It deliberately has none of the player UI** — no timer, no progress bar, no sentence-by-sentence follow-along highlight: to see time or rhythm, watch the finished video; running a timer in the preview page only makes people stare at a stopwatch (measured: leave the tab open and it turns into `204.2s / 6.3s`).
+- **The voiceover UI loads or not based on the data**: clauses + `timings.json` → it lists that slide's voiceover text; clauses but not timing-synced yet → it lists the text only and marks it 「(未对时)」 ("not timing-synced"), so you can look at the HTML before the voiceover exists; no clauses, or `--no-script` → the panel and the voiceover button do not appear at all, and the picture fills the full width.
+- **Layout adapts to the window**: top/bottom bars wrap and overlong topic names get an ellipsis; in narrow windows and on phones the voiceover panel collapses into a bottom drawer, collapsed by default (picture first); phones get touch-bar buttons + swipe-to-turn pages; the overview grid picks its column count from the width, and the height uses `100dvh` (so the phone's address bar cannot cut it off).
+- Why not "just open `slides/*.html`": the delays `--t1/--t2/--t3` and the canvas size are injected by the render pipeline from `timings.json`, while tokens.css holds only placeholder values (`--t2:800ms`), so opening the original file shows "every animation crammed into the first two seconds". The play page generates a **snapshot copy** (`preview/play/<name>.html`), writes the measured delays into `<html style>` (equivalent to pipeline injection, highest precedence) and adds `<base href="../../slides/">` so themes and assets resolve as usual. Measured: on the same slide at t=3.0s the copy's second level is `opacity 0` (not yet entered) while the original file is `opacity 1` (already in) — the copy is what matches the finished video. **When there is no timing yet** it spaces the stages actually used in the HTML evenly (0.3/1.3/2.3s), and a yellow bar at the top of the page honestly marks it 「不是成片时序」 ("not final-video timing").
+- The copy is a snapshot: after editing `slides/` you must re-run; both the page and the copy's head state the real file path and the generation time.
+- When an old project has no `no-fx` rules, the rule is injected into the motion-off copy alone as a fallback and the page says so (otherwise switching over is blank, which gets misread as a "motion-off bug").
+- New `init-project.mjs --upgrade-css`: idempotently adds the new `no-fx` rules to an old project's `tokens.css` (touching only that one file, never `script.json` or anything else). The previous advice for every old project — "regenerate with the new init-project" — was wrong: that needs `--force` and resets `script.json`.
+- New `scripts/nofx-css.mjs`: the single source of truth for the `no-fx` rules (`init-project` writing them into tokens.css and `preview-page`'s fallback injection share one copy, so the CSS cannot drift).
+- `fx-spotlight`'s keyframes gained `opacity: 1` — it is listed as a usable entrance class in this skill's docs, but it only did `clip-path`, so the `opacity:0` base state never came back up → using it made the element permanently invisible (caught by `check-slides`'s item 5b: "the docs say it works, the gate says it doesn't").
+- SKILL.md's Gate 4 gained "hand the play page over for the user to present themselves"; `references/render.md` gained a play-page section (positioning, keys, the three voiceover states, why the copy is mandatory); `authoring.md`'s motion-switch section states that `X` must be used for the comparison before delivery.
+- Tests +27 (99 in total): `preview-page` (injection / merging / no-fx / base ordering / self-contained with no external links / **no timer** / **the word 「配音」 ("dubbing") must not appear** / responsive and touch, the three voiceover states, the equal-interval fallback, containment refusals / idempotent upgrade), `tokens-fx` (asserts of the template that **every non-infinite fx animation's keyframes declare opacity**, that the no-fx rules reset opacity, and that `--upgrade-css` is idempotent without touching other files).- `fetch-official-images.mjs` gained `--url <图片URL>[,...]`: the **scattered image URLs** picked up by inspecting an official site's DOM with the built-in browser now have a way to disk (the docs previously said "then download it" without giving a means, while the script could only do "page URL + index"). It obeys the same discipline (host allowlist, per-hop redirect re-checks, 30MB cap, filename sanitising, no overwriting existing files) **and needs no Playwright** — when a site requires login or lazy-loads on scroll and the script cannot open it, this is the only way to disk. Filename derivation (`imageNameFromUrl`: take the last path segment, prefer the URL's extension, otherwise use the content-type, and fall back to `.bin`) lives with `sanitizeFilename` in `url-policy.mjs`; 8 new test cases (4 filename derivations + 4 CLI refusals/usage). Also, **the duplicate check moved ahead of reading the body** (an existing image is not downloaded in vain, and network flakiness cannot report "download failed" for something that already exists), plus two guard rails: **a 0-byte response fails outright** (no empty files for 404 error pages / hotlink protection) and **redirect loops / self-redirects** produce a diagnosable error (with the redirect chain).
+
+**Chart toolkit v2 (late 2026-09-18, converged through 7 rounds of visual acceptance)**
+
+Origin: the user asked "are the chart templates not elegant enough? is there anything better in the donated library?". The audit conclusion: the donated library's chart layouts are **Chart.js (CDN + canvas)** — long since banned here (unreachable offline, and canvas animation cannot be seeked frame by frame); its hand-written column chart is plainer still and uses the `align-items:flex-end` trap we had already recorded; **there is no motion worth copying**. So the toolkit was rebuilt in-house, with a 1920×1080 render handed to the visual acceptance gate every round, converging after 7 rounds.
+
+- **Three pure-CSS effects that "were not possible before"** (driven by `@property`-registered properties, still frame-seekable): `fx-sweep` (a donut **truly swept out**, replacing the hard-popping `fx-pop`), `fx-count` (a **rolling** number, integers), `fx-grow-w/h` (growing by `width`/`height`); plus `fx-dot` (data points popping out one by one) and the whole set of `.chart*` primitives (`chart-head/title/note/num/plot/cap/row/track/cell/wrap/target`).
+- **The invariant is unchanged**: the static default = the final value (keyframes grow from 0 to the final value), and with motion off everything stops at the completed state, with no extra reset rules needed; `--upgrade-css` now fills in no-fx and the **chart toolkit** together (idempotent, touching only tokens.css).
+- **The elegance disciplines written into the docs** (`references/authoring.md`'s chart section rewritten): same kind, same colour, plus one accent; numbers in a monospaced face and never inside a scaled element; a baseline and **visible** fine gridlines; one chart tells one thing; a zero-value floor; the caption carries the scope and point in time; vertically centred inside the **subtitle safe area**; values in one uniform column; **column height = value ÷ axis maximum, with the axis starting at 0**.
+- **Acceptance caught and fixed 7 real defects** (all of the "the picture is obviously wrong, yet the pipeline reports success" kind): ① `fx-grow-x` **squashed the number inside the bar horizontally**; ② white text on a de-emphasised bar **could not be read**; ③ adding a value outside the bar narrowed the track → **two different rulers in one chart**; ④ the column height percentage was computed against the whole column and anything overflowing was pushed back by `flex-shrink` → **84% and 72% drawn the same height** (data distortion); ⑤ the grid spanned the whole column while the plot area was shorter → readings taken against the grid did not match the labels; ⑥ content was laid out into the **subtitle band** and covered by the subtitles (bottom 84–168px, centred, 73% wide); ⑦ a translucent de-emphasised bar let the gridlines show through, reading as a segmented stack.
+- New chart recipes: bullet (actual vs target), slope (before/after), sparkline (a mini trend beside a number), joining the existing horizontal bars / columns / line / donut / progress bar — 7 recipes in total, plus the motion selection table.
+- Tests +7 (**114** in total / 8 files): `chart-kit` (the motion classes and keyframe opacity, `@property`, "static default = final value", the column chart's three-layer structure and grid position, the row's three fixed-width columns, `--upgrade-css` filling in the toolkit idempotently); the workflow's explicit list was kept in sync.
+
+**Table toolkit (deep into 2026-09-18, after the user asked "what table types are there right now")**
+
+Origin: a survey — a real `<table>` layout **exists only once** (a "quarter / revenue / year-on-year" example), `kpi-grid`/`roadmap`/`comparison` only have a card-based tabular feel, and **tokens.css has no table primitives at all**, with every page writing its own inline styles. Tables were the thinnest part of the template system; filled in:
+
+- **Four table primitives** (`scripts/table-css.mjs` as the single source; init-project writes them into tokens.css, `--upgrade-css` fills them into old projects, idempotent): `.tbl` (data table: a muted header, horizontal rules only, numeric columns `.num` right-aligned and monospaced, `.up/.down` moves driven by tokens, `.key` for the highlighted row, `.sum` for the total row), `.kv` (spec list / time-and-event table), `.matrix` (comparison matrix: ticks and crosses + a `.hi` highlighted column), `.rank` (ranking table: rank + an inline micro-bar + the value, with the bar length likewise = value ÷ axis maximum).
+- **Legibility hard targets are written into the primitives and watched by the tests**: main data at `var(--fs-body)` (30px), headers at `--fs-caption`, row height ≥72px, horizontal rules only — no vertical rules and no zebra striping. Measured lesson: the table body originally used `--fs-caption` (24px) and was a strain to read even on the author's own screen — and the video will also be watched on a phone, so it was raised to the body size.
+- **Docs**: `references/authoring.md` gained a "table toolkit" section (seven disciplines + five recipes + a selection cheat sheet), and the `table` layout row points at it; SKILL.md's symptom table gained a row (tables unreadable on a phone → use the primitives, not the caption size).
+- Two example screens (data table + ranking / comparison matrix + spec list) live in `_probe/tableproj` (an internal verification artefact); geometric assertions: row heights 80/72, numeric columns each right-aligned to the same vertical line, and no content intruding into the subtitle band.
+- Tests +6 (**120** in total / 9 files): `table-kit` (all four primitives present, main data at the body size, row height 72, up/down driven by tokens, **no vertical rules or zebra striping may appear**, `--upgrade-css` filling in the primitives idempotently); the workflow list kept in sync.
+
+**Docs**
+
+- Kickoff alignment's "topic and audience" → "**topic and domain**" (regulated subjects must be asked about the disclaimer and data-source attribution); new `references/compliance.md` (red lines for financial voiceover, the three requisites for numbers, up/down colours flipped per audience, how to write the disclaimer, medical/legal/advertising law, the Gate list); `tokens.css` gained `--up/--down` and `.disclaimer`; `check-slides.mjs` gained the whole-video finance-keyword self-check.
+- **Layer entrance order is no longer fixed to "title first"**: it became "every slide carries at least two information layers on separate stages, with the order decided by the strong-sync principle" (a big number first / the question first / an image first are all legal); what the gate blocks is "only one layer".
+- Added the `roadmap` layout's HTML snippet (it existed in the table but the code block was missing), and fixed the kpi-grid snippet's use of the undefined class `.grid g4` (which silently stacks vertically).
+
+## 1.0.1 — 2026-09-17
+
+**New: domain and compliance (regulated subjects)**
+
+- `references/compliance.md` (new) — how to confirm the domain (subject + audience + whether it is regulated), the three red lines for financial voiceover (no operational advice / no price predictions / no invented data or graphics), the three requisites for numbers (**scope + currency + point in time**), flipping up/down colours per audience (A-shares/Hong Kong = red up, green down), how and where to write the disclaimer (`.disclaimer`, on screen ≥3s, not read aloud and not in the subtitles), key points for medical / legal / government / advertising law, and the Gate closing checklist
+- **Kickoff alignment**'s first batch of questions was upgraded from "topic and audience" to "**topic and domain**": for regulated subjects (financial investment research / healthcare / legal / government policy / marketing performance claims) one more question is mandatory — "do you want a disclaimer and data-source attribution" (default yes) — and the conclusion is recorded in Gate 0 and `research/notes.md`
+- `tokens.css` gained the `--up` / `--down` tokens dedicated to up/down moves (default = `--good` / `--bad`) and the `.disclaimer` / `.disclaimer-box` primitives (purely additive — old projects and existing themes are unaffected)
+- `check-slides.mjs` gained a **whole-video domain self-check**: hitting several finance/investment-research keywords with no disclaimer or sources line → an info-level prompt (a prompt, not an error; ignore it if the user explicitly declined the disclaimer)
+- `research.md` gained "confirm the domain first" and the finance/investment-research pitfalls (year-on-year ≠ quarter-on-quarter, GAAP ≠ non-GAAP, treating an old as-of date as "current"); `authoring.md` gained the disclaimer snippet, switched kpi-grid to `var(--up)/var(--down)` and added a finance row to the theme cheat sheet; `evals` gained 2 entries (finance kickoff confirmation, up/down colours by audience)
+
+## 1.0.0 — 2026-09-17
+
+First public release.
+
+**Pipeline**: kickoff alignment (language/style/subtitles/canvas/voice/asset boundaries) → research → script → TTS → measured timing → images → staged HTML entrance → frame-by-frame rendering → reverse ASR verification — 7 phases and 6 confirmation gates in total.
+
+**Scripts (11, plain Node, no build step)**
+
+| Script | Purpose |
+|---|---|
+| `init-project.mjs` | Generates the project skeleton (directory + tokens.css + slide templates + the script.json contract) |
+| `plan-timings.mjs` | ffprobe-measures each TTS segment → duration, each level's entrance moment, each sentence's onset |
+| `check-timing.mjs` | Silence detection measures each sentence's real onset, compares it with the estimate and can calibrate |
+| `check-theme.mjs` | The WCAG contrast gate for every theme (body/secondary/subtitles/accent) |
+| `check-slides.mjs` | Static checks before rendering (undefined variables / missing images / external resources / `data-stage` without animation / hard-coded colours) |
+| `capture.mjs` | Final-state screenshots or frame-stepping capture, subtitles burned in by default |
+| `build-video.mjs` | Encode → concatenate → align the audio track → mix BGM → mux → self-check + SRT |
+| `asr.mjs` | ASR transcription compared against the script (direct REST calls); character-level timestamps can verify each sentence's onset |
+| `prep-image.mjs` | Image checks and limited cropping (crop up to 20% of the area) |
+| `fetch-official-images.mjs` | List and download candidate asset images from the official site |
+| `tools.mjs` | Multi-anchor discovery of ffmpeg/ffprobe and Node packages |
+
+**Design system**: 13 themes, 17 layouts, the image-frame primitive (`.img-frame`), 14 entrance/ambient animations, staged entrance and the stagger container.
+
+**Language**: Mandarin Chinese / English / Cantonese — the character-count and speech-rate baselines, the subtitle line width and the ASR language check all switch with the language.
+
+**Canvas**: 1920×1080 landscape and 1080×1920 vertical (the subtitle geometry adapts to the canvas ratio).
+
+**Subtitles**: monolingual or bilingual, burned in, plus a separate `out/subs.srt`.
+
+**Sound**: BGM underneath (optional, looped and faded in/out automatically), reverse ASR verification (voice language, consistency of numbers and proper nouns).
+
+**Research**: `references/research.md` — four-tier source grading, the four hard rules (multi-source cross-checking / primary sources first / labelled scope and date / uncertain items stay out of the script), query design, handling contradictions across sources, the notes template, and the search tools available per environment (including mmx search's 10-result cap).
+
+**Two toolchains**: the mcode sandbox uses platform connectors; other Agent environments use mmx-cli (for TTS) and this repo's `asr.mjs` (for ASR).
+
+**Silent failures already fixed (all gated)**
+
+- Entrance delays overridden by the `animation` shorthand → elements entered at 0 seconds (delays moved to variable slots)
+- Undefined CSS variable + transparent text fill → the text becomes completely invisible (a static block in check-slides)
+- Image load failure → a broken icon and no error (capture names it at runtime)
+- Missing clauses → a subtitle-less video shipped silently (capture warns)
+- Overlapping subtitle windows → two adjacent sentences visible at once, looking like "ghosting/typos" (the fade-out now reaches zero inside the window, plus a window self-check)
+- Theme overrides ignored and only the first block taken for a selector of the same name (check-theme's merge semantics corrected)
diff --git a/plugins/Wzdhehe/html2video-for-mcode/LICENSE b/plugins/Wzdhehe/html2video-for-mcode/LICENSE
new file mode 100644
index 00000000..73c5531e
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/LICENSE
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Wzdhehe
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/plugins/Wzdhehe/html2video-for-mcode/README.md b/plugins/Wzdhehe/html2video-for-mcode/README.md
new file mode 100644
index 00000000..772da126
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/README.md
@@ -0,0 +1,281 @@
+**English** | [中文](README.zh-CN.md)
+
+# HTML 2 Video for mcode
+
+Turn a topic, an outline, or a finished script into a **narrated MP4**: HTML slides with staged
+entrance animations, a TTS voiceover, burned-in subtitles, and an ASR pass that verifies the
+voiceover actually says what the script says.
+
+Built for MiniMax Code (mcode), and runnable in any agent host through `mmx-cli`.
+
+## What the user gets
+
+Ask in plain language, get a publishable video. Here is a real request — Chinese is the default
+narration language (English and Cantonese are supported too), so the sample input is quoted as it
+was typed:
+
+> 帮我把这份大纲做成一条 60 秒的中文口播视频:三张关键数字、结尾一句行动号召,用深色科技主题,加中文字幕。
+
+*"Turn this outline into a 60-second Chinese voiceover video: three key numbers, a closing call to
+action, a dark tech theme, Chinese subtitles."*
+
+Result:
+
+```
+my-video/
+├── slides/            8 HTML slides + tokens.css (13 themes, 17 layout recipes)
+├── audio/             8 TTS clips
+├── build/timings.json measured durations + per-layer entrance times
+├── preview/*.png      terminal-state frames
+├── preview/play/      self-contained play page (screen the deck without encoding)
+└── out/
+    ├── final.mp4      ★ the deliverable (1920×1080 or 1080×1920, H.264 + AAC)
+    ├── subs.srt       subtitles for platform upload
+    └── slide-*.mp4    per-slide segments
+```
+
+## Pipeline
+
+One Skill drives a 12-script pipeline (plus eight internal modules — path containment / URL policy,
+`no-fx` rules, chart CSS, table CSS, the **managed CSS region** (rev delimiters + in-place
+replacement), the generated-body template with its content fingerprint, and the scan cap — under
+`skills/html2video-for-mcode/scripts/`):
+
+| Stage | What happens |
+|---|---|
+| 1. Kickoff alignment | Ask about language (Chinese / English / Cantonese), style & brand color, subtitles (none / single / bilingual), canvas (16:9 or 9:16), duration, voice, asset boundaries |
+| 2. Research | Fact-check the topic and record sources before writing |
+| 3. Script | Per-slide narration split into clauses; every slide carries at least two information layers on separate animation stages (order is free) |
+| 4. TTS | Voice synthesis (mcode connector, or `mmx-cli` elsewhere), then measure every clip with ffprobe |
+| 5. Timing | Derive each slide's duration and each animation's entrance time from the **measured** audio — nothing is hand-written |
+| 6. Assets | Official sources first, image framing primitives, compliance manifest |
+| 7. HTML | Staged entrance animations bound to the measured timings; layout recipes for 17 slide types |
+| 8. Render | Deterministic frame-stepping capture (animations land in the video), ffmpeg assembly with subtitles, optional background music. Before spending encode time, `preview-page.mjs` writes a play page so the deck can be screened in a browser |
+| 9. Verify | ASR transcription compared against the script; contrast, theme, and static-slide gates |
+
+## Design decisions worth knowing
+
+- **Nothing about timing is hand-written.** Every slide's duration and every animation's entrance
+  time comes from the measured TTS audio, so "the voiceover finished but the picture is still
+  waiting" cannot happen by construction.
+- **Every slide has at least two information layers** on separate animation stages (which layer
+  enters first is a judgement call — a big number, a question, or an image may lead), so a slide is
+  never just one line of text with nothing to look at. Entrance order is checked statically, not
+  left to hope.
+- **Animations land in the video.** Capture steps frames deterministically instead of
+  screen-recording, so entrance animations are actually rendered rather than frozen.
+- **Regulated topics get compliance handling.** For finance / medical / legal / policy decks the
+  workflow asks about disclaimers and source attribution up front, pins every number to its basis
+  (scope + currency + as-of date), and flips the up/down colour convention to the audience's market
+  (A-shares and Hong Kong read red as up) — see `references/compliance.md`.
+- **Charts are drawn in plain CSS/SVG, and effects can be switched off in one place.** No chart
+  library and no canvas (offline they would not load, and canvas animations cannot be frame-seeked).
+  Adding `no-fx` to the root element — or to any container, for a single slide — turns every
+  entrance and ambient animation off; the renderer then emits static frames and the timing stays
+  intact. The chart toolbox enforces the data-viz basics in the template itself: bars grow by
+  animating `width` (scaling would squash the numbers), column heights resolve against the plot
+  area so they always match their values, the grid shares the bars' scale, and charts stay clear of
+  the subtitle band. Tables ship four shapes (`.tbl` data table, `.kv` spec list, `.matrix` feature
+  matrix, `.rank` ranking with inline bars) with type and row heights sized to survive being watched
+  on a phone.
+- **Rendering is gated.** A static check refuses to render slides with undefined CSS variables,
+  missing images, external resources, or entrance animations without an animation class — the
+  failure modes that otherwise ship a video that looks broken while every script reports success.
+  It also refuses entrance animations whose keyframes never lift the `opacity: 0` base state — those
+  elements would silently stay invisible in the finished video.
+
+## Install
+
+**As a plugin (MiniMax Code):** install `plugins/Wzdhehe/html2video-for-mcode` from the
+[MiniMax-Code-Plugins catalog](https://github.com/MiniMax-AI/MiniMax-Code-Plugins/tree/main/plugins/Wzdhehe/html2video-for-mcode) — that hosted
+directory is the publication unit, and the copy reviewed there is the one users get.
+
+**As a standalone skill (any AgentSkills host):** copy the reviewed skill directory **out of that
+hosted plugin** into your host's skills directory — not from any other source:
+
+```bash
+# from a checkout of the MiniMax-Code-Plugins repository:
+cp -r plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode ~/.claude/skills/
+# project-level
+cp -r plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode <your-project>/.claude/skills/
+```
+
+Then install the two dependencies **in your video project** (not in the skill directory):
+
+```bash
+cd <your-video-project>
+npm i playwright && npx playwright install chromium
+# ffmpeg: winget install Gyan.FFmpeg / brew install ffmpeg / apt install ffmpeg
+#         or: npm i ffmpeg-static ffprobe-static
+```
+
+## Requirements
+
+- **Node.js 18+** (plain ESM, no build step).
+- **ffmpeg / ffprobe** on `PATH`, or `ffmpeg-static` / `ffprobe-static` in the project. Scripts
+  probe `PATH → project node_modules → skill parents → common install locations`.
+- **Playwright Chromium** for capture. Install it in the video project — scripts resolve it from
+  the project directory, the working directory, or the global npm root.
+- **Voice synthesis**, one of: mcode platform connectors, `mmx-cli`
+  (`npm i -g mmx-cli && mmx auth login --api-key sk-...`), or your own TTS written to
+  `audio/<id>.mp3`. Both mcode connectors and `mmx-cli`/REST spend a **MiniMax account's paid,
+  quota-metered balance** — the account is a required, billable service charged per character/minute.
+- **Reverse ASR verification** via `scripts/asr.mjs`, which transcribes the voiceover and compares
+  it against the script (numbers, proper nouns, and spoken language). Defaults to `mmx speech
+  transcribe` (mmx-cli ≥ 1.0.26 — the same login as TTS, no key handling needed); without mmx-cli,
+  `--provider api` + `MINIMAX_API_KEY` calls REST directly. Either path spends the same
+  **paid, quota-metered MiniMax account**; each transcription call is billed by audio duration.
+
+## Quick start
+
+```bash
+node <skill>/scripts/init-project.mjs ./my-video --topic "My topic"
+# fill research/notes.md and script.json (clauses = one line of narration each)
+# synthesize audio/01.mp3 … audio/08.mp3
+node <skill>/scripts/plan-timings.mjs ./my-video     # measure audio → timings.json
+node <skill>/scripts/check-slides.mjs ./my-video     # static gate before rendering
+node <skill>/scripts/capture.mjs ./my-video --mode still
+node <skill>/scripts/preview-page.mjs ./my-video --open   # play page: screen it yourself first
+node <skill>/scripts/capture.mjs ./my-video --mode motion
+node <skill>/scripts/build-video.mjs ./my-video --asr
+```
+
+`SKILL.md` carries the full workflow (7 phases, 6 confirmation gates); `references/` holds the
+authoring rules, asset-sourcing SOP, TTS/timing notes, and rendering internals.
+
+`preview-page.mjs` writes a self-contained play page to `preview/play/index.html` (open it from
+disk, no server) whose only job is **screening the HTML**: arrows (or a swipe) page through the
+slides, `O` is an overview, and `X` switches to a `no-fx` copy — if the picture goes blank, some
+keyframes never lift the `opacity: 0` base state. With animations on, the arrows **step through the entrance
+level by level** (`→` reveals the next level on the spot and only then turns the page) — nothing waits for a
+timer. `T` compares hard cut against cross-dissolve for the slide change, so the transition can be decided by
+ear and eye before rendering; both toggles sit in the bottom bar. There is
+deliberately no timer, progress bar, or karaoke-style highlight: for timing, watch the finished
+video. The narration panel is data-driven (`P` toggles it): it lists the slide's script lines when
+they exist, is skipped entirely when they do not, and `--no-script` hides it outright — so a deck
+whose voiceover is not written yet still previews fine. The layout adapts down to phones (panel
+becomes a collapsible bottom drawer, touch buttons appear). The page loads snapshots with the
+**measured** stage delays injected from `timings.json`, so what you see in the browser matches the
+timing of the final video (opening `slides/*.html` directly does not — those files carry
+placeholder delays; without `timings.json` the page spreads the stages evenly and says so).
+
+## Supported platforms
+
+Windows, macOS, and Linux. All scripts are Node ESM and shell-agnostic. On Windows prefer Git
+Bash / WSL over PowerShell (complex argument lists with non-ASCII paths can misbehave there);
+ffmpeg and Chromium paths are auto-detected.
+
+## Network access
+
+Nothing is contacted unless you invoke the step that needs it:
+
+- `scripts/asr.mjs` — HTTPS `POST` to `https://api.minimaxi.com/v1/speech_to_text`
+  (or `https://api.minimax.io` when `MINIMAX_REGION=global`). Only when you run it. **The API key
+  is sent to these two official hosts only**: any other `--base-url` / `MINIMAX_BASE_URL` is
+  rejected before the request is made, unless you explicitly pass `--allow-any-endpoint`
+  (self-hosted gateway / testing, at your own risk).
+- `scripts/fetch-official-images.mjs` — opens the URL **you** pass (an official site, or a local
+  `file://` page with `--allow-file`) to list and download candidate images, or downloads the exact
+  image URLs you hand it with `--url` (what the in-app browser route uses; no Playwright needed). Targets are checked
+  before any request: loopback, link-local (including cloud metadata `169.254.169.254`), private
+  and CGNAT ranges, dotless hostnames, URLs with embedded credentials, and non-HTTP(S) schemes are
+  all refused, every redirect hop is re-checked the same way, and responses are size-capped
+  (30 MB by default, `--max-mb`).
+- Voice synthesis happens through mcode connectors or `mmx-cli`, which contact MiniMax.
+- Everything else (timing, static checks, capture, encoding, theme validation) is fully offline.
+
+No telemetry, no analytics, no hidden endpoints, no installers, no native binaries.
+
+## Data use
+
+- The narration text you write is sent to the speech service you chose; the produced audio is sent
+  to the ASR service only if you run `scripts/asr.mjs`.
+- Assets you fetch are downloaded into your project's `assets/` directory and must be recorded in
+  `assets/MANIFEST.md` with source and license.
+- **Writes stay inside the project directory you pass in.** Every path derived from `script.json`
+  (slide `id` / `html` / `audio`, `bgm.file`) is validated first: ids must match
+  `^[A-Za-z0-9_-]{1,64}$`, paths must resolve inside the project, and symlink escapes are refused —
+  a hand-edited or prompt-injected `script.json` cannot make the pipeline read, write, or
+  recursively delete anything outside your project. One declared exception: `prep-image.mjs --crop`
+  takes its input and output **explicitly on the command line** (the same trust level as running
+  ffmpeg yourself), so those two paths are not restricted to the project — overwriting an existing
+  output still requires `--force`. Every pipeline write whose path is *derived* (project + id,
+  `preview/`, `build/`, `out/`, `asr.mjs --out`) remains fully contained.
+- **Nothing is overwritten silently.** `init-project.mjs` refuses a non-empty target directory
+  (re-initialising needs `--force`, which only resets its own five generated files);
+  `fetch-official-images.mjs` and `prep-image.mjs` do not clobber existing files without `--force`
+  and keep downloads inside your working directory by default. `--upgrade-css` updates the one
+  **managed `tokens` region** in `tokens.css` — the whole generated body (theme tokens, the `--fs-*`
+  scale, fx keyframes and utilities, `.fx-stagger`), with the three toolkit blocks (no-fx / chart /
+  table) nested inside — by replacing it in place. Stale copies **outside** the region are cleaned
+  **rule by rule**: rules that are verbatim copies of the module text are deleted (they win the CSS
+  cascade and would override the region), while your own rules — one or two overrides sitting next
+  to a copy included — are kept and reported; a `tokens.css.bak` backup is written first. Put your
+  own rules **after** the region, or the region's values win. `--check-css` reports staleness
+  without changing anything.
+- No credentials are stored or embedded: with the default mmx provider the script never touches a
+  key at all (mmx holds its own login); on the `--provider api` path it reads a key from
+  `MINIMAX_API_KEY` or `--api-key` at runtime and never writes it anywhere.
+- `KIT_PROJECT_DIR` is internal plumbing, set automatically by the CLI entry points so shared
+  helpers can locate the project directory; you never need to set it yourself.
+
+## Verification
+
+The Skill ships an executable test suite (`skills/html2video-for-mcode/tests/`, plain
+`node:test` — no extra dependencies):
+
+```bash
+# in the MiniMax-Code-Plugins monorepo (the plugin PR context), from the repository root:
+node --test "plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/*.test.mjs"
+# in a standalone clone of this skill repo, from its root:
+node --test "tests/*.test.mjs"
+```
+
+285 tests in fifteen files: `safe-paths` (malicious slide ids / paths, canary intactness, symlink
+escapes including a **dangling** link that must be caught rather than skipped), `no-clobber` (refusing to overwrite), `endpoint-allowlist` (key never leaves the official
+hosts — plus a local server that proves the gate sits before the request, and provider selection:
+`mmx speech transcribe` default vs `--provider api` fallback, with an mmx-shim integration pair
+asserting the exact argv mapping and tmpdir staging/cleanup), `fetch-policy` (SSRF,
+`file://`, redirect and filename rules), `connect-binding` (DNS fails closed, and the SSRF veto rides the connection's own lookup — egress via `policyGet`, browser traffic through a loopback veto proxy; the rebinding cases assert the target receives **zero requests**), `preview-page` (snapshot timing injection, `base` ordering,
+self-containment, containment refusals, the no-timer rule), `tokens-fx` (every entrance animation in
+the generated `tokens.css` must declare `opacity`, `no-fx` must reset it, `--upgrade-css` is
+idempotent) and `fetch-policy`/`no-clobber` (the `--url` download route obeys the same SSRF,
+redirect, size and no-clobber rules), plus
+`chart-kit` (the chart toolbox is really in the generated `tokens.css`: keyframes, registered
+properties, and "off = final state"), `table-kit` (table primitives keep their legibility rules:
+body-size type, ≥72px rows, token-driven up/down colours, matrix highlight covers the header row
+too), `css-kit` (the managed CSS region: a region that is stale, missing, duplicated, has a broken
+delimiter, or has a leftover copy outside it must be detected and repaired in place — invisible copies
+outside the region win the CSS cascade, which is what the rewrite is about; project-side overrides
+survive rule-by-rule, legacy bare-text files get wrapped in place instead of duplicated, `--check-css`
+exit codes, CRLF files must not false-positive, pathological interleaving resolves in one run,
+oversized inputs are refused — and the two false greens an adversarial re-review then found *inside*
+the new mechanism: a live region **plus a stale copy of the generated body outside it** used to report
+`ok`, and the `ok` note for a project override now says whether it sits before the region (loses the
+cascade) or after it (wins)) and `render-smoke` (init → timings → static gate → capture → build, end
+to end, **plus a pixel-level proof that upgrading a stale project really changes the rendered picture**), plus the capture layout-geometry self-check (content overflow / top-edge clipping named with numbers) and the safe-center scaffold contract. Two newer suites cover the second review round and the delivery-quality fixes: `review-round2` (canary tests proving output containment, a local server proving a rejected destination receives **zero** requests, single-quoted/unquoted HTML attributes, ASR request errors counted as failures, ffmpeg discovered from the project) and `cover-transition` (cover image present, `attached_pic` stream embedded, first frame not black, no black frame at any cut point in both hard-cut and cross-dissolve modes, total duration unchanged) and `review-round3` (the seven functional defects: `check-timing --calibrate` must not write back half-measured boundaries, `--topic <value> <project>` must not treat the value as the project directory, `--json` is position-independent, `--min WxH`, `--get` all-failed exits non-zero, `capture --mode/--dsf` and `grab-frames --at` reject bad values at the entry point, the subtitle-band gate only fires on positioned elements, `--dry-run` prints the plan and exits 0, and a bare-string `transition` must mean the same thing to the gate and the renderer — plus the stale-CSS gates on `check-slides` / `capture` / `build-video`, and the chart-structure gate: percentage-height bars sitting bare in a flex column get flagged (they collapse via flex-shrink; wrap them in `.chart-plot-cell`)) and `subtitles-invalidate` (deleting clauses, or `--no-subs`, and rebuilding must remove the old subtitles from the video — asserted at pixel level; plus keyframe-shape invariants, full-brightness / cross-still-leak checks on the spliced stills, and cover / still-mode-base must-stay-subtitle-free checks (the 2026-09-22 subtitle regressions)). Everything runs with ffmpeg, ffprobe and
+Chromium installed (on `PATH`, or in this repo's / your project's `node_modules` — `npm i
+ffmpeg-static ffprobe-static playwright`). Without them, every test that needs one of them **skips
+with its stated reason** (a tool is absent, or the filesystem forbids creating a junction) — nothing
+pretends to pass, and the pipeline scripts themselves exit with `找不到 ffmpeg` instead of running
+half a pipeline. The scoped workflow
+`.github/workflows/html2video-for-mcode-smoke.yml` — which lives in the **MiniMax-Code-Plugins
+monorepo**) and runs on the plugin PR and on main —
+installs them and runs everything for real.
+
+## Troubleshooting
+
+`SKILL.md` ends with a symptom → cause → fix table covering the failures this pipeline has actually
+hit: silent tails after the voiceover, slides with nothing on them, elements entering at 0 seconds,
+invisible text from undefined CSS variables, entrance elements that never appear at all (keyframes
+that never lift the `opacity: 0` base state), a `no-fx` switch that leaves the slide blank, opening
+`slides/*.html` and finding every animation crammed into the first two seconds, a blank play page,
+broken images, subtitles washed out on dark themes, wrong voice language, mismatched concat
+durations, and finance decks missing a disclaimer or with the up/down colour flipped.
+
+## License
+
+MIT — see `LICENSE`. Portions of the design system (10 themes, image-frame primitives, several
+entrance animations) are adapted from
+[html-ppt-skill](https://github.com/lewislulu/html-ppt-skill) (MIT, Copyright (c) 2026 lewis); the
+full notice and the upstream MIT text are in [`THIRD-PARTY-NOTICES.md`](THIRD-PARTY-NOTICES.md).
diff --git a/plugins/Wzdhehe/html2video-for-mcode/README.zh-CN.md b/plugins/Wzdhehe/html2video-for-mcode/README.zh-CN.md
new file mode 100644
index 00000000..b798f7ca
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/README.zh-CN.md
@@ -0,0 +1,158 @@
+[English](README.md) | **中文**
+
+# HTML 2 Video for mcode
+
+把一句主题、一份大纲或一篇定稿脚本,变成**带口播的成片 MP4**:HTML 幻灯片(分步入场动画)+ TTS 配音 + 烧录字幕,并用 ASR 反向校验"配音是否真的念的是脚本里的话"。
+
+为 MiniMax Code(mcode)定制,也能在其他 Agent 环境里通过 `mmx-cli` 运行。
+
+## 用户能得到什么
+
+用大白话提需求,拿到一条可直接发布的视频:
+
+> 帮我把这份大纲做成一条 60 秒的中文口播视频:三张关键数字、结尾一句行动号召,用深色科技主题,加中文字幕。
+
+产物:
+
+```
+my-video/
+├── slides/            8 张 HTML + tokens.css(13 套主题、17 种版式配方)
+├── audio/             8 段 TTS 音频
+├── build/timings.json 实测时长 + 每个视觉层的入场时刻
+├── preview/*.png      终态预览帧
+├── preview/play/      自包含放映页(不编码也能先过一遍)
+└── out/
+    ├── final.mp4      ★ 交付成片(1920×1080 或 1080×1920,H.264 + AAC)
+    ├── subs.srt       供平台上传的字幕
+    └── slide-*.mp4    逐张分段
+```
+
+## 流水线
+
+一个技能驱动 12 个脚本,外加八个内部模块(路径收监 / URL 策略 / `no-fx` 规则 / 图表 CSS / 表格 CSS / **受管 CSS 区**——rev 定界 + 原地替换 / 生成物模板与内容指纹 / 扫描上限,都在 `skills/html2video-for-mcode/scripts/` 下):
+
+| 阶段 | 做什么 |
+|---|---|
+| 1. 开工对齐 | 询问语言(中文/英语/粤语)、风格与品牌色、字幕(不要/单语/双语)、画布(16:9 或 9:16)、时长、音色、素材边界 |
+| 2. 信息搜集 | 事实性题材先核查并记录来源,再动笔 |
+| 3. 脚本 | 每张 slide 的口播拆成逐句 clauses;每张至少两个信息层、分属不同动画 stage(先后顺序自由) |
+| 4. TTS | 配音合成(mcode connector,其他环境用 `mmx-cli`),随后用 ffprobe 实测每段时长 |
+| 5. 对时 | 每张时长、每层入场时刻**全部由实测音频推出**,不手写任何秒数 |
+| 6. 配图 | 官方素材优先、图片框原语、合规清单登记 |
+| 7. HTML | 分步入场动画绑定实测时刻;17 种版式配方 |
+| 8. 渲染 | 确定性逐帧步进捕获(动画真正进视频)、ffmpeg 合成与字幕、可选 BGM。真正编码前先用 `preview-page.mjs` 出的放映页在浏览器里过一遍 |
+| 9. 校验 | ASR 转写与脚本比对;对比度、主题、渲染前静态闸门 |
+
+## 几个关键设计
+
+- **时序不靠手写**:每张时长与每个动画的入场时刻都来自实测音频,所以"配音念完了画面还在等"在结构上就不可能发生。
+- **每张至少两个信息层**,分属不同动画 stage(哪层先入场是判断题 —— 大数字、设问、图都可以先出),不会出现"只有一行大字"的页面;入场顺序是**静态检查**出来的,不靠运气。
+- **动画真的进视频**:捕获用逐帧步进而不是录屏,入场动画是渲染出来的,不是冻结在终态。
+- **渲染有闸门**:静态检查会拒绝未定义 CSS 变量、图片缺失、外链资源、入场动画缺动画类的页面,也会拒绝"关键帧从不把 `opacity: 0` 基础态抬回来"的入场动画(那类元素会在成片里静默隐形)—— 这些正是"视频看着坏了但每个脚本都报成功"的元凶。
+- **受监管题材走合规流程**:财经 / 医疗 / 法律 / 政务类片子开工就问免责声明与出处标注,每个数字钉住口径(口径 + 币种 + 时点),涨跌色按受众市场翻转(A 股 / 港股是红涨绿跌)—— 见 `references/compliance.md`。
+- **图表纯 CSS/SVG 自绘,动效可一键关**:不引图表库、不用 canvas(离线取不到,canvas 动画也逐帧 seek 不到);根元素(或任意容器,单张生效)加 `no-fx` 即可关掉全部入场与氛围动效,渲染自动走静态帧、时长不变。图表工具箱还把数据可视化的底线写进模板:条形用动画 `width` 生长(缩放会把数字压扁)、柱高相对绘图区算(保证与数值一致)、网格与柱高共用一把尺子、图表避开字幕带。表格有四种形态(`.tbl` 数据表 / `.kv` 规格表 / `.matrix` 对比矩阵 / `.rank` 排名表),字号与行高按"手机上也要看清"定死。
+
+## 安装
+
+**作为插件(MiniMax Code)**:从 [MiniMax-Code-Plugins 目录](https://github.com/MiniMax-AI/MiniMax-Code-Plugins/tree/main/plugins/Wzdhehe/html2video-for-mcode)安装
+`plugins/Wzdhehe/html2video-for-mcode` —— 宿主目录就是发布单元,评审过的那份才是用户拿到的。
+
+**作为独立技能(任意 AgentSkills 宿主)**:从**该宿主插件**里复制评审过的技能目录,不要从其他来源安装:
+
+```bash
+# 从 MiniMax-Code-Plugins 仓库的检出:
+cp -r plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode ~/.claude/skills/
+# 项目级安装
+cp -r plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode <你的项目>/.claude/skills/
+```
+
+然后把两个依赖装到**你的视频项目里**(不是技能目录里):
+
+```bash
+cd <你的视频项目>
+npm i playwright && npx playwright install chromium
+# ffmpeg: winget install Gyan.FFmpeg / brew install ffmpeg / apt install ffmpeg
+#         或: npm i ffmpeg-static ffprobe-static
+```
+
+## 依赖要求
+
+- **Node.js 18+**(纯 ESM,无构建步骤)。
+- **ffmpeg / ffprobe**:在 `PATH` 上,或项目里装 `ffmpeg-static` / `ffprobe-static`。脚本按 `PATH → 项目 node_modules → 技能上两级 → 常见安装位置` 探测。
+- **Playwright Chromium**:截图用。装在视频项目里即可 —— 脚本会从项目目录、工作目录、npm 全局逐个解析。
+- **配音**:三选一 —— mcode 平台 connector、`mmx-cli`(`npm i -g mmx-cli && mmx auth login --api-key sk-...`)、或你自己的 TTS(把音频写到 `audio/<id>.mp3`)。mcode connector 与 `mmx-cli`/REST 消耗的都是 **MiniMax 账号的付费配额** —— 该账号是必需的计费服务,按字数/时长计费。
+- **反向 ASR 校验**:`scripts/asr.mjs` 把配音转写回来与脚本比对(数字、专名、语种)。默认走 `mmx speech transcribe`(mmx-cli ≥ 1.0.26,与配音同一套 mmx 登录,不需要另行配 Key);没有 mmx-cli 时用 `--provider api` + `MINIMAX_API_KEY` 直连 REST。两条路径消耗的都是同一个**付费、按配额计费的 MiniMax 账号**,每次转写按音频时长计费。
+
+## 快速开始
+
+```bash
+node <skill>/scripts/init-project.mjs ./my-video --topic "我的主题"
+# 填 research/notes.md 与 script.json(clauses = 每句口播一行)
+# 合成 audio/01.mp3 … audio/08.mp3
+node <skill>/scripts/plan-timings.mjs ./my-video     # 实测音频 → timings.json
+node <skill>/scripts/check-slides.mjs ./my-video     # 渲染前静态闸门
+node <skill>/scripts/capture.mjs ./my-video --mode still
+node <skill>/scripts/preview-page.mjs ./my-video --open   # 放映页: 用户自己先过一遍
+node <skill>/scripts/capture.mjs ./my-video --mode motion
+node <skill>/scripts/build-video.mjs ./my-video --asr
+```
+
+工作流全文(7 阶段、6 个确认闸门)在 `SKILL.md`;`references/` 放着编写规范、配图 SOP、TTS/对时说明与渲染内幕。
+
+`preview-page.mjs` 生成一个自包含的**放映页** `preview/play/index.html`(双击即看,不需要起服务),
+它只干一件事:**把 HTML 画面放一遍** —— `←` `→`(或触屏左右滑)翻页、`O` 总览、
+`X` 切到同张的 `no-fx` 副本(切过去画面**变空**就说明有关键帧没把 `opacity:0` 的基础态抬回来)。
+动效开时箭头是**逐级入场**(按 → 先出下一级,出完才翻页),不等计时器;`T` 现场对比**硬切 / 交叉溶解**
+(切页方式渲染前就能定)。两个开关都在底栏。
+**刻意不做计时器、进度条、跟读高亮** —— 要看时间或节奏就直接看成片。口播文案面板按数据自动决定:
+有文案就列出(`P` 可开/关),没有就整个不出,`--no-script` 也能强制不要 —— 口播还没做时照样能先看画面。
+布局一路适配到手机(面板收成可收起的底部抽屉,出现触摸按钮)。副本按 `timings.json` **注入实测延迟**,
+所以浏览器里看到的时序就是成片时序(直接打开 `slides/*.html` 不是:那些文件里是占位延迟;没有 timings 时
+页面按等间隔预览并在顶部如实标注)。
+
+## 支持平台
+
+Windows / macOS / Linux。脚本全部是 Node ESM,不依赖特定 shell。Windows 上建议用 Git Bash 或 WSL 而非 PowerShell(非 ASCII 路径 + 复杂参数组合容易出问题);ffmpeg 与 Chromium 路径自动探测。
+
+## 网络访问
+
+**只有你主动执行的那一步才会联网**:
+
+- `scripts/asr.mjs` —— HTTPS `POST` 到 `https://api.minimaxi.com/v1/speech_to_text`(海外套餐设 `MINIMAX_REGION=global` 时走 `https://api.minimax.io`)。仅在你运行时。**API Key 只发这两个官方域**:其他 `--base-url` / `MINIMAX_BASE_URL` 取值会在发起请求之前被拒绝,确需自定义网关要显式加 `--allow-any-endpoint`(自担风险)。
+- `scripts/fetch-official-images.mjs` —— 打开**你传入的**网址(官方网站,或加 `--allow-file` 的本地 `file://` 页面)以列出并下载候选配图;也可以 `--url <图片URL>` 直接下载你指定的图片(内置浏览器取图那条路走的就是它,不需要 Playwright)。**请求之前先过目标校验**:loopback、链路本地(含云元数据 `169.254.169.254`)、私网与 CGNAT 段、无点主机名、带内嵌凭据的 URL、非 HTTP(S) 协议一律拒绝;重定向**逐跳**复用同一策略;响应大小有上限(默认 30MB,`--max-mb` 可调)。
+- 配音合成经由 mcode connector 或 `mmx-cli`,它们会访问 MiniMax。
+- 其余全部离线:对时、静态检查、截图、编码、主题对比度校验。
+
+无遥测、无埋点、无隐藏端点、无安装器、无原生二进制。
+
+## 数据使用
+
+- 你写的口播文本会送到你选择的语音服务;产物音频只在你运行 `scripts/asr.mjs` 时才送到 ASR 服务。
+- 抓取的素材下载到项目的 `assets/` 目录,并需在 `assets/MANIFEST.md` 登记来源与许可。
+- **写入只发生在你传入的项目目录内。** `script.json` 派生出的每个路径(slide 的 `id` / `html` / `audio`、`bgm.file`)都先过校验:id 必须匹配 `^[A-Za-z0-9_-]{1,64}$`,路径必须 resolve 在项目目录内,符号链接逃逸同样拒绝 —— 手改或被注入的 `script.json` 无法让流水线读写或递归删除项目目录之外的东西。**唯一声明过的例外**:`prep-image.mjs --crop` 的输入与输出都由命令行**显式给出**(与你直接跑 ffmpeg 同一信任级),这两个路径不限于项目内 —— 覆盖已存在输出仍需 `--force`。所有**派生**路径的流水线写入(项目 + id、`preview/`、`build/`、`out/`、`asr.mjs --out`)都完整收监。
+- **不会静默覆盖任何东西。** `init-project.mjs` 对非空目标目录直接拒绝(重新初始化需 `--force`,且只重置它自己生成的 5 个文件);`fetch-official-images.mjs` 与 `prep-image.mjs` 覆盖已存在文件需 `--force`,下载默认收在工作目录内。`--upgrade-css` 更新 tokens.css 里**唯一的 tokens 受管区**(技能生成的整段主体:主题令牌 / `--fs-*` 字号 / fx 关键帧与工具类 / `.fx-stagger`,三个工具箱子块嵌在其中)时按 rev 原地替换。区外还留着的旧副本**按条清理**:逐字来自模块原文的规则会被删掉(它们在层叠上会压过受管区),你自己的规则(包括紧挨副本的一两条覆写)保留并报出;写前先备份 `tokens.css.bak`。你自己的规则要写在**受管区之后**,否则受管区的值优先。`--check-css` 只查不改。
+- **不存储、不内嵌任何凭据**:默认 mmx 提供方下脚本完全不接触 Key(mmx 自己持有登录);`--provider api` 路径运行时从 `MINIMAX_API_KEY` 或 `--api-key` 读取,且从不写出到任何文件。
+- `KIT_PROJECT_DIR` 是内部变量,由 CLI 入口自动设置,让共用助手能定位项目目录;用户永远不需要自己设置。
+
+## 验证
+
+技能自带可执行测试(`skills/html2video-for-mcode/tests/`,纯 `node:test`,无额外依赖):
+
+```bash
+# MiniMax-Code-Plugins monorepo(插件 PR 语境), 在仓库根执行:
+node --test "plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/*.test.mjs"
+# 本技能仓的独立 clone, 在仓库根执行:
+node --test "tests/*.test.mjs"
+```
+
+十五个文件共 285 例:`safe-paths`(恶意 slide id / 路径、canary 完好性、符号链接逃逸 —— 含**悬空**链接必须被拦下,而不是当"不存在"跳过)、`no-clobber`(覆盖拒绝)、`endpoint-allowlist`(Key 不离开官方域,并用本地服务器证明闸门在请求之前;外加提供方选择:`mmx speech transcribe` 缺省与 `--provider api` 后备,mmx shim 集成对断言参数映射与临时目录中转/清理)、`fetch-policy`(SSRF、`file://`、重定向与文件名规则、**IPv4-mapped IPv6 十六进制形式**)、`connect-binding`(DNS 失败即拒(fail-closed),SSRF 否决挂到**建连那次** lookup —— 出网走 `policyGet`、浏览器流量走本地否决代理;rebinding 用例以"目标收到 **0 个请求**"为判据)、`preview-page`(快照注入实测延迟、`base` 顺序、自包含无外链、越界拒绝、无计时器、**竖版画布、iframe 强制相对解析**)、`tokens-fx`(模板里每个入场动画的关键帧必须声明 `opacity`、`no-fx` 必须重置基础态、`--upgrade-css` 幂等)、`chart-kit`(模板里真带着图表工具箱:keyframes / 注册属性 /「关动效 = 终态」不变量)、`table-kit`(表格原语的可读性硬指标:正文号、行高 ≥72px、涨跌走令牌、矩阵高亮列覆盖表头)、`css-kit`(**整段受管区**:陈旧 / 缺失 / 重复 / 定界符被改坏 / 区外还留着旧副本, 五种都要检出并原地修好 —— 区外副本会在层叠上压过受管区, 这正是本轮重写的起因; 项目端覆写按条保留、老格式裸文本原地包裹而不重复、`--check-css` 退出码、CRLF 不误报、病态互踩一次到位、超大输入拒绝 —— 外加复查在新机制自己身上揪出的两个假绿灯: 受管区是最新的、但区外还留着一份**主体副本**时以前报 `ok`(现在按主体形状识别并报 `legacy-outside`); 项目覆写的 `ok` 提示会点名它**在受管区之前还是之后**(之前则层叠上输, 之后才赢))与 `render-smoke`(init → 对时 → 静态闸门 → 截图 → 成片全链,**still 复截后静态回退必须点名警告**,外加**升级前后画面真的变了的像素级证明**), 以及 capture 的布局几何自检(溢出/顶部被裁点名数字)与骨架 safe center 契约。`review-round2`(输出收监 canary:项目内符号链接不得写到项目外;本地服务器证明被拒目标收到 0 个请求;单引号属性;ASR 请求失败必须计入失败;项目内 ffmpeg 发现)与 `cover-transition`(封面 + `attached_pic` 流 + 首帧非黑 + 硬切/溶解两种模式下切页点均无黑帧且总时长不变)与 `review-round3`(七个功能缺陷:`--calibrate` 不许写回半份实测、`--topic <值> <项目目录>` 不许把值当目录、`--json` 位置无关、`--min WxH`、`--get` 全失败退非零、`capture --mode/--dsf` 与 `grab-frames --at` 在入口拒绝非法值、字幕带闸门只对定位元素生效、`--dry-run` 打完计划必须退 0、裸字符串 `transition` 在闸门与渲染器里同义;以及 check-slides / capture / build-video 的陈旧 CSS 闸门, 外加图表结构闸门: 百分比柱体裸放 flex column 会被点名(flex-shrink 压扁, 应包进 .chart-plot-cell))与 `subtitles-invalidate`(删掉 clauses 或 `--no-subs` 重建后, 成片里不得再有旧字幕 —— 像素级断言; 外加字幕关键帧形状不变量、静帧满亮/跨句残留判据、以及封面/静帧基底不得烙字幕(2026-09-22 字幕回归);全套用例需要装好 ffmpeg、ffprobe 与 Chromium(`PATH` 上,或本仓/项目 `node_modules` 里 —— `npm i ffmpeg-static ffprobe-static playwright`):缺了它们,需要这些工具的用例会**带着明确原因跳过**(缺工具、文件系统不允许建 junction),不会假装通过;渲染脚本本身则会明确报 `找不到 ffmpeg` 退出,不会假装跑完半条流水线。scoped workflow `.github/workflows/html2video-for-mcode-smoke.yml`(**位于 MiniMax-Code-Plugins monorepo**)会在插件 PR 与 main 上装齐依赖并把全部用例真跑一遍。
+
+## 排错
+
+`SKILL.md` 末尾有一张"症状 → 原因 → 处置"表,覆盖这条流水线真实踩过的坑:配音念完画面还在等、页面只有标题、元素在 0 秒就入场、未定义 CSS 变量导致文字隐形、**元素入场后永远不出现(关键帧没把 `opacity` 抬回来)**、**切了 `no-fx` 画面反而更空**、**双击 `slides/*.html` 发现动画全挤在开头**、**放映页空白/图裂**、图片 broken、深色主题字幕糊底、音色语种不对、拼接后时长不符、**财经片漏了免责声明或涨跌色反了**。
+
+## 许可
+
+MIT —— 见 `LICENSE`。设计系统的一部分(10 套主题、图片框原语、若干入场动画)改编自
+[html-ppt-skill](https://github.com/lewislulu/html-ppt-skill)(MIT,Copyright (c) 2026 lewis);
+完整声明与上游 MIT 原文见 [`THIRD-PARTY-NOTICES.md`](THIRD-PARTY-NOTICES.md)。
diff --git a/plugins/Wzdhehe/html2video-for-mcode/THIRD-PARTY-NOTICES.md b/plugins/Wzdhehe/html2video-for-mcode/THIRD-PARTY-NOTICES.md
new file mode 100644
index 00000000..34f1deae
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/THIRD-PARTY-NOTICES.md
@@ -0,0 +1,65 @@
+# Third-party components and license notices
+
+This skill contains material adapted from third-party open-source projects. As the MIT license requires, the original copyright notices and the license text are reproduced below.
+
+---
+
+## html-ppt-skill
+
+- **Source**: https://github.com/lewislulu/html-ppt-skill
+- **Copyright**: Copyright (c) 2026 lewis <sudolewis@gmail.com>
+- **License**: MIT
+- **Date introduced**: 2026-09-17
+- **Scope of adaptation** (all rewritten to fit this skill's single-page static rendering and audio-visual sync system):
+  - `slides/tokens.css` generated by `scripts/init-project.mjs`:
+    - the color tokens of 10 named themes (`minimal-white` / `swiss-grid` / `corporate-clean` / `editorial-serif` /
+      `magazine-bold` / `tokyo-night` / `catppuccin-mocha` / `nord` / `xiaohongshu-white` / `soft-pastel`)
+      — the variable names are mapped to this skill's naming (`--surface`→`--panel`, `--text-1/2/3`→`--fg/--muted/--fg-3`,
+      `--border`→`--line`, and so on), and the external fonts (Playfair Display / Space Grotesk and others) were replaced with local font stacks
+    - the image-frame primitives `.img-frame` / `.contain` / `.fill` / `.img-scrim` / `.img-cap` / `.img-tag` and their
+      `--img-ratio` / `--img-pos` variables
+    - the keyframes of some entrance animations (`fx-blur` `fx-rise` `fx-pop` `fx-spotlight` `fx-ripple` `fx-glitch`
+      `fx-shimmer` `fx-kenburns` `fx-stagger`) — the trigger mechanism was changed from "JS re-triggering on slide change" to
+      "`animation-delay: var(--tN)` absolute time" so it works with frame-stepping rendering
+    - the typographic roles and container classes `.kicker` / `.eyebrow` / `.lede` / `.card` (with three variants) / `.pill` / `.divider-accent`
+  - Some layout skeletons in `references/authoring.md` were rewritten from the versions of the same name under its `templates/single-page/`
+  - The image-frame usage rules in `references/image-sources.md` follow the design intent of its `.img-frame`
+
+**Not adopted** (incompatible with its design or carrying extra dependencies): `runtime.js` (keyboard navigation / presenter mode / overview),
+the 20 canvas effects under `assets/animations/fx/` (rAF-driven, cannot be frame-seeked),
+`fonts.css` and every external Google Fonts link, the Chart.js / highlight.js CDN dependencies, and the `@media print` pagination logic.
+
+The "page through it once" need is served by this skill's own `scripts/preview-page.mjs` (which generates `preview/play/index.html`),
+and that code has nothing to do with upstream: the upstream runtime.js is built for a live human presentation (presenter window, teleprompter script, timer, BroadcastChannel two-window sync),
+whereas this skill produces a video and has no live speaking step; the play page only replays slide by slide + compares the motion switch + shows the voiceover panel.
+
+### The MIT license text
+
+```
+MIT License
+
+Copyright (c) 2026 lewis <sudolewis@gmail.com>
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+```
+
+---
+
+**Commercial use**: the MIT license permits commercial use, modification, redistribution and sublicensing; the only obligation is to retain the copyright and license notices above.
+The final artefacts this skill produces (video, subtitles, cover, and so on) are the user's own creations and do not inherit the license obligations above.
diff --git a/plugins/Wzdhehe/html2video-for-mcode/plugin.json b/plugins/Wzdhehe/html2video-for-mcode/plugin.json
new file mode 100644
index 00000000..efade5c8
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/plugin.json
@@ -0,0 +1,23 @@
+{
+  "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
+  "name": "html2video-for-mcode",
+  "version": "1.9.19",
+  "description": "Turn a topic, outline, or script into a narrated MP4: HTML slides with staged entrance animations, TTS voiceover, ffmpeg assembly, and ASR verification.",
+  "author": {
+    "name": "Wzdhehe",
+    "url": "https://github.com/Wzdhehe"
+  },
+  "homepage": "https://github.com/MiniMax-AI/MiniMax-Code-Plugins/tree/main/plugins/Wzdhehe/html2video-for-mcode",
+  "repository": "https://github.com/MiniMax-AI/MiniMax-Code-Plugins",
+  "license": "MIT",
+  "keywords": [
+    "video",
+    "html-to-video",
+    "slides",
+    "tts",
+    "voiceover",
+    "subtitles",
+    "ffmpeg",
+    "playwright"
+  ]
+}
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/SKILL.md b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/SKILL.md
new file mode 100644
index 00000000..58527117
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/SKILL.md
@@ -0,0 +1,302 @@
+---
+name: html2video-for-mcode
+description: Turn a script/outline/topic into a finished MP4 with Chinese voiceover (HTML slides + TTS + ffmpeg), tailored to mcode and runnable in other Agent environments with mmx-cli. Trigger when the user wants to turn content into a video, html 转 mp4 (html to mp4), 幻灯片口播视频 (narrated slide video), slides video, narrated video, 一分钟介绍视频 (one-minute intro video), 抖音/视频号竖版视频 (vertical video for Douyin/WeChat Channels); also trigger when they only want to look at the finished HTML slides first, present them once themselves, or preview the animations (with or without the voiceover yet) — the skill generates a self-contained play page, so you don't have to spend minutes rendering a finished video first. Also trigger for 受监管题材 (regulated subjects) — finance/earnings/investment research/healthcare/policy — they need disclaimers and data-scope handling. Also trigger on symptoms — anywhere the frame looks wrong or the timing is off; the full list is in references/symptoms.md.
+---
+
+# HTML 2 Video for mcode: script → narrated video
+
+Turn a topic/outline into a publishable MP4 (default 1920×1080, switchable to 1080×1920 vertical): HTML slides (staged entrance animations) + Chinese TTS narration + ffmpeg assembly + reverse ASR verification.
+
+## The four iron rules (violate any one and the output is guaranteed to be reworked)
+
+1. **All durations come only from ffprobe measurements; never write them by hand.** Each slide's duration = that segment's measured TTS duration + tail padding (default 0.8s). Don't estimate, don't round, don't hard-code.
+2. **TTS before HTML.** Produce the audio first, measure the real durations, work out the entrance moment of every visual layer, then write the HTML. All animation delays are injected through the CSS variables `--t1/--t2/--t3`; never hard-code seconds in the HTML. This is the fundamental cure for "too much silence after the narration ends" and "audio and visuals out of sync".
+3. **Wait for user confirmation at every Gate; skip no steps.** See the workflow below for the Gate list.
+4. **Align before starting; don't just charge ahead.** The first response = **ask each of the questions below, one by one, in question form, then stop and wait for the user's answers**; each item may come with a recommended default, but it must be "ask, then wait" — not "announce the default and treat silence as consent". Only when the user explicitly says "go with the defaults", or leaves an item unanswered, do you adopt the default, and before moving to the next step, restate which defaults were finally adopted. The answers become the acceptance baseline for every subsequent Gate.
+
+   ⚠ Counter-example (hit for real on 2026-09-18 — never again): firing off one message that reads "默认项(不特别说的话就按这个):中文普通话口播 · 中文字幕 · 1920×1080 · 温润男声 · 预计 10–12 张…" ["Defaults (unless you say otherwise, we use these): Mandarin Chinese narration · Chinese subtitles · 1920×1080 · warm male voice · estimated 10–12 slides…"] and then pressing on — that is not asking, it is informing; the user never gets the chance to take a position on each item, which amounts to no alignment at all. The right way: lay it out as a question list (2–3 options + a recommended default per item), say plainly "reply to confirm, or tell me which ones to change", and only move once the answers come back.
+
+| What to ask | Options / defaults |
+|---|---|
+| **Topic and domain** | One sentence on the subject + **which domain it belongs to** (finance/business · tech/product · popular science/education · brand/marketing · culture/history · life/consumer · government/policy); the domain decides layout and information density; full constraints for regulated domains are in the "Sensitive content and disclaimer" row below and in `references/compliance.md` |
+| **Audience** | Who it's for: **domain professionals / general tech-savvy public / management briefing / client demo / internal training** (recommend based on the subject, but have the user confirm). The same subject written for researchers and for the general public is two different scripts — this item decides term density, information density and tone of voice |
+| **Language (must ask)** | What language the narration is in: **Mandarin Chinese `zh`** (default) / **English `en`** / Cantonese `yue` / other BCP-47. This decides 4 things: ① which language the narration script is written in; ② **the voice must match the language** (Chinese uses `Chinese (Mandarin)_*`, English uses `English_*`; a mismatch produces a strange accent, so verify the language with ASR after writing); ③ the characters/speech-rate baseline (Chinese characters/sec vs English words/sec — plan-timings switches automatically); ④ the ASR language header (zh forces Mandarin, so it can catch Cantonese) |
+| **Style and colors (theme color)** | First offer 2–3 candidate themes based on the audience (business `minimal-white`/`swiss-grid`/`corporate-clean`; editorial magazine `editorial-serif`/`magazine-bold`; dark tech `tokyo-night`/`catppuccin-mocha`/`nord`; consumer/lifestyle `xiaohongshu-white`/`soft-pastel`). **Then ask one more question about theme-color preference**: use the theme's built-in primary color, or is there a brand color to specify (if given, override `--accent` per the "custom primary color" section of authoring.md and run check-theme to verify contrast) |
+| **Data and charts** | When the subject contains key numbers, comparisons, proportions or trends (earnings/product data/market comparisons/progress), you **must ask**: ① is there any data worth a slide of its own; ② do you want chart slides, and roughly how many (default suggestion: hard data + a suitable layout → 1–2 slides; pure narrative subjects skip this); ③ style preference: horizontal bars / columns / donut / line / progress bar, or "let the agent pick per the data shape" (selection cheat sheet in the authoring.md chart toolkit — all pure CSS/SVG, works offline); ④ numbers on the chart are also bound by Gate 0 (≥2 independent sources, with scope and timestamp). **Even if the user doesn't bring up charts, proactively offer this option for data-dense subjects** — a string of comparable numbers read out as bullets is worse than one chart |
+| **Subtitles** | No subtitles / monolingual (**same language as the narration**, default) / **bilingual** (main line = narration language, second line `text2` = the other language, e.g. Chinese narration with English subtitles) — decides whether to write text2 and whether to use `--no-subs` |
+| **Sensitive content and disclaimer** | If the subject touches any of these, ask "do you want a disclaimer + a sources-and-scope slide at the end": financial opinions/figures, medical, legal, policy, **negative-event reporting**, personal information/data citations — default is **yes**. Regulated domains (financial investment research/healthcare/legal/government policy/marketing performance claims) follow the full `references/compliance.md` set: a closing `.disclaimer` line (on screen ≥3s), up/down colors flipped per audience, numbers carrying scope + currency + timestamp. Omit only if the user explicitly says no |
+| **Canvas and platform (aspect ratio)** | Landscape 1920×1080 (default, suits Bilibili/official sites) or vertical 1080×1920 (Douyin/WeChat Channels/Xiaohongshu); vertical needs a stacked layout |
+| Approximate duration | Default ~60s (6–10s per slide × 8 slides); short-video platforms can compress it to 30s; offer a range for the user to pick |
+| **TTS voice and speed (rough direction)** | ①**Gender/voice direction**: male / female / neutral × warm / crisp / magnetic / lively; ②**speed**: normal (**default 1.1**, ≈5.3 Chinese characters/sec) / a bit slower / a bit faster. A rough direction is enough — Phase 2 auditions 3 candidate voices, and **plays the same sentence at all three speeds (1.05/1.15/1.25) alongside them so the two are decided together** (see tts-and-timing.md); once decided, write it into `speed` in `script.json` and restate it |
+| **Images and asset boundaries** | First ask **whether to use images at all**: pure typography, no images (number cards + large type + quotes) / allow official images pulled from the web (default) / only user-provided assets. **For negative-event subjects (blow-ups/penalties/lawsuits/controversies), add: do you want announcement or news-report screenshots as illustrations** (default suggestion: yes, 1–2; discipline in image-sources.md path D). The specifics — "which ones to use, what still needs to be supplied" — are left to the Gate 3 asset confirmation |
+
+## Contents and tools
+
+The skill ships with **12 command-line scripts + 8 internal modules** (call them directly by path within the skill directory, passing the project directory as an argument, no copying needed; `tests/` also holds a set of node:test security and smoke tests, auto-discovered from `node --test` at the repo root):
+
+| Script | Purpose |
+|---|---|
+| `scripts/init-project.mjs <project> [--topic "topic name"] [--force] [--upgrade-css] [--check-css]` | Generates the project skeleton: directory + tokens.css + slide templates + the script.json contract. **Refuses to run when the target directory is non-empty** (it would reset 5 generated files); re-initializing requires an explicit `--force`. `--upgrade-css` updates the project's tokens.css to the skill's current version. The **whole generated body** (theme tokens, `--fs-*` scale, fx keyframes/utilities, `.fx-stagger` rules, and the three toolkit blocks no-fx/charts/tables nested inside) lives in one `tokens` **managed region**, so it is **replaced in place** by content rev — no duplicate sections, idempotent, `tokens.css.bak` written first. **Put your own project rules after the managed region**; everything outside it is never touched (rules that duplicate a toolkit are the only exception: a leftover full copy would win the cascade and is cleaned up, its original text staying in the `.bak`). `--check-css` only inspects: stale/missing/duplicated/broken-delimiter/leftover-copy items are reported one by one, and any of them exits 1. A project whose file is still the old "bare text" format but byte-identical to the current version reports **ok** (rendering is already correct), with a note that the next upgrade will wrap it in place |
+| `scripts/plan-timings.mjs <project> [--pacing=<chars per second>]` | ffprobe-measures each TTS segment → per-slide duration, each stage's entrance moment, **each sentence's clause timing** → `build/timings.json`; `--pacing` overrides the speech-rate baseline (for re-estimating when the speed warning fires) |
+| `scripts/check-timing.mjs <project> [--calibrate]` | Silence detection measures each sentence's real onset and compares it with the estimate; `--calibrate` calibrates the timings to the measurements and then re-renders. For slides whose silence-segment count doesn't match and which haven't been calibrated, use `asr.mjs --verify-timing` to get character-level timestamps and measure the sentence onset |
+| `scripts/check-theme.mjs <project>` | Verifies WCAG contrast for all themes — **five pairs**: `--accent-ink/--accent`, `--fg/--bg`, `--muted/--bg`, `--fg-3/--bg`, and `--sub-fg/--sub-bg` (subtitle pill composited onto the background); exit code 1 if any fails; new themes must pass this gate |
+| `scripts/check-slides.mjs <project> [--ids 01,02] [--quiet]` | **Static gate before rendering**: undefined CSS variables, missing images/external resources, `data-stage` without an fx class, fx keyframes that don't include opacity (which would stay invisible forever), use of undefined classes (silently unstyled — info level), **tokens.css managed region out of date — error level, blocks screenshots** (an old CSS still renders, just with stale rules; the fix command is printed), **`.fx-stagger` and `data-stage` on the same slide (which would override entrance timing)**, **absolutely positioned `bottom` falling inside the subtitle band — warning level: burned-in subtitles would cover the caption, and a still does not show it**, hard-coded colors, and a whole-video domain self-check. Any ✗ and don't take screenshots |
+| `scripts/prep-image.mjs --check <image...>` / `--crop <in> <out> [--ratio 16:9] [--anchor ...] [--force]` | Execution helper for the image SOP: checks dimensions and crop risk; crops by anchor (enforcing "crop away ≤20%, never upscale to pad") |
+| `scripts/fetch-official-images.mjs <page URL> [--get 1,3] [--out-dir <dir>] [--min WxH] [--json] [--allow-file] [--max-mb N] [--force]` / `--url <image URL>[,...]` | Lists and downloads candidate images from an official page/local page (`--json` lists candidates without downloading; `--min` filters by size); **when the site needs login or lazy-loads on scroll, use the built-in browser to inspect the image URL, then drop it straight to disk with `--url` (no Playwright needed)**. Intranet and metadata addresses are always rejected (including IPv4-mapped IPv6 forms), every redirect hop is re-checked, and output defaults to inside the working directory |
+| `scripts/capture.mjs <project> [--mode still\|motion] [--ids 01,02] [--no-subs]` | Playwright screenshots. still = a single final-state frame (invalidates that slide's old frame directory); motion = frame-by-frame stepping through the entrance animations. `--dsf 1..4` sets the device scale factor (**default 1**; `--dsf 2` supersamples at 3840×2160 and build-video downsamples with lanczos — sharper text edges, ~4× render time, so use it for the final master). **Both capture and build-video refuse to run when the project's managed region is stale** (they print the issue list and the `--upgrade-css` command, and exit 1); `--allow-stale-css` skips that gate explicitly if you really must render the project's old CSS. **Subtitles are burned in by default** (content comes from clauses); `--no-subs` turns that off; **each sentence's subtitle is also emitted as a static image** (`build/substills/`), and subtitle windows the frame sequence can't cover are spliced in by build-video — otherwise subtitle changes after the animation window would freeze; **slide 1 additionally emits `preview/cover.png`** (title fully visible, no subtitles — used as the cover and the video's first frame) |
+| `scripts/preview-page.mjs <project> [--open] [--no-script]` | Generates the **play page** `preview/play/index.html` (single file, no CDN/server — double-click to view over file://; the snapshot copies read the project's slides/ via `<base href>`, so share the project directory or at least `play/`, not `index.html` alone): **with animations on, ←→/touch swipe step through the entrance level by level** — each press of → reveals the next level (a sub-heading or small chart animates in on the spot), and only after the last level does it turn the page; with animations off it just turns pages. X toggles animations on/off, P toggles narration on/off (all toggles are buttons in the bottom bar, and the label text literally reads "动效开/动效关" "口播开/口播关" — animations on/off, narration on/off; the top bar keeps only the page number), O overview, F fullscreen. **Frame switching is double-buffered with no white flash; UI text adapts between Chinese and English per the `lang` in script.json**. **It does only one thing: "show the frame"** — no timer/progress bar/read-along highlight/replay button (to see timing, watch the finished video). Space/PageDown/PageUp/Home/End/Esc all work (the footer lists every key; pressing →/← at the first/last frame flashes an "already first/last" hint instead of doing nothing). Overview thumbnails travel inside `play/thumbs/` (copied from `preview/*.png` at generation; slides without a capture get a counted placeholder and a terminal warning). External automation should use the **`window.playPage` API** (`state` getter + `go/next/prev/setFx/setNarr/setTrans/overview`) instead of simulating keys and scraping `innerText`. Snapshots inject the measured delays from `timings.json` (the level-by-level animation shares its source with the finished video); the narration panel decides automatically from the data whether to load at all, collapses into a bottom drawer with default-collapsed state in narrow windows/on phones, and the canvas adapts to landscape or vertical |
+| `scripts/build-video.mjs <project> [--asr] [--dry-run] [--transition cut|xfade]` | Encodes each slide → concatenates → aligns the audio track → muxes → self-checks + emits `out/subs.srt`; `--asr` splits the audio **by sentence** + generates a checklist; `--dry-run` only prints the ffmpeg commands that would run (for troubleshooting). **Transitions default to a hard cut** (no black frame between segments; the first segment dissolves in from the cover and the last fades out), and `--transition xfade` or `transition` in `script.json` switches to a 0.4s cross-dissolve (the dissolve eats the extra frames kept at the end of each segment, so the total duration is unchanged). When subtitle stills exist it splices as "frame-sequence segments + subtitle segments" (the segment lengths must sum exactly to the slide's duration) |
+| `scripts/grab-frames.mjs <project> [--ids 05,11] [--at 0.5] [--both]` | **Finished-video frame extraction check**: compute each slide's absolute start from `timings.json` and pull frames from `out/final.mp4` into `build/introspect/` — problems like a caption covered by subtitles, the last-level element missing from the frame, or too little still time at the end are "invisible in a still, only visible in the finished video", so go through them slide by slide before publishing |
+| `scripts/asr.mjs <project> [--provider mmx\|api] [--api-key K] [--verify-timing] [--from <transcript>] [--allow-any-endpoint]` | Calls ASR to transcribe and verifies sentence by sentence whether the audio says what the script says (mismatched numbers/traditional characters count as ✗); `--verify-timing` uses character-level timestamps to measure sentence onset. Defaults to `mmx speech transcribe` (mmx-cli ≥ 1.0.26, same login as TTS — no key handling here); `--provider api` calls REST directly instead, and then the key is only ever sent to official domains |
+
+Internal modules (imported by the scripts above, never run on their own): `tools.mjs` (ffmpeg/ffprobe probing + path jail `safeId/safeRel/inside`), `url-policy.mjs` (ASR endpoint allowlist + SSRF/redirect policy — DNS fails closed, and the veto rides the connection's own lookup: `checkedLookup`/`policyGet` for downloads, a loopback veto proxy for browser traffic — plus download filenames), `nofx-css.mjs` (single source of truth for the `no-fx` rules), `chart-css.mjs` (single source of truth for chart animations and chart primitives), `table-css.mjs` (table primitives: `.tbl/.kv/.matrix/.rank`), `css-kit.mjs` (the **managed-region mechanism**: rev delimiter comments + `findAllBlocks` + in-place replacement + the status vocabulary `ok`/`stale`/`duplicate`/`broken`/`legacy-outside`/`missing`, shared by `--upgrade-css`/`--check-css`/check-slides/preview-page/capture/build-video), `tokens-template.mjs` (the single source of the generated tokens.css body + its content hash `TOKENS_REV`), `limits.mjs` (the 2 MB scan cap that applies to both tokens.css and slide HTML). The whole generated body is wrapped in a `tokens` managed region **with the three toolkit blocks nested inside**, so after any source module changes, `--upgrade-css` propagates exactly into old projects; `BLOCK_IDS` in css-kit is the single list of region ids.
+
+Requirements: Node 18+ (scripts use fileURLToPath for compatibility, so they don't depend on Node 20.11's import.meta.dirname), `npm i playwright && npx playwright install chromium` (inside the project directory). ffmpeg/ffprobe are auto-detected: PATH → node_modules (ffmpeg-static/ffprobe-static) → common install locations; if they can't be found you get item-by-item diagnostics instead of a cryptic error.
+
+## Installing into other Agent environments
+
+This skill is the standard "one folder + SKILL.md at the root" shape (the frontmatter `name` / `description` are already written to spec), so dropping it into the corresponding skills directory is enough for it to be discovered. **Copy the reviewed skill directory out of the hosted plugin — not from any other source**: the hosted directory is the publication unit, and a copy taken from anywhere else has not been reviewed.
+
+```bash
+# from a checkout of the MiniMax-Code-Plugins repository:
+cp -r plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode ~/.claude/skills/        # Claude Code (personal scope) / OpenClaw / other AgentSkills hosts
+# project-scope install
+cp -r plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode <your project>/.claude/skills/
+```
+
+After installing, you only need to add two **dependencies** (in your video project, not in the skill directory):
+
+```bash
+cd <your video project>
+npm i playwright && npx playwright install chromium        # for screenshots
+# ffmpeg, pick one: winget install Gyan.FFmpeg / brew install ffmpeg / apt install ffmpeg
+# or: npm i ffmpeg-static ffprobe-static
+```
+
+**Why the skill still runs when installed elsewhere**: `scripts/tools.mjs` resolves `playwright` through a list of anchors — "the skill's own location → the project directory → the working directory at call time → npm global"; ffmpeg/ffprobe are likewise probed in four tiers (PATH → the project's node_modules → two levels up from the skill → common install locations). So "skill in `~/.claude/skills/`, dependencies in the project" is a supported setup (measured: the skill still renders images even when it lives outside the project tree).
+
+## Runtime: two toolchains (same scripts, only the tool source changes)
+
+The script layer (screenshots / rendering / compositing / validation / images) is completely environment-agnostic; **only TTS, music and ASR depend on platform capabilities**.
+
+**Environment rule (hard): in the mcode sandbox use the left column only — do not install or call mmx-cli there; `mcode-tools` covers TTS, music and ASR. mmx-cli is the fallback toolchain for non-mcode hosts only.**
+
+| Capability | mcode sandbox (preferred) | Other Agent environments (Claude Code / OpenClaw / Cursor, etc.) |
+|---|---|---|
+| TTS synthesis | `mcode-tools connector call connector__matrix__batch_text_to_audio --args '{...}'` (≤10 items per batch, the main path); for a single voice test use `connector__matrix__synthesize_speech` | `mmx speech synthesize --text "第一句口播。" --voice <voice_id> --speed 1.1 --out audio/01.mp3`; voice list via `mmx speech voices` |
+| Writing results to disk | `get_asset_url <node_id>` → download to `audio/<id>.mp3` | `--out` writes straight to disk |
+| BGM music | `connector__matrix__batch_text_to_music` (≤5 items per batch) | ⚠ **mmx-cli has no music generation** → have the user provide a music file (confirm licensing, then register it in MANIFEST), or skip BGM |
+| Reverse ASR verification | `mcode-tools upload_temp_url` + `connector__matrix__listen_audio` | **`node scripts/asr.mjs <project>`** — transcribes via `mmx speech transcribe` (mmx-cli ≥ 1.0.26; same mmx login as TTS, no separate key juggling), with no dependency on mcode and no need to install whisper; `--provider api` + `MINIMAX_API_KEY` is the direct-REST fallback for environments without mmx-cli. **It automatically compares against the expected text in the checklist and back-fills it, and mismatched numbers/traditional characters (Cantonese) are marked ✗ outright**. To measure sentence onset with character-level timestamps: `--verify-timing` |
+| Image assets | Built-in browser inspecting the official site's DOM (preferred; take the image URL to disk with `scripts/fetch-official-images.mjs --url`) / official brand kit | Same image paths as on the left (in this environment use `fetch-official-images`, which ships its own Playwright headless browser for rendering pages); for abstract illustrations use `mmx image generate --prompt "..." --aspect-ratio 16:9 --n 3` (use `9:16` for vertical projects; **abstract concept images only — generating logos / screenshots / real people's faces is forbidden**), then register it per `image-sources.md` |
+| Research | `web_search` / `web_fetch` (built in); for SPA/JS-rendered pages open them in the **built-in browser** to read the body text (usage discipline in references/research.md) | `mmx search "<keyword>"` / `mmx text chat`; for SPA pages render the body text with the project's Playwright **headless browser** (minimal command in research.md) |
+
+**First-time mmx-cli setup** (non-mcode environments): `npm install -g mmx-cli` (TTS works on any version; ASR needs ≥ 1.0.26 for `speech transcribe`) → `mmx auth login --api-key sk-xxx` → verify with `mmx quota`. A 401 is usually a region mismatch: `mmx config set --key region --value cn|global`. On the script side, ASR reuses that same mmx login; set `MINIMAX_API_KEY` (+ `MINIMAX_REGION=cn|global`) only if you force the direct-REST path (`--provider api`).
+
+**Discipline does not change with the environment**: durations still come from ffprobe measurements, subtitles still come from `clauses[]`, voices still have to be auditioned and language-verified (mcode sandbox: `mcode-tools upload_temp_url` + `connector__matrix__listen_audio`; other environments: `asr.mjs` above), and no Gate is skipped. **Search and body-text extraction are not limited to the tools in the table above either**: if other search skills/plugins are installed locally (web search, web reading, etc.), or any headless browser (the project's Playwright counts), use whichever one can get the body text — only the tool changes; source grading, two-source cross-checking and scope annotation are not negotiable.
+
+## Workflow (7 phases · 6 Gates)
+
+```
+kickoff alignment (iron rule 4 — no gate, but it must happen first)
+Phase 0 research      → Gate 0 fact list
+Phase 1 script design → Gate 1 per-slide narration script
+Phase 2 TTS + timing  → Gate 2 audition + duration table
+Phase 3 assets        → Gate 3 asset list + preview + compliance check
+Phase 4 HTML          → Gate 4 final-state screenshots
+Phase 5 render + ASR  → Gate 5 finished video
+Phase 6 delivery
+```
+
+**The order is scaffolding, not advice.** The narration script is final before TTS; TTS durations are measured before you touch assets and HTML; the asset list passes Gate 3 before it goes into the pages; the HTML passes the final-state screenshot before rendering. Changing the narration script = re-run from Phase 2 (TTS is cheap, redoing it costs little; forcing changes onto stale durations is the real disaster). Each Gate presents the user with exactly what its "acceptance-artifact list" specifies, and you never move forward without an OK — even when it looks obvious, confirm.
+
+**To resume/continue an old project, do one thing before starting work**: `node <技能>/scripts/init-project.mjs <project> --check-css`. If it reports anything other than `ok` (stale / missing / duplicate / broken delimiter / leftover copy outside the region), run `--upgrade-css` — it replaces the managed region in place by content rev, leaves everything outside it alone, and backs up `.bak` automatically. The same check now **blocks** `check-slides`, `capture` and `build-video`: an old region still renders, just with stale rules, so a stale project would otherwise ship the wrong picture with no error anywhere. **Note that the upgrade only changes tokens.css** — already-generated `preview/*.png`, `build/frames/` and `out/*.mp4` still hold frames from the old CSS: affected slides need capture re-run (still + motion), preview-page re-run, then build-video. To tell which Phase a project stopped at: `research/notes.md` has content → past Gate 0; `script.json` clauses filled in → Gate 1; `audio/*.mp3` complete → Gate 2; `assets/` complete → Gate 3; `slides/*.html` complete → Gate 4; `build/frames` or `out/*.mp4` present → Gate 5.
+
+### Phase 0 · Research (conditional)
+
+**Full method in `references/research.md`** (source grading, query design, handling contradictions, notes template). Key points:
+
+- The test is simple: if the finished video will contain specific **numbers, dates, names, quotes or attributions** → you must research. Generic/lyrical/creative subjects can skip it, but any sentence that "looks like a fact" still has to be verified.
+- **Four hard rules**: ① key numbers need **at least 2 independent sources** (with only one, use hedged wording or downgrade to an approximation); ② **primary sources first** (official announcements/earnings reports/technical reports/government statistics); second-hand retellings must be traced back to the original; ③ label the **scope and date** (annualized or single quarter? weekly active or monthly active? which currency?); ④ **anything with no traceable source or that can't be determined goes into "uncertain items" — never into the narration script**.
+- **How to get the body text**: try `web_fetch` first; SPA/JS-rendered official sites and investor-relations pages often return an empty shell, so switch to the **mcode built-in browser** (a host capability of the mcode environment) or a **Playwright headless browser** (any environment, already installed in the project) and render before reading the body text (minimal command in references/research.md); prospectuses/earnings reports are PDFs — take the numbers from the PDF text itself. When press releases and media retellings conflict, the official text wins. If other search skills/plugins are installed locally, they're equally usable — tools are open, discipline is not.
+- **Regulated domains** (finance/healthcare/legal/policy/marketing claims → read `references/compliance.md`) add three more constraints: ① financial numbers must carry **scope + currency + timestamp** ("Q2 revenue", not "current revenue"); ② the narration **gives no operational advice, does not predict prices, and uses no absolute language**; ③ the disclaimer and data sources (whether, how and where) are settled with the user at Gate 0, not left to cause rework at the finished-video stage.
+- Output `research/notes.md`: each entry has source URL + scope date + grade + second source; also list "uncertain items" and "content that must not go into the script"; for regulated subjects add a line recording **the confirmed domain + disclaimer wording**.
+- **Gate 0**: the fact list goes to the user — focus on having them confirm **numbers, names and scopes**; for regulated subjects, confirm the disclaimer, data timestamps and up/down colors at the same time.
+
+### Phase 1 · Script design (content volume is controlled here)
+
+Fill it in slide by slide in `script.json` (the contract file that every later script reads):
+
+```json
+{
+  "topic": "OpenAI 一分钟", "voice": "Chinese (Mandarin)_Gentleman",
+  "speed": {"default": 1.1, "first": 1.05, "last": 1.05},   // 试听时与用户定的语速(1.1 ≈ 中文 5.3 字/秒)
+  "transition": {"type": "cut"},                              // 切页方式: cut 硬切(默认) / {"type":"xfade","duration":0.4} 交叉溶解
+  "fps": 30, "width": 1920, "height": 1080,
+  "bgm": "assets/bgm.mp3",
+  "slides": [
+    {"id": "03", "layout": "bullets", "html": "03-lines.html", "audio": "03.mp3",
+     "title": "三大产品线",
+     "clauses": [
+       {"stage": 1, "text": "三大产品线。"},
+       {"stage": 2, "text": "ChatGPT 对话,GPT 系列模型,Sora 视频生成。",
+        "text2": "Chat, models, video generation."}
+     ]}
+  ]
+}
+```
+
+Optional fields: `clauses[].text2` = the second line of bilingual subtitles (omit it for Chinese-only subtitles); `clauses[].say` = the **spoken form** of that sentence for TTS and ASR verification (omit it to speak `text` as written); top-level `bgm` = `"assets/bgm.mp3"` or `{file, volume:0.12, fadeIn:1.5, fadeOut:2.5}` (when set, build-video loops it + fades it in and out underneath automatically).
+
+**Display form vs spoken form** — `text` is what goes on screen, `say` is what gets spoken. On-screen text always keeps the natural written form (`2026年`, `82.3%`, `SWE2.3`); when the broadcast reading differs, keep `text` as is and add `say`: years are read digit by digit (`say: "二零二六年…"` — TTS otherwise reads `2026年` as "两千零二十六年"; plan-timings warns), percentages/decimals speak the units (`82.3%` → `say: "…百分之八十二点三…"`), benchmark/model names follow reading habit (`SWE2.3` → `say: "SWE二点三"`). `say` never reaches the burned-in subtitles or `out/subs.srt` — those always come from `text` (see references/tts-and-timing.md).
+
+**Content-volume hard rules** (detailed version in `references/authoring.md`):
+
+- Chinese narration ≈ 5.3 characters/sec (speed 1.1; ≈4.8 at speed 1.0). Target 6–10 seconds per slide → 25–48 characters of narration; 12–20 characters for the first and last slides. Hard limit is 60 characters per slide; split into two slides if it goes over.
+- **Every slide (except the first and last) must have at least two information layers, mapped to different stages — but the order of the layers is free** (a big number can enter first and the title later; the question first and the answer after; it's decided by the strong-sync principle: a layer hangs on the stage of the sentence that mentions it). Title only, no expansion = a violation, send it back. Same for the narration: at least two sentences per slide, each driving a different layer.
+- Each element of `clauses` is one narration sentence; `stage` declares "when this sentence starts, which visual layer should appear". The number of stages ≈ the number of clauses, one-to-one.
+- There are 17 layouts in total (the original 8 + 9 added: kpi-grid / stat-highlight / table / timeline / roadmap / comparison / flow-diagram / terminal / big-quote); each one's required frame elements and character-count range are in the layout table in `references/authoring.md`; **charts (bars/columns/donut/line/progress) have a pure CSS/SVG cheat sheet — no external chart library needed**.
+- **For data-dense subjects, proactively make charts**: when the narration contains ≥2 comparable numbers, proportions or a trend, prefer the `data-viz` layout + chart toolkit over listing numbers as bullets — whether to chart and in what style was already asked in the "Data and charts" row of kickoff alignment; if the user didn't raise it then, offer the option proactively (charting is built in, don't leave it idle). Numbers must come from verified scopes (Gate 0), and bar heights/lengths are computed from the values (see the nine disciplines in authoring.md).
+- If the user asked for **bilingual subtitles** during kickoff alignment, write `text2` for every sentence here (a translation of the same sentence, no reordering, ≤60 characters).
+
+**Gate 1**: the user reviews the slide-by-slide narration script + layout assignment, presented per slide as "# / layout / frame / **table and animation** / verbatim narration". **The table-and-animation column is a required part of the presentation**: slides that use table primitives must state the table's shape (how many rows and columns, which primitive `.tbl/.matrix/.rank/.kv`, which row/column is highlighted); slides with entrance animations must name the effect and the order (e.g. "three cards fade in one by one, staggered 150ms", "the number counts up to 92.4%", "the bars grow in from the left one after another") — the user has to be able to play the finished video once in their head from this table, instead of seeing it move for the first time at Gate 4. **No Phase 2 until every slide is OK'd.**
+
+### Phase 2 · TTS + measured timing sync
+
+Which toolchain TTS uses is in the "Runtime" comparison table above (mcode uses `connector__matrix__*`, other environments use `mmx speech synthesize`); command templates and retry discipline are in `references/tts-and-timing.md`. Key points: batches ≤10 items; on partial failure, sleep 10–30s and **retry only the failed items**; first and last slides use speed 1.05 (but keep 1.1 for short sentences ≤10 characters — see the measured pitfalls in that file). Output goes to `audio/<id>.mp3`. **The audition step must ask about voice and speed together** (one sentence each from 3 candidate voices + the top-choice voice at all three speeds); only batch-synthesize once the user has decided.
+
+Then sync the timing:
+
+```bash
+node <skill>/scripts/plan-timings.mjs <project>
+```
+
+It does this: ffprobe measures each segment → per-slide duration = measured + tail padding (default 0.8s, adjustable via the slide's `tail` field: tight 0.4 / relaxed 1.2) → each stage's entrance moment = the onset estimated from that sentence's share of the characters − 0.2s (visuals slightly ahead of the voice, which reads as in sync) → each sentence's onset/duration written into `clauses[]` (shared by subtitles, ASR sentence splitting and timing calibration) → writes `build/timings.json` and prints warnings (abnormal speech rate, over 15s, last stage too close to the ending, sentences that would wrap past two subtitle lines — line width scales with the canvas — ≈24 CJK chars per line at 1080 wide, ≈33 at 1920, one-line craft guideline ≈18; two lines are fine).
+
+Optional but recommended (especially when the user has reported "audio and visuals out of sync"): `node <skill>/scripts/check-timing.mjs <project>` measures each sentence's true onset with silence detection and prints an "estimate vs measured" comparison table; if the deviation is large, `--calibrate`, then delete `build/frames/` and re-render. Note that "feels out of sync" is often a design misalignment — a big number/hero image must hang on the stage of **the sentence that mentions it** (strong-sync principle, see authoring.md).
+
+**Gate 2**: the user listens through every narration audio segment + reviews the duration table from plan-timings.
+
+### Phase 3 · Asset collection (compliance is enforced here)
+
+Follow the image SOP in `references/image-sources.md` (core principle: **for images with an incomplete subject / cut off halfway / wrong aspect, search again for the full view first — don't force a crop**). Key points:
+
+1. Write the requirements list (which slides need images, and what kind).
+2. **Priority: official channels (brand kit / simple-icons / official documentation screenshots) → open the official site in the built-in browser or Playwright and inspect the DOM for official assets (highest measured success rate) → Wikipedia CC → pure-typography fallback (number cards + large type + quotes, no images). Negative-event subjects (blow-ups / penalties / lawsuits / controversies) won't have assets on the official site, so switch to **announcement and news-report screenshots**: exchange/regulatory announcements, company statements, or screenshots of mainstream media report pages (top-tier authorities and portals like Tencent News/Sina are both fine; self-media excluded), and the frame must keep the outlet name and date (commands and discipline in image-sources.md path D).** The old image-downloader (Bing) is off by default: on niche Chinese-language subjects it returned irrelevant images 5/5 times.
+3. Add positive keywords when searching for images (panorama/full view/front view/wide shot), and screen with thumbnails before downloading: subject centered, aspect close, no watermark, no unrelated logos, ≥1200px; if one doesn't fit, swap it — don't settle.
+4. Once you have an image, first run `node <技能>/scripts/prep-image.mjs --check <图>` to check dimensions and crop risk; if the subject touches the edge, swap the image.
+5. Compliance-check every asset: no watermark, trademarks only in a fair-use context, photos from licensable sources, screenshots (official docs or news/announcements) must keep their source and date and be registered in MANIFEST.
+6. Save to `assets/`, one line per asset in `assets/MANIFEST.md` (content/source/license).
+7. **Generating logos, screenshots, avatars or QR codes out of thin air is forbidden; putting watermarked images into assets is forbidden; dropping a bare `<img>` is forbidden (it must be wrapped in `.img-frame`).**
+
+**Gate 3** (after asset collection is done and both the facts and candidate images are in hand): give the user the asset-list table + a preview of each asset (thumbnail/description), and ask three things: ① **which ones to use** — have the user pick slide by slide (annotate recommended usage, e.g. "slide 02 with the hero image"); whatever they cross out is dropped; ② **is there anything to add** — links (official pages/report pages/data pages), screenshots, logos, pictures, data files, reference narration scripts can all be handed to the agent directly; **user-provided assets have the highest priority** (they save searching and the source is reliable): images go to `assets/`, pass `prep-image --check`, and are registered in MANIFEST as "source: user-provided"; links are saved to disk per image-sources.md paths A/B/D or used as Gate 0 fact sources; ③ "are the asset sources and licensing fine". **Only write the HTML after the user has signed off.**
+
+### Phase 4 · HTML (staged entrance is implemented here)
+
+One file per slide, in `slides/`, with the filename matching the `html` field in script.json. Conventions:
+
+- The root element `<html data-theme="...">` selects the theme (the initial ones `a|b|c`, plus `minimal-white / swiss-grid / corporate-clean / editorial-serif / magazine-bold / tokyo-night / catppuccin-mocha / nord / xiaohongshu-white / soft-pastel`; theme cheat sheet and contrast verification in authoring.md); import `tokens.css`; the frame container is `.stage`, containing the `.brand` corner badge and the `.slide-num` page number.
+- **Every block that should enter gets `data-stage="1|2|3"` + one fx utility class** (`fx-up/fx-fade/fx-grow/fx-blur/fx-rise/fx-pop/fx-spotlight/fx-ripple/fx-glitch/fx-draw`; ambient ones `fx-pulse/fx-shimmer/fx-kenburns` take no data-stage); don't write delays — the pipeline injects `--t1/--t2/--t3` from `timings.json`. For staggering within a level use the `.fx-stagger` container (baseline `style="--stagger-base:var(--t2)"`, which is also the default) or inline `style="animation-delay:calc(var(--t2) + 150ms)"`.
+- ⚠ **`data-stage` must be used together with an fx class** (the attribute alone with no animation class leaves it stuck at opacity:0 forever); for how delays are implemented see "how stage delays work" in authoring.md, and don't hand-write `animation-delay: var(--tN)` when changing animations.
+- Always wrap images in `.img-frame` (`.contain` for screenshots/charts; `--img-ratio` sets the aspect; `--img-pos` protects the subject; the `.img-cap` caption goes outside the frame).
+- Ambient animations (infinite looping breathing/floating) are allowed, but they must not carry information and get no `data-stage`.
+- **Animations can be switched off in one move**: `<html data-theme="..." class="no-fx">` turns them off for the whole project, adding `no-fx` to any container turns them off for a single slide, and an element with no fx class is simply static. With them off, motion capture automatically degrades to producing the video from static frames; durations and audio-visual sync are unaffected and subtitles work as usual (details in authoring.md "animation switch").
+- Do not use transitions for entrances (the screenshot pipeline can't seek to them); use only `@keyframes`. External Google Fonts are banned (unstable offline) — use a system font stack (tokens.css already configures a CJK fallback).
+- Use only the `assets/` list confirmed at Gate 3; add no unreviewed assets.
+- If you changed the project's `tokens.css` (custom brand color, A-share up/down color flip, subtitle hooks `--sub-bg/--sub-ring`) → first run `node <技能>/scripts/check-theme.mjs <project>` to verify contrast (skippable if the built-in themes are untouched). Discovering a contrast problem at Gate 5, in the finished video, is the most expensive rework there is.
+- ⚠ **After writing the 8 slides, run the static check before screenshotting**: `node <技能>/scripts/check-slides.mjs <project>` — it catches undefined CSS variables (which make text invisible), missing images/external resources, data-stage without an fx class, and hard-coded colors. Any ✗ and don't screenshot: a frame that is correct but "invisible" is the hardest thing to debug.
+
+Once the slides are written, produce the final-state preview (the fastest path, for Gate 4):
+
+```bash
+node <skill>/scripts/capture.mjs <project> --mode still
+node <skill>/scripts/preview-page.mjs <project> --open   # 放映页: 让用户自己过一遍动效
+```
+
+**Gate 4**: walk the user through `preview/<id>.png` slide by slide (layout, text pressure, assets) + **hand the play page `preview/play/index.html` over for the user to present themselves** (are the animations really all appearing, is the speed acceptable, does the frame hold up). Static images can't show animation — failures like "the element never appears", "turning animations off leaves a blank frame", "column heights collapse" only surface when it moves or when you toggle between the two; the play page is a 30-second self-check, so there's no need to wait for the 3–6 minute motion encode. Timing and whether the narration lines up are not judged here (the play page deliberately has no timer) — the finished video and `timings.json` are the source of truth.
+
+**This step must ask the user two questions (since 1.6.0; per iron rule 4, ask as questions + options and stop to wait for answers)** — both can only be decided by "seeing how it moves"; static images and written descriptions can't substitute:
+
+1. **Slide transition**: hard cut (default, recommended) / 0.4s cross-dissolve / custom dissolve duration. Have the user press `T` in the play page to **compare live** (the play page switches between hard cut and dissolve instantly; the dissolve is a 0.4s blend that never passes through black), and once decided, write it into `transition` in `script.json` (`{"type":"cut"}` or `{"type":"xfade","duration":0.4}`).
+2. **Entrance layering**: should the big title and the sub-information **all appear at once**, or be **revealed level by level** (the current `data-stage` default). For "at once", collapse that slide's `data-stage` levels into a single level (title and sub-items on the same stage, or drop the later stages), then re-run that slide's `capture --mode motion` and rebuild; in the play page, verify the collapsed result level by level with "animations on + →".
+
+Once both are decided, restate "final choices: transition = X / entrance = Y". `check-slides.mjs` prints the measured status of both at the end (the transition value + the level count per slide) — ask based on that, not on impressions. **Only enter Phase 5 after the user has signed off.**
+
+### Phase 5 · Render + ASR verification
+
+```bash
+node <skill>/scripts/capture.mjs <project> --mode motion   # entrance animations rendered frame by frame into the video
+node <skill>/scripts/build-video.mjs <project> --asr
+```
+
+- `--mode motion`: frame-by-frame stepping (pause all animations → seek frame by frame → screenshot → encode); the animation window is rendered frame by frame and still segments get the tail frame automatically, so durations are exact. **Subtitles are burned in by default** (content from clauses; the display window = from that sentence's onset to the next sentence's onset; centered at the bottom of the frame; not shown in still previews, only in the finished video; `--no-subs` turns it off). Cost is about 50–200ms per frame — 8 slides × 30fps is roughly 3–6 minutes, so budget for the timeout. In a hurry you can produce the video in still mode (animations don't make it into the video, only fades).
+- ⚠ **Re-running a still capture invalidates that slide's frame directory** (capture's design to prevent stale-frame contamination). If you changed the HTML after Gate 4 and want another look → `capture --mode still --ids <id>`, and **after confirming, you must re-run `--mode motion` for the same ids before build-video**; otherwise that slide falls back to a static image in the video and its animations are silently lost (build-video calls this out with a ⚠).
+- build-video does this automatically: encode each slide (uniform parameters) → concat (automatic fallback re-encode on duration drift) → align the audio track with `apad` using each slide's measured duration → **bed the BGM underneath (only when bgm is configured: loop to fill, fade in/out, voice first; falls back to voice-only if mixing fails)** → mux → ffprobe duration check + full decode self-check, non-zero exit code if it fails; it also outputs `out/subs.srt` (same source and same windows as the burned-in subtitles; bilingual subtitles are generated automatically as two lines from each clause's text2, for platform upload).
+- `--asr`: splits out `asr/part-<id>-<k>.mp3` **by sentence** + generates `asr/checklist.md`. The transcription is compared with the expected text: numbers, years and product names must match; homophones are tolerable. **If a segment's transcription bleeds in the start of the previous sentence, that sentence actually started later than estimated — run check-timing to calibrate.**
+  - mcode: upload with `mcode-tools upload_temp_url`, then hand off to `connector__matrix__listen_audio`.
+  - **Other environments**: `node scripts/asr.mjs <project>` — transcribes via `mmx speech transcribe` (mmx-cli ≥ 1.0.26, same login as TTS), compares and back-fills the checklist automatically; failing items (mismatched numbers/traditional characters) are reported with a non-zero exit code. Without mmx-cli: `MINIMAX_API_KEY=sk-xxx node scripts/asr.mjs <project> --provider api` calls REST directly with the same Key. For more accurate onset times: `--verify-timing`.
+  - For a slide that fails: change the narration or redo that segment's TTS → re-run plan-timings → re-render that slide (just delete the corresponding slide's frame directory).
+
+**Gate 5**: the user reviews the finished video `out/final.mp4` + the ASR verification table, including a subtitle readability check (play it once muted — can the subtitles carry the meaning) and the BGM level (is the voice always clear). **Before publishing, pull real frames from the finished video and look at them slide by slide**: `node <技能>/scripts/grab-frames.mjs <project>` → `build/introspect/` — whether captions/credits are covered by subtitles, whether all the last-level elements made it into the frame, and whether the ending holds still long enough; these three are invisible in both stills and the play page.
+
+### Phase 6 · Delivery
+
+```
+out/final.mp4            # main deliverable
+out/slide-*.mp4          # per-slide segments (publishable on their own)
+out/cover.png            # cover image (canvas size; the video embeds it as attached_pic for thumbnails, this file is for platforms that want a manual upload)
+out/subs.srt             # subtitles for platform upload (same source and windows as the burned-in ones)
+build/timings.json       # single source of truth for timing (read by check-timing / asr / preview-page)
+build/substills/         # one still per subtitle sentence + window manifest (windows the frame sequence cannot cover; build-video concatenates them)
+build/audio-timeline.wav # aligned audio timeline
+preview/*.png  preview/play/index.html   slides/*.html  slides/tokens.css
+audio/*.mp3  assets/(含 MANIFEST.md)  research/notes.md  asr/(校验记录)
+```
+
+(`slides/tokens.css.bak` is the pre-upgrade backup made by `--upgrade-css`; delete it once the finished video checks out.)
+
+## Reference files (read on demand, not all of them)
+
+- `references/authoring.md` — the 17 layout specs (required frame elements for each) + the content-volume table + how to use the entrance system (including how stage delays work) + the **chart toolkit** (chart primitives, the nine elegance disciplines, 7 recipes and animation selection) + the **table toolkit** (four primitives, seven disciplines, five shapes) + theme cheat sheet + vertical notes + copy-pasteable HTML snippets
+- `references/compliance.md` — domain and compliance (required reading for regulated subjects): how to confirm the domain, red lines for financial narration, the three requirements for numbers, flipping up/down colors by audience, how and where to write the disclaimer, key points for medical/legal/advertising law, and the closing checklist
+- `references/research.md` — research (source grading, hard cross-validation rules, query design, handling contradictions, notes template)
+- `references/image-sources.md` — image and asset SOP (the three image paths, positive/negative query terms, two-stage screening, how to use the image frame, hard crop limits, the three-piece visual verification, and an index of common subjects)
+- `references/tts-and-timing.md` — mcode TTS connector commands, retry discipline, the timing algorithm, tail-padding and speech-rate tuning, and the measured voice table
+- `references/render.md` — how rendering works (why frame-by-frame stepping, how fonts are awaited), the subtitle system, cover and slide transitions, BGM mixing, manual ffmpeg commands, and a troubleshooting table
+- `references/symptoms.md` — **the long symptom list** (moved out of the `description`, which the host caps at 1024 characters): symptom → the script/section that usually owns the fix. Reach for it when the user describes a problem rather than asking for a video
+- `THIRD-PARTY-NOTICES.md` — third-party component license notices (10 themes and some CSS primitives adapted from html-ppt-skill, MIT)
+
+## Common symptoms → one-line diagnosis
+
+> The full symptom index (symptom → the section that usually owns the fix, including the ones that only make sense as *trigger* phrases) is `references/symptoms.md`. The `description` field only carries a short pointer to it because the host caps that field at 1024 characters.
+
+| Symptom | Root cause | Action |
+|---|---|---|
+| The frame lingers long after TTS finishes speaking | Durations were estimated rather than measured, or the tail is too large | Re-run plan-timings; adjust that slide's `tail` |
+| The frame has only a title, no expansion | Violates the content-volume hard rule | Add an expansion layer + its clause, go back to Phase 1 |
+| Animations aren't in the video | Still mode was used | Re-render in motion mode |
+| Audio and visuals feel out of sync | One of three: hand-written animation delays / estimation drift / the visual anchor hung on the wrong sentence (strong-sync principle) | First remove the hand-written delays; run check-timing for measured data and --calibrate if the drift is large; if the anchor is on the wrong sentence, fix the stage mapping |
+| The voice is in the wrong language (Cantonese/traditional characters) | The platform's voice labels are mismatched; names can't be trusted | The Phase 1 audition must include ASR voice verification (see the measured voice table in tts-and-timing.md) |
+| The finished video has no subtitles | --no-subs was used, or clauses are missing | capture burns them in by default; confirm timings.json has clauses |
+| Bilingual subtitles wanted | — | Add `text2` to the clause; the frame shows two lines + the SRT gets two lines automatically; keep the second line ≤60 characters |
+| Want background music | — | Configure top-level `bgm`; build-video loops it + fades it in/out underneath automatically (default volume 0.12); ASR verification still runs on the voice-only track |
+| In a hurry / the subject calls for restraint, no animation | — | `<html class="no-fx">` turns off all animation in one move, or add `no-fx` to a single slide's container; the video is then produced from static frames with unchanged durations (see authoring.md) |
+| Total duration is wrong after concatenation | Segments mixed from different encoder parameters | Let build-video encode all segments uniformly; it already falls back to re-encoding automatically |
+| Chinese text shows as boxes | No CJK font on the system | On Linux install fonts-noto-cjk; or switch to an installed font |
+| A "can't find ffprobe/ffmpeg" error | The binaries aren't on PATH | The scripts already auto-detect PATH→node_modules→common locations; install ffmpeg-static or run winget install Gyan.FFmpeg |
+| **Some elements enter at 0 seconds, others on schedule** | The delays in an old tokens.css are overridden by the `.fx-*` shorthand | Switch to the new tokens.css (delays go through `--fx-delay`); rationale in authoring.md |
+| The image subject is cropped out of frame / the image breaks the layout | A bare `<img>`, or cover with the wrong aspect | Wrap in `.img-frame` + `--img-pos` to protect the subject; use `.contain` for screenshots; if the subject touches the edge, re-search the image per the SOP |
+| Subtitles blur into the background on dark themes | The theme doesn't override the subtitle hooks | Add `--sub-bg` (darker) + `--sub-ring: 1px solid rgba(255,255,255,.16)` to that theme; run check-theme to verify |
+| Vertical output wanted (Douyin/WeChat Channels) | — | Set `width:1080, height:1920` in `script.json` and switch the layout to stacked (see the vertical section of authoring.md) |
+| Numbers/text are written but invisible | An undefined CSS variable + `-webkit-text-fill-color: transparent` invalidates the whole background declaration | Run `check-slides.mjs` to locate it, then add the definition or write `var(--x, 默认值)` |
+| A chart/line/some element is completely missing after entering | That fx class's keyframes don't declare opacity — the base state of `[data-stage]` is `opacity:0`, and an animation that only does transform/stroke can't bring it back | Add `opacity:1` to the keyframes' from/to; `check-slides.mjs` reports this directly |
+| Turning on `no-fx` makes the frame emptier instead | Early templates only turned animations off without restoring the `[data-stage]` `opacity:0` base state | The project's tokens.css is outdated: `node scripts/init-project.mjs <project> --upgrade-css` updates the managed region in place (idempotent); when the play page meets an old tokens.css it injects the current body into its **snapshot copies** and says so explicitly — note that a correct preview page does **not** mean the rendered video is current, so upgrade before capture/build-video |
+| The skill's CSS changed (or you used a new class like `.chart-ticks`) but the project doesn't pick it up | The project tokens.css's managed region lags behind the skill's single source (the old check only looked at presence, so after one patch it always reported "no upgrade needed"; duplicate blocks and copies outside the region won the cascade silently) | `node scripts/init-project.mjs <project> --check-css` shows the status; `--upgrade-css` replaces the region in place by content rev (leaving your own rules alone); **check-slides, capture and build-video all refuse to run while it is stale** |
+| Double-clicking `slides/*.html` shows the animation crammed at the start | Animation delays are injected by the pipeline from `timings.json`; tokens.css only holds placeholder values (`--t2:800ms`) | Don't open the original file directly: run `preview-page.mjs` and use its copy (with the measured delays injected), or produce the video with `--mode motion` |
+| Bar heights don't match the labeled values (tall bars squashed) | Bars placed directly in a flex column with a percentage height: the basis is the **whole column** height, and anything overflowing the remaining space is squeezed back by `flex-shrink`, silently distorting it | Wrap the bars in `.chart-plot-cell` (a 1fr row) so the percentage is computed against the plot area; you also get a baseline and equal-height columns for free |
+| The line chart shrinks into a small patch in the middle | The SVG's `viewBox` aspect doesn't match the container, so `preserveAspectRatio` scales the content down and centers it | For full width use `width:100%;height:auto`; for half width or a fixed height, change the viewBox to a ratio close to the container's |
+| A chart/caption is covered by the subtitles | Content was laid out into the subtitle band (bottom 84–168px, centered, 73% wide) | Start the layout container at `padding-bottom: 190px`; see the "subtitle safe area" section in authoring.md |
+| Tables are unreadable on phones | The table text uses the caption size (24px), or row height was squeezed below 72px | Use the `.tbl/.kv/.matrix/.rank` primitives: main data uses the body size (30px) and row height is 72px built in; see "table toolkit" in authoring.md |
+| The iframe in the play page is blank/broken images | The copy's `<base href>` was stripped, or slides/ was moved | Re-run `preview-page.mjs` to regenerate the snapshot; don't hand-edit the original file (the copy is a snapshot, so you must re-run after changing slides) |
+| Images show a broken icon | The file is missing, or the SVG itself is broken (bad XML / depends on external resources / no dimensions) | Use `check-slides.mjs` to check the path; inline the SVG into the HTML; capture also names which image failed to load at render time |
+| A finance video has no disclaimer / the up-down colors are inverted / numbers are challenged on scope | The domain wasn't confirmed at kickoff, so default colors and default wording were used as-is | Read `references/compliance.md`: add a closing `.disclaimer` line (on screen ≥3s), change metric cards to `var(--up)/var(--down)` flipped for the audience's market, and give every number its scope + currency + timestamp |
+| Not sure whether a disclaimer is needed / whether this counts as regulated | The domain wasn't confirmed | Ask "domain + disclaimer or not" in the first batch of kickoff questions; financial investment research, healthcare, legal, policy and marketing performance claims default to yes. `check-slides.mjs` warns when it hits finance keywords but sees no disclaimer line |
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/THIRD-PARTY-NOTICES.md b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/THIRD-PARTY-NOTICES.md
new file mode 100644
index 00000000..34f1deae
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/THIRD-PARTY-NOTICES.md
@@ -0,0 +1,65 @@
+# Third-party components and license notices
+
+This skill contains material adapted from third-party open-source projects. As the MIT license requires, the original copyright notices and the license text are reproduced below.
+
+---
+
+## html-ppt-skill
+
+- **Source**: https://github.com/lewislulu/html-ppt-skill
+- **Copyright**: Copyright (c) 2026 lewis <sudolewis@gmail.com>
+- **License**: MIT
+- **Date introduced**: 2026-09-17
+- **Scope of adaptation** (all rewritten to fit this skill's single-page static rendering and audio-visual sync system):
+  - `slides/tokens.css` generated by `scripts/init-project.mjs`:
+    - the color tokens of 10 named themes (`minimal-white` / `swiss-grid` / `corporate-clean` / `editorial-serif` /
+      `magazine-bold` / `tokyo-night` / `catppuccin-mocha` / `nord` / `xiaohongshu-white` / `soft-pastel`)
+      — the variable names are mapped to this skill's naming (`--surface`→`--panel`, `--text-1/2/3`→`--fg/--muted/--fg-3`,
+      `--border`→`--line`, and so on), and the external fonts (Playfair Display / Space Grotesk and others) were replaced with local font stacks
+    - the image-frame primitives `.img-frame` / `.contain` / `.fill` / `.img-scrim` / `.img-cap` / `.img-tag` and their
+      `--img-ratio` / `--img-pos` variables
+    - the keyframes of some entrance animations (`fx-blur` `fx-rise` `fx-pop` `fx-spotlight` `fx-ripple` `fx-glitch`
+      `fx-shimmer` `fx-kenburns` `fx-stagger`) — the trigger mechanism was changed from "JS re-triggering on slide change" to
+      "`animation-delay: var(--tN)` absolute time" so it works with frame-stepping rendering
+    - the typographic roles and container classes `.kicker` / `.eyebrow` / `.lede` / `.card` (with three variants) / `.pill` / `.divider-accent`
+  - Some layout skeletons in `references/authoring.md` were rewritten from the versions of the same name under its `templates/single-page/`
+  - The image-frame usage rules in `references/image-sources.md` follow the design intent of its `.img-frame`
+
+**Not adopted** (incompatible with its design or carrying extra dependencies): `runtime.js` (keyboard navigation / presenter mode / overview),
+the 20 canvas effects under `assets/animations/fx/` (rAF-driven, cannot be frame-seeked),
+`fonts.css` and every external Google Fonts link, the Chart.js / highlight.js CDN dependencies, and the `@media print` pagination logic.
+
+The "page through it once" need is served by this skill's own `scripts/preview-page.mjs` (which generates `preview/play/index.html`),
+and that code has nothing to do with upstream: the upstream runtime.js is built for a live human presentation (presenter window, teleprompter script, timer, BroadcastChannel two-window sync),
+whereas this skill produces a video and has no live speaking step; the play page only replays slide by slide + compares the motion switch + shows the voiceover panel.
+
+### The MIT license text
+
+```
+MIT License
+
+Copyright (c) 2026 lewis <sudolewis@gmail.com>
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+```
+
+---
+
+**Commercial use**: the MIT license permits commercial use, modification, redistribution and sublicensing; the only obligation is to retain the copyright and license notices above.
+The final artefacts this skill produces (video, subtitles, cover, and so on) are the user's own creations and do not inherit the license obligations above.
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/evals/evals.json b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/evals/evals.json
new file mode 100644
index 00000000..a368dcfc
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/evals/evals.json
@@ -0,0 +1,194 @@
+{
+  "skill_name": "html2video-for-mcode",
+  "evals": [
+    {
+      "id": 0,
+      "name": "full-run-from-outline",
+      "prompt": "帮我把这份大纲做成一条一分钟的中文口播视频,1920×1080:介绍 OpenAI——十年时间、三大产品线、九亿周活、25B 年化营收、92% 财富500强渗透、一句 Altman 引语收尾。",
+      "expected_output": "走完整工作流且顺序正确: 事实性题材先搜集过 Gate 0 → 逐张填 script.json(内容量硬规则: 每张标题层+展开层, 口播 25-48 字)过 Gate 1 → TTS 落 audio/ → plan-timings 实测时长过 Gate 2 → 按 authoring.md 写 8 张 HTML(data-stage+fx-*, 不写死延迟) → capture still 过 Gate 4 → capture motion + build-video --asr → ASR 逐段比对过 Gate 5。全程不手写任何时长数字。",
+      "files": []
+    },
+    {
+      "id": 1,
+      "name": "long-tail-silence",
+      "prompt": "之前用别的方法做的幻灯片视频有个毛病:每页 TTS 念完以后画面还要呆等三四秒才切页,很拖。你这套能解决吗?",
+      "expected_output": "解释根因是时长靠估算而非实测。给出本流水线的解法: TTS 先行 + ffprobe 实测 + slide 时长=实测+tail(默认0.8s), tail 可按紧凑/舒缓调 0.4/1.2; 并说明 stage 入场时刻按字数占比从实测时长推出, 音画同源。",
+      "files": []
+    },
+    {
+      "id": 2,
+      "name": "title-only-slide",
+      "prompt": "我看你出的第 3 张只有一个大标题,口播念了一整句,画面什么都没有,能不能这页先出标题,然后跟着口播出几条细节?",
+      "expected_output": "判定为违反内容量硬规则(画面视觉块≥2 且分属不同 stage)。修正: 补展开层(如 bullets 三条)挂 data-stage=2 + fx-up, 口播拆成引入句(clause stage=1)+展开句(clause stage=2), 入场时刻由 plan-timings 按展开句开口时间自动注入, 不手写延迟。",
+      "files": []
+    },
+    {
+      "id": 3,
+      "name": "tts-partial-failure",
+      "prompt": "batch_text_to_audio 返回 8 条里成功 5 条失败 3 条, 报 rate limit, 怎么重试?",
+      "expected_output": "sleep 10-30s 后只把失败的 3 条重新组一个小批重试, 绝不重发整批; 全部落盘 audio/ 后重跑 plan-timings 再继续。",
+      "files": []
+    },
+    {
+      "id": 4,
+      "name": "asr-mismatch",
+      "prompt": "ASR 校验发现第 5 张转写里 '92%' 听成了 '九二', 数字对不上, 怎么处理?",
+      "expected_output": "text(字幕)保持 '92%' 不动, 给该句加 say 换一个更清晰的读法(如 say:'…增长九成二…' —— 显示/发音分离: 发音形态只进 say, 不上屏; 换词要真的改变发音, 念作'百分之九十二'的同义改写解决不了听错)→ 只重做该段 TTS → 重跑 plan-timings → 删 build/frames/05 与 out/slide-05.mp4 → 该张重 capture + build-video, 不整片重做。若确认是同音字级微小差异可接受, 数字/产品名不一致必须返工。",
+      "files": []
+    },
+    {
+      "id": 5,
+      "name": "asset-gate-browser-route",
+      "prompt": "做一条介绍某中国 AI 公司的视频, 需要他们的 logo、CEO 照片和产品截图。素材这块你怎么弄?",
+      "expected_output": "按优先级执行: 官方渠道(brand kit/simple-icons/官方文档截图)优先; 缺的用内置浏览器或 Playwright 打开官网 inspect DOM 提取官方 CDN 资源(2026-09 实测中国 AI 公司 6/6 相关); image-downloader(Bing)默认不用(中文题材实测系统性返回无关图); 每个素材查合规(水印/商标语境/照片授权/截图出处), 落盘 assets/ 并逐条登记 MANIFEST.md; 素材清单+预览过 Gate 3, 明确问用户来源与授权是否 OK, 过了才写 HTML。禁止凭空造图; 无合适素材降级纯排版。",
+      "files": []
+    },
+    {
+      "id": 6,
+      "name": "subtitle-burn-and-srt",
+      "prompt": "成片需要有字幕(很多人在静音场景看), 还要能传抖音/B站。字幕怎么加?",
+      "expected_output": "不需要在 HTML 里写字幕: capture 默认把 timings.json 的 clauses 逐句烧进画面(底部居中胶囊, 显示窗=该句开口到下句开口, --no-subs 可关), build-video 同时输出 out/subs.srt 供平台上传。单句建议 ≤18 字(可读性惯例; 胶囊实际行宽 ≈24 字@1080 / ≈33 字@1920 —— 随画布缩放, plan-timings 折 3 行起才警告)。Gate 5 用静音播放验收字幕可读性。",
+      "files": []
+    },
+    {
+      "id": 7,
+      "name": "voice-language-verification",
+      "prompt": "音色列表里选了个'新闻女声 Chinese (Mandarin)_News_Anchor', 标着普通话, 直接用可以吗?",
+      "expected_output": "不可以, 标签不可信(该音色 2026-09 实测输出粤语)。Phase 1 试听必须带 ASR 验音: 3 个候选各单条 TTS 同一句探针文案(含数字专名) → listen_audio 转写 → 检查简体/繁体、数字念对; 把实测语种标进试听清单给用户三选一。tts-and-timing.md 附实测音色表可作起点。",
+      "files": []
+    },
+    {
+      "id": 8,
+      "name": "bilingual-subtitles",
+      "prompt": "成片要中英双语字幕, 给海外同事看。怎么加?",
+      "expected_output": "不用改 HTML: script.json 每个 clause 加 text2(同一句的翻译, 不重排语序, ≤60 字符), capture 自动把字幕渲染成两行(中文主行+次行小字号), build-video 的 out/subs.srt 同步双行。plan-timings 对超长 text2 会预警。只做部分句子也可以。",
+      "files": []
+    },
+    {
+      "id": 9,
+      "name": "bgm-mix",
+      "prompt": "想给成片加个背景音乐, 别盖过人声。",
+      "expected_output": "用 connector__matrix__batch_text_to_music 生成无人声、可循环的 BGM 放 assets/bgm.mp3(或用户自备并确认授权), script.json 顶层加 bgm 配置; build-video 自动循环补满+淡入淡出+按 volume(默认 0.12)垫在人声下, 混音失败退回纯人声; ASR 校验仍走纯人声轨不受干扰。Gate 5 验人声清晰度。",
+      "files": []
+    },
+    {
+      "id": 10,
+      "name": "image-sourcing-sop",
+      "prompt": "配一张学校大门的实景图, 搜到的图主体都在画面边缘或者只有半截, 比例也不对。怎么办?",
+      "expected_output": "按 references/image-sources.md 的 SOP: 核心原则是'先重搜全貌图, 别硬裁硬用'。query 加正向词(全景/全貌/正面/远景)排除词(特写/近景/一角/半截); 下载前缩略图筛(主体居中/比例接近/无水印无无关logo/≥1200px), 不合适就换; 拿到图跑 prep-image.mjs --check; 图片一律套 .img-frame, 用 --img-pos 控制主体可见区或 .contain 留白, 不要裸放 img 或硬裁(裁切硬限: 裁掉≤20%、不放大补边); 都不行就换题材(数字卡/大字/引言)。",
+      "files": []
+    },
+    {
+      "id": 11,
+      "name": "vertical-canvas",
+      "prompt": "同一条内容还要出一版竖屏的, 发给视频号。要改多少东西?",
+      "expected_output": "只改 script.json 的 width:1080 / height:1920 —— 渲染管线自动注入 --stage-w/--stage-h, 字幕几何(位置/宽度/字号)跟随画布自动收窄, 不用改代码。但版式要重排: 横屏左右分栏改上下堆叠, 内容边距 ~90px, 标题降到 --fs-h2 量级, 每屏信息量减一档(3条→2条), 图片框用 3/4 或 1/1。竖版注意事项见 authoring.md。",
+      "files": []
+    },
+    {
+      "id": 12,
+      "name": "stage-delay-covered-by-shorthand",
+      "prompt": "成片里有些元素一开页就入场了, 另一些却按口播节奏入场, 时序很乱。是 plan-timings 估错了吗?",
+      "expected_output": "不像估算问题(估算偏差通常 <0.3s)。典型根因是 tokens.css 里入场延迟被覆盖: [data-stage=N] 的 animation-delay 与 .fx-* 的 animation 简写同为 (0,1,0) 优先级, 简写更靠后会把延迟重置为 0 —— 于是只写了内联 calc(var(--t2)+…) 的元素时序正确, 纯靠属性声明的元素 0 秒入场。修法: 用新版 tokens.css(延迟走 shorthand 的变量槽 --fx-delay), 不要把延迟写成独立的 animation-delay。先抽成片帧确认, 再决定要不要动算法。",
+      "files": []
+    },
+    {
+      "id": 13,
+      "name": "non-mcode-environment-via-mmx",
+      "prompt": "我在 Claude Code 里, 没有 mcode 的 connector。这个技能还能跑吗? TTS 和 ASR 怎么办?",
+      "expected_output": "能跑 —— 脚本层(截图/渲染/合成/校验/配图)完全环境无关, 只有 TTS/音乐/ASR 依赖平台。TTS 用 mmx-cli: mmx speech synthesize --text ... --voice <id> --speed 1.1 --out audio/01.mp3(音色 mmx speech voices); ASR 用技能自带 node scripts/asr.mjs <项目>(同一把 API Key 直调 /v1/speech_to_text, 自动比对 checklist, 数字/繁体字不符判 ✗); 音乐生成 mmx-cli 没有 → 用户自备或跳过; 配图可用 mmx image(仅限抽象图, 禁生成 logo/截图/真人头像)。首次配置 npm i -g mmx-cli + mmx auth login, 401 时设 region。纪律不变: 时长实测、字幕来自 clauses、音色要试听验语种、Gate 不跳。",
+      "files": []
+    },
+    {
+      "id": 14,
+      "name": "asr-rest-and-timing-verify",
+      "prompt": "ASR 校验怎么跑? 我想顺便确认动画到底有没有对上口播。",
+      "expected_output": "build-video --asr 先按句切出 asr/part-<id>-<k>.mp3; 然后 MINIMAX_API_KEY=sk-xxx node scripts/asr.mjs <项目> 直调 REST 转写, 自动与 checklist 预期文本比对并回填(数字/繁体字不符 → ✗ 非 0 退出; 超 500s/50MB 自动转码)。要确认对轴: node scripts/asr.mjs <项目> --verify-timing 用接口的字级时间戳(verbose_json + timestamp_level=word)实测每句开口, 与 timings.json 的 clauses[].start 对比; 偏差大就校准后重渲染。也可先用 check-timing.mjs(静音法, 离线免费)。",
+      "files": []
+    },
+    {
+      "id": 15,
+      "name": "english-narration",
+      "prompt": "这条视频要发给海外同事, 口播改成英语。要改哪些地方?",
+      "expected_output": "开工对齐时就该问清语言(铁律 4 的语言行: 中文/英语/粤语/其他)。改动四处: ① script.json 设 lang:'en'; ② 口播稿改写英语并按英文基准控量 —— plan-timings 自动切换(约 14 字符/秒、常见 9–18; 字幕单句建议 ≤42 字符(可读性惯例), 胶囊实际行宽 ≈55 字符@1080 / ≈75@1920, 折 3 行起才警告; 每张约 14–26 词); ③ 音色换成 English_* 前缀(Chinese (Mandarin)_* 配英文会出怪腔调), 做完用 asr.mjs --language en 验语种; ④ 若开双语, 主行=英语口播、text2 放中文。排版、脚本、渲染流程都不用动; 数字写法分离(text 写 900M 上字幕, 加 say:'nine hundred million' 控制发音 —— TTS/ASR 吃 say ?? text, 字幕恒显示形态)。",
+      "files": []
+    },
+    {
+      "id": 16,
+      "name": "subtitle-readability-verification",
+      "prompt": "字幕到底怎么验收? 我怕深色主题下看不清。",
+      "expected_output": "三件套缺一不可: ① 数值关 —— node scripts/check-theme.mjs <项目> 算\"字幕文字 vs 胶囊合成到背景后\"的对比度, <4.5:1 直接不通过(深色主题若不足, 覆写 --sub-bg 更深 + --sub-ring 分隔); ② 帧关 —— 字幕只在成片出现(still 会 finish 掉), 抽 motion 帧 crop 底部条带看: 是否出现/是否被裁/是否与页码重叠(capture 在缺 clauses 时会告警, 不会静默出无字幕片); ③ 静音关(Gate 5)—— 关声看整片, 只靠字幕能否看懂。",
+      "files": []
+    },
+    {
+      "id": 17,
+      "name": "custom-brand-accent",
+      "prompt": "客户给了品牌色 #C2410C, 能用上吗?",
+      "expected_output": "能: 主题主色是令牌。在项目自己的 slides/tokens.css 末尾追加 [data-theme=\"...\"] 覆写 --accent/--accent-ink/--accent-2/--accent-3/--grad(改项目副本不动技能模板), 然后必须跑 check-theme.mjs 验对比度 —— accent-ink 对 accent ≥3:1 才可接受(实测 #C2410C+白字 5.18:1 通过; 浅色如 #FFB84D+白字只有 1.72:1 会被判 ✗ 并退出码 1)。深色主题另看 --sub-bg 字幕胶囊是否仍分得开。",
+      "files": []
+    },
+    {
+      "id": 18,
+      "name": "undefined-css-variable-invisible-text",
+      "prompt": "成片里那个大数字 42% 完全看不见, 但 HTML 里明明写了, 也不报错。为什么?",
+      "expected_output": "未定义 CSS 变量: background:linear-gradient(...,var(--coral-a),var(--coral-b)) 里有个变量没定义 → 整条声明计算为 none; 该元素又用了 -webkit-text-fill-color:transparent(渐变数字的标准写法) → 文字彻底隐形且无任何报错。常见诱因是从外部单页 HTML 抄配色, 变量名不在 tokens.css 里。修法: 项目 tokens.css 补定义, 或写 var(--coral-a, #ED3366) 带默认值; 之后跑 node scripts/check-slides.mjs <项目> 复验(它会静态抓出所有未定义且无 fallback 的变量)。",
+      "files": []
+    },
+    {
+      "id": 19,
+      "name": "broken-image-and-svg-under-file",
+      "prompt": "架构图的 SVG 在截图里显示成 broken image, 是 file:// 的兼容性问题吗?",
+      "expected_output": "不是协议问题 —— 实测 file:// 下 <img src=\"x.svg\"> 正常加载(400px), PNG 同样正常。broken 的真因通常是 SVG 文件自身: XML 有误、依赖外部资源或 Web 字体(inline 到 img 时被禁)、缺 width/height、或下载失败存成了 HTML。建议一律 inline 进 HTML(彻底绕开)。防线有两道: check-slides.mjs 静态查图片路径缺失/外链并提示 SVG 建议 inline; capture 渲染时会点名哪张图没加载出来(以前是静默出 broken 图标)。",
+      "files": []
+    },
+    {
+      "id": 20,
+      "name": "research-hard-rules",
+      "prompt": "这条片子里有个数字\"九亿周活\", 我直接写进口播稿行不行?",
+      "expected_output": "先过 references/research.md 的四条硬规则: ① 关键数字至少 2 个独立来源(只有一个来源就改限定措辞\"官方口径为…\"或降级为约数); ② 优先一手(官方公告/财报/技术报告/政府统计), 二手转述要回溯原文; ③ 标注口径与日期(周活≠月活, 年化≠单季, 币种要写清); ④ 查不到出处或无法判定 → 进 research/notes.md 的\"不确定项\", 绝不进口播稿。数字类断言还要注意时效(优先 12 个月内)与\"最快/第一\"这类需要新来源的表述。Gate 0 让用户确认数字/名称/口径。",
+      "files": []
+    },
+    {
+      "id": 21,
+      "name": "regulated-domain-kickoff",
+      "prompt": "帮我做一条港股上市公司二季度财报的 60 秒口播视频, 中文, 重点讲营收和净利的变化。",
+      "expected_output": "开工对齐除了常规项(语言/风格/字幕/画布/音色/素材边界), 必须先确认领域并问免责: 题材落在财经/投研 → 读 references/compliance.md, 问\"要不要免责声明与数据出处标注\"(默认要), 并按受众确认涨跌色(A股/港股=红涨绿跌, 指标卡改用 var(--up)/var(--down) 并在项目 tokens.css 覆写, 而不是默认的绿涨红跌)。数字必须带三要件: 口径(同比/环比、单季/累计、GAAP/非GAAP)+ 币种(港元)+ 时点(财报披露日, 不能写\"目前营收\")。口播不给操作建议、不预测股价、不用绝对化用语。结尾用 .disclaimer 小字行(停留 ≥3s, 口播不念、不进字幕)。check-slides.mjs 命中财经关键词而全片没有免责行时会给整片级提示。",
+      "files": []
+    },
+    {
+      "id": 22,
+      "name": "finance-updown-color",
+      "prompt": "这条财经视频里我用绿色表示上涨、红色表示下跌, 可以吧?",
+      "expected_output": "要看受众市场。tokens.css 的语义色默认 --good 绿 / --bad 红, 即欧美读法(绿涨红跌); A 股/港股受众是红涨绿跌, 用反了整片观感颠倒。做法: 指标卡一律写 var(--up)/var(--down)(模板已定义 --up:var(--good)/--down:var(--bad)), 需要翻转时在项目 slides/tokens.css 末尾追加 :root{--up:#D92B2B;--down:#12A150}(旧项目缺这两个变量要补定义, check-slides 会报未定义)。同时在 Gate 3 素材确认时问清受众市场, 并把结论记进 research/notes.md。依据见 references/compliance.md。",
+      "files": []
+    },
+    {
+      "id": 23,
+      "name": "chart-animation-and-no-fx",
+      "prompt": "财务页要画营收对比图和一条三年趋势线, 要有点动效; 但老板说这版赶时间, 最好能一个开关把动画全关掉, 画面别变。",
+      "expected_output": "图表全部纯 CSS/SVG 自绘(禁外链 Chart.js/ECharts: 离线取不到 + canvas 动画逐帧 seek 不到)。动效走现成 fx 类且逐帧可 seek: 带标签的条形 fx-grow-w 按 width 生长(数字不变形)、纯色块 fx-grow-x、柱状 fx-grow-y 从底长出、折线 fx-draw 描边画入(SVG 自设 stroke-dasharray:800)、环形/进度 fx-sweep 扫出 + fx-count 数字滚动(见 authoring.md 动效选型表); 错峰用 --fx-delay:calc(var(--t2)+N×120ms)。开关: 根元素加 class=\"no-fx\" 一键全关(单张容器加则只关该张), 关掉后 motion 捕获检测不到动画窗、自动按静态帧出片, 时长与音画同步不变, 字幕照常烧录; 实测动画终态与 no-fx 帧 PSNR 51.7dB(画面一致)。⚠ 两个必踩的坑: ① 只做 transform/描边的关键帧必须显式写 opacity:1, 否则 [data-stage] 的 opacity:0 基础态抬不回来 → 元素永久隐形(check-slides 5b 项会报); ② 柱状图外层容器不能写 align-items:flex-end(列 wrapper 高度塌陷 → 柱子百分比高度变 0), 要用默认 stretch + 列内 justify-content:flex-end。见 references/authoring.md 的\"纯 CSS/SVG 图表\"与\"动效开关\"两节。",
+      "files": []
+    },
+    {
+      "id": 24,
+      "name": "preview-before-render",
+      "prompt": "片子渲染要等好几分钟, 我想先自己把八张连着放一遍看看动效节奏对不对, 别每次都等出片。",
+      "expected_output": "跑 node <技能>/scripts/preview-page.mjs <项目目录> --open → 生成自包含放映页 preview/play/index.html(file:// 双击即看, 零依赖、不需起服务)。页面只做「把 HTML 画面放一遍」: ←→ 或触屏左右滑翻页(动效开时箭头是**逐级步进**: 先出下一级、出完才翻页)、T 现场对比硬切/交叉溶解、X 动效/关动效对照、P 口播文案开/关、O 总览、F 全屏, 另有 Space/PgDn 前进、PgUp 后退、Home/End 跳首末张、Esc 关总览(footer 列全; 首/末再按 →/← 会弹「已是第一张/最后一张」提示, 不再毫无反应)(刻意不做「重播本张」——逐级步进已经能回到任意一级); 刻意不做计时器/进度条/逐句跟读高亮 —— 要看时间或节奏就直接看成片, 预览页里跑计时器只会让人盯秒表(实测标签页放着就变成 204.2s / 6.3s)。口播 UI 按数据自动决定: 有 clauses + 有 timings.json → 列出该张口播文案; 有 clauses 没对时 → 只列文案并标「(未对时)」; 没有 clauses 或 --no-script → 面板与按钮完全不出现, 画面占满整宽。布局自适应: 窄窗口/手机上口播面板收成底部抽屉且默认收起, 手机给触摸条按钮+左右滑动翻页, 总览网格自动列数, 高度用 100dvh。⚠ 关键: 不要直接双击 slides/*.html —— 动画延迟 --t1/--t2/--t3 是管线按 timings.json 注入的, 原文件里只有占位值(--t2:800ms), 看到的是「所有动画挤在开头两秒」的假象; 放映页的副本把实测延迟写进 <html style> 并加 <base href=\"../../slides/\">, 浏览器里的时序才等于成片时序(实测同一张 t=3.0s: 副本第二层 opacity 0 未入场, 原文件已入场)。还没对时则按 HTML 里用到的 stage 等间隔(0.3/1.3/2.3s)预览, 页面顶部黄条如实标注。副本是快照, 改完 slides 必须重跑(两个开关都在底栏, 顶部不再有动效开关)。总览缩略图随 play/thumbs/ 携带(生成时从 preview/*.png 拷入, 没跑 capture 的张次用占位并在终端计数告警); 外部自动化不要模拟键盘再抠 innerText —— 页面暴露 window.playPage(state/go/next/prev/setFx/setNarr/setTrans/overview)编程接口。X 切过去若画面变空, 说明有关键帧没把 opacity:0 基础态抬回来; 老项目(受管区落后)先跑 init-project.mjs <项目> --upgrade-css 补规则(幂等, 不碰 script.json); 放映页只是快照兜底, **成片仍用项目里的旧 CSS**, 所以升级要在 capture/build-video 之前做。",
+      "files": []
+    },
+    {
+      "id": 25,
+      "name": "chart-elegance",
+      "prompt": "你这几张图表看着有点业余, 能不能做得更讲究一点? 图表库不给用(离线), 但只要 CSS/SVG。",
+      "expected_output": "用 tokens.css 里的图表工具箱(`.chart*` 原语 + `fx-grow-w/h · fx-sweep · fx-count · fx-dot`;老项目先 `init-project --upgrade-css` 补齐)。优雅不是加特效, 是九条纪律: ①同类同色、只强调关键一条(其余 `.dim` 用不透明同色系; 彩虹色条形=业余感来源); ②数值等宽(`.chart-num`)且绝不放进被缩放的元素(`scaleX` 会把数字拉扁 → 带标签的条用 `fx-grow-w` 按 `width` 生长); ③有基线与**看得见**的细网格(网格画在 `.chart-plot-cell` 上, 与柱高共用一把尺子, 顶边即轴上限); ④一图只讲一件事(`.chart-note` 写在标题行右侧); ⑤零值/极小值保底; ⑥图注写口径与时点(`.chart-cap`)。另有四条版面纪律: **避开字幕安全区**(1080 高下底部 84–168px、居中 73% 宽留给字幕, 容器 `padding-bottom:190px` 起步); 在安全区内**垂直居中**(别堆在顶部); 数值统一右对齐同一列(都用 `.chart-val.out`); **柱高 = 数值 ÷ 轴上限、轴从 0 起**, 并把轴上限写进标题(凭手感写百分比 = 数据失真)。动效选型: 条形 `fx-grow-w`(带标签)/`fx-grow-x`(纯色块); 柱状 `fx-grow-y`(标签在外)/`fx-grow-h`(标签在内); 折线 `fx-draw` + `fx-dot`(标记规则统一: 全标或只标末点); 环形 `fx-sweep`(**真扫出**, 不用 `fx-pop` 硬弹)+ 中心 `fx-count`; 错峰 `--fx-delay:calc(var(--t2) + N×120ms)`。禁外链图表库(离线取不到、canvas 动画逐帧 seek 不到)、禁 AI 生图当图表、数值必须来自已核实口径。可加图表类型: bullet(实际 vs 目标, `.chart-target` 虚线)、slope(前后对比)、sparkline(数字旁迷你趋势)。",
+      "files": []
+    },
+    {
+      "id": 26,
+      "name": "table-forms",
+      "prompt": "这个季度的数据我想用表格讲, 但我们的模板里表格好像只有一种样式?",
+      "expected_output": "表格有五种形态, 都走 tokens.css 的原语(老项目先 `init-project --upgrade-css` 补): ① 数据表 `.tbl`(多期/多主体对比: 表头弱化、数值列 `.num` 右对齐 + 等宽数字、涨跌 `.up/.down` 走 `var(--up)/var(--down)` 并按受众市场翻转、当前行 `.key` 高亮、合计行 `.sum`); ② 规格表 `.kv`(无表头, 左标签右值 —— 参数/条款); ③ 对比矩阵 `.matrix`(行维度 × 列方案, 单元格只放勾叉, `.hi` 高亮本方案列); ④ 排名表 `.rank`(名次 + 内联微缩条 + 数值, 条长 = 数值 ÷ 轴上限); ⑤ 时间事件表(用 `.kv` 两列)。七条纪律: 主数据正文号 ≥30px(视频要在手机上看; 实测 caption 24px 吃力)、行高 ≥72px、一页一表 ≤6 行、数值右对齐等宽、涨跌走令牌、只画横线不画竖线不要斑马纹、高亮只给一行、表注写口径与时点。表格不要和图表放同一页(一页一件事)。",
+      "files": []
+    }
+  ]
+}
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/authoring.md b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/authoring.md
new file mode 100644
index 00000000..307d1bcf
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/authoring.md
@@ -0,0 +1,616 @@
+# HTML Design and Content Volume Guidelines
+
+## Core principle: every slide is "one claim + one set of elaboration"
+
+The two most common failures in a finished video can both be stopped at the authoring stage:
+
+1. **A title with no elaboration** —— the audience watches one big line for 7 seconds; the information content is zero.
+2. **All the text dumped onto the screen at once** —— no rhythm; by the time the voiceover reaches the second half of the clause, the visuals have long since run out of anything to say.
+
+The fix is a mandatory three-layer structure: **title layer + elaboration layer + visual anchor (optional), with the three layers on different stages**. The voiceover and the frame are two voices of the same sentence.
+
+**The entrance order of the layers is not fixed —— the narrative decides it; title-first is only the default.** The criterion is the strong-sync principle: whichever layer a given voiceover clause mentions is the stage that layer hangs on. Entirely legal variants:
+
+- **Big number in first, title out later** (the default shape of stat-highlight): the first voiceover clause slams the number; number at s1, explanation and title at s2;
+- **Question out first, answer next**: question in large type s1 → answer/data s2 → title gathers it up s3;
+- **Image in first, text lands after**: main image s1 → keyword annotation s2 (common for image-led subjects).
+
+What must always be blocked is never "the title does not come first", but **only one layer**: the whole slide stops on the title/a single block of content, with no second information layer pushing it forward.
+
+## Content volume table (hard rules)
+
+Chinese voiceover ≈ 5.3 chars/sec (speed 1.1; 4.8 at 1.0), target 6–10s per slide. **English at ~14 chars/sec (about 150 words/min)** —— the character counts in the table below are the Chinese measure; for English projects read them halved, in "words": 14–26 words per slide, hard ceiling 34 words (plan-timings switches the baseline and the warning band automatically by `lang`).
+
+| Layout | Purpose | Required on the frame (any one missing is a violation) | Voiceover length | clauses structure |
+|---|---|---|---|---|
+| title-hero | Opening claim | Large title + one subtitle line + logo/theme corner mark | 12–20 | 1 sentence |
+| statement | Single-point assertion | Assertion in large type (s1) + supporting small type (s2) + visual anchor (s3 optional) | 18–28 | 2 sentences |
+| bullets | Parallel points | Title (s1) + 3–4 points, each ≤14 chars (s2) | 25–40 | 2 sentences |
+| compare | Before/after contrast | Title (s1) + left and right columns with ≥1 item each + divider (s2) | 22–35 | 2 sentences |
+| data-viz | Data persuasion | Title (s1) + ≥2 number cards/charts (s2) | 25–40 | 2 sentences |
+| code | Demonstrating a call | Title (s1) + code block ≥3 lines + highlighted comment (s2) | 25–40 | 2 sentences |
+| quote | Quotation transition | Quotation in large type (s1) + attribution/title (s2) | 20–32 | 2 sentences |
+| closing | Close | One closing line + logo/CTA | 12–20 | 1 sentence |
+
+Decision rules (written for the executor, for self-check):
+
+- Apart from title-hero/closing, **the frame carries ≥2 visual blocks on different stages**; "just a title + page number" is sent straight back. **The order of the layers is free** (the number may precede the title, the question may precede the answer); what is wrong is "only one layer", not "the title is not at s1".
+- **Strong-sync principle (a lesson measured in 2026-09): the visual anchor of each stage (big number/keyword/main image) must be written into the voiceover clause that triggers it, never left in the previous clause.** Counter-example: the very first voiceover clause is "三亿人在用" ("three hundred million people use it"), but the "3 亿" number card hangs on stage 2 —— when the audience hears "三亿" the frame does not react, and by the time the number card enters, the voiceover has already moved to the next clause; what the viewer feels is "the animation does not line up". Correct example: the first clause only sets up ("它有三个你可能用过的产品" — "it has three products you have probably used"), "3 亿" opens the second clause, and the number card hangs on stage 2 → sound and picture slam "3 亿" down together, strong sync.
+- The mapping between voiceover clauses and stages: **the visuals of stage k enter the moment clause k starts speaking**. Number of stages ≈ number of clauses; any extra visual layers get explicit times via `stageTimes` in script.json.
+- Hard ceiling of 60 characters of voiceover for a single slide; over that means the slide is doing the work of two, so split it. In subtitle scenarios, prefer one line per sentence — the pill fits **≈24 CJK characters per line at 1080 width, ≈33 at 1920** (it scales with the canvas; one-line craft guideline ≈18; capture burns subtitles as one wrapped block: two lines still read fine, `plan-timings` warns at three or more); when bilingual, the second line (text2) is ≤60 characters.
+- Numbers and proper nouns are read in Chinese in the voiceover ("九亿", not "900000000"), and only the frame shows Arabic numerals. (English projects are the opposite: the voiceover writes the English reading "nine hundred million" and the frame shows 900M.)
+- Bilingual subtitles (optional): supplying `text2` for each clause automatically produces two lines (main line in Chinese + second line in a smaller size); text2 is the translation of the same clause, so do not reorder the wording.
+
+## The staged entrance system (how to write the HTML)
+
+tokens.css already ships this built in, and there are only four rules:
+
+1. A block that should enter: `data-stage="1|2|3"` + an fx utility class (`fx-up/fx-fade/fx-grow/fx-draw`).
+2. **Write no delay in seconds at all**. `--t1/--t2/--t3` are injected by the render pipeline from measured TTS durations; opening the file directly in a browser uses the placeholder defaults in tokens (0/0.8/2.0s).
+3. Staggering within the same layer (e.g. three bullets appearing one after another): inline `style="animation-delay:calc(var(--t2) + 150ms)"`, +300ms for the second block, and so on.
+4. Ambient animations (`fx-pulse` and other infinite loops) may only be decorative (glows, breathing dots); **give them no data-stage and let them carry no information** —— they take no part in duration calculation and may be frozen at an arbitrary phase in a still frame.
+
+Prohibitions: no transition for entrance effects (the pipeline cannot seek to a transition state frame by frame); no orchestrating with JS timers (setTimeout-driven layering never happens in frame-stepping mode); no external Google Fonts for type (they render as plain boxes offline) —— use the system font stack in tokens.css.
+
+### Three hard checks when writing HTML (run `node <skill>/scripts/check-slides.mjs <project>` the moment you finish)
+
+All three are silent failures hit during 2026-09 testing —— the frame is visibly wrong, yet the pipeline reports success the whole way through:
+
+1. **Variables must be defined first, or carry a fallback.** In `background: linear-gradient(135deg, var(--coral-a), var(--coral-b))`, if even one variable is undefined the whole declaration computes to `none`; and if that element also uses `-webkit-text-fill-color: transparent` (the standard way to write gradient numbers), **the text becomes completely invisible without raising any error**. This is easiest to hit when borrowing colours from another single-page HTML: the borrowed variable names (e.g. `--coral-*`) are not in tokens.css. Either add the definitions to the project tokens.css, or write `var(--coral-a, #ED3366)` to supply a default.
+2. **Images use relative paths and the files must exist**; for `.svg` it is best to **inline it into the HTML**. Measured: under file://, `<img src="x.svg">` loads fine (400px), so a broken icon is usually not a protocol problem but the SVG itself: malformed XML, a dependency on external resources/web fonts, missing width/height, or a failed download saved as HTML. Inlining solves all of it at once.
+3. **Do not copy an external single page's colours/variables over wholesale.** Copying the layout structure is fine, but colours must land on tokens variables (`--accent` / `--good` / `--grad` …); hard-coded hex/rgb bleeds through when you switch theme.
+
+## Theme and visual tone
+
+`<html data-theme="...">`:
+
+- **a · off-white + warm orange** (default, general business): bg `#FAFAF7`, accent `#FF5B2E`, serif large titles. Narrative, opinion, people-oriented.
+- **b · dark + green** (tech feel): bg `#0E0F12`, accent `#10A37F`, high-contrast data. Product, AI, developer-oriented.
+- **c · black-and-white + blue** (minimal and premium): bg `#FFFFFF`, accent `#1F6FEB`. Finance, reports, serious material.
+
+Type discipline: at most two levels on screen at once per slide (title + one secondary); body text ≥30px —— anything smaller is inviting a reviewer's complaint. Safe margin: content ≥96px from the edge (except brand/slide-num).
+
+⚠ **Subtitle safe area (at 1080 height)**: the subtitle pill occupies **the bottom 84–168px, centred, about 73% wide** (40px type + 12px padding top and bottom, 7.8% from the bottom); **do not put body text, values, charts or figure captions in this band** —— otherwise the subtitles land directly on top of the content in the finished video and neither can be read (caught by visual acceptance on 2026-09-18: an example page had its figure caption at the bottom and the subtitle covered it, producing "ghosted text"). The fix: start from `padding-bottom: 190px` on the layout container and keep only brand/slide-num at the bottom. Pad the **top** as well (the scaffold `.layout` ships `padding: 120px 160px 190px` + `justify-content: safe center` — keep the `safe`: under plain `center`, content that outgrows the area is pushed **above** the canvas and clipped there, measured: the kicker landed at y≈17px after a 125px overflow; and never push content down with `margin-top` inside a centered layout — margins join the content height and make the overflow worse). **Portrait (1080×1920)**: the same formulas computed against the canvas —— ≈149px from the bottom (7.8% × 1920), type squeezed to 30px by `--sub-scale`, the pill band at roughly **150–215px from the bottom and about 787px wide centred**; start from `padding-bottom: 240px` on the layout container, and since subtitle lines are narrower, keep the per-sentence character count tighter than in landscape.
+
+## Asset acquisition (Phase 3 · Gate 3)
+
+Process: list the requirements (which slides need images, and what) → source by the priority order below → compliance self-check → write to disk under `assets/` and register in `MANIFEST.md` → **present Gate 3 to the user**.
+
+**Sourcing priority (step down level by level, never force it):**
+
+1. **Official channels (first choice)**: logos from the official brand kit or simple-icons (curl-able); product screenshots taken from official help/docs; data from official reports. Strongest compliance; for factual subjects try to stop at this level.
+2. **Extract from the official site with a built-in browser / Playwright (measured and recommended, 2026-09)**: open the official site or official news page with the agent's built-in browser or Playwright, inspect the DOM and extract `<img>`/`<source>` assets (verify the src is an official CDN domain, then write to disk with `fetch-official-images.mjs --url <URL> --out-dir assets` —— the same host/size/filename validation set). Measured on Chinese AI-company subjects: 6/6 all relevant, far better than image-search engines. If the user chose "official assets only / pure layout" during kickoff alignment, skip this whole level.
+3. **Pure-layout fallback (the fallback that is always available)**: large type + number cards + colour + the expressive power of the layout itself. Better to omit than to fake —— when no suitable asset exists, use this level; the frame still holds up.
+
+> ⚠ **image-downloader (Bing image search) is off by default.** Measured 2026-09-17: for Chinese AI-company subjects, 5/5 keywords returned entirely irrelevant images (Bing returned "猜你喜欢" (recommended-for-you) cards and the script's regex swallowed the lot). Unless the subject is an English-language mass-market brand and you are willing to filter every image by hand, do not take this route.
+
+**Compliance self-check (run through it for every asset):**
+
+- No watermarks. For a watermarked candidate: use a different image, or crop inside the watermark; if you cannot crop it out, do not use it.
+- Trademarks/logos are used only in the legitimate reference context of "talking about that brand", never as decorative filler.
+- Photos must come from licensable sources (official press images, CC-licensed, free stock libraries); for celebrity/people photos with no verifiable licence, prefer a text-only quotation card.
+- Screenshots state their source (official doc name + date).
+
+**Registration and use:**
+
+- `assets/MANIFEST.md` gets one line per asset: filename / content / source URL or channel / licence. Present Gate 3 to the user together with the asset previews and ask explicitly "are the sources and licences OK?".
+- **Never generate logos, screenshots, avatars or QR codes out of thin air; never let a watermarked image into the assets.**
+- **Images must be framed; never place a bare `<img>`**: use `.img-frame` (ratio and cropping belong to the frame) / `.img-frame.contain` (mandatory for screenshots, charts and images containing text; no cropping, centred with letterboxing) / `--img-ratio` (frame ratio; 16:9 is not required) / `--img-pos: top|center|bottom` (controls the visible part of the subject, **the preferred replacement for hard cropping**) / `.img-scrim` (darken to carry white text) / `.img-cap` (figure caption, written outside the frame) / `.img-tag` (corner tag). The full SOP for choosing and cropping images is in `references/image-sources.md`.
+- Reference images in HTML with a relative path `../assets/xxx.png`; capture waits for images to finish loading (4s cap each, skipped if they do not load —— so assets must be local first; never link an external image).
+
+## Layout snippets (copy-ready)
+
+The snippets below build on tokens.css plus each slide's own `<style>` (layout skeletons such as `.layout/.cols/.cards/.num/.cap` are defined inline per slide and do not come from tokens.css); they list only the structure inside `<main>`; the outer `.stage`, `.brand`, `.slide-num` follow `_template.html`.
+
+**bullets**
+```html
+<main class="layout">
+  <h1 class="fx-up" data-stage="1">三大产品线</h1>
+  <ul>
+    <li class="fx-up" data-stage="2" style="animation-delay:calc(var(--t2))">ChatGPT — 对话</li>
+    <li class="fx-up" data-stage="2" style="animation-delay:calc(var(--t2) + 150ms)">GPT 系列 — 模型</li>
+    <li class="fx-up" data-stage="2" style="animation-delay:calc(var(--t2) + 300ms)">Sora — 视频生成</li>
+  </ul>
+</main>
+```
+
+**compare**
+```html
+<main class="layout">
+  <h1 class="fx-up" data-stage="1">三年,高中生 → 研究员</h1>
+  <div class="cols">
+    <div class="fx-grow" data-stage="2" style="animation-delay:calc(var(--t2))">
+      <h2>2023 · GPT-3.5</h2><p>会聊天,常出错</p>
+    </div>
+    <div class="fx-grow" data-stage="2" style="animation-delay:calc(var(--t2) + 200ms)">
+      <h2>今天 · GPT-5.6</h2><p>可托付研究级任务</p>
+    </div>
+  </div>
+</main>
+```
+
+**data-viz**
+```html
+<main class="layout">
+  <h1 class="fx-up" data-stage="1">九亿人每周在用</h1>
+  <div class="cards">
+    <div class="fx-grow" data-stage="2" style="animation-delay:calc(var(--t2))">
+      <div class="num">9亿</div><div class="cap">周活跃用户</div>
+    </div>
+    <div class="fx-grow" data-stage="2" style="animation-delay:calc(var(--t2) + 180ms)">
+      <div class="num">$25B</div><div class="cap">年化营收</div>
+    </div>
+    <div class="fx-grow" data-stage="2" style="animation-delay:calc(var(--t2) + 360ms)">
+      <div class="num">92%</div><div class="cap">财富500强渗透</div>
+    </div>
+  </div>
+</main>
+```
+
+**code**
+```html
+<main class="layout">
+  <h1 class="fx-up" data-stage="1">一行调用</h1>
+  <pre class="fx-fade" data-stage="2"><code>const answer = await openai.responses.create({
+  model: "gpt-5.6",
+  input: "帮我总结这份报告",
+});
+console.log(answer.output_text);</code></pre>
+</main>
+```
+
+**quote**
+```html
+<main class="layout">
+  <blockquote class="fx-up" data-stage="1">“我们想打造的是比人类更聪明的工具,而不是替代人类。”</blockquote>
+  <p class="fx-fade" data-stage="2">— Sam Altman, OpenAI CEO</p>
+</main>
+```
+
+**title-hero / closing**: large title centred + one line of subtitle/CTA, a single stage is enough; closing may add a logo with `fx-grow`.
+
+Layout CSS (`.layout/.cols/.cards/.num/.cap` etc.) is defined inline in each slide's `<style>`; shared values (colours, type sizes, easing) must come from tokens variables and must never be hard-coded.
+
+---
+
+# Additional layouts (9 added 2026-09-17)
+
+The original 8 layout names are unchanged (existing projects' `layout` values stay valid); the following are incremental options. **Do not use the same layout on two adjacent slides.**
+
+| Layout | Purpose | Required on the frame (any one missing is a violation) | Voiceover length | clauses |
+|---|---|---|---|---|
+| kpi-grid | A set of metrics | Title (s1) + 3–4 metric cards with change and semantic colour (s2). **Use `var(--up)/var(--down)` for change; for A-share/HK audiences override to red-up/green-down (see compliance.md)** | 25–40 | 2 sentences |
+| stat-highlight | One number decides everything | Huge number (≥200px, `.gradient-text` allowed) (s1) + one line of explanation (s2) | 15–28 | 2 sentences |
+| table | Multi-row comparison | Title (s1) + ≥3-row data table with right-aligned values (s2). **Use the `.tbl` primitive; the five forms and disciplines are in "table toolkit"** | 25–40 | 2 sentences |
+| timeline | Development history | Title (s1) + ≥4 time points (axis + labels) (s2) | 25–40 | 2 sentences |
+| roadmap | Phase planning | Title (s1) + 3–4 columns (NOW/NEXT/LATER and the like) (s2) | 22–35 | 2 sentences |
+| comparison | Before/after, pros/cons | Title (s1) + left and right columns with a vertical divider (s2) | 22–35 | 2 sentences |
+| flow-diagram | Process / pipeline | Title (s1) + 3–5 nodes + arrows (s2) | 22–35 | 2 sentences |
+| terminal | Command-line demo | Title (s1) + terminal window (traffic lights + ≥3 lines) (s2) | 25–40 | 2 sentences |
+| big-quote | Large quotation transition | Quotation in large type (s1) + attribution (s2) | 20–32 | 2 sentences |
+
+```html
+<!-- kpi-grid: 指标卡 + 语义色涨跌(涨跌一律走 --up/--down, 财经题材按受众翻转这两个即可, 见 compliance.md)
+     注意: 布局几何内联在容器 style 里, 不要用没定义过的 .grid/.g4 之类的类(会静默竖排) -->
+<div class="fx-stagger" style="--stagger-base:var(--t2);display:grid;grid-template-columns:repeat(4,1fr);gap:var(--sp-4)">
+  <div class="card"><p class="eyebrow">REVENUE</p>
+    <div style="font-size:var(--fs-h2);font-family:var(--font-display)">1248K</div>
+    <p style="color:var(--up)">↑ 38% YoY</p></div>
+  <div class="card"><p class="eyebrow">RETENTION</p>
+    <div style="font-size:var(--fs-h2);font-family:var(--font-display)">74%</div>
+    <p style="color:var(--warn)">→ 持平</p></div>
+  <div class="card"><p class="eyebrow">NPS</p>
+    <div style="font-size:var(--fs-h2);font-family:var(--font-display)">62</div>
+    <p style="color:var(--up)">↑ 9 pts</p></div>
+  <div class="card"><p class="eyebrow">CHURN</p>
+    <div style="font-size:var(--fs-h2);font-family:var(--font-display)">3.1%</div>
+    <p style="color:var(--down)">↓ 0.4 pts</p></div>
+</div>
+
+<!-- 免责声明行: 受监管题材(财经/医疗/法律/政策)必出; 小字、不抢视觉、停留 ≥3s、口播不念 -->
+<p class="disclaimer fx-fade" data-stage="2">
+  本视频仅为信息分享,不构成任何投资建议。数据来源:公司 2026 年 Q2 财报(8 月披露),截至 2026-06-30。
+</p>
+
+<!-- stat-highlight: 单数字撑满视觉 -->
+<div class="fx-pop" data-stage="1" style="font-size:220px;line-height:1;font-weight:900;letter-spacing:-.05em">
+  <span class="gradient-text">92</span><span class="gradient-text">%</span>
+</div>
+<h3 class="fx-fade" data-stage="2" style="margin-top:var(--sp-3)">的准备工作被省掉</h3>
+
+<!-- table: 数值右对齐, 行高要够(1080p 下 ≥72px) -->
+<table class="fx-fade" data-stage="2" style="width:100%;border-collapse:collapse;font-size:var(--fs-caption)">
+  <thead><tr style="color:var(--fg-3);text-align:left">
+    <th style="padding:var(--sp-2) 0">季度</th>
+    <th style="text-align:right">营收</th><th style="text-align:right">同比</th></tr></thead>
+  <tbody>
+    <tr style="border-top:1px solid var(--line)"><td style="padding:var(--sp-3) 0">Q1</td>
+      <td style="text-align:right">3.2 亿</td><td style="text-align:right;color:var(--good)">+18%</td></tr>
+    <tr style="border-top:1px solid var(--line)"><td style="padding:var(--sp-3) 0">Q2</td>
+      <td style="text-align:right">4.1 亿</td><td style="text-align:right;color:var(--good)">+27%</td></tr>
+  </tbody>
+</table>
+
+<!-- timeline: 横轴 + 时间点; 整轴 fx-fade, 每个时间点在容器 fx-stagger 里逐个上浮(口播点到哪个年份, 哪个冒出来) -->
+<div class="fx-fade" data-stage="2" style="display:flex;gap:var(--sp-4);margin-top:var(--sp-5)">
+  <div class="fx-stagger" style="--stagger-base:var(--t2);display:flex;gap:var(--sp-4);width:100%">
+    <div style="flex:1;border-top:3px solid var(--accent);padding-top:var(--sp-3)">
+      <p class="eyebrow">2021</p><p>成立</p></div>
+    <div style="flex:1;border-top:3px solid var(--accent);padding-top:var(--sp-3)">
+      <p class="eyebrow">2023</p><p>首个产品</p></div>
+    <div style="flex:1;border-top:3px solid var(--accent);padding-top:var(--sp-3)">
+      <p class="eyebrow">2026</p><p>上市</p></div>
+  </div>
+</div>
+
+<!-- roadmap: NOW/NEXT/LATER 三列(阶段规划)。当前阶段用 .card-accent 突出; 4 列时 repeat(4,1fr) 并把要点字号压到 var(--fs-tiny) -->
+<div class="fx-stagger" style="--stagger-base:var(--t2);display:grid;grid-template-columns:repeat(3,1fr);gap:var(--sp-4);margin-top:var(--sp-5)">
+  <div class="card card-accent">
+    <p class="eyebrow" style="color:var(--accent)">NOW · 2026 Q3</p>
+    <p style="margin-top:var(--sp-2)">开放平台公测</p>
+    <p style="margin-top:var(--sp-1);color:var(--muted)">长上下灰度放量</p>
+  </div>
+  <div class="card">
+    <p class="eyebrow" style="color:var(--fg-3)">NEXT · Q4</p>
+    <p style="margin-top:var(--sp-2)">多模态 API 定价</p>
+    <p style="margin-top:var(--sp-1);color:var(--muted)">企业私有化版本</p>
+  </div>
+  <div class="card">
+    <p class="eyebrow" style="color:var(--fg-3)">LATER · 2027</p>
+    <p style="margin-top:var(--sp-2)">Agent 应用商店</p>
+    <p style="margin-top:var(--sp-1);color:var(--muted)">海外节点</p>
+  </div>
+</div>
+
+<!-- comparison: 左右两栏 + 竖分隔线 -->
+<div style="display:grid;grid-template-columns:1fr 1px 1fr;gap:var(--sp-5);align-items:center">
+  <div class="fx-grow" data-stage="2">
+    <p class="eyebrow">三年前</p><h3>像高中生</h3><p style="color:var(--muted)">会聊天,常出错</p></div>
+  <div style="background:var(--line);height:70%"></div>
+  <div class="fx-grow" data-stage="2" style="animation-delay:calc(var(--t2) + 200ms)">
+    <p class="eyebrow">今天</p><h3>像研究员</h3><p style="color:var(--muted)">可托付研究级任务</p></div>
+</div>
+
+<!-- flow-diagram: 节点 + 箭头(用 ▸ 字符, 不引 SVG) -->
+<div class="fx-fade" data-stage="2" style="display:flex;align-items:center;gap:var(--sp-3);margin-top:var(--sp-5)">
+  <div class="card card-soft">脚本</div><span class="accent" style="font-size:var(--fs-h3)">▸</span>
+  <div class="card card-soft">声音</div><span class="accent" style="font-size:var(--fs-h3)">▸</span>
+  <div class="card card-accent">成片</div>
+</div>
+
+<!-- terminal: 红黄绿灯头 + 等宽字体 -->
+<div class="fx-fade" data-stage="2" style="background:#0d1117;border-radius:var(--radius);overflow:hidden">
+  <div style="display:flex;gap:8px;padding:14px 18px;background:#161b22">
+    <span style="width:14px;height:14px;border-radius:50%;background:#ff5f57"></span>
+    <span style="width:14px;height:14px;border-radius:50%;background:#febc2e"></span>
+    <span style="width:14px;height:14px;border-radius:50%;background:#28c840"></span>
+  </div>
+  <pre style="margin:0;padding:var(--sp-4);color:#c9d1d9;font-family:var(--font-mono);font-size:var(--fs-caption);line-height:1.6"><code>$ node build-video.mjs ./proj --asr
+<span style="color:#7ee787">✓ 时长 62.27s / 解码零错</span></code></pre>
+</div>
+
+<!-- big-quote: 全幅引语, 衬线加大 -->
+<blockquote class="fx-blur" data-stage="1"
+  style="font-family:var(--font-display);font-size:var(--fs-h1);line-height:1.3;max-width:1500px">
+  “我们想打造的是比人类更聪明的工具,<span class="accent">而不是替代人类</span>。”
+</blockquote>
+```
+
+---
+
+# Chart toolkit (how to draw data-viz)
+
+Charts are **all hand-drawn**, and three bottom lines never change:
+
+1. **No external chart libraries** (Chart.js / ECharts etc. via CDN): unreachable in the offline sandbox, a violation of the no-external-links iron rule, and canvas animation cannot be seeked frame by frame (it renders static/blank).
+2. **The numbers must come from a measure already verified in research/notes.md** —— a chart "draws numbers that have been verified", it is not decoration; a measure you cannot draw (missing as-of point / mixed currencies) goes back to Phase 0 first.
+3. **Never use AI-generated images as charts** (`image generate` is limited to abstract concept images): data graphics must be traceable to a source, and generated images cannot be audited.
+
+The chart primitives and motion effects all live in `tokens.css` (`.chart*` plus `fx-grow-w/h · fx-sweep · fx-count · fx-dot`). **For existing projects run `node scripts/init-project.mjs <project> --upgrade-css` first to add them** (idempotent), otherwise the new recipes half-die silently: bars do not grow, rings do not sweep, numbers do not roll up.
+
+## The nine disciplines of elegance (more important than adding effects)
+
+1. **Same kind, same colour + one emphasis**: bars/columns in the same series get **the same colour**; the one to highlight uses `--accent` (the `.chart-bar` default) and the rest get `.dim` (automatically weakened into the same hue by `color-mix`). A rainbow bar chart is the most typical source of "amateur feel". **Do not use white text for the value on a dimmed bar** —— a translucent fill with white text cannot be read (measured: judged "unreadable" by visual acceptance); either rely on `.chart-bar.dim .chart-val` switching to ink automatically, or put the value outside the bar: `<span class="chart-val out chart-num">` (after the track, right-aligned, ink colour).
+2. **Never put the number inside an element that gets scaled**: `fx-grow-x` uses `scaleX`, so a number inside the bar is **squashed horizontally** (measured); for bars with labels use `fx-grow-w` (animates `width`, no distortion). Always wrap values in `.chart-num` (monospace + `tabular-nums`) so multi-digit numbers do not jitter while rolling up.
+3. **Have a baseline and fine gridlines**: a column chart takes its baseline from the bottom edge of `.chart-plot-cell` (columns must stay inside the plot container —— if a percentage height is computed against the whole column, anything exceeding the remaining space gets pushed back by `flex-shrink`, and **the column height no longer equals the value**; measured: 84% and 72% were drawn the same height) + the grid is drawn inside the plot area **and must be visible on the page** (use `--line-strong`; if the gridlines are too faint, or only show up where a column covers them, the reader cannot read the scale —— measured: judged fail), and the top edge line is the axis maximum. **For a column chart the target value is fine written in the note on the title line** —— drawing another dashed line at the top crowds the value label on the tallest column and gets read as a minus sign (measured: flagged as a defect); only track charts like bullet use the `.chart-target` vertical tick line. Horizontal bar charts give every track `.chart-track.grid` (shared 25/50/75% vertical ticks); the left edge of the track already carries the zero-axis vertical line. **Make the rows fixed-width using the three grid columns of `.chart-row` (label / track / value slot)** —— otherwise the track of a "row with an outside value" gets squeezed narrower, and one chart ends up with two different rulers and bar lengths that cannot be compared (measured: caught by visual acceptance). Axis labels use `--fs-tiny` + `--fg-3`. Most of a chart's refinement comes from these "invisible lines".
+4. **One chart tells one thing**: use `.chart-note` to write that sentence on the right of the title line ("同比 +38%" = YoY +38%, "距目标 0.2pp" = 0.2pp from target); do not make the audience do the arithmetic.
+5. **Give zero and tiny values a floor**: add `min-height:4px` to `.chart-bar` (or let the `.chart-target` dashed line catch it), otherwise "0" looks like "no data".
+6. **The figure caption states the measure and the as-of date** (`.chart-cap`): mandatory for regulated subjects (see compliance.md), and recommended for other subjects too —— the credibility of a number comes from "when and how it was calculated".
+7. **Pin the title to the same position at the top of the safe area and centre the chart block in the remaining space**: the chart block should occupy the safe area (about 825px usable at 1080 height); do it by keeping a fixed top margin before the `h1`, using `.layout{display:flex;flex-direction:column}` and wrapping the chart block in a layer with `flex:1;display:flex;align-items:center` —— that way the title sits at the same y on every page (the title does not jump when the video plays through) while the chart is centred in the leftover space (measured: titles 170px apart across two pages was judged fail). Bar spacing/height can be enlarged (use `--bar-h` for bar height; 56–64px looks better). A packed top half and a large empty bottom half reads as "the whole page is unbalanced" in the finished video (measured twice by visual acceptance).
+8. **Align values into a single column**: every value in a chart is right-aligned in the same column (all using `.chart-val.out`); white text inside a bar is only for the case where "there is exactly one emphasis bar and it is not side by side with other rows", otherwise the four values split into two columns and the misalignment is obvious at a glance (visual acceptance measurement).
+9. **Column height must be computed from the value, and the axis starts at 0**: writing `height:58%` by feel makes the column height out of proportion to the labelled number —— this is not an aesthetic issue, it is **data distortion** (truncating the vertical axis is the same class of error). Do it this way: fix the axis maximum first (e.g. 4.0%), column height = value ÷ axis maximum (3.1/4.0 = 77.5%), write the axis maximum into the title or the note ("axis 0–4.0%"), and mark the tick values on the left with `.chart-ticks` —— writing only "axis 0–4.0%" with no ticks at all leaves the audience unable to verify the column heights (measured: flagged as a defect). Visual acceptance measurement: the drawing of three columns did not match their labels and was judged fail outright.
+
+Everything else is as before: colours go through tokens (up/down `var(--up)/var(--down)`, flipped by audience for financial subjects), grid `var(--line)`, values `var(--fg)`, legend `--muted`. **Motion is part of the chart**, and **the whole set switches off with one flag** (see "Motion switches").
+
+## Choosing motion effects (frame-by-frame seek determinism unchanged, still pure @keyframes)
+
+| Graphic | Which one | Why |
+|---|---|---|
+| Bar (with label) | `fx-grow-w` + inline `--w:86%` | animates `width`, the number is not squashed |
+| Bar (flat colour block) | `fx-grow-x` | transform is cheaper, and there is no text anyway |
+| Column (label outside) | `fx-grow-y` | the label is a sibling node, unaffected by scaling |
+| Column (label inside) | `fx-grow-h` + inline `--h:70%` | as above, avoids squashing the number |
+| Line | `fx-draw` (set `stroke-dasharray:800` on the element) + `fx-dot` for the points | stroke drawing is inherently seekable; data points pop in one by one |
+| Donut / gauge | `fx-sweep` + inline `--p-to:62` | the conic gradient recomputes frame by frame from the registered property, so it **really sweeps** (no need for a hard `fx-pop`) |
+| Big number | `fx-count` + inline `--n-to:924` | pure CSS counter roll-up; **integers only** (for decimals split the fields or keep it static) |
+
+Staggering is always `--fx-delay:calc(var(--t2) + N×120ms)` (120–150ms between data bars looks best; past 200ms the audience starts waiting).
+
+## Recipes
+
+```html
+<!-- ① 横向条形图(排名/对比首选): 标签在条外, 数值统一右对齐到同一列(都用 .chart-val.out); 关键一条用默认色, 其余 .dim;
+     行宽 = 数值 ÷ 轴上限(这里轴上限 ≈ 2.16 亿 = 1.86/0.86, 所以 1.17→54%、0.67→31%、0.22→10.2%), 别凭手感写百分比;
+     一句话用 .chart-note 挂在标题行; 图注写口径+时点 -->
+<div class="chart" style="margin-top:var(--sp-5)">
+  <div class="chart-head">
+    <span class="chart-title">2026 Q2 各产品线营收</span>
+    <span class="chart-note">A 产品同比 +38%</span>
+  </div>
+  <div class="chart-row">
+    <span class="chart-lbl">产品 A</span>
+    <div class="chart-track">
+      <div class="chart-bar fx-grow-w" data-stage="2" style="--w:86%"><span class="chart-val chart-num">1.86 亿</span></div>
+    </div>
+  </div>
+  <div class="chart-row">
+    <span class="chart-lbl">产品 B</span>
+    <div class="chart-track grid">
+      <div class="chart-bar dim fx-grow-w" data-stage="2" style="--w:54%;--fx-delay:calc(var(--t2) + 120ms)"></div>
+    </div>
+    <span class="chart-val out chart-num fx-fade" data-stage="2" style="--fx-delay:calc(var(--t2) + 120ms)">1.17 亿</span>
+  </div>
+  <p class="chart-cap">口径: 集团合并报表 · 单位: 人民币 · 截至 2026-06-30</p>
+</div>
+
+<!-- ② 柱状图(≤5 个时点): .chart-plot.grid 给细网格, 数值放柱外, 逐个错峰 -->
+<div class="chart" style="margin-top:var(--sp-5)">
+  <!-- 柱高 = 数值 ÷ 轴上限, 轴必须从 0 起: 3.1/3.5/3.8 对 4.0% 轴 → 77.5% / 87.5% / 95%。
+       把轴上限写进标题或 note, 别让读者猜比例尺(凭手感写百分比 = 数据失真) -->
+  <div class="chart-head"><span class="chart-title">付费转化率 · 周(轴 0–4.0%)</span><span class="chart-note">目标 4.0%</span></div>
+  <div style="display:flex;gap:var(--sp-3);align-items:flex-start">
+    <div class="chart-ticks" style="--cols-h:300px"><span>4%</span><span>3%</span><span>2%</span><span>1%</span><span>0</span></div>
+  <div class="chart-cols" style="--cols-h:300px">
+    <div class="chart-col">
+      <div class="chart-plot-cell grid">
+        <div class="chart-bar-wrap" style="height:77.5%">
+          <span class="chart-val chart-num fx-fade" data-stage="2">3.1%</span>
+          <div class="chart-bar dim fx-grow-y" data-stage="2"></div>
+        </div>
+      </div>
+      <span class="chart-lbl fx-fade" data-stage="2">W08</span>
+    </div>
+    <div class="chart-col">
+      <div class="chart-plot-cell grid">
+        <div class="chart-bar-wrap" style="height:87.5%">
+          <span class="chart-val chart-num fx-fade" data-stage="2" style="--fx-delay:calc(var(--t2) + 120ms)">3.5%</span>
+          <div class="chart-bar dim fx-grow-y" data-stage="2" style="--fx-delay:calc(var(--t2) + 120ms)"></div>
+        </div>
+      </div>
+      <span class="chart-lbl fx-fade" data-stage="2" style="--fx-delay:calc(var(--t2) + 120ms)">W10</span>
+    </div>
+    <div class="chart-col">
+      <div class="chart-plot-cell grid">
+        <div class="chart-bar-wrap" style="height:95%">
+          <span class="chart-val chart-num fx-fade" data-stage="2" style="--fx-delay:calc(var(--t2) + 240ms)">3.8%</span>
+          <div class="chart-bar fx-grow-y" data-stage="2" style="--fx-delay:calc(var(--t2) + 240ms)"></div>
+        </div>
+      </div>
+      <span class="chart-lbl fx-fade" data-stage="2" style="--fx-delay:calc(var(--t2) + 240ms)">W12</span>
+    </div>
+  </div>
+  </div>
+  <p class="chart-cap">口径: 注册后 7 日内付费 · 来源: 内部指标平台 · 截至 2026-08-15</p>
+</div>
+
+<!-- ③ 折线图(>5 个时点): inline SVG, 描边画入 + 数据点逐个弹出; 面积用同色低透明度渐变 -->
+<svg class="fx-draw" data-stage="2" viewBox="0 0 900 380" style="width:100%;height:auto" fill="none">
+  <defs><linearGradient id="g1" x1="0" y1="0" x2="0" y2="1">
+    <stop offset="0" stop-color="var(--accent)" stop-opacity=".22"/>
+    <stop offset="1" stop-color="var(--accent)" stop-opacity="0"/>
+  </linearGradient></defs>
+  <path d="M40,330 L215,290 L390,300 L565,180 L740,120 L875,88" stroke="var(--line)" stroke-width="1"/>
+  <path d="M40,330 L215,290 L390,300 L565,180 L740,120 L875,88 L875,360 L40,360 Z" fill="url(#g1)"/>
+  <polyline points="40,330 215,290 390,300 565,180 740,120 875,88"
+    stroke="var(--accent)" stroke-width="5" stroke-linecap="round" stroke-linejoin="round" stroke-dasharray="800"/>
+  <g fill="var(--accent)">
+    <circle class="fx-dot" data-stage="2" cx="565" cy="180" r="9" style="--fx-delay:calc(var(--t2) + 500ms)"/>
+    <circle class="fx-dot" data-stage="2" cx="875" cy="88" r="9" style="--fx-delay:calc(var(--t2) + 800ms)"/>
+  </g>
+  <text x="875" y="60" text-anchor="end" fill="var(--fg)" font-size="26" class="chart-num">96.4</text>
+</svg>
+<!-- fx-draw 的 dasharray 要 ≥ 线长; 数据点用 fx-dot(自带轻微过冲弹入)。**标记规则要统一**: 要么每个点都标、要么只标末点(带数值);
+     只给中间某个点标点、又没有对应数值, 视觉验收会读成"没标全"。
+     ⚠ SVG 的 viewBox 宽高比要和容器一致, 否则 preserveAspectRatio 会把图形缩到中间一小块(实测: 半个版面里图形只占 40%);
+     满宽用 width:100%;height:auto; 半宽/固定高就要把 viewBox 改成接近容器的比例, 别靠 height 硬压 -->
+
+<!-- ④ 环形: fx-sweep 真扫出(--p-to = 目标百分比), 中心数字 fx-count 滚动(--n-to = 整数) -->
+<div class="fx-sweep" data-stage="2" style="--p-to:62;position:relative;width:340px;height:340px;border-radius:50%;
+     background:conic-gradient(var(--accent) 0 calc(var(--pv) * 1%), var(--panel-2) 0)">
+  <div style="position:absolute;inset:52px;border-radius:50%;background:var(--bg);
+              display:flex;flex-direction:column;align-items:center;justify-content:center;gap:4px">
+    <span class="fx-count chart-num" data-stage="2" style="font-size:var(--fs-h1);--n-to:62"></span>
+    <span style="color:var(--muted);font-size:var(--fs-tiny)">市场份额(%)</span>
+  </div>
+</div>
+<!-- ⚠ 中心数字交给 fx-count 后不要再手写 "62"(计数器会自己填); 小数(62.4)不支持, 保持静态或拆两段。
+     fx-count 在入场前是 opacity:0 —— 否则画面会停在"灰色的 0", 被读成"份额 0%/没数据"(视觉验收实测) -->
+
+<!-- ⑤ bullet(实际 vs 目标, 汇报最实用): 一条轨道 + **目标竖刻度线**(标签贴刻度) + 实际条 + 条外数值
+     ⚠ 目标标签别贴轨道右端 —— 会被读成"刻度在最右边"; 也别只给轨道加淡色分界, 看不见等于没有 -->
+<div class="chart-row" style="margin-top:var(--sp-4)">
+  <span class="chart-lbl">年度目标达成</span>
+  <div class="chart-track" style="height:26px">
+    <div class="chart-bar fx-grow-w" data-stage="2" style="--w:78%"></div>
+    <div class="chart-target" style="left:88%"><span>目标 88%</span></div>
+  </div>
+  <span class="chart-val out chart-num fx-fade" data-stage="2">78%</span>
+</div>
+
+<!-- ⑥ slope(前后对比): 两列点 + 斜线, 比双柱更省版面; 主体用 --accent, 对照组用 --line-strong
+     ⚠ 纵坐标按数值比例摆点, 而且**要单独画一条 0 基线**: 让对照组那条线兼任基线, 会出现"1M 画在 2M 之上"(实测被判 fail);
+        "1M→2M 的间距"和"2M→10M 的间距"必须与数值差成比例, 否则"涨了 10 倍"在图上读不出来(实测被判 fail);
+     对照组要如实画: 两年持平就画**水平线**, 并且**两端都标数值**, 标签别压在另一条线上 —— 只标一端、线却明显倾斜, 会被读成"数据不明/自相矛盾"(实测被视觉验收判 fail) -->
+<svg class="fx-draw" data-stage="2" viewBox="0 0 720 320" style="width:100%;height:auto" fill="none">
+  <text x="0" y="24" fill="var(--fg-3)" font-size="22">三年前</text>
+  <text x="720" y="24" text-anchor="end" fill="var(--fg-3)" font-size="22">今天</text>
+  <path d="M60,250 L660,90" stroke="var(--accent)" stroke-width="5" stroke-linecap="round" stroke-dasharray="800"/>
+  <path d="M60,150 L660,200" stroke="var(--line-strong)" stroke-width="4" stroke-linecap="round" stroke-dasharray="800"/>
+  <circle class="fx-dot" data-stage="2" cx="60" cy="250" r="10" fill="var(--accent)"/>
+  <circle class="fx-dot" data-stage="2" cx="660" cy="90" r="10" fill="var(--accent)" style="--fx-delay:calc(var(--t2) + 600ms)"/>
+  <circle class="fx-dot" data-stage="2" cx="60" cy="150" r="8" fill="var(--line-strong)"/>
+  <circle class="fx-dot" data-stage="2" cx="660" cy="200" r="8" fill="var(--line-strong)" style="--fx-delay:calc(var(--t2) + 600ms)"/>
+  <text x="60" y="288" fill="var(--fg)" font-size="24" class="chart-num">62</text>
+  <text x="660" y="72" text-anchor="end" fill="var(--accent)" font-size="30" class="chart-num">92</text>
+</svg>
+
+<!-- ⑦ sparkline(数字旁的迷你趋势, 不占版面): 只有线, 无坐标轴, 用来暗示趋势 -->
+<span style="display:inline-flex;align-items:center;gap:12px">
+  <span class="chart-num" style="font-size:var(--fs-h2);color:var(--fg)">96.4</span>
+  <svg class="fx-draw" data-stage="2" viewBox="0 0 160 48" style="width:160px;height:48px" fill="none">
+    <polyline points="4,40 30,34 56,36 82,22 108,18 134,10 156,6"
+      stroke="var(--accent)" stroke-width="3" stroke-linecap="round" stroke-linejoin="round" stroke-dasharray="300"/>
+    <!-- 颜色跟主体走(--accent); 只有"涨/跌"本身是信息时才用 --up/--down。末端补个小圆点收口 -->
+  </svg>
+</span>
+```
+
+Quick selection guide: ranking/comparison → horizontal bars; time trend (≤5 points) → columns, (>5 points) → line; share (≤3 slices) → donut; actual vs target → bullet; before/after → slope; trend beside a number → sparkline. **Only one chart per slide** (if you place two, you must rank them or split the slide), and the figure caption is always required.
+
+⚠ The whole chart block (including the figure caption) must land **above the subtitle safe area**: at 1080 height the bottom 84–168px is reserved for subtitles, so leave `padding-bottom: 190px` on the layout container.
+
+---
+
+# Table toolkit (the five forms of a data table)
+
+Tables are the easiest thing to make ugly: not enough row height, numbers not right-aligned, change colours used wrongly, gridlines everywhere. The primitives all live in `tokens.css` (`.tbl / .kv / .matrix / .rank`); **for existing projects run `node scripts/init-project.mjs <project> --upgrade-css` first** (idempotent), otherwise none of these classes take effect.
+
+## Table disciplines (seven)
+
+1. **Body type size for the main data, row height ≥72px**: the body of `.tbl`/`.kv`/`.matrix`/`.rank` is already `var(--fs-body)` (readable on both 1080p and phones), and only the table header drops to `--fs-caption`; row height is built in at 72px —— do not shrink the type size or row height just to squeeze in one more row (measured: a 24px table is unreadable on a phone).
+2. **One table per page, 6 rows maximum**: beyond that, split the page or keep only the top 5 + "others". A table page is naturally dense in information; squeeze it further and nobody reads it.
+3. **Right-aligned values + monospace** (`.num`/`.val`): decimal-point alignment is the main source of "professional feel"; numbers use `tabular-nums` so they do not jitter when rolling or switching.
+4. **Change colours go through tokens** (`.up`/`.down` → `var(--up)/var(--down)`): financial subjects flip them by audience (A-shares/HK: red up, green down), see compliance.md; and for non-financial subjects do not invent colours other than green/red.
+5. **Horizontal rules only, no vertical rules, no zebra striping**: the header underline of `.tbl` plus thin lines between rows is enough; zebra stripes and a full grid make the frame dirty.
+6. **Highlight exactly one row** (`.key`: accent vertical bar on the left + 8% background), usually "us / the current phase"; the total row uses `.sum` (thicker top border). Do not use either on several rows at once.
+7. **The table caption states the measure and the as-of date** (the same `.chart-cap` treatment as charts): as with charts, the credibility of a number comes from "when and how it was calculated".
+
+## The five forms and their recipes
+
+```html
+<!-- ① 数据表(最常用: 多期/多主体对比, 带涨跌) —— 表头弱化, 数值右对齐, 涨跌走令牌 -->
+<table class="tbl fx-fade" data-stage="2">
+  <thead><tr><th>季度</th><th class="num">营收</th><th class="num">同比</th><th class="num">毛利率</th></tr></thead>
+  <tbody>
+    <tr><td>2026 Q1</td><td class="num">3.2 亿</td><td class="num up">+18%</td><td class="num">61.4%</td></tr>
+    <tr class="key"><td>2026 Q2</td><td class="num">4.1 亿</td><td class="num up">+27%</td><td class="num">63.8%</td></tr>
+    <tr><td>2026 Q3</td><td class="num">3.9 亿</td><td class="num down">-4%</td><td class="num">62.1%</td></tr>
+    <tr class="sum"><td>合计</td><td class="num">11.2 亿</td><td class="num up">+14%</td><td class="num">62.4%</td></tr>
+  </tbody>
+</table>
+<p class="chart-cap">口径: 集团合并报表 · 单位: 人民币 · 截至 2026-09-30</p>
+
+<!-- ② 规格表(.kv: 无表头, 左标签右值) —— 参数/条款/清单这类"一对一条目"用它, 比两列表格干净 -->
+<table class="kv fx-fade" data-stage="2">
+  <tr><td>上下文窗口</td><td>10,000,000 token</td></tr>
+  <tr><td>单次最长输出</td><td>128,000 token</td></tr>
+  <tr><td>定价(输入 / 输出)</td><td>¥0.8 / ¥2.4 每百万 token</td></tr>
+  <tr><td>上线时间</td><td>2026-08-15</td></tr>
+</table>
+
+<!-- ③ 对比矩阵(.matrix: 行是维度、列是方案, 只放勾叉; .hi 高亮"我们"那一列, th/td 均可) -->
+<table class="matrix fx-fade" data-stage="2">
+  <thead><tr><th>能力</th><th>方案 A</th><th class="hi">方案 B(本方案)</th><th>方案 C</th></tr></thead>
+  <tbody>
+    <tr><td>离线可用</td><td class="no">✗</td><td class="hi yes">✓</td><td class="no">✗</td></tr>
+    <tr><td>成本可控</td><td class="no">✗</td><td class="hi yes">✓</td><td class="yes">✓</td></tr>
+    <tr><td>交付周期 ≤ 2 周</td><td class="yes">✓</td><td class="hi no">✗</td><td class="yes">✓</td></tr>
+  </tbody>
+</table>
+<p class="chart-cap">口径: 内部评测 · 2026-09 · 勾叉按"是否满足该维度硬指标"判定</p>
+
+<!-- ④ 排名表(.rank: 名次 + 微缩条 + 数值) —— 表格里嵌条形, 比纯数字一眼看出差距 -->
+<table class="rank fx-fade" data-stage="2">
+  <tr><td class="no">1</td><td>产品 A</td><td class="bar"><span class="fx-grow-w" data-stage="2" style="--w:100%"></span></td><td class="val">1.86 亿</td></tr>
+  <tr><td class="no">2</td><td>产品 B</td><td class="bar"><span class="dim fx-grow-w" data-stage="2" style="--w:63%;--fx-delay:calc(var(--t2) + 120ms)"></span></td><td class="val">1.17 亿</td></tr>
+  <tr><td class="no">3</td><td>产品 C</td><td class="bar"><span class="dim fx-grow-w" data-stage="2" style="--w:36%;--fx-delay:calc(var(--t2) + 240ms)"></span></td><td class="val">0.67 亿</td></tr>
+</table>
+<!-- 微缩条的宽度同样 = 数值 ÷ 轴上限(见"柱高必须由数值算出来"那条) -->
+
+<!-- ⑤ 时间事件表(时间 × 事件两列) —— 比 timeline 版式更适合"事件多、要点长"的题材 -->
+<table class="kv fx-fade" data-stage="2">
+  <tr><td>2021</td><td>团队成立, 首个原型</td></tr>
+  <tr><td>2023</td><td>产品上线, 首批 1 万用户</td></tr>
+  <tr><td>2026 Q2</td><td>通过聆讯, 递交港股上市申请</td></tr>
+</table>
+```
+
+Quick selection guide: multi-period/multi-entity comparison → ① data table; one-to-one entries → ② spec table (or ⑤ time-event table); ticking options across several plans → ③ comparison matrix; ranking to show the gap → ④ ranking table; totals/subtotals → ① + `.sum`; price/score tiers → ① with `.key` highlighting the current tier. **Do not put a table and a chart on the same page** (one thing per page); the voiceover for a table page must still "read the table" —— never treat the table as a backdrop.
+
+## Motion switches (three granularities)
+
+| Granularity | How | Effect |
+|---|---|---|
+| **Whole project** | `<html data-theme="..." class="no-fx">` (or `.stage.no-fx`) | all entrance/ambient motion off, elements go straight to their final state; motion capture detects no animation window and outputs static frames automatically, with duration and A/V sync unaffected; subtitles are still burned in |
+| **Single slide** | add `no-fx` to that slide's root element (tokens.css rules match by descendant, so any container works) | only this slide has no motion, the rest behave normally |
+| **Single element** | give it no fx class and no data-stage | that element renders statically |
+
+When to use: you need a quick output under time pressure, the subject calls for restraint (government/legal/obituary material), or the user explicitly says "no animation". It can be put to the user as one of the kickoff alignment questions; the default is everything on.
+
+**Chart motion is inside the switch range too**: with it off, bars/columns are directly at their final width, the donut is a complete sector, and numbers are at their final value (their static defaults are written as the final value, so no extra reset rules are needed).
+
+**Verify before delivery**: on the screening page from `preview-page.mjs`, pressing `X` gives the "motion / motion off" comparison —— if it **goes blank** when switched off, some keyframes failed to lift `opacity:0` back; the two versions must look identical to be clean. Older projects (tokens.css generated before this rule) have no `no-fx` rules: run `node scripts/init-project.mjs <project dir> --upgrade-css` first (idempotent), otherwise `<html class="no-fx">` silently stops working and the page goes completely blank instead.
+
+---
+
+# Theme quick reference (audience → theme)
+
+| Scenario | Theme |
+|---|---|
+| Business reporting / fundraising / earnings | `minimal-white` `swiss-grid` `corporate-clean` (plus the original `a`, off-white with warm orange) |
+| Finance / investment research (regulated) | the same three (high data density, restrained colour); **once chosen, handle the disclaimer line and the change colours per `compliance.md` —— do not just pick a theme and start running** |
+| Editorial / magazine / narrative | `editorial-serif` `magazine-bold` |
+| Tech / AI / developer (dark) | `tokyo-night` `catppuccin-mocha` `nord` (plus the original `b`, dark tech green) |
+| Consumer / Xiaohongshu / lifestyle | `xiaohongshu-white` `soft-pastel` (plus the original `c`, minimal blue, a general white-background option) |
+
+Before settling on a theme, run `node <skill>/scripts/check-theme.mjs <project>`: it computes the contrast of every theme; do not use a theme that fails the gate (especially the subtitle pill contrast on dark themes).
+
+## Custom accent colour (when the user supplies a brand colour)
+
+A theme's accent colour is a token, so just override it —— append to the end of the **project's own** `slides/tokens.css` (edit the project copy, never the skill template):
+
+```css
+/* 品牌色覆写: 以 software 主题为底, 只换强调色三兄弟 */
+[data-theme="minimal-white"] {
+  --accent: #E8582A;          /* 品牌主色 */
+  --accent-ink: #FFFFFF;      /* 压在主色上的文字色, 写完必须验 */
+  --accent-2: #F08A5C;        /* 渐变/次级用色 */
+  --accent-3: #B8431E;        /* 深色态 */
+  --grad: linear-gradient(135deg, #E8582A, #F08A5C 55%, #B8431E);
+}
+```
+
+**Hard requirement**: run `node <skill>/scripts/check-theme.mjs <project>` immediately after the change —— `--accent-ink` against `--accent` must be ≥3:1 (below 4.5:1 it may only be used for large type/graphics, never small text). For dark themes also take a quick look at whether `--sub-bg` (the subtitle pill) still separates from the background.
+
+# Vertical canvas (1080×1920)
+
+The `width/height` in `script.json` decides the canvas; the render pipeline injects `--stage-w/--stage-h` into the page and the subtitle geometry narrows with it automatically. **But the layout must be reflowed**: left/right columns that work in landscape get cramped in portrait, so switch to vertical stacking.
+
+- Content margin goes from 96px to ~90px (portrait is narrower, so the usable area is smaller)
+- Title size drops to the `--fs-h2` (44px) order of magnitude —— in portrait, 44px reads like ~78px in landscape
+- Reduce the information per screen by one notch: 3 bullets in landscape → 2 in portrait
+- Use `--img-ratio:3/4` or `1/1` for image frames; they fit portrait composition better than 16/9
+- The subtitle position is decided automatically by `--stage-h`; no manual tuning needed
+
+# ⚠ Must-know: how stage delay is implemented (read before changing anything)
+
+The entrance delay is **not** a standalone `animation-delay: var(--tN)` declaration; it is written into the variable slot of the `animation` shorthand on `.fx-*`:
+
+```css
+[data-stage="2"] { --fx-delay: var(--t2); }                                    /* 只设变量 */
+.fx-rise { animation: fx-rise .9s var(--ease-out) var(--fx-delay, 0ms) both; } /* 变量进 shorthand */
+```
+
+**Why**: `[data-stage="2"]` and `.fx-rise` both have (0,1,0) specificity; if you split this into two declarations, the `.fx-*` shorthand that appears later in the file resets `animation-delay` back to 0 —— every stage then enters at 0 seconds, while elements carrying an inline `calc(var(--t2)+150ms)` are still correct, producing a hard-to-diagnose failure where "some elements have weird timing and some are fine" (confirmed by measurement on 2026-09-17).
+
+**Therefore**: ① do not hand-write `animation-delay: var(--tN)` (use `--fx-delay` or an inline calc instead); ② when defining a custom entrance animation, the delay slot must be written as `var(--fx-delay, 0ms)`; ③ to pin a container's base time, use `style="--stagger-base:var(--t3)"`.
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/compliance.md b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/compliance.md
new file mode 100644
index 00000000..a544abc3
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/compliance.md
@@ -0,0 +1,92 @@
+# Domains and Compliance (required reading for regulated topics)
+
+**When to read this**: at kickoff alignment, first settle **which domain** this video belongs to. If the topic falls in a **regulated domain** — finance/investment, healthcare/medical, legal, government policy, marketing efficacy claims — read this section to the end before you start writing.
+
+It is not legal advice; it turns "the things this kind of video always gets challenged on" into process checkpoints: whether a disclaimer line is required, what every number must carry, and which way the up/down colours point. Get these three wrong and the picture gets reworked no matter how good it looks.
+
+**Name your jurisdiction, then apply its rules.** Regulated domains exist in every market; the specific regulator does not. Illustrative only — US: SEC / FTC / FDA; EU & UK: FCA / ESMA / EMA; China: CSRC / NMPA / SAMR; anywhere else: check the local regulator. So the question to put to the user is never "which regulator?" but **"which market is the audience in?"** — the answer fixes the disclaimer wording, the data rules and the colour convention for the whole video.
+
+## 1. Domain confirmation (part of the first batch of kickoff questions)
+
+**The two questions to ask** (never assume an answer — put both into Gate 0 alongside the data definitions):
+
+| Question | Options / default |
+|---|---|
+| Which domain is this video in? | Finance/business · Technology/product · Science/education · Brand/marketing · Culture/history · Lifestyle/consumer · Government/policy — the domain decides what the visuals must carry (disclaimer line, data provenance, units), how far the narration may go, and which up/down colour convention applies |
+| Disclaimer and source attribution required? | **Required by default** in regulated domains; pure opinion/brand/lifestyle may skip it. A "no" must still be stated explicitly by the user and recorded in the Gate 0 baseline and `research/notes.md` |
+
+**Why you must ask**: a disclaimer is a **content attribute**, not a legal opinion — it does not remove liability, but when it is missing in a finance or medical context it will almost certainly be demanded later, and the rework lands in Phase 5 (the final cut), the most expensive place for it.
+
+## 2. Finance / investment research (the class that goes wrong most often)
+
+### 2.1 Three red lines for the narration
+
+1. **No trading recommendations, no price predictions.** Never write "worth watching / buy the dip / could rise to / sure profit / room to double". To express an expectation, attribute it to a subject: "one institution's target-price range is …", with the institution and the publication date.
+2. **Don't state opinions as facts.** "The most popular fund", "growth is a certainty" must carry a subject (the company says so / the agency's rating) or be rewritten as a verifiable statement (see `research.md` hard rule 5).
+3. **Don't invent numbers or charts.** Candlesticks, earnings screenshots and rankings must all come from official sources; **no data graphic may be AI-generated** (global hard rule — see `image-sources.md`).
+
+### 2.2 Every number must carry three things
+
+| Requirement | Example |
+|---|---|
+| **Definition / scope** | annualised / single quarter / cumulative; GAAP / non-GAAP; year-over-year / quarter-over-quarter; adjusted-price basis |
+| **Currency and units** | USD / EUR / CNY / HKD; millions / billions — always state which |
+| **Date or period** | "as of the 2026-06-30 close", "Q2 2026 results (reported in August)" |
+
+Financial data has a **reporting lag**: don't write "revenue is currently X", write "Q2 revenue of X as reported by the company in August". If second-by-second data such as a share price or an exchange rate must appear, write "as of the YYYY-MM-DD close" and cite the source, and **state the data date in the disclaimer line** (the video stays online for years; the number will be stale).
+
+**Provenance rule**: at least two independent sources for every key claim, each with its date and its definition/scope; state currency and units; and keep what a source reported separate from your own inference — never let an inference ride along in the same sentence as a reported figure.
+
+### 2.3 Up/down colours: this convention differs by region
+
+The semantic colours in `tokens.css` default to `--good` green / `--bad` red — the US/EU reading (green up, red down). **Greater China audiences (CN/HK/TW) read the opposite: red up, green down** — flip this one by accident and the whole video reads inverted. What to do:
+
+```css
+/* 追加在项目 slides/tokens.css 末尾: A股/港股受众 = 红涨绿跌 */
+:root { --up: #D92B2B; --down: #12A150; }
+```
+
+(That comment means: append to the end of the project's `slides/tokens.css` — A-share / Hong Kong audiences = red up, green down.)
+
+- Metric cards always write **`var(--up)` / `var(--down)`** (never `var(--good)` / `var(--bad)` directly); the template already defines `--up: var(--good); --down: var(--bad)`, so an older project missing those two variables just adds the line above (`check-slides.mjs` reports undefined variables).
+- Markets that read green as up (US/EU and others) keep the default — do not flip them.
+- **Ask which market the audience is in while confirming assets at Gate 3**, and settle it together with the colour convention. Default to the audience's own convention, and stay consistent within one video.
+
+### 2.4 Disclaimer: how to write it, where to put it
+
+- **Use a fixed template, don't improvise.** English: "Not investment advice. Data as of <date>, source: <source>." Chinese: "本视频仅为信息分享,不构成任何投资建议。数据来源:<来源>,截至 <日期>。" (Medical follows the same shape: not medical advice — for diagnosis and treatment consult a licensed physician.)
+- **Placement**: the `closing` slide carries a small `.disclaimer` line (bottom or bottom-right), **on screen ≥3 seconds** and legible; a long video may repeat the same small line on the `title-hero` or on the first data page.
+- **The narration does not read the disclaimer aloud** (and it is not burned into the subtitles): the compliance text is carried by the picture, and the voiceover keeps its pace. If the user insists on reading it, read it — but it stays out of the subtitles.
+- A disclaimer **does not replace verification**: get a number wrong and ten disclaimer lines will not save it.
+
+## 3. Healthcare / medical
+
+- No efficacy claims, no diagnosis, no recommending a specific drug or treatment; if you must mention one, describe "what the authoritative guideline / the approved label says", with version and date.
+- Disclaimer line: "This video is general information, not medical advice; consult a licensed physician for diagnosis and treatment."
+- Source ranking follows the finance standard (guidelines first, then the drug regulator — FDA / EMA / NMPA — or the pharmacopoeia, then peer-reviewed papers); **second-hand social-media retellings are not acceptable**.
+
+## 4. Legal / government / policy
+
+- Quote provisions **verbatim + effective date**; distinguish "currently in force" from "draft for comment".
+- No case-specific interpretation ("in your situation you could argue…" → "the provision reads…, individual cases need a licensed lawyer").
+- For policy and statistics cite **official releases only**, and watch for revision notes (see `research.md`).
+
+## 5. Brand / marketing (advertising claims)
+
+- **Absolute wording is a minefield**: best / no.1 / top / only / national-level / 100% effective / never — none of it without an authoritative basis or a substantiation document the user supplies.
+- Efficacy claims need substantiation ("3x more efficient" → state the test conditions and the source); without substantiation, make the claim qualitative.
+- Third-party trademarks, likenesses and endorsement relationships: **never use without authorisation**; comparative claims must be verifiable and must state the comparator and the basis.
+
+## 6. Closing checklist (run once at Gate 0 and once at Gate 5)
+
+- [ ] Domain confirmed, and the **audience market named** — the rules applied are that market's (US: SEC/FTC/FDA; EU & UK: FCA/ESMA/EMA; China: CSRC/NMPA/SAMR; elsewhere: the local regulator)
+- [ ] **The user explicitly said yes / no to a disclaimer** (recorded in the notes)
+- [ ] Every number carries definition / scope + currency / units + date or period (finance)
+- [ ] No trading recommendations, no price predictions, no absolute wording
+- [ ] Up/down colours match the audience market (Greater China = red up / green down; US/EU = green up / red down) and use `var(--up)` / `var(--down)`
+- [ ] If confirmed: the disclaimer is on screen, stays ≥3s, and is not in the narration or the subtitles
+- [ ] All data graphics come from official sources, none AI-generated
+
+> **Regional example — mainland China** (the market this skill was first written for): regulators CSRC / NMPA / SAMR; financial and medical disclaimers in Chinese, as above; A-share and Hong Kong audiences read red = up, so apply the flip in §2.3; advertising claims fall under the Advertising Law, whose absolute-wording list (best / no.1 / top / only / national-level / 100% effective / never) is treated as a hard prohibition. In any other market, substitute that market's regulator and wording rules — the process above does not change.
+
+Related: `research.md` (source ranking and definitions), `authoring.md` (layout and the `.disclaimer` snippet), `image-sources.md` (asset compliance).
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/image-sources.md b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/image-sources.md
new file mode 100644
index 00000000..abb189ab
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/image-sources.md
@@ -0,0 +1,132 @@
+# Image & Asset Sourcing SOP
+
+> Source: distilled from end-to-end testing on 2026-09-17 (hard user requirement). The core principle in one sentence:
+> **When an image is incomplete / cut off / badly proportioned, first search for one that shows the whole subject — do not force-crop it and do not force-use it.**
+
+## 1. Image sourcing paths (by priority)
+
+| Path | How | Use for | Tested |
+|---|---|---|---|
+| **A · Official channels** | Official brand kit / simple-icons (curl works); take product shots from official help/docs; use official reports for data | logos, product shots, data | ✅ Strongest compliance; for factual topics stay at this tier whenever possible |
+| **B · Direct from the official site (highest success rate)** | `node <skill>/scripts/fetch-official-images.mjs <official-site-URL>` lists candidate images (including CSS background images and lazy-loaded ones) → pick indexes → `--get 1,3 --out-dir assets` to download. **When the site needs login / scroll-loading, or the script cannot open it, use the agent's built-in browser to inspect the DOM, grab the image URLs, then `--url <URL>[,<URL>...] --out-dir assets` to write them to disk** (the same host / redirect / size / filename validation, and no Playwright needed); the tool flags suspicious "off-site domain" assets, and can feed straight into `prep-image.mjs --check` | brand logos, product shots, hero images, CEO portraits, news images | ✅ tested on minimaxi.com, 6/6 all relevant |
+| **C · Wikipedia Commons / CC** | Search historical figures, landmarks, public institutions | historical topics | ✅ download once, reuse forever; zero compliance risk |
+| **D · Screenshots of announcements and authoritative coverage** | **Negative-event topics** (collapse / penalty / lawsuit / controversy): the official site will never carry this material, so A/B are guaranteed to fail — the correct source is **the primary carrier of the event itself**: exchange filings, regulatory documents, company official statements / apology letters > media coverage pages (Tier 3 media are acceptable from either region — international: Reuters, AP, Bloomberg, FT, WSJ, BBC, CNBC, The Verge; Chinese: Xinhua, CCTV, The Paper 澎湃, Caixin, Tencent News 腾讯新闻, Sina, NetEase — self-media excluded). Open with the built-in browser or Playwright and **screenshot** (command below) | collapse, regulatory penalty, lawsuit, scandal / controversy | ✅ reporting-based citation; the screenshot must keep the outlet name and date visible |
+| ~~Legacy · Bing image search~~ | image-downloader skill | — | ❌ for niche Chinese-language topics 5/5 were irrelevant (anti-bot interstitial pages + "you may also like" recommendation cards), **off by default** |
+
+**Discipline when inspecting with the built-in browser** (take this route when the site needs interaction or the script cannot open it):
+
+1. **Once you have the image URL, do not "save as" or let the browser download it directly**: hand it to `--url` so it goes through the host allow-list / per-hop redirect re-check / 30MB cap / filename sanitisation / no-overwrite discipline.
+2. **Do not treat "opening a new page" as a refresh**: in the host browser `navigate(replaceCurrentTab)` often opens a new tab (site data held in IDB will look like it "vanished into thin air"); to reload, use reload / F5.
+3. **Batch actions belong in scripts, not on the browser's back**: clicking 15+ times in a row or bulk downloading triggers the site's human verification; leave image search / bulk download to the existing tools (see "Tier C" below).
+4. **Take only assets on official domains**: before downloading, check whether the `src` found by inspect is on the official CDN; images from non-official domains are never used (compliance and copyright).
+
+**On the legacy Bing route**: this skill ships **no** Bing crawler of its own (shipping one would mean shipping a known-failing path). If image-downloader happens to be installed, use it **only when** the topic is an English-language mainstream brand or a generic real-world scene and you are willing to hand-filter every image; for Chinese-language topics, niche companies and the China context, always take path B. **Do not write a throwaway crawler just to fetch an image** — start with the existing tools, and if they come up empty, fall back to pure layout.
+
+**Screenshot discipline for Path D** (negative-event topics, news/announcement screenshots):
+
+1. **Source tiering**: exchange filings / regulatory documents > company official statements (including apology letters and clarification announcements) > media coverage — **rank by primary-ness and editorial accountability, not by country**: Tier 1 primary documents (SEC EDGAR filings, HKEX / SSE / SZSE announcements, regulatory notices, FDA / FTC decisions) outrank Tier 2 company statements, which outrank Tier 3 mainstream media; and within Tier 3, international outlets (Reuters, AP, Bloomberg, FT, WSJ, BBC, CNBC, The Verge) and Chinese outlets (Xinhua, CCTV, The Paper 澎湃, Caixin, Tencent News 腾讯新闻, Sina, NetEase) are **equally qualified sources** — do not insist on a single "most authoritative" one. What is actually excluded is only self-media, aggregator accounts and content farms (unverifiable, often agenda-driven).
+2. **Screenshots must retain verifiable information**: the outlet/institution name, the headline and the publication date must be inside the frame — a screenshot with the source information cropped out is unusable. Close cookie banners and recommendation feeds before shooting; capture the article body area only.
+3. **Saving and logging**: store screenshots in `assets/`, and log one MANIFEST line per file: "content / source URL / publication date / licence (fair use for reporting)"; in the HTML use `.img-frame.contain` + `.img-cap` to state the source and date (e.g. "Image: XX News report · 2026-09").
+4. **Negative financial topics are also bound by compliance.md**: still add the closing disclaimer line, give no trading advice in the narration and no price predictions; third-party logos / people appearing in screenshots follow the existing trademark / portrait rules.
+5. Screenshot command (wait for lazy-loading, then capture the article body area; the same applies to the host's built-in browser screenshot):
+
+```bash
+node -e "const {chromium}=require('playwright');(async()=>{
+  const b=await chromium.launch(); const p=await b.newPage({viewport:{width:1280,height:900}});
+  await p.goto(process.argv[1],{waitUntil:'domcontentloaded'});
+  await p.waitForTimeout(2500);                                   // wait for lazy-loading
+  const art=(await p.$('article'))||(await p.$('main'))||await p.$('body');
+  await art.screenshot({path:process.argv[2]});
+  await b.close();})()" "https://news-site.example/article" assets/news-01.png
+```
+
+Abstract information (numbers, comparisons, quotes) **does not need an image**: use the layout itself (number card / big type / pull quote) as the information layer — the least effort and the most refined solution.
+
+## 2. Filtering at the search stage (mandatory, before downloading)
+
+**Query design**
+- ✅ Add positive words: **panorama / full view / front-on / wide shot / official**
+- ❌ Add exclusion words: **close-up / macro / corner of / half-cut**
+- For people add `official portrait`; for products add `press photo` / `official`
+
+**Glance over the thumbnails first, ask four questions:**
+1. Is the subject centred and prominent? (subject hugging the edge → replace immediately)
+2. Is the ratio close to 16:9? (target ratio first, avoid cropping)
+3. Any watermark / unrelated logo / someone else's face?
+4. Resolution ≥1200px? (anything that will look soft on screen is out)
+
+**If one image does not fit, keep searching the next one — do not settle.**
+
+## 3. Second check once you have the image (mandatory)
+
+```bash
+node <skill>/scripts/prep-image.mjs --check assets/xxx.png
+```
+
+It reports: dimensions, ratio, how much would be cut to reach 16:9, whether the resolution is too low, and a reminder that "if the subject is at the edge, replace the image first".
+
+**How to judge subject position**: eyeball the thumbnail — which band is the subject in, **top / middle / bottom**? At the edge (y near 0 or h) → **drop this one and re-search**; in the middle but the ratio is off → a small crop is fine, or just use the frame's letterbox mode (see §4).
+
+## 4. Prefer solving it with the "frame", not by cropping
+
+`.img-frame` (built into tokens.css) **hands ratio and cropping to the frame, while the image just fills it**:
+
+```html
+<!-- Cover image: cropped to the frame ratio; --img-pos decides which part of the subject is kept -->
+<figure class="img-frame fx-grow" data-stage="2" style="--img-ratio:16/10; --img-pos:top">
+  <img src="../assets/photo.png" alt="description">
+  <span class="img-tag">on site</span>
+</figure>
+<p class="img-cap">The caption goes outside the frame, not inside it (the frame will crop it)</p>
+
+<!-- Screenshots/charts/images with text: use contain, no cropping, centred with letterbox -->
+<figure class="img-frame contain" style="--img-ratio:16/9">
+  <img src="../assets/shot.png" alt="screenshot">
+</figure>
+```
+
+- **Hard rule: every image must sit in a frame; a bare `<img>` is forbidden**. A bare image blows out the layout at its own ratio.
+- `--img-ratio` can be whatever you want (16/10 for a steady look, 1/1 square, 16/9 widescreen) — **16:9 is not required**.
+- `--img-pos: top|center|bottom` = controls the visible part of the subject via object-position — this is the **first-choice replacement for hard cropping**.
+- Screenshots / charts / images carrying text **must** use `.contain`: cropping loses information.
+- `.img-scrim` darkens the bottom so white text stays readable over a photo; `.img-cap` is the caption, `.img-tag` the corner badge.
+
+## 5. Cropping (last resort, with hard limits)
+
+When 5 searches all disappoint, consider in order: ① **change the subject** (switch to a number card / big type / quote); ② **change the layout** (contain letterboxing also looks refined); ③ **find an officially licensed image**; ④ only then crop.
+
+```bash
+node <skill>/scripts/prep-image.mjs --crop assets/in.png assets/out.png --ratio 16:9 --anchor bottom
+```
+
+**Hard crop limits** (enforced by the script): the target ratio must not require upscaling to pad the edges; the cropped area must be **≤20%** — beyond that it refuses outright and tells you to switch to `contain`.
+
+After cropping you **must** eyeball whether the subject is intact — better to switch to `--img-pos` than to crop into the subject.
+
+## 6. The three-step visual verification (mandatory after any HTML change)
+
+1. **Re-capture**: `node <skill>/scripts/capture.mjs <project> --mode still`, then look at `preview/<id>.png`
+2. **Check the subject position**: is the subject centred and slightly above middle? Has half of it been cropped away?
+3. **Pull an actual frame from the finished render**: `ffmpeg -ss <timestamp> -i out/final.mp4 -frames:v 1 frame.png` — you **must check a frame from the finished render**, not just what the capture pipeline produced (capture and final render can differ in fades and subtitle timing)
+
+## 7. Compliance and logging (consistent with authoring.md)
+
+- Every asset lands in `assets/` and is logged in `assets/MANIFEST.md`: filename / content / source URL or channel / licence
+- No watermarks (if there is one, replace the image; cropping out a watermark requires confirming the cropped area is ≤20% and the subject is intact)
+- Trademarks/logos only in the context of "talking about that brand", never as decoration
+- Photos must come from a licensable source; a person's photo with no traceable licence is better replaced by a plain text quote card
+- Screenshots must state their origin (official document name + date)
+- **Never fabricate logos, screenshots, portraits or QR codes out of thin air**
+- Assets must be saved locally before being referenced in the HTML (capture only waits 4s; external images that do not load in time are skipped)
+
+## 8. Image sourcing advice by topic
+
+| Topic | Preferred path | Notes |
+|---|---|---|
+| Chinese AI companies / Hong Kong IPOs | B (official-site DOM) | Bing fails systematically; official hero/product pages carry the most complete assets |
+| Overseas tech companies | A (official brand kit) + B | curl logos straight from simple-icons |
+| Universities / campuses | B (school official site, "official account" images) | real-scene images must be "front-on / full view"; the subject easily hugs the edge |
+| Historical figures / old photos | C (Wikipedia CC) | mind the copyright term |
+| Collapse / regulatory penalty / lawsuit / scandal | D (announcement and coverage screenshots) | official sites carry no such material; screenshot mainstream media coverage or the announcement page and keep the outlet name and date visible; negative financial topics are also bound by compliance.md |
+| Earnings data / comparisons | no image | number card + table layout |
+| Abstract concepts (growth / connection / future) | no image | gradient + big type + icons |
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/render.md b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/render.md
new file mode 100644
index 00000000..6ce5296e
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/render.md
@@ -0,0 +1,187 @@
+# Render Pipeline and Troubleshooting
+
+## Why capture doesn't "wait a few seconds and then screenshot"
+
+A fixed sleep is the most fragile kind of wait: on a slow machine the shot is taken before the animation has finished and the fonts have swapped; on a fast machine it just waits for nothing. capture.mjs is entirely event-driven:
+
+- **Fonts**: wait for each `link[rel=stylesheet]` load/error → explicitly `load()` every face in `document.fonts` (`font-display: swap` postpones the download, and without an active load the font may never be fetched) → `fonts.ready` → two rAFs so layout settles on the real glyphs. Hard cap of 8s overall.
+- **Images**: wait for all `<img>` to be complete, 4s cap per image, skip anything that won't load (which is why assets must be local first).
+- **Page load**: `domcontentloaded`, no waiting for `load`/`networkidle` (a single hanging external resource would burn a dead wait into the result).
+
+## still mode (final-state screenshots)
+
+All finite animations jump straight to their final state with `Animation.finish()` (fill-mode forwards holds it), infinite ambient animations keep running naturally, and the screenshot is "what the viewer sees when this slide is finished". Speed: 1–2s per slide.
+
+## motion mode (animations really go into the video)
+
+Principle = deterministic frame-by-frame stepping:
+
+1. `pause()` all animations and pin them to `currentTime = 0`.
+2. Every frame, set **every animation**'s `currentTime` uniformly to `i / fps`. CSS animations carry their own `animation-delay`, and currentTime is absolute time including the delay, so per-layer timing is naturally correct — which is also why you must not hand-write delays in the HTML and must let the pipeline inject `--t1/--t2/--t3`.
+3. `page.screenshot()` each frame into a PNG sequence.
+4. At encode time `tpad=stop_mode=clone` clones the last frame to fill the whole slide duration: only the animation window is rendered frame by frame, still segments cost nothing.
+
+Shoot only the animation window (longest animation endTime + 0.25s) rather than the full slide duration; 8 slides × 30fps usually needs only 600–1200 frames, about 2–6 minutes (50–200ms/frame). Budget it into the timeout, don't kill it midway. Slides with long entrance choreography scale linearly (frames = Σ per-slide animation windows × fps — measured: 5 × 15s ≈ 2250 frames, 6–8 min) — for quick passes drop `fps` to 24 in `script.json`, trim the choreography, or shoot in batches (`--ids 01,02`, then `--ids 03,04`; frame dirs are per-slide and accumulate).
+
+Limitation (which is what decides when to fall back to still): all page motion must be driven by CSS `@keyframes` / WAAPI. `<video>` elements, `setTimeout` choreography and rAF physics cannot be seeked — this layout spec already disables them, so you normally won't hit this.
+
+## Subtitle system (capture burns them in automatically, on by default)
+
+Subtitles are not written in the HTML and are not the author's concern: capture reads `clauses[]` from `timings.json` (single source of truth) and injects them before the page loads — one `.kit-sub` element per sentence, bottom-centered, pill background + text color taken from theme hooks (`--sub-bg` / `--sub-fg` / `--sub-ring`, defaulting to dark translucent + white text, legible on both light and dark themes); display window = this sentence's onset → next sentence's onset, built as a **full-duration percentage keyframe animation** that works naturally under frame-by-frame seeking, and hidden automatically after `finish()` in still mode (so Gate 4 static previews show no subtitles; subtitle acceptance happens at Gate 5 on the final video).
+
+**All geometry follows the canvas variables**, so neither landscape nor portrait needs code changes:
+- position `bottom = --stage-h × 0.0778` (= 84px at 1920×1080, clear of slide-num)
+- width `max-width = --stage-w × 0.729` (= 1400px at 1920, narrows automatically in portrait)
+- font size/padding `× --sub-scale` (capture computes it from canvas width, narrowing to a 0.75 floor in portrait to stay readable)
+
+On dark themes always check subtitle contrast: run `node scripts/check-theme.mjs <项目>`, which computes the contrast of "subtitle text vs the pill composited onto the background". Dark themes are advised to use `--sub-bg: rgba(0,0,0,.58)` plus `--sub-ring: 1px solid rgba(255,255,255,.16)` for visual separation.
+
+### How to verify subtitle readability (the trio, none optional)
+
+1. **Numeric gate** — run `node scripts/check-theme.mjs <项目>`: it computes "subtitle text vs the pill composited onto the background" contrast, and <4.5:1 is an immediate fail (this is the only gate that automatically stops "subtitles smeared into a dark background").
+2. **Frame gate** — subtitles only appear in the final video (a still preview finishes them away), so you have to pull motion frames to look:
+   ```bash
+   node scripts/capture.mjs <项目> --mode motion --ids 01   # 先出帧
+   # 抽某句开口之后的一帧(例: 第 2 句 3.4s → 取 3.6s)
+   ffmpeg -y -v error -i build/frames/01/f00108.png -vf "crop=iw:0.22*ih:0:0.70*ih" sub-check.png
+   ```
+   **Subtitle changes after the animation window (1.5.0)**: subtitles are full-duration percentage animations, while the frame sequence only covers the animation window — so the remainder used to be padded with the last frame and **every later subtitle change was frozen in the video** (and on the no-fx/static path, `finish()` parks the subtitle animation at `opacity: 0`, so nothing was burned at all). capture now emits one still per sentence (`build/substills/<id>/s<k>.png`) plus a manifest (`build/substills/<id>.json`, recording `framesCover` and each window), and build-video composes `[frame sequence] + [one still per uncovered window]` — the parts must sum to the slide duration, which the duration self-check enforces.
+
+Check three things: whether the subtitles appear, whether they are cropped, and whether the bottom overlaps the page number. If capture finds a slide without clauses it warns outright (it will no longer silently produce a subtitle-less video).
+3. **Mute gate (Gate 5)** — watch the whole video with the sound off: can you follow it on subtitles alone? This is the hearing-impaired viewpoint, and also the acceptance standard for platform auto-captions.
+
+Implementation details (so you know where to look when troubleshooting): the subtitle animation is explicitly excluded from the "animation window" computation (`.kit-sub` does not participate in animEnd), otherwise this full-duration animation would stretch frame-by-frame capture to the whole slide duration and waste several times the frames. When a clause carries `text2` it automatically becomes two lines (second line `.kit-sub-2`, smaller size and lower opacity), no configuration needed. `build-video` additionally emits `out/subs.srt`, from the same source and with the same windows as the burned-in subtitles (two lines when bilingual), for platform upload. A single sentence that would wrap past two pill lines — the pill fits ≈24 CJK chars per line at 1080 wide, ≈33 at 1920 (it scales with the canvas; one-line craft guideline ≈18 CJK / ≈42 English) — or text2 >60 characters, gets a warning from plan-timings.
+
+## The visual verification trio (mandatory after any HTML change)
+
+Don't fool yourself, all three steps are required:
+
+1. **Re-screenshot**: `node scripts/capture.mjs <项目> --mode still --ids <改动的张>`, look at `preview/<id>.png`. ⚠ still **invalidates that slide's frame directory** (to prevent stale frames from polluting it) — once the visuals are confirmed, if you want to rebuild the video you must re-run `--mode motion` for the same ids, otherwise build-video falls back to this slide's static image and the animation is silently lost (it does call out a ⚠, but by the time you see the warning you have already wasted one encode run)
+2. **Check subject placement**: is the image subject in the upper-center of the frame? Is it cropped in half? (image sourcing SOP in `image-sources.md`)
+3. **Pull actual frames from the final video**: `ffmpeg -ss <时刻> -i out/final.mp4 -frames:v 1 frame.png` — you **must check actual frames from final.mp4**, don't only look at the screenshot pipeline's output (the final video's fades and subtitle timing differ from the screenshots)
+
+On attributing stage timing: first get measured data with `check-timing.mjs`, then pull final-video frames to confirm; if "some elements enter at 0s while others follow the timing", that is tokens.css delays being overridden by the `.fx-*` shorthands (see authoring.md for the implementation principle), not estimation error.
+
+## Play page (play the HTML visuals through once, without waiting for encoding)
+
+`node scripts/preview-page.mjs <项目> [--open] [--no-script]` → `preview/play/index.html`, a single file, zero dependencies, double-click to view over file://.
+
+**Positioning: this page does exactly one thing — "look at the visuals"** — level-by-level entrance, slide navigation, motion on/off comparison, overview; the narration text is a bonus. It **deliberately has none of the player UI**: no timer (a ticking `X s / Y s`), no progress bar, no sentence-by-sentence follow-along highlight, no replay button — to see time/rhythm, just watch the final video; running a timer in the preview page only makes people stare at a stopwatch (measured: leave the tab open for a while and it reads `204.2s / 6.3s`, which is meaningless).
+
+| Action | Effect |
+|---|---|
+| `→` / space / swipe left on touch | **Motion on: reveal level by level first** — each press brings out the next entrance level (a second-level heading or a small chart animates in on the spot), and only once they are all out does it advance to the next slide; motion off: advance directly. Level-by-level implementation: the copy's head carries a pre-boot script that, given `?s=k`, sets passed levels to their final state, the current level entering at 0ms, and future levels staying hidden |
+| `←` / swipe right on touch | Motion on: step back level by level first (stepping back doesn't replay animation, it goes straight to final state), and at level 1 it goes back one slide (the previous slide entirely in final state); motion off: go back directly |
+| `X` (or the bottom bar's "motion on/motion off" button) | **Motion on / motion off comparison**: turning motion off switches to `preview/play/<name>.nofx.html` (root element carrying `no-fx`, the whole slide in final state). A frame going blank = some keyframe never lifted `opacity:0` back up; the two versions matching = the motion setup is clean |
+| `P` (or the bottom bar's "narration on/narration off" button) | Narration text panel on / off |
+| `O` / `F` | Overview (preview/*.png thumbnails, click to jump, landing on the full final state) / fullscreen |
+
+The toggles all live in the **bottom bar button strip** (shown on desktop too, not just touch): the button text states the current status directly ("motion on/motion off", "narration on/narration off", the off state lit in a warning color), while the top bar keeps only the topic and page number. Slide changes and level steps use a **double-buffered iframe**: the new frame only swaps in for display once it has finished loading in the hidden frame, so there is no white flash. UI text adapts between Chinese and English per `script.json`'s `lang` (English projects get a fully English button set, no more mixed Chinese).
+
+**The narration UI loads or not based on data** (not on the user remembering to turn it off):
+
+| Case | Page behavior |
+|---|---|
+| Has clauses + has `timings.json` | Lists this slide's narration text on the right (a bottom drawer on narrow screens) |
+| Has clauses, not timing-synced yet | Lists the text, with the title marked "(not timing-synced)" — so you can look at the HTML first even before the narration exists |
+| No clauses, or `--no-script` | No panel and no narration button at all, the visuals fill the full width |
+
+**Layout adapts to the window**: top/bottom bars can wrap (overlong topic names ellipsis-truncated), on narrow windows and phones the narration panel collapses into a bottom drawer and is **collapsed by default** (visuals first), the bottom button strip automatically fills the whole row, touch lets you swipe left/right to change slides, and the overview grid picks its column count from the width. It uses `100dvh` rather than `100vh` so the address bar doesn't cut off a slice on phones.
+
+**Why generate a copy instead of opening `slides/*.html` directly**: animation delays (`--t1/--t2/--t3`) and canvas size are injected by the render pipeline from `timings.json` (capture goes through `addInitScript`), and the HTML and tokens.css only hold placeholder defaults. Double-click the original file and `--t2` is 800ms instead of the measured 4.9s — every entrance animation crowds into the first two seconds and the rhythm you see is completely different from the final video (measured: on the same slide at t=3.0s, level two is still opacity 0 in the copy while it is already visible in the original file). The copy writes these values into `<html style>` (equivalent to pipeline injection, highest precedence) and adds `<base href="../../slides/">` so tokens.css and `../assets/*` resolve as usual.
+
+**When there is no `timings.json` yet** (narration not done): the copy spaces the stages actually used in the HTML at equal intervals (0.3 / 1.3 / 2.3s), and a yellow bar at the top of the page honestly notes "not final-video timing". At least you can see the entrance order clearly — with placeholder values all animations crowd into 2 seconds, which is the truly unreadable version.
+
+The copy is a **snapshot**: after editing `slides/` you must re-run `preview-page.mjs`, otherwise the play page still shows the old visuals (the copy's head comment states the real file path). The play page burns no subtitles and carries no sound — subtitle acceptance is still based on the final video.
+
+When the project's tokens.css lags behind the skill's current version (the `tokens` managed region is stale, missing, duplicated, has a broken delimiter, or a leftover copy sits outside it), the script injects the skill's **current generated body** into the snapshot copies as a fallback: the no-fx rules go only into the motion-off copy, the rest into both (otherwise the old rules would render the new drawing approach broken), and the page says so honestly — including that **the rendered video still uses the project's old CSS**. For a legacy "bare text" file that is already byte-identical to the current version nothing is injected (the rendering is already right). To fix it at the root run `init-project --upgrade-css` **before** capture / build-video, which now refuse to run while the region is stale.
+
+## Cover and transitions (1.6.0)
+
+**Cover (thumbnail)**: `capture.mjs` emits an extra `preview/cover.png` for slide 1 (all base animations at final state, subtitles hidden = the fully revealed title design), and `build-video` does three things:
+
+1. **First-segment cover dissolve** — when encoding segment 1, overlay the cover and fade it out over 0.25s → **frame 0 is the complete cover**, no longer a black frame (previously every segment carried `fade=t=in` and the first frame was fully black; users' measured feedback: "the thumbnail people receive is pure black"). Adds no duration.
+2. **Embedded attached_pic** — at mux time add a track via `-disposition:v:1 attached_pic` (the main video stays `-c:v copy`, no re-encode): file managers / most players / some IMs read it directly as the thumbnail. Note that with an embedded cover you **must not pass `-shortest`** (it would cut the output to that frame's length; the audio track itself is already aligned to the total duration).
+3. **Export `out/cover.png`** (scaled to canvas size) for manual upload to Bilibili/YouTube and similar platforms.
+
+**Transitions (slide changes)**:
+
+| Mode | Effect | Cost |
+|---|---|---|
+| `cut` (default) | **Hard cut** between segments, no black pass; the first segment keeps its cover dissolve and the last segment keeps its fade-out ending | Zero (still `-c copy` concatenation) |
+| `xfade` | 0.4s cross-dissolve between adjacent segments (both frames half-transparent at once, not a black-pass transition) | One full video re-encode |
+
+Usage: `"transition": {"type":"xfade","duration":0.4}` in `script.json`, or CLI `--transition xfade` to override temporarily. **Total duration is unchanged**: under xfade each segment first `tpad`s an extra `duration` seconds of tail frames, the dissolve eats `(n-1)×duration`, and after concatenation it still equals `timings.total` (the self-check verifies this). On the play page, `T` lets you **compare the two slide transitions live** (the dissolve is also a 0.4s overlay in the play page), so the user can decide by feel.
+
+> Previously it was "fade to black 0.3s at the end of each segment + fade in from black 0.25s on the next" ≈ 0.55s of pure black — users' measured feedback: "every big slide change goes through a black screen". Neither hard cut nor dissolve passes through black.
+
+**Three visual gates** (run automatically when build-video finishes; any failure exits non-zero): cover is embedded / first-frame brightness above black level (not a black frame) / no black frames around each slide-change point.
+
+## BGM mixing (optional)
+
+When build-video detects a `bgm` config in script.json, it inserts a mixing step after audio alignment and before mux:
+
+```
+ffmpeg -y -i build/audio-timeline.wav -stream_loop -1 -i assets/bgm.mp3 \
+  -filter_complex "[1:a]aresample=44100,aformat=channel_layouts=mono,volume=0.12,\
+afade=t=in:st=0:d=1.5,afade=t=out:st=<total-2.5>:d=2.5[bg];\
+[0:a][bg]amix=inputs=2:duration=first:normalize=0[out]" \
+  -map "[out]" -c:a pcm_s16le build/audio-mix.wav
+```
+
+- `-stream_loop -1` loops a short BGM to fill the whole video; `duration=first` converges to the length of the voice track.
+- `normalize=0` is mandatory: amix by default splits the volume by input count, which **halves the voice**; normalize=0 keeps the voice at its original level and BGM is controlled only by `volume`.
+- The `normalize` option requires ffmpeg ≥ 4.4; on failure the script warns and falls back to voice only (it does not abort the render).
+- How to verify: `ffmpeg -i out/final.mp4 -af volumedetect -f null -` to see whether mean/max are raised versus the voice-only track (with BGM they should be); `silencedetect` should no longer report silence in the gaps between sentences (the voice gaps are filled by BGM).
+
+## Encoding and concatenation parameters (for manual troubleshooting)
+
+Single slide (frame sequence):
+
+```bash
+ffmpeg -y -framerate 30 -i build/frames/01/f%05d.png \
+  -vf "tpad=stop_mode=clone:stop_duration=99,fade=t=in:st=0:d=0.25,fade=t=out:st=<D-0.35>:d=0.3" \
+  -t <D> -c:v libx264 -pix_fmt yuv420p -preset medium -crf 20 -movflags +faststart out/slide-01.mp4
+```
+
+- `-framerate` must come before `-i` (an image sequence has no timestamps of its own; this defines them).
+- `yuv420p` is mandatory, otherwise some players show green/black.
+- `+faststart` moves moov to the head of the file, so streaming plays on click.
+
+Concatenation: all segments are encoded uniformly by build-video (same encoder, same parameters) and stitched losslessly and instantly with the concat demuxer `-c copy`; after stitching ffprobe verifies, and a total-duration deviation >0.25s automatically falls back to concat filter re-encoding (mixing in foreign segments gives inconsistent timebases that turn 8s into 35s, and copy can't rescue it).
+
+Audio alignment: each segment gets `aresample=44100,aformat=channel_layouts=mono,apad=whole_dur=<该张实测时长>` to pad silence before concat. **Don't use the adelay+concat approach** (the concat filter ignores adelay's offset and all speech piles up at the start).
+
+**Look at the commands before running**: `node scripts/build-video.mjs <项目> --dry-run` prints every ffmpeg command it would execute without actually executing it — when you want to manually verify a slide's parameters, or suspect an encoding-parameter problem rather than an asset problem, run it first and take the commands as-is to the command line to experiment.
+
+## Self-check standards
+
+Before build-video ends it enforces: the final video's ffprobe duration differs from the timings total by ≤0.25s; `ffmpeg -v error -i final.mp4 -f null -` decodes the whole file with zero errors. If either fails, the exit code is non-zero — when you see non-zero, don't deliver.
+
+## Troubleshooting table
+
+| Symptom | Cause | Action |
+|---|---|---|
+| Final video has no sound | Audio wav not generated / mux failed | `ffprobe build/audio-timeline.wav` to check duration; re-run build-video |
+| All speech piled at the start | Used an adelay approach | Use this script's apad approach |
+| Duration explodes after concatenation | Foreign-encoded segments mixed in | Re-encode all segments uniformly with build-video; it already falls back automatically |
+| Chinese renders as boxes | No CJK font on the system | Linux: `apt install fonts-noto-cjk`; or switch the font stack to an installed font |
+| Font ugly first, then correct (FOUT) | External font + font waiting not going through capture | Localize assets/fonts; make sure you use capture.mjs rather than manual screenshots |
+| Animation not in the video | Used still mode | Re-run that slide with `--mode motion` |
+| Animation layer timing wrong | Animation delays hand-written in the HTML | Delete them, rely on --t1/--t2/--t3 injection |
+| chromium fails to launch | Browser not installed | `npx playwright install chromium` (on Linux CI add `--with-deps`) |
+| A slide's segment length is wrong | Frame directory is from old timings | Delete `build/frames/<id>/` and re-run capture; after changing TTS you must re-run plan-timings |
+| Backslash paths break concat on Windows | — | build-video already converts to forward slashes; be careful when hand-assembling list.txt |
+| Some elements enter at 0s, others on time | Old tokens.css: delays overridden by the `.fx-*` shorthands | Switch to the new tokens.css (`--fx-delay`); see authoring.md for the implementation principle |
+| Image subject cropped out of frame | Bare `<img>` or cover with the wrong ratio | Wrap in `.img-frame` + `--img-pos`; for screenshot-type images switch to `.contain`; if the subject hugs the edge, re-search per the SOP |
+| Subtitles smear into the background on dark themes | Theme doesn't override the subtitle hooks | Add `--sub-bg` (darker) + `--sub-ring`; `check-theme.mjs` computes this contrast |
+| spawnSync returns exit -5 or hangs under PowerShell | Chinese paths + complex argument combinations easily trigger this in PowerShell | Run scripts via Git Bash / WSL instead, or inject the binary directory into `$env:PATH` first (this suite's scripts always work fine under Bash) |
+| Numbers/text invisible | Undefined CSS variable voids the whole background + transparent text | `check-slides.mjs` (static) catches undefined variables; capture names the images that failed to load |
+| Broken image icon | Missing file or invalid SVG | Same as above; inlining SVGs into the HTML is always the most reliable |
+
+## Environment differences notes
+
+- **ffmpeg/ffprobe detection**: every script uniformly goes PATH → project/repo node_modules (ffmpeg-static/ffprobe-static) → common install locations (winget Links / scoop / C:\ffmpeg\bin); when not found they list the checks one by one and exit with code 2 (distinct from 1 for business errors).
+- **Windows**: `python`/`node` naming, path separators — this whole suite is implemented in Node, no bash arrays, no octal traps, it just runs. ffmpeg: `winget install Gyan.FFmpeg` or `npm i ffmpeg-static ffprobe-static`.
+- **Linux sandbox (mcode)**: chromium is installed by `npx playwright install chromium`; a Chinese font package is needed; `--with-deps` supplies shared libraries.
+- Supersampling: capture `--dsf 2` emits 3840×2160 frames, and build-video detects the size mismatch and auto-downsamples with lanczos, giving sharper text edges; render time is about ×4, so use it only for the final master. `--dsf` accepts 1–4 and **defaults to 1**; anything else (0, 2.5, `abc`) is rejected at the entry point with a clear message.
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/research.md b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/research.md
new file mode 100644
index 00000000..9abb3cda
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/research.md
@@ -0,0 +1,118 @@
+# Research and fact collection (Phase 0 reference)
+
+Factual content goes wrong in a very concentrated set of ways: **mixed definitions/metric bases, distortion through second-hand retelling, quotes attributed to the wrong person or occasion, stale information**. This section turns "looking something up" from a single search into a reviewable process.
+
+## When research is mandatory
+
+**The test is simple**: if the finished video contains specific **numbers, dates, names, quotes or attribution relationships** → research is mandatory.
+
+- Companies / products / funding / data / history / people / policy: **mandatory**
+- Generic, lyrical, creative or purely opinion pieces: can be skipped, but any sentence that *looks like a fact* still has to be verified
+- The user has already supplied a final script: still review it and mark the factual points that "could be misread or challenged"
+
+## Source tiering (decides whether a source can enter the script)
+
+| Tier | Examples | Use |
+|---|---|---|
+| **Primary** | Official announcements, financial reports, prospectuses, official blogs and technical reports, government / international-organisation statistics, original papers | Use directly; definitions and numbers defer to it |
+| **Authoritative secondary** | Mainstream financial media, industry research reports, the **original sources cited by** Wikipedia | Usable, but key numbers should be traced back to primary wherever possible |
+| **Weak sources** | Self-media, aggregator sites, AI summary pages, unsigned "compiled materials" | **Leads only**, must be traced back to primary; discard if it cannot be |
+| **Unusable** | Forum posts, unsourced screenshots, images/tables of unknown origin | Do not enter the script |
+
+## Hard rules (miss one and it goes back)
+
+1. **Every key number needs at least 2 independent sources.** With only one source, use hedged wording in the script ("per the official figure…") or downgrade it to an approximation.
+2. **Primary first.** Numbers taken from secondary retellings must be traced back to the original text — the longer the retelling chain, the more the number gets processed.
+3. **Label the definition and the date.** Is "revenue" annualised or quarterly? In what currency? Is "users" weekly actives or monthly actives? As of when? Write it clearly in the notes, and carry it into the script where it belongs ("in the first half of this year").
+4. **Recency**: prefer sources from the last 12 months; assertions like "fastest / first / largest / the first ever" especially need a recent source.
+5. **Separate fact from narrative**: "900M weekly actives" (checkable) vs "most popular" (an opinion). Opinions need a subject ("the company itself says…") or must be a quote.
+6. **Quotes verbatim + name and occasion stated**; translated quotes marked "(translated)".
+7. **Anything uncertain stays out of the script**: if the source cannot be found, or sources conflict irreconcilably → write it into the notes as an "uncertain item", keep it out of the narration.
+8. **Sources must be re-checkable**: every item carries a URL or document name + date, so Gate 0 and later review can trace it.
+
+## Query design
+
+- **Entity + metric + time**: `MiniMax 2026 H1 revenue prospectus`; for a Chinese-market entity, the same query in Chinese — `MiniMax 2026 上半年 营收 招股书`
+- **Query in both Chinese and English**: for Chinese-language topics look for the official Chinese announcement; for international topics go to the English original (second-hand translation distorts)
+- **Go straight for the carrier word**: to find numbers, search "financial report / prospectus / annual report / technical report / statistical bulletin" — that hits primary sources faster than searching the entity name
+- **Prefer official domains**: guess the official document path or the official news page first, get the primary source and only then look for secondary interpretation
+
+## What to do when sources conflict
+
+1. Compare **definition and date** first — most "conflicts" are actually different definitions (weekly vs monthly actives, annualised vs single quarter)
+2. Then compare **source tier** (primary > authoritative secondary > weak)
+3. Still conflicting: take the primary / more recent one, and record the disagreement in the notes
+4. Cannot decide on the spot: **do not use that number in the script**, or turn it into a qualitative statement
+
+## Search tools by environment
+
+| Environment | Search | Notes |
+|---|---|---|
+| mcode | `web_search` / `web_fetch` (built in) | `web_fetch` pulls the body text, good for tracing primary pages |
+| Other hosts | `mmx search "keyword"` / `mmx text chat` | **~10 results max, no pagination** — good for "finding a primary entry point"; deeper verification still needs fetching the page or opening the official page |
+| mcode general fallback | built-in browser, open the official page | The same technique as image path B; the official page is always the most reliable |
+| Any environment | **other search skills/plugins installed on this machine** | Use whichever one can actually pull the body text (the tools are open, the discipline is not); try it once to confirm result quality and count before depending on it; source tiering, two-source cross-checking and definition labelling all still apply |
+
+## Fetching: how to actually get official sites / press releases / prospectuses
+
+Search tools only "find the entry point" — **you fetch the body text yourself**. Choose the method by how the page renders:
+
+| Page type | Method | Notes |
+|---|---|---|
+| Static pages / documents / announcement pages | `web_fetch <URL>` | Gets the body text directly, fastest |
+| **SPA / JS-rendered pages** (many corporate sites, investor relations pages) | **mcode built-in browser** (mcode environment) or a **headless Playwright browser** (any environment), render first then take the body text | `web_fetch` often returns nothing but an empty shell HTML (a shell with no text). This skill already depends on Playwright, so reuse it (headless render, nobody has to watch it); any other headless browser automation skill installed on the machine also works, as long as it gets the body text. When using the **mcode built-in browser**: do not use `navigate(replaceCurrentTab)` as a refresh (it opens a new tab), throttle when clicking 15+ times in a row (it triggers human verification), and when you have image URLs, write them to disk via `fetch-official-images.mjs --url`. Command: |
+
+```bash
+# render, then take the body text (works for SPA / lazy-loaded official sites)
+node -e "const {chromium}=require('playwright');(async()=>{
+  const b=await chromium.launch(); const p=await b.newPage();
+  await p.goto(process.argv[1],{waitUntil:'domcontentloaded'});
+  await p.waitForTimeout(2500);                       // wait for async content
+  console.log(await p.evaluate(()=>document.body.innerText));
+  await b.close();})()" "https://company.example/newsroom/some-post"
+```
+
+- For **images/charts** on the same page: `node <skill>/scripts/fetch-official-images.mjs <URL>` (it already opens and renders the page and lists candidate images)
+- **PDF** (prospectuses, financial reports, annual reports, white papers): download directly, then extract the text from the PDF; **numbers defer to the PDF original text**, never to a media retelling
+- Pages that need login / captcha: do not force your way through automation — switch to a public official page or ask the user for the file
+
+## Discipline for using press releases / announcements
+
+**Where to look**: `/news`, `/newsroom`, `/press`, `/investor`, `/ir`, `/announcements`; Chinese companies usually keep columns named "news centre / investor relations / announcements". Exchange announcements — SEC EDGAR for US filings, HKEXnews for Hong Kong, SSE / SZSE (CNINFO) for mainland China, plus regulator decisions such as FDA / FTC — are more authoritative than the company's own site.
+
+**Four rules of discipline**
+
+1. **Identify the publishing entity**: the official original text (company / exchange / institution) ≠ a media retelling. On conflict, **the official original text wins**.
+2. **Identify the publication date and the definition**: numbers in a press release almost always carry a definition ("as of June 2026", "annualised", "unaudited") → **copy the definition verbatim**, never shorten it to a bare number.
+3. **Beware headline exaggeration**: "first / largest / disruptive" mostly appears in media headlines rather than the original text; before writing it into the narration, go back to the original and check whether it really says that.
+4. **Trace syndicated copies**: when the same wire story is republished by several outlets, find the original publisher; quote the original text, not the republished page.
+
+```markdown
+## Core facts
+| Fact | Value/wording | Source (URL or document name) | Definition date | Tier | Second source |
+|---|---|---|---|---|---|
+| Weekly active users | 900M | openai.com/… (official announcement) | 2026-02 | Primary | Reuters 2026-03 |
+
+## Uncertain items (not entering the script)
+- Valuation definitions disagree across reports (800B / 852B), and there is no primary source → do not use
+
+## Content that must not enter the script
+- Unconfirmed rumours, controversial evaluations, unverifiable personal relationships
+```
+
+When presenting at Gate 0, **focus the user on confirming the numbers, names and definitions**, rather than reading through the whole table.
+
+## Confirm the domain first (extra step for regulated topics)
+
+At kickoff alignment, if the topic falls under **finance/investment research, healthcare, legal, government policy, or marketing performance claims**, then in addition to the definition discipline in this section you must also follow `compliance.md`: confirm **whether a disclaimer and data attribution are required** (by default they are), complete every number with the three requisites "definition + currency + point in time", and settle up/down colours by the audience's market (red-up/green-down vs green-up/red-down). This step must reach a conclusion at Gate 0 — do not leave it to the finished video.
+
+## Pitfalls by topic
+
+| Topic | Pitfall | What to do |
+|---|---|---|
+| Company / funding / IPO | Media mix up "valuation / revenue / funding amount"; currency and annualisation basis unclear | Defer to the prospectus or official announcement; state the currency and the period |
+| Finance / investment research (stocks, funds, macro data) | Opinions written as facts, stale points in time presented as "currently", mixed definitions (YoY ≠ QoQ, GAAP ≠ non-GAAP); wrong up/down colours for the audience | Every number carries **definition + currency + point in time**; price figures say "as of the YYYY-MM-DD close"; no trading advice and no price predictions; disclaimer and up/down colours per `compliance.md` |
+| Models / technical benchmarks | Leaderboard versions differ, scores are not comparable | Defer to the official technical report + state the leaderboard name and version |
+| History / people | Quote sources unclear, second-hand retellings distort | Defer to authoritative biographies/archives; every quote must have an original source |
+| Policy / statistics | Using old-version data, ignoring revisions | Defer to the original government or international-organisation release, watch for revision notes |
+| Product features | Calling a "beta / gated rollout" a "live feature" | Defer to the current state in the official documentation, and date it |
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/symptoms.md b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/symptoms.md
new file mode 100644
index 00000000..6bc85b1b
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/symptoms.md
@@ -0,0 +1,33 @@
+# Symptom index (when to reach for this skill, and where to look)
+
+This file holds the long symptom list that used to live in the SKILL.md `description` (the host caps that field at 1024 characters). The description now says "also trigger on symptoms"; **this is the list**.
+
+Use it two ways: as trigger examples (the user describes one of these and you reach for this skill) and as a first routing hint (the right column is where the fix usually lives).
+
+| Symptom as the user describes it | Where it usually comes from |
+|---|---|
+| TTS leaves too long a silence after it finishes speaking | `tail` too large or the pause is in the audio itself → re-cut that TTS segment (`references/tts-and-timing.md`) |
+| the frame has only a title with no detail expansion | the slide violates the "≥2 information layers" rule → add layers, then re-run Gate 1 (`references/authoring.md`) |
+| some elements enter at 0 seconds | the fx delay was overridden — check `--fx-delay` / `.fx-stagger` (`references/authoring.md`, and `check-slides.mjs` item 5d) |
+| entrance animations aren't rendered into the video | motion frames missing — `capture --mode motion` for those ids, then rebuild (still capture invalidates the frame dir) |
+| audio and visuals are out of sync | re-run `plan-timings.mjs`, then `check-timing.mjs` / `asr.mjs --verify-timing` (`references/tts-and-timing.md`) |
+| an image's subject is cropped out of frame | re-pick or re-crop the asset — `prep-image.mjs --check`, and see `references/image-sources.md` §2-4 |
+| subtitles are muddy on a dark background | contrast below 4.5:1 — `check-theme.mjs`, then theme hooks `--sub-bg` / `--sub-fg` / `--sub-ring` (`references/render.md`) |
+| the total duration after concatenation is wrong | `build-video.mjs` duration self-check — usually stale frames or a wrong `timings.json` |
+| fonts flicker or show as boxes | missing glyphs / web fonts — use system CJK fonts and inline or local assets (`references/render.md`) |
+| a chart or some element never appears after entering (the keyframes don't bring opacity back) | the keyframes never lift `opacity` — `check-slides.mjs` item 5b names it; fix the keyframes (`references/authoring.md`) |
+| switching to no-fx makes the frame emptier instead | old tokens.css: `init-project.mjs --upgrade-css` (the no-fx reset is a managed block) |
+| double-clicking slides/*.html to watch the animation shows everything crammed at the start | that is the placeholder delay — open the play page `preview/play/index.html` instead (`references/render.md`) |
+| or a finance video forgot the disclaimer or has the up/down colors inverted | `references/compliance.md`: disclaimer line + the audience market's up/down convention |
+
+## More symptoms worth triggering on
+
+- A slide that reads fine as a still but falls apart once it moves (elements overlapping, things never appearing) → the play page is the 30-second check: `preview-page.mjs`.
+- The thumbnail of the shared file is black, or every page change flashes black → `references/render.md` "Cover and transitions".
+- The voice is too fast/slow, or a Chinese voice came out Cantonese → the audition step plus ASR verification (`references/tts-and-timing.md`).
+- Stale styling after the skill updated (`new classes have no effect`) → `init-project.mjs --check-css` / `--upgrade-css`; see `references/render.md`.
+
+> When a symptom points at a script, run it with `--quiet`/`--ids` to narrow first; every diagnostic script prints the exact next command on failure.
+| a script exits 0 with no output at all (printed nothing, did nothing) | you **imported** the CLI instead of running it — `node -e "import(…)"` never runs `main` by design; run `node <script> …` instead. (Before 1.9.9 an entry reached through a symlinked path could also silently no-op — the entry guard compared raw paths) |
+| Playwright in a cloud sandbox cannot find its browser (`chrome-headless-shell` path errors) | partial browser install: `npx playwright install chromium` lays down `chromium-*/chrome-linux/chrome` while some runtimes look for `chromium_headless_shell-*/…/chrome-headless-shell` — copy+rename the binary as a stopgap (seen on mavis-like sandboxes) |
+| `git clone` from GitHub fails with a certificate/SSL error in a locked-down sandbox | fetch the archive instead: `curl -o repo.zip https://codeload.github.com/<owner>/<repo>/zip/refs/heads/main` (seen on mavis-like sandboxes) |
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/tts-and-timing.md b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/tts-and-timing.md
new file mode 100644
index 00000000..6a52f0d5
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/references/tts-and-timing.md
@@ -0,0 +1,170 @@
+# TTS Integration and the Timing Sync Algorithm
+
+## Why TTS comes before the HTML (the biggest difference between this pipeline and common practice)
+
+The common order is "write the HTML first, then dub it", which forces you to **estimate** each slide's duration and **guess** the animation delays; the typical symptoms it produces are: TTS finishes speaking and the visuals sit there dumbly for several seconds (over-estimated), or an unfolding layer appears only when the narration has already moved elsewhere (guessed wrong).
+
+This pipeline goes the other way: script finalized → run TTS immediately → ffprobe measurement → compute every stage's entrance moment → only then touch the HTML. There are no hard-coded seconds anywhere in the HTML; all delays are injected by capture through `--t1/--t2/--t3`. Change the voice, change the speech rate, change one word, and you only need to re-run TTS and plan-timings — the HTML usually needs not a single character changed.
+
+## Timing sync algorithm (plan-timings.mjs internal logic)
+
+For each slide:
+
+1. `D` = the segment's TTS duration measured by ffprobe.
+2. The k-th narration sentence's onset ≈ `D × (character count of the first k−1 sentences ÷ total character count)` (Chinese speech rate is distributed evenly enough across sentences; the error is usually <0.3s).
+3. Stage s's entrance moment = the onset of the earliest sentence mapped to that level − 0.2s (the visuals lead slightly, which reads as synced; broadcast convention).
+4. Final video duration = `ceil((D + tail) × fps) / fps`, aligned to the frame grid; `tail` defaults to 0.8s.
+
+Character counting ignores whitespace and includes punctuation (TTS pauses naturally at punctuation, so including it is actually more accurate).
+
+Timing data has three layers, each one measurable and traceable; when you suspect audio-visual desync, **get the data first, then act**:
+
+1. **Estimate** (plan-timings): the character-share estimate above. For normal speech-rate samples the error is usually <0.3s.
+2. **Silence measurement** (`node scripts/check-timing.mjs <项目目录>`): TTS has 0.2s+ pauses between sentences, and the script uses ffmpeg silencedetect to measure each sentence's real onset directly, outputting an "estimate vs measured" comparison table; `--calibrate` writes the measured boundaries back into timings.json (it only calibrates slides where the silence-segment count and the sentence count match exactly), after which delete `build/frames/` and re-run capture and build-video.
+3. **ASR review** (build-video `--asr`): cut the audio by sentence and transcribe each part, verifying both content (numbers/proper nouns) and boundaries — if a part's transcript bleeds in the start of the previous sentence, that sentence's real onset is later than the estimate, so go back to layer 2 and calibrate.
+
+Also note that "it feels out of sync" is often a **design-level misplacement** rather than algorithmic error: a big number / hero image must be attached to the stage of **the sentence that mentions it** (see the strong-sync principle in authoring.md); draw the first sentence's highlight inside stage 2 and no matter how accurate the algorithm is, the audience feels the disconnect.
+
+## Whitespace tuning
+
+Whitespace = each slide's `tail` (narration end to slide change). Default **0.8s**: below 0.5s the audience feels chased, above 1.5s it is "finished speaking, staring blankly".
+
+- Tight overall pacing (fast-consumption style): add `"tail": 0.4` to each slide in script.json.
+- Unhurried delivery (brand film): `"tail": 1.2`, with 1.5 for the first and last slides.
+- Per-slide tweak: change only that slide's `tail` and re-run plan-timings; no need to redo TTS.
+
+Speech rate: **1.1** overall (≈5.3 Chinese chars/sec; ≈4.8 at 1.0); the first and last slides at 1.05 are slightly slower for a sense of ceremony. **This must be asked of the user together with the voice during the audition step** (see below); don't just decide it yourself and move on. If plan-timings warns that the speech rate deviates from 3–6.5 chars/s, first check whether speed is set wrong or the character count exceeds the limit — don't rush to re-estimate with --pacing.
+
+⚠ **Measured gotcha: `speed<1` (e.g. 0.95) on short sentences (≤10 chars) may increase rather than decrease the duration** — the TTS engine appears to pad silence on short sentences, making the duration longer instead (found in end-to-end measurement on 2026-09-17). So: if you want 0.95 for the first/last slides, don't make that slide's narration too short; or just keep short sentences at 1.0.
+
+⚠ **Precondition for stage timing**: entrance delays rely on the `--fx-delay` mechanism (see authoring.md for the implementation principle). If the page uses an old tokens.css (one that writes delays as a separate `animation-delay: var(--tN)`), the delay is overridden by the `.fx-*` shorthands, showing up as "some elements enter at 0s" — this affects the feel more than estimation error (usually <0.3s) does, so rule it out first when troubleshooting audio-visual desync.
+
+## TTS commands in the mcode environment
+
+Single synthesis / batch synthesis (≤10 per batch):
+
+```bash
+mcode-tools connector call connector__matrix__batch_text_to_audio --args '{"requests": [
+  {"text": "第一句口播。", "voice_id": "Chinese (Mandarin)_Gentleman", "speed": 1.1, "output_file": "01.mp3"},
+  {"text": "……", "voice_id": "…", "speed": 1.05, "output_file": "08.mp3"}
+]}'
+```
+
+Take `success_items[].node_id` from the result and fetch each download link to disk under `audio/`:
+
+```bash
+mcode-tools get_asset_url <node_id>   # 得到 URL 后 curl 下载为 audio/<id>.mp3
+```
+
+Look up voice candidates with `connector__matrix__get_voice_list`; when aligning at kickoff, first ask the user about the **voice direction (gender + timbre: warm male / crisp female / neutral, magnetic, lively)** — a rough direction is enough, the audition fixes the rest.
+
+**⚠ One audition, two questions (1.6.0): pitch AND pace.** The audition must ask the user *both* which voice to use *and* whether the pace is right — the pace is the complaint users actually report ("the default is a bit slow"), and it cannot be judged from written settings. So besides the three candidate voices, synthesize **the same probe line in the chosen voice at three speeds (1.05 / 1.15 / 1.25)** and hand over all of them with the list. After the user answers, write the decision into `script.json` (`speed.default` = the chosen one, first/last slides 1.05) and restate "voice = X / speed = Y" before moving on. `plan-timings` then checks the measured rate against that value and warns if the two disagree.
+
+**⚠ Voice names are not trustworthy, the audition must verify the voice with ASR (2026-09 measured lesson)**: platform labels get mismatched — `Chinese (Mandarin)_News_Anchor` is labeled a Mandarin news female voice but actually generates **Cantonese** (ASR transcripts show traditional characters such as "來自/從"). Picking a voice by name alone = picking blind. The correct flow:
+
+1. Take 3 candidate voices, each synthesizing one fixed probe line as a single TTS call (containing numbers and proper nouns, e.g. "三家科技公司,去年营收三点五亿美元。") — generate them **individually**, to avoid batch pairing mismatches (the three speed variants are also single calls);
+2. Transcribe and verify each part (on mcode use `listen_audio`; in other environments `node scripts/asr.mjs --file probe.mp3 --language zh`), checking: the transcript is **Simplified** Mandarin (traditional characters appearing = Cantonese/another language, discard), numbers read correctly, no obvious swallowed characters;
+3. Annotate the "measured language" and the speed of each sample into the audition list, hand the audio to the user, and ask the two questions together: **"① which voice? ② which pace — 1.05 / 1.15 / 1.25 (or none of them, give me a direction)"** (this step is in Phase 1, don't skip it).
+
+**Trick to force Mandarin**: pass the `language: zh` header in the call (the script's `--language zh`) and the model decodes as Mandarin; if traditional characters still appear in the result (e.g. 來/從/聲/畫), you can basically conclude that this voice actually outputs Cantonese. To confirm Cantonese in the opposite direction, use `--language yue`.
+
+Measured usable Mandarin voices (2026-09-17, mcode matrix):
+
+| voice_id | label name | measured |
+|---|---|---|
+| `Chinese (Mandarin)_Gentle_Senior` | 温柔学姐 (gentle senior) | ✅ Mandarin, crisp |
+| `Chinese (Mandarin)_Sweet_Lady` | 甜美女声 (sweet female voice) | ✅ Mandarin, friendly (**the primary choice in the 2026-09-17 measurement**) |
+| `Chinese (Mandarin)_Crisp_Girl` | 清脆少女 (crisp young girl) | ✅ Mandarin, clear |
+| `Chinese (Mandarin)_Mature_Woman` | 傲娇御姐 (tsundere mature sister) | ✅ Mandarin, storytelling feel |
+| `Chinese (Mandarin)_Wise_Women` | 阅历姐姐 (worldly-wise sister) | ✅ Mandarin |
+| `Chinese (Mandarin)_Male_Announcer` | 播报男声 (announcer male voice) | ✅ Mandarin (usable) |
+
+| voice_id | label name | measured |
+|---|---|---|
+| `Chinese (Mandarin)_News_Anchor` | 新闻女声 (news female voice) | ❌ actually outputs **Cantonese**, do not use |
+
+(The voice library drifts, so this table is only a starting point; every new project still goes through the three-step ASR voice verification above.)
+
+## Language and voice (must ask at kickoff alignment)
+
+**The narration language is fixed at kickoff alignment**, written into `script.json`'s `lang` (default `zh`). It drives four things:
+
+1. **What language the narration script** is written in (`clauses[].text`); when bilingual, `text2` holds the **other language** (main line = narration language).
+2. **The voice must match the language** — Chinese uses `Chinese (Mandarin)_*`, English uses `English_*`, Cantonese uses the corresponding Cantonese voices. A mismatch shows up as "accented / mispronounced / language drift", and must be verified with ASR once written.
+3. **Character-count and speech-rate baselines**: `plan-timings` switches automatically by `lang` — Chinese 5.3 chars/s @speed 1.1 (4.8 at 1.0; typically 3–6.5), English ~14 chars/s (typically 9–18, about 150 words/min); the single-line subtitle craft guideline is 18 chars for Chinese / 42 for English, while the pill itself fits ≈24 CJK chars per line at 1080 width and ≈33 at 1920 (it scales with the canvas) — plan-timings warns at three or more pill lines.
+4. **The ASR language header**: `asr.mjs` reads `lang` and automatically passes `language: zh|en|yue`. Chinese projects check "whether traditional characters show up" (= Cantonese), English projects check "whether Chinese characters are mixed in".
+
+**Display form vs spoken form (`clauses[].say`)**: `text` is the on-screen/subtitle form and always keeps the natural written form — Arabic numerals and original orthography on screen (`2026年`, `82.3%`, `SWE2.3`), never spelled-out digits (`八十二点三` on screen is wrong; that belongs in `say`). When the broadcast reading differs from the written form, **feed TTS the optional `clauses[].say`** and keep `text` untouched: years digit-by-digit (`say: "二零二六年…"`; TTS reads `2026年` as "两千零二十六年" — plan-timings warns on this), percentages with units (`82.3%` → `say: "…百分之八十二点三…"`), benchmark names by habit (`SWE2.3` → `say: "SWE二点三"`). **The TTS input and the ASR checklist expectation are always `say ?? text`**; the burned-in subtitles and `out/subs.srt` always come from `text`. `say` applies to the main line only (the `text2` second line is display-only).
+
+**Changing language takes only three steps**: change `script.json`'s `lang` → swap in a matching `voice_id` and redo TTS → if bilingual, add/replace `text2` for each sentence. Neither the layout nor the script logic needs to change.
+
+English voice examples (both mcode and mmx-cli track platform updates, so measured results are authoritative):
+
+```bash
+# mcode: 列全部音色后挑 English_* 前缀的
+mcode-tools connector call connector__matrix__get_voice_list --args '{}'
+# mmx-cli: 同样先列再选
+mmx speech voices
+# mmx-cli 示例(官方文档中出现过的英语男声)
+mmx speech synthesize --text "Three companies, three point five billion dollars." --voice English_magnetic_voiced_man --speed 1.1 --out audio/probe.mp3
+```
+
+Voice verification must also carry the language header: `--language en` (English) / `--language zh` (Mandarin) / `--language yue` (Cantonese).
+
+
+
+## BGM (optional, off by default)
+
+Configure once and it beds the whole video automatically: add `"bgm": "assets/bgm.mp3"` at the top level of script.json (or the object form `{file, volume, fadeIn, fadeOut}`). build-video will: loop it to fill the whole video, fade in over 1.5s, fade out over 2.5s at the tail, and hold it under the voice per `volume` (default **0.12**, i.e. -18.4dB); if mixing fails it automatically falls back to voice only and warns. **ASR verification always switches to the voice-only track**, so BGM doesn't interfere with transcription.
+
+- Volume calibration: 0.08 (very restrained, storytelling) ~ 0.15 (lively, fast-consumption); the listening standard = every word of the voice is clear, and listening to the audio track alone with the picture off is effortless. Watch Gate 5 during final-video acceptance.
+- BGM sources: ① `connector__matrix__batch_text_to_music` (generated on the mcode platform, ≤5 per batch, no copyright risk, the first choice); ② music the user provides (licensing must be confirmed, registered in assets/MANIFEST.md). Don't rip music off video sites. **⚠ mmx-cli has no music generation command**, so non-mcode environments can only use ② (or skip BGM).
+- Generation advice: pick the style per subject matter (tech / warm / light), and **require "no vocals, no strong melodic hook, loopable"**; put the resulting file at `assets/bgm.mp3`, and listen to it once on its own to confirm there are no jarring section changes (the loop point must be clean).
+- Requires ffmpeg ≥ 4.4 (amix's normalize option); older versions automatically skip BGM rather than erroring out.
+
+## Bilingual subtitles (optional)
+
+Adding `text2` to a clause in script.json takes effect automatically: the on-screen subtitles become two lines (Chinese main line + second line in a smaller size), and `out/subs.srt` becomes two lines in sync. No HTML changes and no extra parameters needed.
+
+- `text2` rules: it is the translation of the same sentence, **do not reorder the words**; second line ≤60 characters (plan-timings warns when exceeded); use the target language's punctuation conventions.
+- Doing only some sentences is also fine (each sentence is independent); before delivery confirm the language pairing of the bilingual subtitles (ask clearly at Gate 1 alignment whether it is "Chinese-English" or something else).
+
+**Retry discipline for partial failures**: the upstream is very sensitive to resending an entire failed batch (rate limit). After a failure, sleep 10–30s and **retry only the failed entries** (pick the failures out of requests and form a small new batch); don't resend the whole batch.
+
+Immediately after TTS is written to disk:
+
+```bash
+node scripts/plan-timings.mjs <项目目录>
+```
+
+Give the user the output duration table and warnings together for Gate 2. If a slide gets a speech-rate warning at this point, prefer changing the narration word count (back to Phase 1) or adjusting that slide's speed and redoing that segment — don't move on with warnings left standing.
+
+## ASR reverse verification flow
+
+build-video `--asr` cuts `asr/part-<id>-<k>.mp3` **by sentence** (sentence k of slide id) and generates `checklist.md`. Pick either path:
+
+**Path A · mcode sandbox**
+1. `mcode-tools upload_temp_url asr/part-01-1.mp3` to get a public URL.
+2. `mcode-tools connector call connector__matrix__listen_audio --args '{"audio_info": {"url": "<URL>"}}'`.
+3. Compare the transcript against the "expected narration" in the checklist: numbers, years and product names must match exactly; homophone/punctuation differences are acceptable.
+
+**Path B · mmx-cli (any environment, the same login as TTS — mmx-cli ≥ 1.0.26 ships `speech transcribe`)**
+
+```bash
+node scripts/asr.mjs <项目目录>          # 默认走 mmx speech transcribe, 不需要配 Key → 逐句转写 → 自动比对 → 回填 asr/checklist.md
+```
+
+**Path C · direct REST (fallback for environments without mmx-cli, the same single API Key)**
+
+```bash
+export MINIMAX_API_KEY=sk-xxx          # 与 mmx-cli 同一把; 海外套餐加 MINIMAX_REGION=global
+node scripts/asr.mjs <项目目录> --provider api
+```
+
+- Both paths hit the same backend, share the same limits, and differ only in who holds the credentials; `--provider mmx|api` forces one explicitly (the default auto-picks mmx when ≥ 1.0.26 is on PATH).
+- The script performs the **comparison verdict** automatically: numbers mismatched, or traditional characters appearing (suspected Cantonese) → ✗ and the process exits non-zero; low similarity → ⚠ pending review.
+- If the audio exceeds the API limits (>500s or >50MB) it is automatically transcoded with ffmpeg to mono 16k mp3 before upload; no manual handling needed.
+- If you already have transcripts (e.g. produced with whisper in another environment) you can feed them straight in for comparison (no network needed): `node scripts/asr.mjs <项目> --from transcripts.json` (JSON shaped like `{"part-01-1": "文本"}`).
+- If you want **word-level timestamps to measure each sentence's onset** (more accurate than silence detection; the API supports `timestamp_level: word`): `node scripts/asr.mjs <项目> --verify-timing`, which outputs an "estimate vs ASR-measured" comparison table; if the deviation is large, adjust `clauses[].start` in `timings.json` per the results and re-render.
+
+For slides that don't pass: change the narration or redo that segment's TTS → re-run plan-timings → delete `build/frames/<id>/` and `out/slide-<id>.mp4` → re-run capture (that slide) and build-video. Don't redo the whole video.
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/asr.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/asr.mjs
new file mode 100644
index 00000000..a52dd5b7
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/asr.mjs
@@ -0,0 +1,338 @@
+#!/usr/bin/env node
+// html2video-for-mcode · 语音识别(ASR):两条路径 —— ①mmx cli ≥1.0.26 的 `speech transcribe`
+// (缺省首选; 与 TTS 同一套 mmx 登录, Key 全程不经本脚本); ②直连 MiniMax REST 接口(--provider api,
+// 与 mmx 同一把 API Key; 不走 mcode connector, 也不依赖本机 whisper)。
+//
+// 用法:
+//   node asr.mjs <项目目录>                     转写 asr/part-*.mp3, 与 checklist 的预期文本比对并回填
+//   node asr.mjs <项目目录> --verify-timing     用段/字级时间戳实测每句开口时刻, 对比 timings.json(比静音检测更准)
+//   node asr.mjs --file <音频> [--format json|verbose_json|srt|vtt] [--timestamp word] [--language zh]
+//            [--provider mmx|api] [--out <项目内相对路径>](落盘转写结果; 已存在需 --force; 绝对路径/越出项目目录一律拒绝)
+//
+// 提供方: --provider mmx|api 显式指定; 缺省时 PATH 上有 mmx ≥1.0.26 就走 mmx, 否则走 api。
+//   mmx 路径: 输入音频按原路径**只读**交给 CLI; 转写输出先落 os.tmpdir() 专属中转文件, 再由本脚本
+//   写进受监路径(CLI 拿不到任何写回路径), 且子进程环境里剥掉 MINIMAX_API_KEY/MINIMAX_BASE_URL。
+//   api 路径 Key 来源(按序): --api-key sk-xxx → 环境变量 MINIMAX_API_KEY
+//   api 区域(按序): --base-url → 环境变量 MINIMAX_BASE_URL → MINIMAX_REGION(cn=api.minimaxi.com / global=api.minimax.io)
+// 端点安全(仅 api 路径发 Key): Key 只发上面两个官方域; 其他 --base-url/MINIMAX_BASE_URL 一律拒绝, 自建网关需显式 --allow-any-endpoint。
+// 接口约束(官方文档, 两路径同一后端同一约束): wav/aiff/flac/m4a/mp3/aac/opus/ogg; 时长 ≤500s; 大小 ≤50MB。
+//   超限时本脚本会用 ffmpeg 自动转成单声道 16k mp3 再传(识别率不受影响)。
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+import { spawnSync } from 'node:child_process';
+import { flagValue, positionals, probeDuration, requireTool, safeId, safeOut, stripControlChars, validateTimingsIds } from './tools.mjs';
+import { assertAsrEndpoint, PolicyError } from './url-policy.mjs';
+
+const argv = process.argv.slice(2);
+// 位置参数与取值型 flag 的清单统一在 tools.mjs(单一来源): 每个脚本自己维护一份就会漂
+// (2026-09-18 复查: 三个脚本各有一份, 而 tools 那份还误把布尔开关当取值型)
+const positional = positionals(argv);
+// 项目目录要早定义: --file 单文件模式在上面就分流了, 而工具发现(requireTool)要用它
+const dir = path.resolve(positional[0] ?? '.');
+
+const KEY = flagValue(argv, '--api-key') || process.env.MINIMAX_API_KEY || '';
+const REGION = process.env.MINIMAX_REGION || 'cn';
+// 端点白名单: API Key 只发官方 MiniMax 域; 换端点必须显式 --allow-any-endpoint(危险项)。
+// 防的是环境变量/提示词把 MINIMAX_BASE_URL 偷换成收集器后凭证外发。
+let BASE;
+try {
+  BASE = assertAsrEndpoint(flagValue(argv, '--base-url') || process.env.MINIMAX_BASE_URL
+    || (REGION === 'global' ? 'https://api.minimax.io' : 'https://api.minimaxi.com'),
+  { allowAny: argv.includes('--allow-any-endpoint') });
+} catch (e) {
+  if (e instanceof PolicyError) {
+    console.error(`✗ ASR 端点被拒绝: ${e.message}\n  这把 Key 只允许发往官方端点。自定义网关请确认安全后显式加 --allow-any-endpoint`);
+    process.exit(1);
+  }
+  throw e;
+}
+const LANG = flagValue(argv, '--language', '');          // zh / yue / en ... 空=混合识别
+const FORMAT = flagValue(argv, '--format', 'json');
+const TS = flagValue(argv, '--timestamp', '');           // '' | sentence | word
+
+// ── 提供方选择: mmx cli(≥1.0.26 speech transcribe) 或 直连 REST ──
+const PROVIDER = flagValue(argv, '--provider', '');
+if (PROVIDER && PROVIDER !== 'mmx' && PROVIDER !== 'api') {
+  console.error(`✗ --provider 只认 mmx|api(收到 ${PROVIDER})。mmx = mmx cli ≥1.0.26 的 speech transcribe; api = 直连 REST`);
+  process.exit(1);
+}
+// Windows 上 node 不能直接 spawn npm 全局命令(裸名 ENOENT、.cmd 无 shell 是 EINVAL),
+// 只能经 cmd.exe /c。引用不能只看空格: 合法文件名里的 & | < > ( ) ^ 是 cmd 控制符,
+// 不加引号会被截断/改义 —— 命中任一元字符就整段引用, 引号本身双写。
+// (残留限制: cmd 在引号内仍展开 %VAR%; 文件名恰含同名环境变量时 mmx 会拿到错路径然后响亮失败 ——
+//  用户自己的机器自己的路径, 不是信任边界, 记录在案即可。)
+const IS_WIN = process.platform === 'win32';
+// 单一来源: 所有 mmx 子进程(版本探测与 transcribe)统一从这里 spawn —— env 一律剥掉
+// MINIMAX_*(mmx 用自己的登录; Key 无论如何不流向它, 探测也不例外 —— 1.9.2 修正: 1.9.1 只在
+// transcribe 一处剥, 探测子进程仍继承, 声明过头), 引用规则也只有这一份。
+const mmxRun = (args, opts = {}) => {
+  const env = { ...process.env, ...(opts.env || {}) };
+  delete env.MINIMAX_API_KEY;
+  delete env.MINIMAX_BASE_URL;
+  const common = { encoding: 'utf8', windowsHide: true, ...opts, env };
+  if (IS_WIN) {
+    // Windows 上必须经 cmd(mmx 是 npm 的 .cmd shim, Node 不允许无 shell 执行 .cmd)。
+    // **双引号拦不住 cmd 的 %VAR% 展开** —— 引号内的 %NAME% 照样展开, 变量值里若带 "
+    // 还能把整条命令拆开(2026-10-10 独立审计 L2; 旧注释把它说成"错路径、响亮失败", 说轻了)。
+    // 命令行层面没有可用的转义手段, 所以 fail closed: 明确拒绝并给出可执行的提示。
+    const bad = args.find(a => /%/.test(String(a)));
+    if (bad) {
+      return { status: 1, stdout: '', stderr: `参数里不能有 % (Windows cmd 会在引号内展开 %VAR%, 无法安全转义): ${bad}\n   请把项目目录或文件名里的 % 去掉再重跑` };
+    }
+  }
+  return IS_WIN
+    ? spawnSync(process.env.ComSpec || 'cmd.exe', ['/d', '/s', '/c', ['mmx', ...args].map(a => /[\s"&|<>^()%]/.test(a) ? `"${a.replace(/"/g, '""')}"` : a).join(' ')], common)
+    : spawnSync('mmx', args, common);
+};
+// 探测不能靠 --help(未知子命令的 --help 也退 0), 认版本号: ≥1.0.26 才有 speech transcribe
+let mmxProbe;
+const mmxUsable = () => {
+  if (mmxProbe === undefined) {
+    const r = mmxRun(['--version']);
+    const m = r.status === 0 ? String(r.stdout || '').match(/(\d+)\.(\d+)\.(\d+)/) : null;
+    mmxProbe = !!(m && (+m[1] > 1 || (+m[1] === 1 && (+m[2] > 0 || +m[3] >= 26))));
+  }
+  return mmxProbe;
+};
+const USE_MMX = PROVIDER === 'mmx' ? true : (!PROVIDER && mmxUsable());
+if (PROVIDER === 'mmx' && !mmxUsable()) {
+  console.error('✗ --provider mmx 但 PATH 上的 mmx 没有 speech transcribe(要 ≥1.0.26): npm i -g mmx-cli@latest\n  (或改用 --provider api + MINIMAX_API_KEY 直连 REST)');
+  process.exit(1);
+}
+
+// --from <file.json>: 不打网络, 直接用已有转写结果(JSON: { "part-01-1.mp3": "文本", ... }
+// 或 { "part-01-1": "文本" })做比对与回填 —— 例如你在别的环境用 whisper 转写完再进来核对。
+const FROM = (() => {
+  const f = flagValue(argv, '--from');
+  if (!f) return null;
+  if (!fs.existsSync(f)) { console.error(`✗ 找不到 --from 文件: ${f}`); process.exit(1); }
+  return JSON.parse(fs.readFileSync(f, 'utf8'));
+})();
+const fromLookup = name => {
+  if (!FROM) return null;
+  const base = name.replace(/\.[^.]+$/, '');
+  return FROM[name] ?? FROM[base] ?? FROM[path.basename(name)] ?? FROM[path.basename(base)] ?? null;
+};
+
+if (!USE_MMX && !KEY && !FROM) {
+  console.error([
+    '✗ ASR 两条路径都没就绪, 二选一:',
+    '  1) mmx cli(≥1.0.26, 推荐, 与 TTS 同一套登录): npm i -g mmx-cli@latest && mmx auth login --api-key sk-xxx',
+    '  2) 直连 REST(后备): export MINIMAX_API_KEY=sk-xxx 或 node asr.mjs <项目> --api-key sk-xxx',
+    '     海外套餐加 MINIMAX_REGION=global(或 --base-url https://api.minimax.io)',
+    '  查额度与身份: mmx quota / mmx auth status',
+  ].join('\n'));
+  process.exit(2);
+}
+
+// 超限自动转码成单声道 16k mp3(500s/50MB 硬限制)
+function prepareForUpload(file) {
+  const st = fs.statSync(file);
+  const FFMPEG = requireTool('ffmpeg', dir);
+  const dur = probeDuration(requireTool('ffprobe', dir), file);   // 探测收进 tools.mjs 单一实现(第十三轮 review 去重)
+  if (st.size <= 50 * 1024 * 1024 && (!Number.isFinite(dur) || dur <= 500)) return { file, tmpDir: null };
+  // 转码临时文件放 os.tmpdir() 专属目录(1.7.7 复查: 此前写在输入文件旁边 —— 输入在项目内时,
+  // 预置在旁边的文件符号链接(.asr-<名>.16k.mp3)会被 ffmpeg -y 写穿; 临时产物本就不该落项目)。
+  const tmpDir = fs.mkdtempSync(path.join(os.tmpdir(), 'asr-16k-'));
+  const tmp = path.join(tmpDir, 'audio.16k.mp3');
+  const r = spawnSync(FFMPEG, ['-y', '-v', 'error', '-i', file, '-ac', '1', '-ar', '16000', '-c:a', 'libmp3lame', '-b:a', '128k', tmp], { encoding: 'utf8', windowsHide: true });
+  if (r.status !== 0) { fs.rmSync(tmpDir, { recursive: true, force: true }); console.error(`✗ 转码失败: ${file}`); process.exit(1); }
+  console.log(`  (超限已转码: ${path.basename(file)} → 单声道 16k)`);
+  return { file: tmp, tmpDir };
+}
+
+async function transcribe(file, { format = FORMAT, ts = TS, language = LANG } = {}) {
+  const { file: up, tmpDir } = prepareForUpload(file);
+  try {
+    // mmx 路径: 输入按原路径只读传给 CLI; 输出落 os.tmpdir() 专属目录, 由本脚本读回再写受监路径 ——
+    // CLI 拿不到写回路径。子进程环境剥掉 MINIMAX_*: mmx 用自己的登录, Key 无论如何不流向它。
+    if (USE_MMX) {
+      const mmxTmp = fs.mkdtempSync(path.join(os.tmpdir(), 'asr-mmx-'));
+      try {
+        const doc = path.join(mmxTmp, format === 'srt' ? 'out.srt' : format === 'vtt' ? 'out.vtt' : 'out.json');
+        const args = ['speech', 'transcribe', '--file', up, '--model', 'asr-1.0', '--response-format', format, '--out', doc];
+        if (ts) args.push('--timestamp-level', ts);
+        if (language) args.push('--language', language);
+        const r = mmxRun(args, { timeout: 180000 });
+        if (r.error || r.status !== 0) {
+          const t = String((r.stderr || '') + (r.stdout || '')).trim();
+          throw new Error(`mmx speech transcribe 失败${r.error ? `(${r.error.code || r.error.message})` : `(exit ${r.status})`}: ${t.slice(0, 300)}\n  → 检查 mmx auth login / 网络; 老版本没有该子命令: npm i -g mmx-cli@latest(≥1.0.26)`);
+        }
+        if (!fs.existsSync(doc) || !fs.statSync(doc).size) throw new Error('mmx speech transcribe 没有产出结果文件(--out)');
+        const raw = fs.readFileSync(doc, 'utf8');
+        if (format === 'srt' || format === 'vtt') return { text: raw, raw };
+        const j = JSON.parse(raw);
+        return { text: j.text ?? '', duration: j.duration, segments: j.segments, n_speakers: j.n_speakers, raw: j };
+      } finally {
+        fs.rmSync(mmxTmp, { recursive: true, force: true });
+      }
+    }
+    const fd = new FormData();
+    fd.append('model', 'asr-1.0');
+    fd.append('response_format', format);
+    if (ts) fd.append('timestamp_level', ts);
+    fd.append('file', new Blob([fs.readFileSync(up)], { type: 'audio/mpeg' }), path.basename(up));
+    const headers = { Authorization: `Bearer ${KEY}` };
+    if (language) headers.language = language; // 'zh' 强制普通话; 'yue' 确认粤语
+    let res;
+    try {
+      res = await fetch(`${BASE}/v1/speech_to_text`, { method: 'POST', headers, body: fd, signal: AbortSignal.timeout(180000) });
+    } catch (e) {
+      throw new Error(`网络请求失败(${BASE}): ${e.message}\n  → 检查网络/代理; 海外套餐加 MINIMAX_REGION=global`);
+    }
+    const raw = await res.text();
+    if (!res.ok) {
+      let hint = '';
+      if (res.status === 401) hint = ' → Key 与 region 是否匹配?(国内用 cn, 海外用 global)';
+      if (res.status === 402) hint = ' → 额度不足, mmx quota 看余额';
+      if (res.status === 429) hint = ' → 限流, sleep 10–30s 后重试';
+      if (res.status === 422) hint = ' → 音频含敏感内容';
+      throw new Error(`HTTP ${res.status}${hint}\n${raw.slice(0, 300)}`);
+    }
+    if (format === 'srt' || format === 'vtt') return { text: raw, raw };
+    const j = JSON.parse(raw);
+    return { text: j.text ?? '', duration: j.duration, segments: j.segments, n_speakers: j.n_speakers, raw: j };
+  } finally {
+    if (tmpDir) fs.rmSync(tmpDir, { recursive: true, force: true });
+  }
+}
+
+// ── 文本比对:数字必须一致(硬要求), 出现繁体字/相似度过低则判不通过 ──
+const norm = s => String(s || '').replace(/[\s，。、；：！？,.;:!?"'“”‘’()（）\-—…]/g, '');
+// 常见"仅繁体"用字(简体写法不同者)。启发式:命中任意一个即提示人工复核 ——
+// 我们的口播稿一律简体, 识别结果出现繁体通常意味着音色实际输出粤语/其他语种。
+const TRAD_ONLY = '們來從這個時後裡為與對發現經濟億萬說國學會語寫讀聽東車馬門開關長張聲畫點無龍書當雖應該積極幾號樣麼將帶鐵銀錢樂藥歡機場頭題類數體腦鋼護讓認識記計討論詞課實寶確網絡線組織統較轉輪農業產業務員團圖園圓遠邊過進運遊達適選遺醫鐘錯鏡陽陰隨險隱雙雜難靈靜頁順預領頻願風飛飯飲館駕騎驗黨讀寫個們';
+function checkText(expected, got, lang = 'zh') {
+  const e = norm(expected), g = norm(got);
+  const digits = [...new Set((expected.match(/\d+(\.\d+)?%?/g) || []))];
+  const missingDigits = digits.filter(d => !g.includes(d));
+  const issues = [];
+  if (missingDigits.length) issues.push(`数字对不上: ${missingDigits.join(' ')}`);
+  let langBad = false;
+  if (lang === 'zh') {
+    // 期望普通话: 出现"仅繁体"用字通常意味着音色实际输出粤语/其他语种
+    const trad = [...new Set([...g].filter(c => TRAD_ONLY.includes(c)))];
+    if (trad.length) { issues.push(`疑似非普通话(出现繁体字 ${trad.slice(0, 4).join('')})`); langBad = true; }
+  } else if (lang === 'en') {
+    // 期望英语: 转写里混入成片中文说明是语种错了
+    const cjk = [...new Set([...g].filter(c => /[\u4e00-\u9fff]/.test(c)))];
+    if (cjk.length) { issues.push(`疑似非英语(出现中文字符 ${cjk.slice(0, 4).join('')})`); langBad = true; }
+  }
+  // 相似度: 1 - 编辑距离/最长长度
+  const m = e.length, n = g.length;
+  let prev = Array.from({ length: n + 1 }, (_, j) => j);
+  for (let i = 1; i <= m; i++) {
+    const cur = [i, ...Array(n).fill(0)];
+    for (let j = 1; j <= n; j++) cur[j] = Math.min(prev[j] + 1, cur[j - 1] + 1, prev[j - 1] + (e[i - 1] === g[j - 1] ? 0 : 1));
+    prev = cur;
+  }
+  const sim = m && n ? 1 - prev[n] / Math.max(m, n) : 0;
+  if (sim < 0.5) issues.push(`相似度过低 ${sim.toFixed(2)}`);
+  else if (sim < 0.7) issues.push(`相似度偏低 ${sim.toFixed(2)}`);
+  const hard = missingDigits.length > 0 || langBad || sim < 0.5;
+  return { sim, issues, verdict: hard ? '✗' : (sim < 0.7 ? '⚠' : '✓') };
+}
+
+// ── 模式 1: 单文件转写 ──
+if (flagValue(argv, '--file')) {
+  const f = flagValue(argv, '--file');
+  if (!fs.existsSync(f)) { console.error(`✗ 不存在: ${f}`); process.exit(1); }
+  // --out 收监 + 不覆盖(1.7.5 复查③): 此前是任意路径裸写、静默覆盖。校验必须放在 transcribe
+  // 之前 —— 路径非法/覆盖拒绝不该等一次网络请求之后才报(测试也因此不需要网络)。
+  const outRaw = flagValue(argv, '--out');
+  let outPath = null;
+  if (outRaw) {
+    if (path.isAbsolute(outRaw)) { console.error(`✗ --out 必须是项目目录内的相对路径(收到绝对路径): ${outRaw}`); process.exit(1); }
+    outPath = safeOut(dir, outRaw);
+    if (fs.existsSync(outPath) && !argv.includes('--force')) {
+      console.error(`✗ --out 已存在, 不覆盖: ${path.relative(dir, outPath)}(要覆盖加 --force)`);
+      process.exit(1);
+    }
+  }
+  let r;
+  try { r = await transcribe(f); } catch (e) { console.error(`✗ ${e.message}`); process.exit(1); }
+  if (outPath) fs.writeFileSync(outPath, FORMAT === 'srt' || FORMAT === 'vtt' ? r.text : (r.raw ? JSON.stringify(r.raw, null, 2) : r.text));
+  if (FORMAT === 'srt' || FORMAT === 'vtt') console.log(r.text);
+  else console.log(r.text + (r.duration ? `\n(duration ${r.duration}s${r.n_speakers ? `, ${r.n_speakers} speakers` : ''})` : ''));
+  process.exit(0);
+}
+
+// ── 模式 2/3: 项目目录 ──
+const asrDir = path.join(dir, 'asr');
+if (!fs.existsSync(asrDir)) { console.error(`✗ 找不到 ${asrDir} — 先跑 build-video.mjs <项目> --asr 生成按句切分的音频`); process.exit(1); }
+const timingsPath = path.join(dir, 'build', 'timings.json');
+const timings = fs.existsSync(timingsPath) ? JSON.parse(fs.readFileSync(timingsPath, 'utf8')) : null;
+if (timings) validateTimingsIds(timings);
+
+// 模式 3: 用 ASR 时间戳实测句开口, 校对比 timings 估算
+if (argv.includes('--verify-timing') && timings) {
+  let drift = 0, n = 0, worst = null;
+  console.log('ASR 实测句开口(字/段级时间戳) vs plan-timings 估算:\n');
+  for (const t of timings.slides) {
+    if (!Array.isArray(t.clauses) || t.clauses.length < 2) continue;
+    const tid = safeId(t.id);
+    const audio = path.join(dir, 'audio', `${tid}.mp3`);
+    if (!fs.existsSync(audio)) { console.warn(`- 跳过 ${t.id}: 缺音频`); continue; }
+    const r = await transcribe(audio, { format: 'verbose_json', ts: 'word' });
+    const units = r.segments || [];
+    const rows = [];
+    for (let k = 1; k < t.clauses.length; k++) {
+      const needle = norm(t.clauses[k].text).slice(0, 4); // 用该句前 4 字做锚
+      const hit = units.find(u => norm(u.text).includes(needle));
+      const est = t.clauses[k].start;
+      if (hit) { drift += Math.abs(hit.start - est); n++; if (!worst || Math.abs(hit.start - est) > Math.abs(worst.diff)) worst = { id: t.id, k: k + 1, diff: +(hit.start - est).toFixed(2) }; }
+      rows.push(`    第${k + 1}句开口 估算 ${est}s | ASR 实测 ${hit ? hit.start.toFixed(2) + 's' : '(锚点没找到, 换更长的锚词)'} | 差 ${hit ? (hit.start - est).toFixed(2) : '-'}s`);
+    }
+    console.log(`  ${t.id}:`);
+    console.log(rows.join('\n'));
+  }
+  if (n) console.log(`\n平均偏差 ${(drift / n).toFixed(2)}s, 最大 ${worst ? worst.id + ' 第' + worst.k + '句 ' + worst.diff + 's' : '-'}`);
+  console.log('偏差普遍 >0.3s 时: 用 check-timing.mjs --calibrate(静音法)或用本结果手工改 timings.json 的 clauses[].start。');
+  process.exit(0);
+}
+
+// 模式 2: 批量转写 + 与 checklist 对比回填
+const parts = fs.readdirSync(asrDir).filter(f => /^part-.*\.(mp3|wav|m4a|aac|opus|ogg)$/i.test(f)).sort();
+if (!parts.length) { console.error(`✗ ${asrDir} 里没有 part-* 音频`); process.exit(1); }
+// 项目语种: script.json 的 lang 决定默认识别语言(zh 强制普通话 / en / yue ...), 也决定语种校验策略
+const scriptJsonPath = path.join(dir, 'script.json');
+const projectLang = fs.existsSync(scriptJsonPath) ? (JSON.parse(fs.readFileSync(scriptJsonPath, 'utf8')).lang ?? 'zh') : 'zh';
+const useLang = LANG || projectLang;
+const ckPath = safeOut(dir, 'asr', 'checklist.md');
+let ck = fs.existsSync(ckPath) ? fs.readFileSync(ckPath, 'utf8') : '';
+const results = [];
+for (const p of parts) {
+  const f = path.join(asrDir, p);
+  try {
+    const r = FROM ? { text: fromLookup(p) ?? '' } : await transcribe(f, { language: useLang });
+    // 转写文本来自不可信音频内容: 控制字节能改终端标题/染色, bidi 覆盖能让比对与阅读都被骗
+    // (2026-10-10 审计 L4)。判定的输入用净化后的文本, 写回与打印也用它。
+    r.text = stripControlChars(r.text);
+    const rowRe = new RegExp(`\\|\\s*asr/${p.replace(/[.*+?^${}()|[\]\\]/g, '\\$&')}\\s*\\|([^|]*)\\|([^|]*)\\|([^|]*)\\|([^|]*)\\|`);
+    const mm = ck.match(rowRe);
+    const expected = mm ? mm[2].trim() : '';
+    const verdict = expected ? checkText(expected, r.text, projectLang) : null;
+    results.push({ p, text: r.text, expected, verdict });
+    if (mm) {
+      // 函数替换: 转写文本里的 $& / $1 否则会被当替换模式展开, 弄脏 checklist 行
+      ck = ck.replace(rowRe, () => `| asr/${p} |${mm[1]}|${mm[2]}| ${r.text} | ${verdict ? verdict.verdict + (verdict.issues.length ? ' ' + verdict.issues.join('; ') : '') : ''} |`);
+    }
+    console.log(`${verdict ? verdict.verdict : '·'} ${p}  ${r.text}${verdict && verdict.issues.length ? '\n    ⚠ ' + verdict.issues.join('; ') : ''}`);
+  } catch (e) {
+    console.error(`✗ ${p}: ${e.message}`);
+    results.push({ p, error: e.message });
+  }
+}
+if (ck) { fs.writeFileSync(ckPath, ck); console.log(`\n已回填 asr/checklist.md(转写 + 判定)`); }
+const bad = results.filter(r => r.verdict?.verdict === '✗').length;
+const warn = results.filter(r => r.verdict?.verdict === '⚠').length;
+// 请求失败(鉴权/限流/网络)此前只记进 results 不进统计 → "1 通过 / 0 不通过 / 退出 0" 把闸门放过去了
+// (二审 P2): 它们必须单独计数、不算通过, 且整轮非零退出
+const errored = results.filter(r => r.error);
+console.log(`\n完成: ${results.length - bad - warn - errored.length} 通过 / ${warn} 待复核 / ${bad} 不通过${errored.length ? ` / ${errored.length} 请求失败` : ''}`);
+if (bad || errored.length) {
+  if (bad) console.error('不通过项: 数字或语种不符 —— 改口播或重做该段 TTS, 重跑 plan-timings 与该张渲染。');
+  if (errored.length) console.error(`请求失败 ${errored.length} 段(鉴权/限流/网络): ${errored.map(r => r.p).join(' ')} — 这些段没有校验结论, 修好网络或 Key 后重跑; 不能当通过。`);
+  process.exit(1);
+}
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/build-video.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/build-video.mjs
new file mode 100644
index 00000000..1e378e78
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/build-video.mjs
@@ -0,0 +1,432 @@
+#!/usr/bin/env node
+// html2video-for-mcode · 组装成片: 单张编码 → 拼接 → 音轨对位 → mux → 自检; --asr 按句切分校验 + 出 SRT 字幕。
+// 用法: node build-video.mjs <项目目录> [--asr] [--dry-run] [--allow-stale-css]
+import fs from 'node:fs';
+import path from 'node:path';
+import { spawnSync } from 'node:child_process';
+import { XFADE_DEFAULT_DUR, assertConcatPathSafe, stripControlChars, buildSrt, flagValue, positionalDir, probeDuration as ffprobeDuration, probeSize as ffprobeDims, readTransition, requireFreshCss, requireTool, safeId, safeOut, safeRel, validateScriptPaths, validateTimingsIds } from './tools.mjs';
+
+const argv = process.argv.slice(2);
+const dir = positionalDir(argv);
+const DRY = argv.includes('--dry-run');
+const WANT_ASR = argv.includes('--asr');
+
+// 入口闸门(与 capture 同一道): 受管块落后即停 —— 否则成片用的是旧 CSS, 全程没有任何报错
+requireFreshCss(dir, { who: 'build-video', allowStale: argv.includes('--allow-stale-css') });
+
+process.env.KIT_PROJECT_DIR = dir;
+
+const run = (cmd, args, label) => {
+  if (DRY) { console.log(`[dry-run] ${path.basename(cmd)} ${args.join(' ')}`); return ''; }
+  const r = spawnSync(cmd, args, { encoding: 'utf8', windowsHide: true });
+  if (r.status !== 0) {
+    console.error(`✗ ${label || path.basename(cmd)} 失败:\n${(r.stderr || '').split('\n').slice(-8).join('\n')}`);
+    process.exit(1);
+  }
+  return r.stdout ?? '';
+};
+
+const script = JSON.parse(fs.readFileSync(path.join(dir, 'script.json'), 'utf8'));
+validateScriptPaths(script, dir); // script.json 是 agent 可编辑文件: id/audio/bgm 派生路径先收监
+const timings = JSON.parse(fs.readFileSync(path.join(dir, 'build', 'timings.json'), 'utf8'));
+validateTimingsIds(timings);
+// W/H 会拼进 ffmpeg filter 字符串(scale=...), 强转整数并限范围, 防字符串注入
+const W = clampDim(script.width, 1920, 'width'), H = clampDim(script.height, 1080, 'height');
+function clampDim(v, dflt, name) {
+  const n = Number(v ?? dflt);
+  if (!Number.isInteger(n) || n < 16 || n > 16384) {
+    console.error(`✗ script.${name} 非法: ${JSON.stringify(v)} — 需要 16–16384 的整数`);
+    process.exit(1);
+  }
+  return n;
+}
+const fps = clampNum(timings.fps ?? script.fps, 30, 'fps', 1, 240, 'timings.fps/script.fps');
+const total = timings.total;
+const abs = p => path.resolve(p).replace(/\\/g, '/');
+// 同 clampDim, 但允许小数、不带默认值兜底(bgm.volume/fade 这类会拼进 ffmpeg -filter_complex 的数值都用它)
+function clampNum(v, dflt, name, min, max, where = `script.${name}`) {
+  const n = Number(v ?? dflt);
+  if (!Number.isFinite(n) || n < min || n > max) {
+    console.error(`✗ ${where} 非法: ${JSON.stringify(v)} — 需要 ${min}–${max} 的数值(该值会拼进 ffmpeg 参数)`);
+    process.exit(1);
+  }
+  return n;
+}
+// BGM 数值全部在编码开始前定下来: 坏配置要立刻失败, 不能等渲完几分钟才报(这批值都会拼进 -filter_complex)
+const bgmCfgRaw = script.bgm ? (typeof script.bgm === 'string' ? { file: script.bgm } : script.bgm) : null;
+const bgmNum = bgmCfgRaw ? {
+  vol: clampNum(bgmCfgRaw.volume, 0.12, 'volume', 0, 4, 'bgm.volume'),
+  fi: clampNum(bgmCfgRaw.fadeIn, 1.5, 'fadeIn', 0, 30, 'bgm.fadeIn'),
+  fo: clampNum(bgmCfgRaw.fadeOut, 2.5, 'fadeOut', 0, 30, 'bgm.fadeOut'),
+} : null;
+
+// 工具发现放在全部配置校验之后: 配置写错就是配置写错, 与这台机器上有没有 ffmpeg 无关。
+// 反过来的话, 少装了 ffmpeg 的人拿到的是"找不到 ffmpeg"(退出 2), 真正要报的 width/bgm/fps 非法被吞掉。
+const FFMPEG = requireTool('ffmpeg', dir);
+const FFPROBE = requireTool('ffprobe', dir);
+
+// ffprobe 探测收进 tools.mjs 单一实现(第十三轮 review 去重); 这里保留本地薄封装, 调用点不动。
+const probeDur = f => ffprobeDuration(FFPROBE, f);
+const probeSize = f => ffprobeDims(FFPROBE, f);
+
+fs.mkdirSync(safeOut(dir, 'out'), { recursive: true });
+fs.mkdirSync(safeOut(dir, 'build'), { recursive: true });
+
+// ── 1. 单张编码 ─────────────────────────────────────────────
+// 转场(1.6.0): 默认"硬切" —— 段间不再淡出到黑再淡入(那会在每次切页留 ≈0.55s 纯黑,
+// 实测反馈"每一个大页切换过程会经过黑屏")。首段用封面溶解进入(帧 0 = 完整封面, 不再是黑帧),
+// 末段保留结尾淡出收尾;中间段之间是硬切。xfade 溶解见 TRANSITION。
+// 解析统一走 tools.readTransition(): --transition <值> / "--transition=<值>" / script.transition 三个来源,
+// 以及裸字符串与 {type,duration} 两种写法, 闸门(check-slides)与这里必须得到完全一致的结论(D6)。
+const TRANSITION = (() => {
+  const a = argv.find(x => x.startsWith('--transition='));
+  const cli = flagValue(argv, '--transition', null);
+  const v = (a ? a.split('=')[1] : cli) ?? script.transition ?? null;
+  // CLI(或裸字符串)只给了类型时, 时长仍取 script.transition.duration —— 老行为:
+  // {duration:1.5} + --transition xfade = 1.5s; 不能因为换了解析函数就悄悄退回 0.4s
+  const merged = typeof v === 'string' && script.transition && typeof script.transition === 'object'
+    ? { type: v, duration: script.transition.duration ?? XFADE_DEFAULT_DUR }
+    : v;
+  try {
+    // 报错要点名来源: --transition 写错时不该说成 script.transition 写错
+    return readTransition(merged, { where: (a || cli) ? '--transition' : 'script.transition' });
+  } catch (e) {
+    console.error(`✗ ${e.message}`);
+    process.exit(1);
+  }
+})();
+const coverPath = path.join(dir, 'preview', 'cover.png');
+const hasCover = fs.existsSync(coverPath);
+if (TRANSITION.type === 'xfade') console.log(`  转场: 交叉溶解 ${TRANSITION.dur}s(段尾各多留 ${TRANSITION.dur}s, 叠化后总时长不变)`);
+else console.log('  转场: 硬切(段间不淡出到黑;首段封面溶解、末段淡出收尾)');
+
+const segs = [];
+const segDurs = [];
+const nSeg = timings.slides.length;
+for (const [segIdx, t] of timings.slides.entries()) {
+  const tid = safeId(t.id);
+  const D = t.duration;
+  const isFirst = segIdx === 0, isLast = segIdx === nSeg - 1;
+  // xfade: 除末段外每段多留 transition.dur 的尾帧(叠化时被吃掉, 总时长仍等于 timings.total)
+  const tailHold = TRANSITION.type === 'xfade' && !isLast ? TRANSITION.dur : 0;
+  const fadeOut = Math.max(0, D - 0.35);
+  // 淡出只留给末段收尾;淡入一律不用(首段走封面溶解,其余硬切)
+  const fades = isLast ? `fade=t=out:st=${fadeOut.toFixed(3)}:d=0.3` : '';
+  const fdir = path.join(dir, 'build', 'frames', tid);
+  const firstFrame = path.join(fdir, 'f00000.png');
+  const png = path.join(dir, 'preview', `${tid}.png`);
+  const seg = safeOut(dir, 'out', `slide-${tid}.mp4`);
+  const common = ['-c:v', 'libx264', '-pix_fmt', 'yuv420p', '-preset', 'medium', '-crf', '20', '-movflags', '+faststart'];
+  const stillPng = k => path.join(dir, 'build', 'substills', tid, `s${k}.png`);
+  // 首段的"封面溶解": 封面淡出叠在画面上 → 帧 0 = 完整封面, 0.25s 内溶解进入入场动画(不加时长)。
+  // (2026-09-18 复查 E5: 这里原本是个带 input/chain 两个字段的结构体, 两个字段都是死代码 ——
+  //  真正管用的判断在 encode() 里, 改成布尔, 少一处"看起来会用"的误导。)
+  const useCoverOverlay = isFirst && hasCover;
+
+  // 字幕时间轴(二审 P1): 帧序列只覆盖动画窗, 之后的字幕变化要靠 capture 逐句截的静态图拼上来;
+  // 静态/no-fx 路径没有帧序列, 整张就由"基底图 + 逐句字幕图"拼成(否则整片一帧字幕都没有)。
+  // 复查补(1.7.0): 清单必须与当前这张对齐才用 —— 否则删掉 clauses/改过时长之后, 陈旧清单会把
+  // 早已不存在的字幕拼回成片(capture 已按张清理, 这里再做一次防御性核对, 因为清单是磁盘文件)
+  let subStills = [];
+  const manPath = path.join(dir, 'build', 'substills', `${tid}.json`);
+  if (fs.existsSync(manPath)) {
+    let man = null;
+    try { man = JSON.parse(fs.readFileSync(manPath, 'utf8')); } catch { /* 坏清单按无处理 */ }
+    const clauseCount = Array.isArray(t.clauses) ? t.clauses.length : 0;
+    const durOk = man && Number.isFinite(man.duration) && Math.abs(man.duration - D) <= 0.05;
+    // 清单的 framesCover 必须等于"当前帧序列的实际时长"(无帧序列 = 0)。时长与句序都对、
+    // 但动画窗重截过(帧数变了)的清单, 会把字幕静帧从错误的时刻拼进成片 —— B2 点名的第三项。
+    const nFramesNow = fs.existsSync(fdir) ? fs.readdirSync(fdir).filter(f => /^f\d+\.png$/.test(f)).length : 0;
+    const coverNow = nFramesNow / fps;
+    const coverOk = man && Number.isFinite(man.framesCover) && Math.abs(man.framesCover - coverNow) <= 0.05;
+    if (man && !durOk) console.warn(`⚠ ${tid}: 字幕清单时长 ${man.duration}s ≠ 当前 ${D}s — 忽略该清单(重跑 capture 生成)`);
+    if (man && durOk && !coverOk) console.warn(`⚠ ${tid}: 字幕清单帧覆盖 ${man.framesCover}s ≠ 当前 ${coverNow}s — 忽略该清单(重跑 capture 生成)`);
+    if (man && durOk && coverOk && clauseCount === 0) console.warn(`⚠ ${tid}: 当前没有 clauses 但存在字幕清单 — 忽略(重跑 capture 会清掉)`);
+    if (man && durOk && coverOk && clauseCount > 0) {
+      subStills = (man.stills ?? []).filter(s2 => Number.isFinite(s2.start) && Number.isFinite(s2.end) &&
+        s2.start >= -0.001 && s2.end <= man.duration + 0.05 &&
+        s2.end - s2.start > 0.02 && Number.isInteger(s2.k) && s2.k >= 0 && s2.k < clauseCount &&
+        fs.existsSync(stillPng(s2.k)));
+    }
+  }
+
+  // 统一走 filter_complex, 标签式组装: 基底 → (首段: 封面溶解) → (末段: 淡出)
+  // 这样三种编码路径(帧序列+字幕段 / 帧序列 / 静态图)共用同一段收尾逻辑, 不会再出现
+  // "-vf 里拼空字符串" 那类边界(硬切时 fades 为空)。
+  const DUR = D + tailHold;                       // 实际编码时长(xfade 时多留尾帧)
+  const encode = (inputs, baseGraph, baseLabel, label) => {
+    const graph = [baseGraph];
+    let cur = baseLabel;
+    if (useCoverOverlay) {
+      const ovIdx = inputs.length;
+      graph.push(`[${ovIdx}:v]scale=${W}:${H}:flags=lanczos,setsar=1,fade=t=out:st=0:d=0.25[ov]`);
+      graph.push(`[${cur}][ov]overlay=format=auto[merged]`);
+      cur = 'merged';
+      inputs = [...inputs, { args: ['-loop', '1', '-i', coverPath] }];
+    }
+    if (fades) { graph.push(`[${cur}]${fades}[fin]`); cur = 'fin'; }
+    const args = ['-y', ...inputs.flatMap(i => i.args)];
+    run(FFMPEG, [...args, '-filter_complex', graph.join(';'), '-map', `[${cur}]`, '-t', DUR.toFixed(4), ...common, seg], label);
+  };
+
+  // [逐张素材] → concat 段的组装(两处调用: 帧序列+字幕段 / 静态图+字幕段)。
+  // 2026-09-18 复查 E5: 这段拼装原先是复制两份的(只有"每段各自探尺寸"vs"整张一个前缀"这点差别),
+  // 差别用 preOf 一个回调表达, 于是"末段补齐到整张时长"这条关键规则只有一处实现。
+  const SCALE = `scale=${W}:${H}:flags=lanczos,`;
+  const encodeConcat = (parts, { preOf, label }) => {
+    const sum = parts.reduce((a, p) => a + p.dur, 0);
+    parts[parts.length - 1].dur += (D + tailHold) - sum;   // 末段补齐: 吃掉取整误差, 总长严格 = D+tailHold
+    const chains = parts.map((p, i) => `[${i}:v]${preOf(p)}setsar=1,fps=${fps}[v${i}]`).join(';');
+    const cat = parts.map((_, i) => `[v${i}]`).join('') + `concat=n=${parts.length}:v=1:a=0[cat]`;
+    const inputs = parts.map(p => ({ args: p.kind === 'frames'
+      ? ['-framerate', String(fps), '-t', p.dur.toFixed(4), '-i', p.src]
+      : ['-loop', '1', '-framerate', String(fps), '-t', p.dur.toFixed(4), '-i', p.src] }));
+    encode(inputs, `${chains};${cat}`, 'cat', label);
+  };
+
+  if (fs.existsSync(firstFrame)) {
+    const nFrames = fs.readdirSync(fdir).filter(f => /^f\d+\.png$/.test(f)).length;
+    const framesDur = nFrames / fps;
+    if (subStills.length) {
+      // [帧序列段] + [每句字幕段]: 段长由字幕窗口决定, 末段补齐到整张时长(吃掉取整误差)
+      const parts = [
+        { kind: 'frames', src: path.join(fdir, 'f%05d.png'), dur: framesDur },
+        ...subStills.map(s2 => ({ kind: 'still', src: stillPng(s2.k), dur: s2.end - s2.start })),
+      ];
+      // 帧序列与字幕静帧可能尺寸不同(dsf 超采样), 所以逐段各自探尺寸
+      const preOf = p => { const sz = probeSize(p.src); return sz && (sz[0] !== W || sz[1] !== H) ? SCALE : ''; };
+      encodeConcat(parts, { preOf, label: `编码 ${tid} (帧序列 ${framesDur.toFixed(1)}s + ${subStills.length} 段字幕)` });
+    } else {
+      const sz = probeSize(firstFrame);
+      const pre = [`tpad=stop_mode=clone:stop_duration=${(DUR + 1).toFixed(3)}`];
+      if (sz && (sz[0] !== W || sz[1] !== H)) pre.push(`scale=${W}:${H}:flags=lanczos`); // dsf 2 超采样降采
+      encode([{ args: ['-framerate', String(fps), '-i', path.join(fdir, 'f%05d.png')] }],
+        `[0:v]${pre.join(',')},setsar=1,${'fps=' + fps}[base]`, 'base', `编码 ${tid} (帧序列)`);
+    }
+  } else if (fs.existsSync(png)) {
+    // still 复截会把该张帧目录作废(capture 的防旧帧污染设计), 回退静态图 = 该张动画不进视频;
+    // "改完 HTML 跑 still 复看再直接 build-video"极易踩进且此前零提示, 必须点名怎么补
+    console.warn(`⚠ ${tid} 无帧序列, 用 preview/${tid}.png 静态图出片(该张动画不进视频) — 若应有动画: node scripts/capture.mjs <项目目录> --mode motion --ids ${tid} 后重建`);
+    const sz = probeSize(png);
+    const pre = [];
+    if (sz && (sz[0] !== W || sz[1] !== H)) pre.push(`scale=${W}:${H}:flags=lanczos`);
+    pre.push('setsar=1', `fps=${fps}`);
+    if (subStills.length) {
+      const head = subStills[0].start;                 // 第一句开口前: 无字幕的基底图
+      const parts = [
+        ...(head > 0.02 ? [{ kind: 'still', src: png, dur: head }] : []),
+        ...subStills.map(s2 => ({ kind: 'still', src: stillPng(s2.k), dur: s2.end - s2.start })),
+      ];
+      // 这一支里所有段都是同一尺寸的静态图, 所以共用一个前缀
+      const preOf = () => (sz && (sz[0] !== W || sz[1] !== H) ? SCALE : '');
+      encodeConcat(parts, { preOf, label: `编码 ${tid} (静态图 + ${subStills.length} 段字幕)` });
+    } else {
+      encode([{ args: ['-loop', '1', '-i', png] }],
+        `[0:v]${pre.join(',')}[base]`, 'base', `编码 ${tid} (静态图)`);
+    }
+  } else {
+    console.error(`✗ ${tid} 既无帧序列也无 preview/${tid}.png — 先运行 capture.mjs`);
+    process.exit(1);
+  }
+  segs.push(seg);
+  segDurs.push(DUR);
+  const d = probeDur(seg);
+  if (d != null && Math.abs(d - DUR) > 0.2) console.warn(`⚠ ${tid} 段长 ${d.toFixed(2)}s ≠ 预期 ${DUR.toFixed(2)}s`);
+}
+
+// ── 2. 拼接视频 ──────────────────────────────────────────────
+// cut(默认): 同参数段无损 copy 拼接(段间是硬切, 不再经过黑场)。
+// xfade: 段尾已各多留 dur 秒尾帧, 这里用 xfade 叠化 —— 叠化吃掉 (n-1)×dur, 总时长仍等于 timings.total。
+const noaudio = safeOut(dir, 'build', 'video-noaudio.mp4');
+let vd = null;
+if (TRANSITION.type === 'xfade' && segs.length > 1) {
+  const inputs = segs.flatMap(s => ['-i', s]);
+  let fc = '', cur = '[0:v]', acc = 0;
+  for (let i = 1; i < segs.length; i++) {
+    acc += segDurs[i - 1] - TRANSITION.dur;
+    const out = i === segs.length - 1 ? '[v]' : `[x${i}]`;
+    fc += `${cur}[${i}:v]xfade=transition=fade:duration=${TRANSITION.dur}:offset=${acc.toFixed(3)}${out};`;
+    cur = out;
+  }
+  run(FFMPEG, ['-y', ...inputs, '-filter_complex', fc.slice(0, -1), '-map', '[v]',
+    '-c:v', 'libx264', '-pix_fmt', 'yuv420p', '-preset', 'medium', '-crf', '20',
+    '-r', String(fps), '-movflags', '+faststart', noaudio], '拼接(xfade 溶解)');
+  vd = probeDur(noaudio);
+} else {
+  const listFile = safeOut(dir, 'build', 'concat.txt');
+  // 守卫在 tools.assertConcatPathSafe(单源, 可单测): 见那里的 av_get_token 约定说明
+  try {
+    segs.forEach(s => assertConcatPathSafe(abs(s)));
+  } catch (e) {
+    console.error(`✗ ${e.message}`);
+    process.exit(1);
+  }
+  fs.writeFileSync(listFile, segs.map(s => `file '${abs(s)}'`).join('\n') + '\n');
+  run(FFMPEG, ['-y', '-f', 'concat', '-safe', '0', '-i', listFile, '-c', 'copy', noaudio], '拼接(copy)');
+  vd = probeDur(noaudio);
+}
+if (vd == null || Math.abs(vd - total) > 0.25) {
+  console.warn(`⚠ 拼接时长 ${vd?.toFixed(2) ?? '?'}s 偏离预期 ${total.toFixed(2)}s — 回退 concat filter 重编码`);
+  const inputs = segs.flatMap(s => ['-i', s]);
+  const fc = segs.map((_, i) => `[${i}:v]`).join('') + `concat=n=${segs.length}:v=1:a=0[v]`;
+  run(FFMPEG, ['-y', ...inputs, '-filter_complex', fc, '-map', '[v]',
+    '-c:v', 'libx264', '-pix_fmt', 'yuv420p', '-preset', 'medium', '-crf', '20',
+    '-r', String(fps), '-movflags', '+faststart', noaudio], '拼接(重编码)');
+  vd = probeDur(noaudio);
+}
+
+// ── 2b. 封面(1.6.0): 导出 out/cover.png 供平台上传; 成片里内嵌 attached_pic ──
+const coverOut = hasCover ? safeOut(dir, 'out', 'cover.png') : null;
+if (hasCover) {
+  const szc = probeSize(coverPath);
+  run(FFMPEG, ['-y', '-i', coverPath, ...(szc && (szc[0] !== W || szc[1] !== H) ? ['-vf', `scale=${W}:${H}:flags=lanczos`] : []),
+    '-frames:v', '1', coverOut], '导出封面 out/cover.png');
+  console.log('  封面: out/cover.png(画布尺寸)+ 成片内嵌 attached_pic(文件管理器/IM 的缩略图读它)');
+} else {
+  console.warn('⚠ 缺 preview/cover.png(封面图)—— 重跑 capture 会为第 1 张生成; 现在成片没有内嵌封面');
+}
+
+// ── 3. 音轨对位(每段按实测时长补静音到整张, 再顺序拼接) ──────
+const audioInputs = [];
+const chains = [];
+timings.slides.forEach((t, i) => {
+  const tid = safeId(t.id);
+  const audioRel = script.slides.find(s => s.id === tid)?.audio ?? `${tid}.mp3`;
+  const a = safeRel(path.join(dir, 'audio'), audioRel, { where: `slides[${tid}].audio` });
+  if (!fs.existsSync(a)) { console.error(`✗ 缺音频 ${a}`); process.exit(1); }
+  audioInputs.push(a);
+  chains.push(`[${i}:a]aresample=44100,aformat=channel_layouts=mono,apad=whole_dur=${t.duration.toFixed(4)}[s${i}]`);
+});
+const audioWav = safeOut(dir, 'build', 'audio-timeline.wav');
+run(FFMPEG, ['-y', ...audioInputs.flatMap(a => ['-i', a]),
+  '-filter_complex', [...chains, `${chains.map((_, i) => `[s${i}]`).join('')}concat=n=${chains.length}:v=0:a=1[out]`].join(';'),
+  '-map', '[out]', '-c:a', 'pcm_s16le', audioWav], '音轨对位');
+
+// ── 4. BGM(可选): script.json 里配 bgm 即自动垫底 — 循环补齐、淡入淡出、人声优先 ──
+let audioFinal = audioWav;
+if (script.bgm) {
+  const cfg = bgmCfgRaw;
+  const { vol, fi, fo } = bgmNum;
+  const bgmPath = safeRel(dir, cfg.file, { where: 'bgm.file' }); // 拒绝绝对路径与越界(原来的 path.resolve 会整体逃逸)
+  if (!fs.existsSync(bgmPath)) {
+    console.warn(`⚠ 配置了 bgm 但找不到 ${bgmPath} — 跳过, 只出人声`);
+  } else {
+    const mixWav = safeOut(dir, 'build', 'audio-mix.wav');
+    const fc = `[1:a]aresample=44100,aformat=channel_layouts=mono,volume=${vol},`
+      + `afade=t=in:st=0:d=${fi},afade=t=out:st=${Math.max(0, total - fo).toFixed(3)}:d=${fo}[bg];`
+      + `[0:a][bg]amix=inputs=2:duration=first:normalize=0[out]`;
+    const r = spawnSync(FFMPEG, ['-y', '-i', audioWav, '-stream_loop', '-1', '-i', bgmPath,
+      '-filter_complex', fc, '-map', '[out]', '-c:a', 'pcm_s16le', mixWav], { encoding: 'utf8', windowsHide: true });
+    if (r.status === 0) {
+      audioFinal = mixWav;
+      console.log(`  BGM: ${cfg.file} @音量 ${vol} (淡入 ${fi}s / 淡出 ${fo}s, 自动循环补满)`);
+    } else {
+      console.warn('⚠ BGM 混音失败(amix normalize 需 ffmpeg ≥ 4.4) — 已跳过, 只出人声');
+    }
+  }
+}
+
+// ── 5. mux ──────────────────────────────────────────────────
+// 封面走 attached_pic: 主视频流仍然 copy(不重编码), 封面是单帧 PNG 流 ——
+// 文件管理器 / 多数播放器 / 部分 IM 的缩略图直接读它, 发给别人时不再是黑首帧。
+// 注意: 内嵌封面时不能带 -shortest(会把输出截到那一帧的长度), 音轨本身已按总时长对齐。
+const final = safeOut(dir, 'out', 'final.mp4');
+run(FFMPEG, hasCover
+  ? ['-y', '-i', noaudio, '-i', audioFinal, '-i', coverPath, '-map', '0:v:0', '-map', '1:a:0', '-map', '2:v:0',
+    '-c:v:0', 'copy', '-c:a', 'aac', '-b:a', '192k', '-c:v:1', 'png', '-disposition:v:1', 'attached_pic',
+    '-movflags', '+faststart', final]
+  : ['-y', '-i', noaudio, '-i', audioFinal, '-map', '0:v:0', '-map', '1:a:0',
+    '-c:v', 'copy', '-c:a', 'aac', '-b:a', '192k', '-shortest', '-movflags', '+faststart', final],
+  'mux 成片');
+
+// ── 6. 自检 ─────────────────────────────────────────────────
+// --dry-run 什么都没真编码, 产物不存在 → 自检必然全红。以前它会照跑并退出 1, 于是"看一眼计划"
+// 这个用法总是在 CI/agent 里报失败(2026-09-18 复查 D6 时踩到)。dry-run 下只打印计划, 直接成功。
+if (DRY) {
+  console.log('[dry-run] 已打印完整计划(未编码、未写产物), 跳过自检');
+  process.exit(0);
+}
+const fd = probeDur(final);
+const okDur = fd != null && Math.abs(fd - total) <= 0.25;
+const decode = spawnSync(FFMPEG, ['-v', 'error', '-i', final, '-f', 'null', '-'], { encoding: 'utf8', windowsHide: true });
+const okDecode = decode.status === 0;
+console.log(`\n成片: ${abs(final)}`);
+console.log(`  时长 ${fd?.toFixed(2) ?? '?'}s / 预期 ${total.toFixed(2)}s ${okDur ? '✓' : '✗'}`);
+console.log(`  全量解码 ${okDecode ? '✓ 无错误' : '✗ ' + (decode.stderr || '').slice(0, 300)}`);
+
+// 画面级自检(1.6.0): 首帧是不是黑、切页处有没有黑帧、封面在不在 —— 这三类此前全靠人眼,
+// 而"发给别人看到黑缩略图""每次切页黑屏"恰恰是用户实测反馈出来的。
+const meanLuma = t => {
+  const o = spawnSync(FFMPEG, ['-hide_banner', '-ss', String(t), '-i', final, '-frames:v', '1',
+    '-vf', 'signalstats,metadata=print:file=-', '-f', 'null', '-'], { encoding: 'utf8', windowsHide: true });
+  const m = /YAVG=([0-9.]+)/.exec((o.stdout || '') + (o.stderr || ''));
+  return m ? parseFloat(m[1]) : null;
+};
+const BLACK = 16;                                     // 视频黑电平 ~16(YAVG); 低于它按"黑帧"处理
+const gates = [];
+// ① 首帧: 不能是黑帧(封面溶解后应为完整封面)
+if (!hasCover) {
+  const l0 = meanLuma(0.05);
+  gates.push({ ok: l0 == null || l0 > BLACK, msg: `首帧亮度 ${l0?.toFixed(0) ?? '?'}(应 > ${BLACK}) —— 分享时的缩略图就是它` });
+} else {
+  gates.push({ ok: true, msg: '封面已内嵌(attached_pic), 缩略图不依赖首帧' });
+  const l0 = meanLuma(0.05);
+  gates.push({ ok: l0 == null || l0 > BLACK, msg: `首帧亮度 ${l0?.toFixed(0) ?? '?'}(应 > ${BLACK}, 封面溶解)` });
+}
+// ② 切页处不能有黑帧(硬切/溶解都不该经过黑场)
+let blackCuts = 0;
+let cum2 = 0;
+for (let i = 0; i < timings.slides.length - 1; i++) {
+  cum2 += timings.slides[i].duration;
+  for (const dt of [-0.12, 0, 0.12]) {
+    const l = meanLuma(Math.max(0, cum2 + dt));
+    if (l != null && l <= BLACK) { blackCuts++; break; }
+  }
+}
+gates.push({ ok: blackCuts === 0, msg: blackCuts ? `${blackCuts} 处切页检出黑帧(转场不应经过黑场; 检查 transition 与 fade 设置)` : '切页无黑帧' });
+for (const g of gates) console.log(`  ${g.ok ? '✓' : '✗'} ${g.msg}`);
+if (!okDur || !okDecode || gates.some(g => !g.ok)) process.exit(1);
+
+// ── 7. SRT 字幕文件(与烧录字幕同源同窗, 供平台上传用) ────────
+// 生成逻辑在 tools.buildSrt(单一来源): clause 文案是不可信输入, 换行会在成品里伪造成
+// 额外 cue —— 净化与编号都在那一处, 测试也能直接断言。
+const srt = buildSrt(timings);
+if (srt.count) {
+  fs.writeFileSync(safeOut(dir, 'out', 'subs.srt'), srt.text);
+  console.log(`  字幕 out/subs.srt (${srt.count} 条, 与画面烧录字幕同源)`);
+}
+
+// ── 8. ASR 按句切分 + 校验清单 ──────────────────────────────
+if (WANT_ASR) {
+  fs.mkdirSync(safeOut(dir, 'asr'), { recursive: true });
+  for (const old of fs.readdirSync(path.join(dir, 'asr')).filter(f => f.startsWith('part-'))) {
+    // 注意顺序: 先逐个 safeOut 删掉旧切分(预置的叶子文件符号链接在这里就会被拒),
+    // 之后 :433 的 ffmpeg 写 part-<id>-<n>.mp3 才不会有"顺着预置链接写出去"的口子。
+    fs.rmSync(safeOut(dir, 'asr', old)); // 清掉上一轮的旧切分, 避免新旧混淆
+  }
+  const lines = ['# ASR 反向校验(按句切分)', '',
+    '每段 = 一句口播。逐段上传转写(mcode: upload_temp_url → connector__matrix__listen_audio), 与"预期文本"比对:',
+    '数字/年份/产品名必须完全一致; 同音字与标点差异可接受。同时核对: 若某段转写混入了上一句的开头,',
+    '说明该句实际开口比估算晚 → 跑 node scripts/check-timing.mjs 做静音实测校准。',
+    '不过关的 slide: 改口播或重做该段 TTS → 重跑 plan-timings → 删 build/frames/<id>/ 与 out/slide-<id>.mp4 后重建。', '',
+    '| part | 全片时间 | 预期文本 | ASR 转写 | 通过? |', '|---|---|---|---|---|'];
+  let count = 0;
+  let asrCum = 0;                 // 全片时间轴在本段的累计起点(SRT 那份累计在 tools.buildSrt 里)
+  for (const t of timings.slides) {
+    const tid = safeId(t.id);
+    const clauses = Array.isArray(t.clauses) && t.clauses.length ? t.clauses : [{ start: 0, text: t.script }];
+    clauses.forEach((c, i) => {
+      const end = t.clauses[i + 1]?.start ?? t.duration;
+      const part = path.join(dir, 'asr', `part-${tid}-${i + 1}.mp3`);
+      run(FFMPEG, ['-y', '-ss', (asrCum + c.start).toFixed(3), '-t', (end - c.start).toFixed(3), '-i', audioWav,
+        '-ac', '1', '-ar', '16000', '-c:a', 'libmp3lame', '-b:a', '128k', part], `ASR 切分 ${tid}-${i + 1}`);
+      // 预期文本取发音形态(say ?? text): 音频里说的是 say; 显示形态(text)只进字幕/srt
+      lines.push(`| asr/part-${tid}-${i + 1}.mp3 | ${(asrCum + c.start).toFixed(1)}s | ${stripControlChars(c.say ?? c.text)} |  |  |`);
+      count++;
+    });
+    asrCum += t.duration;
+  }
+  fs.writeFileSync(safeOut(dir, 'asr', 'checklist.md'), lines.join('\n') + '\n');
+  console.log(`  ASR 素材与清单已生成: asr/ (${count} 句, 每句独立切分)`);
+}
+console.log('\n完成 ✅');
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/capture.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/capture.mjs
new file mode 100644
index 00000000..6aff6e10
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/capture.mjs
@@ -0,0 +1,366 @@
+#!/usr/bin/env node
+// html2video-for-mcode · 画面采集: HTML → 终态 PNG(still) / 逐帧 PNG 序列(motion)
+// 用法: node capture.mjs <项目目录> [--mode still|motion] [--ids 01,03] [--dsf 1-4] [--no-subs] [--allow-stale-css]
+// 依赖: 项目目录或其上层 node_modules 里有 playwright, 且已装 chromium。
+import fs from 'node:fs';
+import path from 'node:path';
+import { pathToFileURL } from 'node:url';
+import { flagValue, loadPackage, positionalDir, requireFreshCss, safeId, safeOut, safeRel, subPillCss, subScale, subtitleKeyframes, subtitleWindows, validateScriptPaths, validateTimingsIds } from './tools.mjs';
+import { startVetoProxy } from './url-policy.mjs';
+
+const argv = process.argv.slice(2);
+const dir = positionalDir(argv);
+// 参数先验后用: 写错的 --mode/--dsf 以前会一路带到 playwright(page.setViewportSize(scale=NaN) /
+// 走错分支静默出静态图), 报错点离真正的原因很远 —— 这里是"离用户最近"的地方, 直接说清。
+const mode = flagValue(argv, '--mode', 'still');
+if (mode !== 'still' && mode !== 'motion') {
+  console.error(`✗ --mode ${JSON.stringify(mode)} 非法: 只能是 still(终态图) 或 motion(逐帧序列)`);
+  process.exit(1);
+}
+const DSF_RAW = flagValue(argv, '--dsf', '1');
+const dsf = Number(DSF_RAW);
+if (!Number.isInteger(dsf) || dsf < 1 || dsf > 4) {
+  console.error(`✗ --dsf ${JSON.stringify(DSF_RAW)} 非法: 取 1–4 的整数(deviceScaleFactor, 2 = 2 倍图; 过大在 4K 画布上会爆内存)`);
+  process.exit(1);
+}
+const SUBS = !argv.includes('--no-subs'); // 字幕默认烧录
+const idsFilter = flagValue(argv, '--ids', '') ? flagValue(argv, '--ids', '').split(',').map(s => s.trim()) : null;
+
+// 入口闸门: tokens.css 受管块落后就停 —— 旧 CSS 会照常出图, 全程不报错(见 tools.mjs 的说明)
+requireFreshCss(dir, { who: 'capture', allowStale: argv.includes('--allow-stale-css') });
+
+const script = JSON.parse(fs.readFileSync(path.join(dir, 'script.json'), 'utf8'));
+validateScriptPaths(script, dir); // script.json 是 agent 可编辑文件: id/html 等派生路径先收监再使用
+const timingsPath = path.join(dir, 'build', 'timings.json');
+if (!fs.existsSync(timingsPath)) {
+  console.error('✗ 缺 build/timings.json — 先运行 plan-timings.mjs');
+  process.exit(1);
+}
+const timings = JSON.parse(fs.readFileSync(timingsPath, 'utf8'));
+validateTimingsIds(timings);
+
+// 画布值会进 playwright viewport(与 preview-page 的 CSS/JS、build-video 的 ffmpeg filter 同一信任级),
+// 先验后用; 放在 playwright 加载之前 —— 无 playwright 的 CI 上这道门也测得到(第八轮复查 F3)
+const cvW = Number(script.width ?? 1920), cvH = Number(script.height ?? 1080);
+if (!Number.isInteger(cvW) || cvW < 16 || cvW > 16384 || !Number.isInteger(cvH) || cvH < 16 || cvH > 16384) {
+  console.error(`✗ script.width/height 非法: ${JSON.stringify(script.width)} × ${JSON.stringify(script.height)} — 需要 16–16384 的整数(与 preview-page / build-video 同一道门)`);
+  process.exit(1);
+}
+
+const slides = script.slides.filter(s => !idsFilter || idsFilter.includes(s.id));
+const firstId = script.slides[0]?.id;   // 封面图取自成片第 1 段(与 --ids 无关)
+if (!slides.length) { console.error('✗ 没有匹配的 slide'); process.exit(1); }
+// 部分匹配不再静默少做(2026-09-28 Mavis 交接): --ids 里没命中的 id 点名告警 + 打印本次实际
+// 生效列表。顺带覆盖 PowerShell 裸逗号列表吃前导零的形态(03,05 → 3,5 匹配不上零填充 id)。
+if (idsFilter) {
+  const hitIds = new Set(slides.map(s => s.id));
+  const miss = idsFilter.filter(id => !hitIds.has(id));
+  if (miss.length) console.warn(`⚠ --ids 里没有对应 slide 的 id, 已忽略: ${miss.join(', ')}(注意 PowerShell 裸写 03,05 会被当数字数组吃掉前导零 —— 加引号 --ids "03,05" 或逐个传)`);
+  console.log(`本次处理: ${[...hitIds].join(', ')}`);
+}
+
+const playwright = await loadPackage('playwright', { projectDir: dir });
+if (!playwright) {
+  console.error('✗ 未找到 playwright(已按 项目目录 / 调用目录 / npm 全局 逐个找过)。在项目目录执行:\n  npm i playwright\n  npx playwright install chromium');
+  process.exit(1);
+}
+
+// 等字体 + 图片就绪: 样式表 <link> load → 每个 @font-face load() → fonts.ready → 图片 → 2×rAF。8s 硬上限。
+async function waitAssets(page) {
+  await page.evaluate(() => new Promise(resolve => {
+    let settled = false;
+    const done = () => { if (!settled) { settled = true; resolve(); } };
+    setTimeout(done, 8000);
+    const raf = () => new Promise(r => requestAnimationFrame(() => r()));
+    (async () => {
+      try {
+        const links = [...document.querySelectorAll('link[rel="stylesheet"]')];
+        await Promise.all(links.map(l => new Promise(r => {
+          if (l.sheet) return r();
+          l.addEventListener('load', r, { once: true });
+          l.addEventListener('error', r, { once: true });
+        })));
+        await Promise.all([...document.fonts].map(f => f.load().catch(() => {})));
+        await document.fonts.ready;
+        const imgs = [...document.images].filter(i => !i.complete);
+        if (imgs.length) await Promise.race([
+          Promise.all(imgs.map(i => new Promise(r => {
+            i.addEventListener('load', r, { once: true });
+            i.addEventListener('error', r, { once: true });
+          }))),
+          new Promise(r => setTimeout(r, 4000)),
+        ]);
+        await raf(); await raf();
+      } catch { /* 尽力而为 */ }
+      done();
+    })();
+  }));
+}
+
+// 渲染 Chromium 也进否决隧道(2026-09-28 第 23 轮收口): slide HTML 理论上本地生成、外链由
+// check-slides 的 error 闸门拦 —— 但闸门不是连接期; 万一被绕过, 私网/元数据地址在这里被
+// 建连期否决(公共外网照常可达, file:// 页不经代理)。
+const veto = await startVetoProxy();
+const browser = await playwright.chromium.launch({ headless: true, proxy: { server: `http://127.0.0.1:${veto.port}` } });
+const context = await browser.newContext({
+  viewport: { width: script.width ?? 1920, height: script.height ?? 1080 },
+  deviceScaleFactor: dsf,
+});
+fs.mkdirSync(safeOut(dir, 'preview'), { recursive: true });
+
+// ── 封面图(1.6.0): 第 1 张另出一张"标题全露、无字幕"的终态图 preview/cover.png ──
+// 用途: build-video 把它内嵌为 attached_pic(文件管理器/播放器/多数 IM 的缩略图都读它),
+// 并导出 out/cover.png 供平台手动上传;首段还用它做 0.25s 溶解过渡, 免得成片首帧是黑的。
+async function captureCover(page, { sid, firstId }) {
+  if (sid !== firstId) return false;
+  await page.evaluate(() => {
+    // 封面必须无字幕(平台自己会叠字): 显式写内联隐藏, 不依赖动画 —— captureSubStills 的
+    // 钉死会把字幕动画整个撤掉, 靠 currentTime=0 藏字幕对已钉元素无效(2026-09-22 实测封面烙图)
+    document.querySelectorAll('.kit-sub').forEach(el => { el.style.animation = 'none'; el.style.opacity = '0'; });
+    document.getAnimations().forEach(a => {
+      try { a.finish(); } catch { /* 无限氛围动画 finish 会抛, 忽略 */ }   // 基底走终态
+    });
+  });
+  await page.evaluate(() => new Promise(r => requestAnimationFrame(() => requestAnimationFrame(r))));
+  await page.screenshot({ path: safeOut(dir, 'preview', 'cover.png') });
+  return true;
+}
+
+// ── 字幕时间轴(二审 P1): 字幕是"全时长百分比动画", 而逐帧只覆盖动画窗, 其余靠 tpad 冻尾帧 ——
+// 动画窗之后的字幕变化永远进不了画面(实测: SRT 第 4 句在 5.3s, 6.0s 的帧还显示第 2 句);
+// 静态/no-fx 路径更是一帧都不带字幕(finish() 把字幕动画跳到 opacity:0)。
+// 解法: 帧序列之外, 再为每个"帧覆盖不到的字幕窗口"截一张**终态基底 + 该句字幕**的静态图,
+// 由 build-video 按窗口时长把它们拼在帧序列之后(清单见 build/substills/<id>.json)。
+// 每张的字幕静帧都是**派生数据**: 重截该张时先整目录清掉, 否则旧句子会阴魂不散 ——
+// `--no-subs`、删掉 clauses、句子变少时, 残留的 s<k>.png 会被 build-video 拼回成片
+// (2026-09-18 复查实测: 只作废 build/frames 不够)。
+function invalidateSubStills(sid) {
+  fs.rmSync(safeOut(dir, 'build', 'substills', sid), { recursive: true, force: true });
+  fs.rmSync(safeOut(dir, 'build', 'substills', sid + '.json'), { force: true });   // 清单一起清, 免得陈旧清单配新图
+}
+
+async function captureSubStills(page, { sid, t, framesCover }) {
+  invalidateSubStills(sid);                       // 先清旧图; 本张若无需字幕, 清完即保持空
+  if (!SUBS || !Array.isArray(t.clauses) || !t.clauses.length) return 0;
+  const stills = [];
+  for (let k = 0; k < t.clauses.length; k++) {
+    const start = t.clauses[k].start;
+    const end = t.clauses[k + 1]?.start ?? t.duration;
+    const from = Math.max(start, framesCover);
+    if (end - from <= 0.02) continue;                 // 这段已被帧序列覆盖(逐帧 seek 时字幕本身就是对的)
+    await page.evaluate(({ k }) => {
+      const subs = document.querySelectorAll('.kit-sub');
+      // 每张截图共用同一页(逐句字幕图、封面、终态图之间): 每一轮都必须把**全部**字幕重新钉一遍 ——
+      // 目标钉满亮、其余钉隐藏。不能只钉目标: 钉 = 停动画 + 写内联样式, 上一轮被钉过的元素已无
+      // 动画可拨回 0, 内联样式会漏进后续截图(2026-09-22 实测: 下一张静帧里叠出上一句的深色底板补丁)。
+      subs.forEach((el, j) => {
+        el.style.animation = 'none';
+        el.style.opacity = j === k ? '1' : '0';
+      });
+      document.getAnimations().forEach(a => {
+        try { a.finish(); } catch { /* 无限氛围动画 finish 会抛, 忽略 */ }   // 基底一律终态
+      });
+    }, { k });
+    await page.evaluate(() => new Promise(r => requestAnimationFrame(() => requestAnimationFrame(r))));
+    const out = safeOut(dir, 'build', 'substills', sid, `s${k}.png`);
+    fs.mkdirSync(path.dirname(out), { recursive: true });
+    await page.screenshot({ path: out });
+    stills.push({ k, start: from, end });
+  }
+  fs.mkdirSync(safeOut(dir, 'build', 'substills'), { recursive: true });
+  fs.writeFileSync(safeOut(dir, 'build', 'substills', sid + '.json'),
+    JSON.stringify({ fps: timings.fps ?? 30, duration: t.duration, framesCover, stills }, null, 2));
+  return stills.length;
+}
+
+let done = 0;
+for (const s of slides) {
+  const sid = safeId(s.id); // 递归删帧目录/写 preview 都由 id 派生, 在源头再拦一次
+  const htmlPath = safeRel(path.join(dir, 'slides'), s.html ?? `${sid}.html`, { where: `slides[${sid}].html` });
+  if (!fs.existsSync(htmlPath)) { console.warn(`- 跳过 ${s.id}: 缺 ${htmlPath}`); continue; }
+  const t = timings.slides.find(x => x.id === s.id);
+  if (!t) { console.warn(`- 跳过 ${s.id}: timings 里无此张`); continue; }
+
+  const page = await context.newPage();
+  // 在页面任何样式生效前, 注入: 实测的 stage 延迟(--t1/--t2/--t3)、画布尺寸(--stage-w/h)、字幕缩放(--sub-scale)。
+  // HTML/tokens.css 里只写占位默认值, 真值一律由管线给 —— 这样改 TTS 或改画布都不用动 HTML。
+  await page.addInitScript(payload => {
+    const apply = () => {
+      const el = document.documentElement;
+      if (!el) return;
+      for (const [k, v] of Object.entries(payload.stages)) el.style.setProperty('--t' + k, `${Math.max(0, Math.round(v * 1000))}ms`);
+      el.style.setProperty('--stage-w', payload.w + 'px');
+      el.style.setProperty('--stage-h', payload.h + 'px');
+      el.style.setProperty('--sub-scale', String(payload.subScale));
+    };
+    apply();
+    document.addEventListener('DOMContentLoaded', apply, { once: true });
+  }, {
+    stages: t.stages ?? {},
+    w: script.width ?? 1920,
+    h: script.height ?? 1080,
+    // 字幕字号随画布宽度缩放, 竖版 (1080 宽) 收窄到 0.75 下限保证可读
+    subScale: subScale(script.width),   // 几何唯一来源 tools.subScale(clamp(W/1920, 0.75, 1.25))
+  });
+
+  // 字幕(默认开, --no-subs 关): 内容取自 timings.clauses(单一数据源, 不在 HTML 里重写),
+  // 显示窗 = 该句开口 → 下句开口。做成百分比关键帧动画: 逐帧 seek 天然工作, still 模式 finish() 后自动隐藏。
+  // 必须在 goto 之前 addInitScript 才会生效。
+  if (SUBS && Array.isArray(t.clauses) && t.clauses.length) {
+    // 胶囊样式整体来自 tools.subPillCss(单一来源): 含 width:max-content —— 没有它,
+    // left:50% 的可用宽度只有画布一半, max-width 永远够不到, 字幕提前一半折行(Mavis A/B 实测)
+    await page.addInitScript(({ clauses, durMs, css }) => {
+      const mount = () => {
+        const st = document.createElement('style');
+        st.textContent = css;
+        (document.head || document.documentElement).appendChild(st);
+        const host = document.querySelector('.stage') || document.body;
+        // 幂等: 若 mount 因任何原因被触发多次, 先清掉上一次的挂载, 避免字幕元素累积
+        document.querySelectorAll('.kit-sub').forEach(el => el.remove());
+        clauses.forEach((c, i) => {
+          const d = document.createElement('div');
+          d.className = 'kit-sub';
+          d.style.animation = `kit-sub-${i} ${durMs}ms linear both`;
+          const l1 = document.createElement('div');
+          l1.textContent = c.text;
+          d.appendChild(l1);
+          if (c.text2) { // 双语第二行(可选): clause 里给 text2 即自动两行
+            const l2 = document.createElement('div');
+            l2.className = 'kit-sub-2';
+            l2.textContent = c.text2;
+            d.appendChild(l2);
+          }
+          host.appendChild(d);
+        });
+      };
+      if (document.readyState === 'loading') document.addEventListener('DOMContentLoaded', mount, { once: true });
+      else mount();
+    }, { clauses: t.clauses, durMs: t.duration * 1000, css: subPillCss() + subtitleKeyframes(t.clauses, t.duration) });
+
+    // 字幕窗口自检(与页面关键帧同一来源 tools.subtitleWindows): 任意时刻最多一条字幕可见。
+    // 淡入/淡出都收在各自窗口内 —— 真正的重叠只有一种: 上一句窗口过短被 a+0.5 顶出(prevB > a)。
+    // (2026-09-22 修: 旧判据拿"已是淡出终点"的 prevB 再叠一次淡出长度, 双重计数 → 恒真误报。)
+    const wins = subtitleWindows(t.clauses, t.duration);
+    for (let i = 0; i < wins.length; i++) {
+      const { a, b, isLast } = wins[i];
+      if (!isLast && b - a < 0.6) console.warn(`⚠ ${s.id}: 第 ${i + 1} 句字幕窗口仅 ${(t.duration * (b - a) / 100).toFixed(2)}s, 可能会一闪而过`);
+      if (i > 0) {
+        const prevB = wins[i - 1].b;
+        if (prevB > a + 0.001) {
+          console.warn(`⚠ ${s.id}: 第 ${i} 句与第 ${i + 1} 句字幕会同时可见(重叠 ${(t.duration * (prevB - a) / 100).toFixed(2)}s)`);
+        }
+      }
+    }
+  } else if (SUBS) {
+    console.warn(`⚠ ${s.id}: 本张没有 clauses, 不会烧录字幕 —— 若这条片要有字幕, 回 Phase 2 重跑 plan-timings`);
+  }
+
+  await page.goto(pathToFileURL(htmlPath).href, { waitUntil: 'domcontentloaded' });
+  await waitAssets(page);
+
+  // 图片兜底告警: waitAssets 是"尽力而为", 加载不动的图会让画面出 broken 图标却一路报成功。
+  // 这里显式报出来(静态路径缺失查 check-slides.mjs; 这里覆盖"文件在但 SVG/图片本身坏了"的情况)。
+  const brokenImgs = await page.evaluate(() =>
+    [...document.images].filter(i => !i.complete || i.naturalWidth === 0).map(i => (i.getAttribute('src') || '').slice(0, 60)));
+  if (brokenImgs.length) {
+    console.warn(`⚠ ${s.id}: ${brokenImgs.length} 张图没渲染出来 → ${brokenImgs.slice(0, 3).join(' | ')}`);
+    console.warn('   先跑 check-slides.mjs 定位; SVG 类资源建议 inline 进 HTML(依赖外部资源/XML 有误/缺 width-height 都会 broken)');
+  }
+
+  // 布局几何自检(2026-09-22 实测踩坑): 内容总高超过可用区时, 普通 flex center 把溢出推到画布
+  // 上方之外裁掉(实测 kicker 被顶到 y≈17px、溢出 ≈125px, 而截图目检被当成"没渲染"放过)。
+  // 静态文本查不出"内容太高", 这里实测并点名数字; 骨架 .layout 的 safe center + 上下 padding
+  // (见 _template.html)让顶部永不被裁, 但溢出本身仍值得报。
+  const geo = await page.evaluate(() => {
+    const lay = document.querySelector('.layout') || document.querySelector('.stage');
+    if (!lay) return null;
+    const kids = [...lay.children].filter(el => {
+      const st = getComputedStyle(el);
+      return st.position !== 'absolute' && st.visibility !== 'hidden' && el.getBoundingClientRect().height > 0;
+    });
+    return {
+      scroll: lay.scrollHeight,
+      client: lay.clientHeight,
+      top: kids.length ? Math.min(...kids.map(el => el.getBoundingClientRect().top)) : null,
+    };
+  });
+  if (geo && (geo.scroll > geo.client + 2 || (geo.top != null && geo.top < 24))) {
+    const over = geo.scroll - geo.client;
+    console.warn(`⚠ ${s.id}: 布局几何 — 内容总高 ${geo.scroll}px${over > 2 ? ` > 可用 ${geo.client}px(溢出 ≈${over}px, 普通 flex center 会把它推到画布上方裁掉)` : ''}${geo.top != null && geo.top < 24 ? `; 顶部元素被推到 y≈${Math.round(geo.top)}px(贴顶/被裁)` : ''} — 减内容/降字号/缩绘图区; 骨架 .layout 是 safe center + 上下 padding(见 _template.html), 别退回普通 center、也别用 margin-top 推内容(会加进内容总高)`);
+  }
+
+  // 帧目录是派生数据: 本次只要产出的是静态图(still 或 无动画), 旧 motion 帧一律作废,
+  // 否则 build-video 会优先用残留帧, 把过时动画混进成片(切 no-fx 后重渲染时必踩)
+  // safeOut: 输出侧也要防"项目内某段是符号链接"——rmSync(recursive) 会穿透符号链接删到项目外(二审 P1)
+  const invalidateFrames = () => fs.rmSync(safeOut(dir, 'build', 'frames', sid), { recursive: true, force: true });
+
+  if (mode === 'still') {
+    // 直接跳到所有有限动画的终态(finish), 无限氛围动画保持运行, 截图即终态。
+    await page.evaluate(() => {
+      document.getAnimations().forEach(a => { try { a.finish(); } catch { /* infinite */ } });
+      // 终态图是"无字幕基底"(build-video 拼在第一句开口前): 显式藏字幕 —— 依赖关键帧终值
+      // 恰好为 0 太脆(1.9.4 的末句形状曾把最后一句烙进基底), 两端隐藏契约另有 unit 测试背书
+      document.querySelectorAll('.kit-sub').forEach(el => { el.style.animation = 'none'; el.style.opacity = '0'; });
+    });
+    await page.evaluate(() => new Promise(r => requestAnimationFrame(() => requestAnimationFrame(r))));
+    await page.screenshot({ path: safeOut(dir, 'preview', `${sid}.png`) });
+    // 静态路径没有帧序列 → 字幕窗口全部覆盖不到, 逐句出静态图交给 build-video 拼
+    const nStill = await captureSubStills(page, { sid, t, framesCover: 0 });
+    const cover = await captureCover(page, { sid, firstId });
+    invalidateFrames();
+    console.log(`✓ ${sid} 终态截图 → preview/${sid}.png${nStill ? ` + 字幕图 ${nStill} 张(静态路径, 由 build-video 拼时段)` : ''}${cover ? ' + 封面 preview/cover.png' : ''}`);
+  } else {
+    // 逐帧步进: 全部动画暂停在 0, 每帧统一 seek 到 t, 截图。CSS 动画自带 delay, seek 是绝对时间, 时序天然正确。
+    await page.evaluate(() => {
+      document.getAnimations().forEach(a => { try { a.pause(); a.currentTime = 0; } catch { /* ignore */ } });
+    });
+    const meta = await page.evaluate(() => {
+      let end = 0, count = 0;
+      for (const a of document.getAnimations()) {
+        count++;
+        const el = a.effect?.target;
+        if (el?.closest?.('.kit-sub')) continue; // 字幕是全时长百分比动画, 不参与动画窗计算
+        let ct; try { ct = a.effect.getComputedTiming(); } catch { continue; }
+        if (Number.isFinite(ct.endTime)) end = Math.max(end, ct.endTime / 1000);
+      }
+      return { count, animEnd: end };
+    });
+    if (meta.animEnd <= 0.05) {
+      // 页面没有任何有限动画: 静态页, 单帧即全部信息, 走 still 路径(字幕同样逐句出图)
+      await page.screenshot({ path: safeOut(dir, 'preview', `${sid}.png`) });
+      const nStill = await captureSubStills(page, { sid, t, framesCover: 0 });
+      const cover = await captureCover(page, { sid, firstId });
+      invalidateFrames();
+      console.log(`✓ ${sid} 无动画, 静态截图 → preview/${sid}.png${nStill ? ` + 字幕图 ${nStill} 张` : ''}${cover ? ' + 封面 preview/cover.png' : ''}`);
+      await page.close(); done++; continue;
+    }
+    const fps = timings.fps ?? 30;
+    const windowS = Math.min(t.duration, meta.animEnd + 0.25); // 动画窗口逐帧, 其余靠字幕图/尾帧补
+    const frames = Math.max(1, Math.ceil(windowS * fps));
+    const fdir = safeOut(dir, 'build', 'frames', sid);
+    fs.rmSync(fdir, { recursive: true, force: true });
+    fs.mkdirSync(fdir, { recursive: true });
+    const t0 = Date.now();
+    for (let i = 0; i < frames; i++) {
+      const ms = (i / fps) * 1000;
+      await page.evaluate(ms => {
+        document.getAnimations().forEach(a => { try { a.currentTime = ms; } catch { /* ignore */ } });
+      }, ms);
+      await page.screenshot({ path: path.join(fdir, 'f' + String(i).padStart(5, '0') + '.png') });
+      if (i > 0 && i % 60 === 0) console.log(`  ${s.id}: ${i}/${frames} 帧 (${((Date.now() - t0) / 1000).toFixed(0)}s)`);
+    }
+    fs.copyFileSync(path.join(fdir, 'f' + String(frames - 1).padStart(5, '0') + '.png'),
+      safeOut(dir, 'preview', `${sid}.png`));
+    // 动画窗之后的字幕变化: 逐句出静态图(framesCover = 帧序列实际时长, 这批窗口交给它之后的拼接)
+    const nStill = await captureSubStills(page, { sid, t, framesCover: frames / fps });
+    const cover = await captureCover(page, { sid, firstId });
+    console.log(`✓ ${sid} ${frames} 帧 @${fps}fps (动画窗 ${windowS.toFixed(1)}s / 成片 ${t.duration.toFixed(1)}s) → build/frames/${sid}/${nStill ? ` + 字幕图 ${nStill} 张` : ''}${cover ? ' + 封面 preview/cover.png' : ''} + preview/${sid}.png`);
+  }
+  await page.close();
+  done++;
+}
+
+await context.close();
+await browser.close();
+await veto.close();   // 先关浏览器再关代理(关代理会掐掉仍在途的隧道)
+console.log(`完成: ${done}/${slides.length} 张 (mode=${mode})`);
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/chart-css.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/chart-css.mjs
new file mode 100644
index 00000000..9b4ec73b
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/chart-css.mjs
@@ -0,0 +1,98 @@
+// html2video-for-mcode · 图表动效与图表原语的唯一来源。
+// init-project 把它写进 tokens.css 的受管块; `--upgrade-css` 按内容 rev 原地更新(见 css-kit.mjs)。
+// 老项目缺这段时新配方会静默半死(条形不生长 / 环形不扫出 / 数字不滚动)。
+
+export const CHART_CSS = `/* ── 图表动效(数据图形专用) ─────────────────────────────────────
+   静态默认值一律 = 终值, keyframes 从 0 长到终值 —— 这样关掉动效(no-fx)时画面自然
+   停在完成态, 不需要额外复位规则。目标值由内联 --w / --p-to / --n-to 给。
+   三个注册属性(@property)让"扫出/宽度生长/数字滚动"都能被逐帧 seek(纯 CSS 动画)。
+   为什么不用 scaleX 给带标签的条形: scaleX 会把条内的数字横向拉扁(2026-09-18 实测),
+   带标签用 fx-grow-w(动画 width), 只有纯色块才用 fx-grow-x。 */
+@property --pv { syntax: '<number>'; initial-value: 0; inherits: false; }
+@property --cnt { syntax: '<integer>'; initial-value: 0; inherits: false; }
+
+.fx-grow-w { width: var(--w, 100%); animation: fx-grow-w-kf .9s var(--ease-out) var(--fx-delay, 0ms) both; }
+@keyframes fx-grow-w-kf { from { width: 0; opacity: 1; } to { width: var(--w, 100%); opacity: 1; } }
+
+.fx-grow-h { height: var(--h, 100%); animation: fx-grow-h-kf .9s var(--ease-out) var(--fx-delay, 0ms) both; }
+@keyframes fx-grow-h-kf { from { height: 0; opacity: 1; } to { height: var(--h, 100%); opacity: 1; } }
+
+/* 环形/仪表: --p-to 是目标百分比(0–100); 锥面渐变随 --pv 逐帧重算 */
+.fx-sweep { --pv: var(--p-to, 100); animation: fx-sweep-kf 1.1s var(--ease-out) var(--fx-delay, 0ms) both; }
+@keyframes fx-sweep-kf { from { --pv: 0; opacity: 1; } to { --pv: var(--p-to, 100); opacity: 1; } }
+
+.fx-count { --cnt: var(--n-to, 0); counter-reset: fxcount var(--cnt); animation: fx-count-kf 1s var(--ease-out) var(--fx-delay, 0ms) both; }
+.fx-count::after { content: counter(fxcount); }
+/* 数字滚动: 只支持整数。入场前 opacity:0 —— 否则画面会停在"灰色的 0", 读成"没数据"(2026-09-18 judge 抓到) */
+@keyframes fx-count-kf { 0% { --cnt: 0; opacity: 0; } 15% { opacity: 1; } 100% { --cnt: var(--n-to, 0); opacity: 1; } }
+
+/* 数据点: 折线/散点上的小圆点逐个弹出(配合 --fx-delay 错峰) */
+.fx-dot { animation: fx-dot-kf .45s cubic-bezier(.22,1.3,.36,1) var(--fx-delay, 0ms) both; }
+@keyframes fx-dot-kf { from { opacity: 0; transform: scale(0); } to { opacity: 1; transform: scale(1); } }
+
+/* 图表原语(数据图形统一用它, 别再各写一套内联样式) —— 配色纪律: 同类同色,
+   关键一条用 --accent, 其余用弱化的同类色(color-mix); 涨跌用 --up/--down(受众翻转向 compliance.md) */
+.chart { display: flex; flex-direction: column; gap: var(--sp-3); }
+.chart-head { display: flex; align-items: baseline; gap: var(--sp-3); }
+.chart-title { font-size: var(--fs-caption); color: var(--muted); }
+.chart-num { font-family: var(--font-mono); font-variant-numeric: tabular-nums; }
+.chart-note { margin-left: auto; font-size: var(--fs-tiny); color: var(--accent); border: 1px solid var(--line);
+  border-radius: 999px; padding: 2px 10px; }
+.chart-plot { position: relative; }
+.chart-plot.grid { background-image: repeating-linear-gradient(to top, var(--line) 0 1px, transparent 1px 25%); }
+/* 用 grid 固定三列(标签 / 轨道 / 数值位) —— 有没有条外数值, 轨道都一样长,
+   否则"加了数值的行"轨道会被挤窄, 同一张图里出现两把尺子, 条长失去可比性(2026-09-18 视觉验收抓到) */
+.chart-row { display: grid; grid-template-columns: var(--lbl-w, 200px) minmax(0, 1fr) var(--val-w, 96px);
+  align-items: center; gap: var(--sp-3); }
+.chart-lbl { flex: 0 0 200px; text-align: right; color: var(--muted); font-size: var(--fs-caption); }
+.chart-track { position: relative; flex: 1 1 auto; width: 100%; height: var(--bar-h, 34px); background: var(--panel-2);
+  border-radius: var(--radius-sm); overflow: hidden; }
+/* 纵向细网格: 每条轨道共享 25/50/75% 刻度, 观感上就是"同一把尺子"(比只给浅灰轨道精致) */
+.chart-track.grid { background-image: repeating-linear-gradient(to right, var(--line) 0 1px, transparent 1px 25%); }
+/* 零轴: 轨道左缘一条 1px 竖线, 让"从这里开始量"有明确起点 */
+.chart-track::before { content: ""; position: absolute; left: 0; top: 0; bottom: 0; width: 1px;
+  background: var(--line-strong); opacity: .5; z-index: 1; }
+.chart-bar { height: 100%; background: var(--accent); display: flex; align-items: center;
+  justify-content: flex-end; padding-right: 12px; overflow: hidden; }
+.chart-bar.dim { background: var(--accent);      /* 弱化条用不透明同色系: 半透明会让格子线透出来, 被误读成分段/堆叠 */
+  background: color-mix(in srgb, var(--accent) 30%, var(--bg)); }
+.chart-bar .chart-val { color: var(--accent-ink); font-size: var(--fs-tiny); white-space: nowrap; }
+/* 弱化条(半透明同色系)上压白字读不出来 —— 条内数值自动改墨色; 也可以直接把数值放条外(.out) */
+.chart-bar.dim .chart-val { color: var(--fg); }
+.chart-val.out { color: var(--fg); font-size: var(--fs-caption); }
+.chart-cols { display: flex; align-items: stretch; justify-content: center; gap: var(--sp-4); height: var(--cols-h, 380px); }
+/* 三层结构, 每一层的职责单一:
+   .chart-col       = [绘图区 1fr] + [轴标签 auto]
+   .chart-plot-cell = 绘图区本身: 底边 = 基线, 网格也画在这里(所以网格与柱高同一把尺子)
+   .chart-bar-wrap  = 一根柱的高度基准(style="height:77.5%" = 数值 ÷ 轴上限), 数值标签挂它顶上
+   ⚠ 柱高必须相对绘图区算: 直接把柱子放进 flex 列按百分比设高, 基准是整列、超出剩余空间会被
+   flex-shrink 压回去(实测 84% 与 72% 画成一样高); 网格若铺在整列上、柱区却更矮, 读者按网格
+   读出的数值就与标注不符(两条都实测被视觉验收判 fail) */
+.chart-col { flex: 0 1 200px; display: grid; grid-template-rows: minmax(0, 1fr) auto; gap: 8px; }
+.chart-plot-cell { position: relative; display: flex; flex-direction: column; justify-content: flex-end;
+  min-height: 0; border-bottom: 1px solid var(--line-strong); }
+.chart-plot-cell.grid { border-top: 1px solid var(--line-strong);   /* 顶边 = 轴上限, 让"轴 0–X"看得见 */
+  background-image: repeating-linear-gradient(to top, var(--line-strong) 0 1px, transparent 1px 25%); }
+/* 目标线有两种朝向, 都按比例尺定位:
+   ① 柱状图里是横线(bottom:100% = 轴上限); ② bullet 轨道里是**竖线**(left:88% = 目标位),
+     标签必须贴在刻度旁边 —— 贴轨道右端会被读成"刻度在最右边"(实测被视觉验收抓到) */
+.chart-plot-cell .chart-target { position: absolute; left: 0; right: 0; border-top: 1px dashed var(--line-strong); }
+.chart-track .chart-target { position: absolute; top: -5px; bottom: -5px; border-left: 2px dashed var(--fg-3); }
+.chart-track .chart-target span { position: absolute; top: -1.5em; right: 4px; white-space: nowrap;
+  font-size: var(--fs-tiny); color: var(--fg-3); }
+.chart-bar-wrap { position: relative; display: flex; flex-direction: column; justify-content: flex-end; min-height: 0; }
+.chart-bar-wrap .chart-val { position: absolute; bottom: 100%; left: 0; right: 0; text-align: center; margin-bottom: 6px; }
+.chart-bar-wrap .chart-bar { width: 100%; height: 100%; border-radius: var(--radius-sm) var(--radius-sm) 0 0; padding: 0; }
+.chart-col .chart-val { color: var(--fg); align-self: stretch; text-align: center; }  /* 与柱下轴标签同一轴线 */
+.chart-col .chart-lbl { flex: 0 0 auto; text-align: center; }
+/* 左侧刻度列: 与绘图区等高(底部留出轴标签那一行的高度), 让"轴 0–X%"在图上可核验 */
+.chart-ticks { display: flex; flex-direction: column; justify-content: space-between; align-items: flex-end;
+  height: calc(var(--cols-h, 380px) - 40px); font-size: var(--fs-tiny); color: var(--fg-3); padding-right: var(--sp-2); }
+.chart-target { position: absolute; left: 0; right: 0; border-top: 1px dashed var(--line-strong); }
+.chart-target span { position: absolute; right: 0; top: -1.4em; font-size: var(--fs-tiny); color: var(--fg-3); }
+.chart-cap { font-size: var(--fs-tiny); color: var(--fg-3); }`;
+
+// 判定落在承重的那条规则上: 没有 fx-sweep 的关键帧, 新配方就会静默失效
+export function hasChartKit(css) {
+  return /@keyframes\s+fx-sweep-kf\s*\{/.test(String(css ?? '')) && /@property\s+--pv\s*\{/.test(String(css ?? ''));
+}
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/check-slides.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/check-slides.mjs
new file mode 100644
index 00000000..766f6f5f
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/check-slides.mjs
@@ -0,0 +1,376 @@
+#!/usr/bin/env node
+// html2video-for-mcode · 渲染前静态检查: 把"画面静默出错但流水线报成功"这类问题挡在截图之前
+// 用法: node check-slides.mjs <项目目录> [--ids 01,02] [--quiet]
+//
+// 检查项(全部来自实测踩过的坑):
+//   1. 未定义的 CSS 变量 —— var(--x) 没有 fallback 且 tokens.css/内联样式都没定义 →
+//      整条声明计算为 none; 若同时用了 -webkit-text-fill-color:transparent, 文字会完全隐形
+//   2. 图片文件不存在 / 引用了 SVG —— SVG 在 file:// 实测能加载(400px), 但依赖外部资源、
+//      XML 有误、缺 width/height 或下载失败存成 HTML 时都会显示 broken 图标, 建议 inline
+//   3. 硬编码颜色(#hex / rgb() / hsl()) —— 换主题时会串色, 且说明配色没落到 token
+//   4. 外链资源(http/https 的 src/href) —— 离线沙箱会失败 + 引入 FOUT 风险
+//   5. data-stage 没配 fx-* 类 —— 元素会永远停在 opacity:0(.fx-stagger 也不兜底: 新模板的
+//      stagger 规则带 :not([data-stage]), 带 data-stage 就必须自己配 fx 类/自己管入场)
+//   5b. fx 类的关键帧不含 opacity —— 同上, 基础态 opacity:0 抬不回来, 元素永远隐形
+//      (只做 transform/描边的动画必须显式写 opacity:1; 2026-09-18 实测踩过)
+//   5c. class 用了 tokens.css 与本张 <style> 都没有的类 —— 元素静默无样式(.grid g4 / <p class="dim">
+//      那类), 提示级; 纯语义钩子可忽略
+//   5d. .fx-stagger 与 data-stage 同张共存 —— 旧 tokens.css 的 stagger nth-child 延迟特异度更高,
+//      会覆盖容器内带 data-stage 子元素的时刻(提前几秒冒头, 2026-09-18 实测踩过), 提示确认布局
+//   5e. 绝对定位元素落在字幕带(bottom < 168px@1080) —— 成片字幕会压住它, still 预览看不出
+//      (终态不一定显示字幕), 只有抽成片帧才见(2026-09-18 实测踩过, 最贵的一类)
+//   6. tokens.css 落后于技能当前版本(受管区 rev 不一致/缺失/重复/定界符被改坏/区外残留副本) ——
+//      **错误级, 阻断截图(退出 1)**: 不升级的话新版类(如 .chart-ticks)会静默无样式, 旧规则按层叠
+//      压过新版("改了 CSS 但项目里没生效"的根因, 见 css-kit.mjs)。注意本闸门**没有** --allow-stale-css,
+//      要放行只有两条路: 升级它, 或者不跑这条检查(要旧 CSS 出片请用 capture / build-video 的该旗标)
+//   7. 整片级领域自查 —— 命中多个财经/投研关键词却全片没有免责或出处行 → 提示(不是错误),
+//      提醒确认领域与免责口径(见 references/compliance.md); 用户明确不要免责时可忽略
+import fs from 'node:fs';
+import path from 'node:path';
+import { canonicalPath, expectedTokens, flagValue, inside, positionalDir, readTransition, safeId, safeRel, subBandTop } from './tools.mjs';
+import { kitStatuses } from './css-kit.mjs';
+import { MAX_SCAN_BYTES } from './limits.mjs';   // 对 tokens.css 与 HTML 都生效, 单一来源(见 limits.mjs)
+
+// tokens.css 里的 @keyframes 名 → 是否声明了 opacity(用来判断"动画能否把基础态抬回可见")
+function opacityAwareKeyframes(css) {
+  const map = new Map();
+  for (const m of css.matchAll(/@keyframes\s+([\w-]+)\s*\{/g)) {
+    // 从头扫到配平的右花括号, 拿到该关键帧块
+    let i = m.index + m[0].length - 1, depth = 0, end = i;
+    for (; end < css.length; end++) {
+      if (css[end] === '{') depth++;
+      else if (css[end] === '}') { depth--; if (depth === 0) break; }
+    }
+    map.set(m[1], /opacity\s*:/.test(css.slice(i, end)));
+  }
+  return map;
+}
+// .fx-x { animation: <kf-name> ... } → fx 类名到关键帧名的映射
+function fxClassKeyframes(css) {
+  const map = new Map();
+  for (const m of css.matchAll(/\.(fx-[\w-]+)\s*(?:,[^{]*)?\{([^}]*)\}/g)) {
+    const anim = /animation\s*:\s*([\w-]+)/.exec(m[2]);
+    if (anim) map.set(m[1], anim[1]);
+  }
+  return map;
+}
+
+const argv = process.argv.slice(2);
+const dir = positionalDir(argv);
+const QUIET = argv.includes('--quiet');
+const idsFilter = flagValue(argv, '--ids', '') ? flagValue(argv, '--ids', '').split(',').map(s => s.trim()) : null;
+
+const scriptPath = path.join(dir, 'script.json');
+if (!fs.existsSync(scriptPath)) { console.error(`✗ 找不到 ${scriptPath}`); process.exit(1); }
+const script = JSON.parse(fs.readFileSync(scriptPath, 'utf8'));
+const slidesDir = path.join(dir, 'slides');
+const tokensPath = path.join(slidesDir, 'tokens.css');
+if (!fs.existsSync(tokensPath)) { console.error(`✗ 找不到 ${tokensPath}`); process.exit(1); }
+const tokensRaw = fs.readFileSync(tokensPath, 'utf8');
+if (tokensRaw.length > MAX_SCAN_BYTES) { console.error(`✗ tokens.css 超过 ${MAX_SCAN_BYTES / 1e6}MB, 拒绝扫描(正常项目 ≈15KB)`); process.exit(1); }
+const tokensCss = tokensRaw.replace(/\/\*[\s\S]*?\*\//g, '');
+
+// tokens.css 里定义的变量
+const defined = new Set([...tokensCss.matchAll(/(--[\w-]+)\s*:/g)].map(m => m[1]));
+// tokens.css 选择器里出现过的类名(存在即可命中, 不分裸类/复合 —— 那交给 5c 的提示与人工判断)
+const definedClasses = new Set([...tokensCss.matchAll(/\.(-?[A-Za-z_][\w-]*)/g)].map(m => m[1]));
+// 不在 tokens.css / 本张 <style> 里、但合法的类: no-fx 是放映页开关, kit-sub* 是 capture 烧字幕时注入的
+const CLASS_ALLOW = new Set(['no-fx', 'kit-sub', 'kit-sub-2']);
+// fx 类 → 关键帧名, 以及关键帧是否声明 opacity(供 5b 检查: 动画必须能把基础态 opacity:0 抬回 1)
+const fxKeyframes = fxClassKeyframes(tokensCss);
+const kfOpacity = opacityAwareKeyframes(tokensCss);
+// 渲染管线/浏览器在运行时注入或框架自带的变量, 不需要在 tokens.css 里定义
+const RUNTIME_VARS = new Set([
+  '--t1', '--t2', '--t3', '--fx-delay', '--stagger-base',           // capture 注入 / 入场延迟
+  '--stage-w', '--stage-h', '--sub-scale',                          // capture 注入的画布与字幕
+  '--sub-bg', '--sub-fg', '--sub-ring',                             // 字幕钩子(有内置默认值)
+  '--img-ratio', '--img-pos',                                       // .img-frame 的局部变量(有默认值)
+  '--deck-scale', '--logo-size', '--logo-opacity', '--logo-inset-x', '--logo-inset-y', '--logo-print',
+]);
+
+const slides = script.slides.filter(s => !idsFilter || idsFilter.includes(s.id));
+let errors = 0, warns = 0;
+const report = [];
+const deckText = [];   // 整片文本(HTML 去标签 + 口播稿), 用于第 6 项领域自查
+
+for (const s of slides) {
+  const file = safeRel(slidesDir, s.html ?? `${safeId(s.id)}.html`, { where: `slides[${s.id}].html` });
+  if (!fs.existsSync(file)) {
+    // 未写的 slide 只是"还没做", 截图时本就会跳过 —— 记提示而不是错误, 避免把告警训成噪音
+    report.push({ id: s.id, level: 'warn', msg: `未写: ${path.relative(dir, file)}(截图时会跳过)` });
+    warns++;
+    continue;
+  }
+  const html = fs.readFileSync(file, 'utf8');
+  if (html.length > MAX_SCAN_BYTES) {
+    report.push({ id: s.id, level: 'error', msg: `HTML 有 ${Math.round(html.length / 1e4) / 100}MB, 超过 ${MAX_SCAN_BYTES / 1e6}MB 上限(正常 ≈10KB) — 拒绝扫描, 请检查文件是否被误写` });
+    errors++;
+    continue;
+  }
+  // 注释不是代码: 注释里的 `--x:` 曾经能顶替变量定义、注释里的 var() 又会变成"用到它"的
+  // 假需求(2026-10-10 审计 M3: `<!-- --c-fake: 1 -->` + `var(--c-fake)` 能让未定义检查全绿)。
+  // 未闭合的 `<!--` 也要兜住, 否则尾部整段注释仍会被扫。
+  const htmlCode = html.replace(/<!--[\s\S]*?-->/g, ' ').replace(/<!--[\s\S]*$/, ' ');
+  const localDefs = new Set([...htmlCode.matchAll(/(--[\w-]+)\s*:/g)].map(m => m[1]));
+  deckText.push({ id: s.id, text: html.replace(/<[^>]*>/g, ' ') });
+
+  // 本张的 CSS 区域(<style> 块 + style 属性), 供 fx 类与外链两处检查共用
+  const cssRegions = [
+    ...[...html.matchAll(/<style[^>]*>([\s\S]*?)<\/style>/gi)].map(m => m[1]),
+    ...[...html.matchAll(/(?<![-\w])style\s*=\s*(?:"([^"]*)"|'([^']*)')/gi)].map(m => m[1] ?? m[2] ?? ''),
+  ].join('\n');
+  const localFx = fxClassKeyframes(cssRegions);
+  const localKfOpacity = opacityAwareKeyframes(cssRegions);
+
+  // 1. 未定义变量(只报没有 fallback 的)
+  for (const m of htmlCode.matchAll(/var\(\s*(--[\w-]+)\s*(,)?/g)) {
+    const name = m[1], hasFallback = !!m[2];
+    if (hasFallback || defined.has(name) || localDefs.has(name) || RUNTIME_VARS.has(name)) continue;
+    report.push({ id: s.id, level: 'error', msg: `变量 ${name} 未定义且无 fallback → 该声明会失效(文字可能直接隐形)` });
+    errors++;
+  }
+
+  // 2. 图片: 文件必须存在; SVG 建议 inline。除词法判定外, 已存在的文件再按**真实路径**比一次 ——
+  //    纯词法 inside() 挡不住"slides/ 下的目录段是 junction/符号链接"这种二阶越界读(词法上
+  //    引用还在项目里, 真实路径已经在项目外)。2026-10-10 审计 L1: 旧注释声称防了二阶, 实际没防。
+  for (const m of html.matchAll(/<img\b[^>]*\bsrc\s*=\s*["']([^"']+)["']/g)) {
+    const src = m[1];
+    if (/^(https?:)?\/\//.test(src) || src.startsWith('data:')) continue; // 外链单独在下面报
+    const resolved = path.resolve(path.dirname(file), src);
+    if (path.isAbsolute(src) || !inside(dir, resolved)) {
+      report.push({ id: s.id, level: 'error', msg: `图片引用越出项目目录: ${src} — 素材必须先落 assets/ 再引用` });
+      errors++;
+      continue;
+    }
+    if (fs.existsSync(resolved) && !inside(canonicalPath(dir), canonicalPath(resolved))) {
+      report.push({ id: s.id, level: 'error', msg: `图片引用经链接指向项目外: ${src} — 真实路径已越出项目目录(素材必须先落 assets/ 再引用)` });
+      errors++;
+      continue;
+    }
+    if (!fs.existsSync(resolved)) {
+      report.push({ id: s.id, level: 'error', msg: `图片不存在: ${src}(会渲染成 broken 图标)` });
+      errors++;
+    } else if (/\.svg$/i.test(src)) {
+      report.push({ id: s.id, level: 'warn', msg: `引用了 SVG 文件: ${src} — file:// 下能加载, 但 SVG 若有 XML 错误/依赖外部资源/缺 width-height 就会 broken; 架构图类建议 inline 进 HTML` });
+      warns++;
+    }
+  }
+
+  // 3. 硬编码颜色(排除 tokens 引用本身)
+  const hexes = [...html.matchAll(/#[0-9a-fA-F]{3,8}\b/g)].map(m => m[0]);
+  const rgbs = [...html.matchAll(/\b(rgba?|hsla?)\(/g)].map(m => m[1]);
+  if (hexes.length || rgbs.length) {
+    report.push({ id: s.id, level: 'warn', msg: `硬编码颜色 ${hexes.length + rgbs.length} 处(${[...hexes.slice(0, 3), ...rgbs.slice(0, 2)].join(' ')}…) — 换主题会串色, 请改成 var(--...)` });
+    warns++;
+  }
+
+  // 4. 外链资源(2026-10-10 审计 M4 收紧)。旧规则只认 `src|href="https://…"` 的双引号形式,
+  //    于是大写 SRC=、未加引号、协议相对 //host/x.png、以及 CSS 里的 @import/url(...) 全部漏过。
+  //    只在**属性**与**CSS 区域**里找, 散文里的可见 URL 不算(幻灯片正文本来就常写链接)。
+  const isRemote = v => /^(?:[a-z][a-z0-9+.-]*:)?\/\//i.test(v) || /^file:/i.test(v);
+  for (const m of html.matchAll(/(?<![-\w])(?:src|href|poster|data-src|data-href|srcset)\s*=\s*(?:"([^"]*)"|'([^']*)'|([^\s"'<>`]+))/gi)) {
+    const raw = m[1] ?? m[2] ?? m[3] ?? '';
+    for (const piece of raw.split(/\s*,\s*/)) {
+      if (!piece || !isRemote(piece)) continue;
+      report.push({ id: s.id, level: 'error', msg: `外链资源: ${piece.slice(0, 60)} — 离线沙箱取不到, 且字体会 FOUT` });
+      errors++;
+      break;
+    }
+  }
+  for (const m of cssRegions.matchAll(/(?:url\(\s*['"]?|@import\s+['"])([^'")\s]+)/gi)) {
+    if (!isRemote(m[1])) continue;
+    report.push({ id: s.id, level: 'error', msg: `CSS 外链: ${m[1].slice(0, 60)}(@import / url()) — 离线沙箱取不到` });
+    errors++;
+  }
+
+  // 5. data-stage 没配 fx-* 类(容器/子元素都一样: stagger 规则带 :not([data-stage]), 不再兜底)
+  for (const m of html.matchAll(/<[^>]*\bdata-stage\s*=\s*(?:"[^"]*"|'[^']*'|[^\s"'<>`]+)[^>]*>/gi)) {
+    const tag = m[0];
+    const cls = /(?<![-\w])class\s*=\s*(?:"([^"]*)"|'([^']*)'|([^\s"'<>`]+))/i.exec(tag);
+    const list = cls ? (cls[1] ?? cls[2] ?? cls[3] ?? '').split(/\s+/).filter(Boolean) : [];
+    const fx = list.filter(x => x.startsWith('fx-'));
+    if (!fx.length) {
+      report.push({ id: s.id, level: 'warn', msg: `有 data-stage 但没有 fx-* 类: ${tag.slice(0, 70)}… — 元素会永远停在 opacity:0(.fx-stagger 不兜底带 data-stage 的元素; 给它配一个 fx-up/fx-fade, 或去掉 data-stage 让 stagger 管)` });
+      warns++;
+      continue;
+    }
+    // 只认"真有动画声明"的 fx 类(2026-10-10 审计 M2): class="fx-notreal" 这类只借前缀就能
+    // 过关, 元素其实不会动 —— 于是 data-stage 的入场时刻是空的。
+    const declared = fx.filter(c => c === 'fx-stagger' || fxKeyframes.has(c) || localFx.has(c));
+    if (!declared.length) {
+      report.push({ id: s.id, level: 'warn', msg: `fx 类没有对应动画声明: ${fx.map(c => '.' + c).join(' .')} — 元素不会动(data-stage 的入场时刻是空的); 补 @keyframes/animation, 或改用 tokens.css 里已有的 fx-* 类` });
+      warns++;
+    }
+  }
+
+  // 5b. fx 类的关键帧不改 opacity → 基础态 opacity:0 抬不回来, 元素永远隐形
+  for (const m of html.matchAll(/<[^>]*\bdata-stage\s*=\s*(?:"[^"]*"|'[^']*'|[^\s"'<>`]+)[^>]*>/gi)) {
+    const tag = m[0];
+    const cls = /(?<![-\w])class\s*=\s*(?:"([^"]*)"|'([^']*)'|([^\s"'<>`]+))/i.exec(tag);
+    if (!cls) continue;
+    const list = (cls[1] ?? cls[2] ?? cls[3] ?? '').split(/\s+/).filter(Boolean);
+    if (list.includes('fx-stagger')) continue;
+    for (const c of list.filter(x => x.startsWith('fx-'))) {
+      // tokens.css 与**本张 CSS** 都要查(旧实现只看 tokens.css, 本地自定义 fx 类整体漏检)
+      const kf = fxKeyframes.get(c) ?? localFx.get(c);
+      if (!kf) continue;
+      const aware = (kfOpacity.get(kf) ?? localKfOpacity.get(kf));
+      if (aware === false) {
+        report.push({ id: s.id, level: 'error', msg: `.${c} 的关键帧 ${kf} 没声明 opacity, 而 [data-stage] 基础态是 opacity:0 → 该元素入场后永远不可见; 在关键帧里补 opacity:1` });
+        errors++;
+      }
+    }
+  }
+
+  // 5c. 类名有据可查: class="..." 里的每个类, tokens.css / 本张 <style> / 白名单至少要有一处,
+  //     否则该元素静默无样式(.grid g4 / <p class="dim"> 那类坑; 提示级, 纯语义钩子可忽略)
+  const localCss = [...html.matchAll(/<style[^>]*>([\s\S]*?)<\/style>/gi)].map(m => m[1]).join('\n');
+  const localClasses = new Set([...localCss.matchAll(/\.(-?[A-Za-z_][\w-]*)/g)].map(m => m[1]));
+  const used = new Set();
+  for (const m of html.matchAll(/\bclass\s*=\s*["']([^"']*)["']/g)) {
+    for (const c of m[1].split(/\s+/)) if (c) used.add(c);
+  }
+  const missing = [...used].filter(c => !definedClasses.has(c) && !localClasses.has(c) && !CLASS_ALLOW.has(c));
+  if (missing.length) {
+    report.push({ id: s.id, level: 'warn', msg: `类在 tokens.css 与本张 <style> 里都无定义: .${missing.join(' .')} — 元素会静默无样式; 若只是无样式的语义钩子可忽略, 否则补规则或改用已有类(若缺的类来自图表/表格工具箱, 先跑 node scripts/init-project.mjs <项目> --check-css 看工具箱新旧)` });
+    warns++;
+  }
+
+  // 5d. .fx-stagger 与 data-stage 同张共存(2026-09-18 实测踩坑): 旧 tokens.css 的 stagger
+  //     nth-child 延迟特异度(0,2,0)高于 fx-* 类(0,1,0), 会覆盖容器内带 data-stage 子元素的
+  //     入场时刻(实测提前 3 秒冒头)。新模板已用 :not([data-stage]) 根治; 这里做布局确认提示
+  if (/\bfx-stagger\b/.test(html) && /\bdata-stage\s*=/.test(html)) {
+    report.push({ id: s.id, level: 'warn', msg: '本张同时用了 .fx-stagger 与 data-stage — 确认带 data-stage 的元素不在 .fx-stagger 容器内(旧 tokens.css 的 stagger 延迟会覆盖它的时刻, 元素提前冒头; capture --mode motion 打印的"动画窗"比预期短就是信号)。根治: 新模板的 stagger 规则带 :not([data-stage]); 老项目给 tokens.css 里 .fx-stagger > * 九条选择器补 :not([data-stage]), 或把带 stage 的块挪出容器' });
+    warns++;
+  }
+
+  // 5e. 绝对定位元素落在字幕带(2026-09-18 实测踩坑, 最贵的一类): 成片字幕占底部一条带, 图注/
+  //     落款写 bottom:96px 会被字幕压住 —— still 预览看不出(终态不一定显示字幕), 只有抽成片帧才见。
+  //     **提示级**(字幕之外的合法贴底元素可忽略; 与 SKILL.md 的措辞一致)。
+  //     2026-09-18 复查 D5 收紧两处误报: ① 前加边界, 别把 padding-bottom / margin-bottom 当 bottom;
+  //     ② 要求同一条规则里有 position: absolute|fixed —— 流内元素写 bottom 不产生位移, 报了是噪音。
+  //     2026-09-28 Mavis 交接: 常量 168 按单行几何写死 —— 单行实测 170px, 折行后 230px(62px 碰撞区
+  //     永远不报, 08/11 实际压住内容而闸门全绿)。现按 tools.subBandTop 实算: 有 text2 的张用双行档;
+  //     .kit-sub 修 width:max-content 后恒单行, 但闸门按最坏形状算(长句极端时仍可能折行)。
+  const bandTop = subBandTop(Number(script.height ?? 1080), Number(script.width ?? 1920),
+    { bilingual: (s.clauses ?? []).some(c => c.text2) });
+  const cssRules = [
+    ...[...html.matchAll(/<style[^>]*>([\s\S]*?)<\/style>/gi)]
+      .flatMap(m2 => [...m2[1].matchAll(/\{([^{}]*)\}/g)].map(m3 => m3[1])),   // style 块: 按规则体切(含 @media 内层)
+    ...[...html.matchAll(/style\s*=\s*"([^"]*)"/gi)].map(m2 => m2[1]),          // 行内 style 整体就是一条规则
+  ];
+  const inBand = [];
+  for (const rule of cssRules) {
+    if (!/position\s*:\s*(absolute|fixed)/i.test(rule)) continue;
+    for (const m2 of rule.matchAll(/(?<![-\w])bottom\s*:\s*([\d.]+)\s*px/g)) {
+      const v = parseFloat(m2[1]);
+      if (Number.isFinite(v) && v < bandTop) inBand.push(v);
+    }
+  }
+  if (inBand.length) {
+    report.push({ id: s.id, level: 'warn', msg: `绝对定位元素的 bottom:${inBand[0]}px 等 ${inBand.length} 处落在字幕带(底部 0–${bandTop}px @${script.height ?? 1080}) — 成片字幕会压住它, still 预览看不出; 图注/落款挪到 bottom ≥ ${bandTop}px 或放回流内(padding-bottom 会兜住), 验收用 grab-frames.mjs 抽成片实帧` });
+    warns++;
+  }
+
+  // 5e-bis. 流内贴底启发式(2026-09-28 Mavis 交接: 5e 只扫绝对定位, 流内贴底完全不在射程 ——
+  //     margin-top:auto / 前置 flex:1 空占位把内容顶到底部, 字幕带检查看不见, 静帧也看不见,
+  //     只有 grab-frames 抽成片帧才见)。启发式只提示"去复核", 不定罪(合法贴底元素存在)。
+  const flowBottom = /margin-top\s*:\s*auto/.test(html) || /flex\s*:\s*1[^;}]*;\s*\/\*?\s*占位/.test(html);
+  if (flowBottom) {
+    report.push({ id: s.id, level: 'warn', msg: `本张有流内贴底写法(margin-top:auto / flex:1 占位) — 字幕带闸门只扫绝对定位, 流内贴底是否被成片字幕压住, 用 grab-frames.mjs 抽成片实帧复核(still 预览看不出)` });
+    warns++;
+  }
+
+  // 5f. 柱状图结构(2026-09-22 域外 agent 实测踩坑): 百分比柱体(值基准高度)直接放 flex column 会被
+  //     flex-shrink 压扁 —— 实测 84%/72% 画成同高, 而截图阶段"柱子没出来"极易被当成"还没渲染"放过。
+  //     规范形态: .chart-col 的 grid 1fr 行(.chart-plot-cell)包住柱体(authoring.md 图表节)。
+  //     提示级; 三合取判据(柱状语义类 + <100% 高度 + 张内有 flex-direction:column 且全张无
+  //     chart-plot-cell) —— 非 flex column 的百分比高度是合法用法, 不该误报。
+  const barTags = [...html.matchAll(/<[a-z][a-z0-9]*\b[^>]*>/gi)].map(m2 => m2[0]).filter(t => {
+    const cls = /class\s*=\s*"([^"]*)"/i.exec(t)?.[1] ?? '';
+    const sty = /style\s*=\s*"([^"]*)"/i.exec(t)?.[1] ?? '';
+    const h = /\bheight\s*:\s*([\d.]+)\s*%/i.exec(sty);
+    return h && parseFloat(h[1]) < 100 && /(bar|pillar|plot|gauge|stack)/i.test(cls);
+  });
+  if (barTags.length && !/chart-plot-cell/.test(html) && /flex-direction\s*:\s*column/i.test(html)) {
+    const sample = (/class\s*=\s*"([^"]*)"/i.exec(barTags[0])?.[1] ?? '').split(/\s+/)[0];
+    report.push({ id: s.id, level: 'warn', msg: `百分比柱体 ${barTags.length} 处(如 .${sample})没包 .chart-plot-cell — 柱体直接放 flex column 会被 flex-shrink 压扁(实测 84%/72% 画成同高, "柱子没出来/高度对不上"多半是它): 按 authoring.md 图表节用 .chart-col > .chart-plot-cell.grid > .chart-bar-wrap{height:N%} 结构; 若这些百分比高度不在 flex column 里可忽略本条` });
+    warns++;
+  }
+}
+
+// 6. tokens.css 受管块落后——rev 与技能当前不一致 / 缺失 / 重复 / 坏定界 / 区外残留副本
+//    → **错误级, 阻断截图**: 源头模块改了、项目没升级时画面照出, 只是新类静默无样式、旧规则按层叠
+//    压过新版 —— 这正是"改了 CSS 但没生效"的根因(2026-09-18 定案)。判据与 capture / build-video
+//    入口的 requireFreshCss 完全一致(同一份 kitStatuses + 同一份期望内容), 这里是截图前的静态闸门。
+const cssNotes = [];
+for (const st of kitStatuses(tokensRaw, { expected: expectedTokens() })) {
+  if (st.status === 'ok') continue;
+  cssNotes.push(`tokens.css 的${st.label}: [${st.status}] ${st.detail} → node scripts/init-project.mjs <项目目录> --upgrade-css 原地更新受管区; 否则新版类会静默无样式`);
+  errors++;
+}
+
+// 7. 受监管领域自查(整片级): 像财经/投研内容却没见免责或出处标注 → 提示, 不是错误
+for (const s of slides) {
+  const cls = (s.clauses ?? []).map(c => `${c.text ?? ''} ${c.text2 ?? ''}`).join(' ');
+  if (cls.trim()) deckText.push({ id: s.id, text: cls });
+}
+const FIN_WORDS = ['投资建议', '股价', '涨跌', '涨停', '跌停', '营收', '净利润', '毛利率', '市值', '市盈率',
+  '招股', '财报', '年报', '季报', '基金', '收益率', '汇率', 'A股', '港股', '美股', '募资', '估值', '分红'];
+const hits = new Map();   // 词 → 首次出现的 slide id
+for (const { id, text } of deckText) {
+  for (const w of FIN_WORDS) if (!hits.has(w) && text.includes(w)) hits.set(w, id);
+}
+const hasDisclaimerLine = deckText.some(({ text }) => /不构成|仅供参考|风险提示|免责/.test(text));
+const deckNotes = [];
+if (hits.size >= 2 && !hasDisclaimerLine) {
+  const words = [...hits.keys()];
+  deckNotes.push(`像财经/投研内容(命中 ${words.slice(0, 5).join('、')}${words.length > 5 ? ' 等' : ''}, 见 ${hits.get(words[0])} 等张), 但整片没找到免责/出处行 — 若题材受监管(财经/医疗/法律/政策/营销宣称), 结尾补 .disclaimer 行、确认涨跌色(A股=红涨绿跌)、并给数字补口径+币种+时点, 见 references/compliance.md; 用户已明确说不要免责则忽略本条`);
+  warns++;
+}
+
+// 8. Gate 4 的预览参数(1.6.0): 切页方式与入场层级 —— 这两项正是"HTML 做完给用户预览模拟"时要问的,
+//    在这里打印出来, 让 agent 拿着实测数字去问, 而不是凭印象。
+const transition = script.transition ?? null;
+const transLabel = (() => {
+  if (transition == null) return '未设置(老项目: 成片按硬切出; 想改就在 script.json 写 transition)';
+  try {
+    return readTransition(transition).label;   // 与 build-video 同一解析(裸字符串/对象都一样, D6)
+  } catch (e) {
+    return `✗ ${e.message}`;
+  }
+})();
+const levelLine = slides
+  .map(s => {
+    if (!s.html) return null;
+    const f = safeRel(slidesDir, s.html, { where: `slides[${s.id}].html` });
+    if (!fs.existsSync(f)) return null;
+    const h = fs.readFileSync(f, 'utf8');
+    const stages = new Set([...h.matchAll(/\bdata-stage\s*=\s*["'](\d+)["']/g)].map(m => Number(m[1])));
+    return `${s.id}=${stages.size || 1} 级`;
+  })
+  .filter(Boolean)
+  .join(' · ');
+if (!QUIET) {
+  console.log(`\n预览参数(Gate 4 要问用户的两项): 转场 = ${transLabel}`);
+  console.log(`  入场层级 = ${levelLine}(>1 级 = 逐级揭示; 用户要"大标题与子信息一次性出现"时, 把该张各级 data-stage 收敛为同一级再重跑 capture)`);
+}
+
+if (!QUIET) {
+  const bySlide = new Map();
+  for (const r of report) { if (!bySlide.has(r.id)) bySlide.set(r.id, []); bySlide.get(r.id).push(r); }
+  for (const s of slides) {
+    const rs = bySlide.get(s.id);
+    if (!rs) { console.log(`✓ ${s.id}`); continue; }
+    console.log(`${rs.some(r => r.level === 'error') ? '✗' : '⚠'} ${s.id}`);
+    for (const r of rs) console.log(`    ${r.level === 'error' ? '✗' : '⚠'} ${r.msg}`);
+  }
+  for (const n of cssNotes) console.log(`✗ tokens.css: ${n}`);
+  for (const n of deckNotes) console.log(`⚠ 整片: ${n}`);
+}
+console.log(`\n静态检查: ${slides.length} 张 · ✗ ${errors} 项错误 / ⚠ ${warns} 项提示`);
+if (errors) {
+  console.error('错误项会让画面对但"看不见/缺内容", 或用的是旧 CSS: 未定义变量补定义或加 fallback; 图片补齐或改 inline SVG; 外链改本地; 受管块落后按上面的 --upgrade-css 命令更新。');
+  process.exit(1);
+}
+console.log('可以进入截图(无阻塞项)✓');
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/check-theme.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/check-theme.mjs
new file mode 100644
index 00000000..107a5fee
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/check-theme.mjs
@@ -0,0 +1,93 @@
+#!/usr/bin/env node
+// html2video-for-mcode · 主题对比度校验(WCAG 2.1)
+// 用法: node check-theme.mjs <项目目录|tokens.css 路径>
+// 判据: accent-ink/accent ≥3:1(强调色只用于大字与图形), fg/bg ≥4.5:1,
+//       muted/bg ≥3:1(次级文本 ≥24px 属大字), 字幕 fg 对"胶囊合成到背景后"的颜色 ≥4.5:1。
+// 有 ✗ 时退出码 1 —— 新增或修改主题必须过这道闸。
+import fs from 'node:fs';
+import path from 'node:path';
+
+const arg = process.argv.slice(2).find(a => !a.startsWith('--')) ?? '.';
+const file = arg.endsWith('.css') ? path.resolve(arg) : path.join(path.resolve(arg), 'slides', 'tokens.css');
+if (!fs.existsSync(file)) { console.error(`✗ 找不到 ${file}`); process.exit(1); }
+const css = fs.readFileSync(file, 'utf8').replace(/\/\*[\s\S]*?\*\//g, ''); // 去注释, 避免注释里的色值干扰
+
+function parseBlock(selector) {
+  const esc = selector.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+  const re = new RegExp(esc + '\\s*\\{([^}]*)\\}', 'g');
+  const out = {};
+  let m, found = false;
+  // 收集所有同名块并按顺序合并: 项目在 tokens.css 末尾追加的品牌色覆写要能覆盖主题自带值
+  while ((m = re.exec(css))) {
+    found = true;
+    for (const decl of m[1].split(';')) {
+      const i = decl.indexOf(':');
+      if (i < 0) continue;
+      const k = decl.slice(0, i).trim();
+      if (k.startsWith('--')) out[k] = decl.slice(i + 1).trim();
+    }
+  }
+  return found ? out : null;
+}
+
+function color(v) {
+  if (!v) return null;
+  v = v.trim();
+  let m = v.match(/^#([0-9a-f]{3})$/i);
+  if (m) return [17 * parseInt(m[1][0], 16), 17 * parseInt(m[1][1], 16), 17 * parseInt(m[1][2], 16), 1];
+  m = v.match(/^#([0-9a-f]{6})/i);
+  if (m) return [parseInt(m[1].slice(0, 2), 16), parseInt(m[1].slice(2, 4), 16), parseInt(m[1].slice(4, 6), 16), 1];
+  m = v.match(/^rgba?\(\s*([\d.]+)\s*,\s*([\d.]+)\s*,\s*([\d.]+)\s*(?:,\s*([\d.]+)\s*)?\)$/i);
+  if (m) return [+m[1], +m[2], +m[3], m[4] === undefined ? 1 : +m[4]];
+  return null;
+}
+const lum = c => {
+  const f = x => { x /= 255; return x <= 0.03928 ? x / 12.92 : Math.pow((x + 0.055) / 1.055, 2.4); };
+  return 0.2126 * f(c[0]) + 0.7152 * f(c[1]) + 0.0722 * f(c[2]);
+};
+const ratio = (a, b) => {
+  const l1 = lum(a), l2 = lum(b);
+  const hi = Math.max(l1, l2), lo = Math.min(l1, l2);
+  return (hi + 0.05) / (lo + 0.05);
+};
+const blend = (fg, bg) => (fg[3] >= 1 ? fg : [0, 1, 2].map(i => fg[i] * fg[3] + bg[i] * (1 - fg[3])).concat(1));
+
+const base = parseBlock(':root');
+if (!base) { console.error('✗ tokens.css 里找不到 :root 块'); process.exit(1); }
+const themeNames = [...new Set(['a', ...[...css.matchAll(/\[data-theme="([^"]+)"\]\s*\{/g)].map(m => m[1])])];
+
+const CHECKS = [
+  { label: 'accent-ink / accent', fg: '--accent-ink', bg: '--accent', err: 3.0, warn: 4.5, note: '强调色上的文字(仅大字/图形)' },
+  { label: 'fg / bg', fg: '--fg', bg: '--bg', err: 4.5, warn: 7.0, note: '正文' },
+  { label: 'muted / bg', fg: '--muted', bg: '--bg', err: 3.0, warn: 4.5, note: '次级文本(≥24px)' },
+  { label: 'fg-3 / bg', fg: '--fg-3', bg: '--bg', err: 2.0, warn: 3.0, note: '三级文本/图注' },
+  { label: 'sub-fg / 字幕底', fg: '--sub-fg', bg: '--sub-bg', err: 4.5, warn: 7.0, note: '字幕(胶囊合成到背景后)', composite: true },
+];
+
+let fail = 0, warnCount = 0;
+console.log(`校验 ${path.relative(process.cwd(), file)} · ${themeNames.length} 套主题\n`);
+for (const name of themeNames) {
+  // 主题 a 的底来自 :root, 但项目仍可在 tokens.css 里追加 [data-theme="a"] 覆写品牌色 —— 也要合并
+  const ov = parseBlock(`[data-theme="${name}"]`) ?? {};
+  const t = { ...base, ...ov };
+  const bg = color(t['--bg']) ?? color(base['--bg']);
+  const rows = [];
+  for (const c of CHECKS) {
+    const fg = color(t[c.fg]), rawBg = color(t[c.bg]);
+    if (!fg || !rawBg) { rows.push(`  ${c.label.padEnd(18)} —(缺令牌)`); continue; }
+    const cmpBg = c.composite ? blend(rawBg, bg) : rawBg;
+    const r = ratio(fg, cmpBg);
+    const mark = r < c.err ? '✗' : r < c.warn ? '⚠' : '✓';
+    if (mark === '✗') fail++;
+    if (mark === '⚠') warnCount++;
+    rows.push(`  ${mark} ${c.label.padEnd(18)} ${r.toFixed(2)}:1   ${c.note}`);
+  }
+  console.log(`── ${name} ${'─'.repeat(Math.max(0, 34 - name.length))}`);
+  console.log(rows.join('\n'));
+}
+console.log(`\n✗ ${fail} 项不达标, ⚠ ${warnCount} 项偏弱。`);
+if (fail) {
+  console.error('不达标的项: accent 上的文字请改用大字/图形, 或调整 --accent-ink; 字幕对比不足请加深 --sub-bg(深色主题可加 --sub-ring 做视觉分隔)。');
+  process.exit(1);
+}
+console.log('全部通过 ✓');
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/check-timing.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/check-timing.mjs
new file mode 100644
index 00000000..54b63717
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/check-timing.mjs
@@ -0,0 +1,109 @@
+#!/usr/bin/env node
+// html2video-for-mcode · 对时校验: 用静音检测实测每句口播的真实开口时刻, 与 plan-timings 的字数估算对比。
+// 用法: node check-timing.mjs <项目目录> [--calibrate]
+//   默认只输出对比表; --calibrate 把实测边界写回 build/timings.json 的 clauses[].start 与 stages{}。
+// 原理: TTS 在句间(。!?;)会留 0.2s+ 的停顿 → ffmpeg silencedetect 找句间静音 → 边界后即下一句开口。
+import fs from 'node:fs';
+import path from 'node:path';
+import { spawnSync } from 'node:child_process';
+import { matchBoundaries } from './tools.mjs';
+import { positionalDir, requireTool, safeId, safeOut, safeRel, validateScriptPaths, validateTimingsIds } from './tools.mjs';
+
+const argv = process.argv.slice(2);
+const dir = positionalDir(argv);
+const CALIBRATE = argv.includes('--calibrate');
+const LEAD = 0.2;
+const NOISE = '-38dB', MIN_SIL = 0.18; // 句间停顿检测阈值
+
+process.env.KIT_PROJECT_DIR = dir;
+const FFMPEG = requireTool('ffmpeg', dir);
+
+const timingsPath = safeOut(dir, 'build', 'timings.json');   // 写回走收监(读不到时下面会报缺文件)
+if (!fs.existsSync(timingsPath)) { console.error('✗ 缺 build/timings.json — 先运行 plan-timings.mjs'); process.exit(1); }
+const timings = JSON.parse(fs.readFileSync(timingsPath, 'utf8'));
+validateTimingsIds(timings);
+const script = JSON.parse(fs.readFileSync(path.join(dir, 'script.json'), 'utf8'));
+validateScriptPaths(script, dir);
+
+function detectGaps(file, tts) {
+  const r = spawnSync(FFMPEG, ['-i', file, '-af', `silencedetect=noise=${NOISE}:d=${MIN_SIL}`, '-f', 'null', '-'], { encoding: 'utf8', windowsHide: true });
+  const text = (r.stderr || '') + (r.stdout || '');
+  const starts = [...text.matchAll(/silence_start:\s*([\d.]+)/g)].map(m => parseFloat(m[1]));
+  const ends = [...text.matchAll(/silence_end:\s*([\d.]+)/g)].map(m => parseFloat(m[1]));
+  const gaps = [];
+  for (let i = 0; i < starts.length; i++) {
+    const s = starts[i], e = ends[i] ?? tts;
+    if (s > 0.15 && e < tts - 0.15) gaps.push({ start: s, end: e });
+  }
+  return gaps;
+}
+
+// 句间边界匹配: 静音段里混着句内逗号停顿, 不能假设段数=句数-1。
+// 实现已挪 tools.matchBoundaries(单一来源, 发音形态口径 say ?? text —— 第 25 轮: 按 text
+// 对齐时 say/text 标点不一致会精确模式错位, --calibrate 写回错的 start)。
+
+const report = [];
+let calibratable = 0;
+for (const t of timings.slides) {
+  if (!t.clauses || t.clauses.length < 2) continue;
+  const tid = safeId(t.id);
+  const audioPath = safeRel(path.join(dir, 'audio'), script.slides.find(s => s.id === tid)?.audio ?? `${tid}.mp3`, { where: `slides[${tid}].audio` });
+  if (!fs.existsSync(audioPath)) { console.warn(`- 跳过 ${tid}: 缺音频`); continue; }
+  const gaps = detectGaps(audioPath, t.tts);
+  const { method, meas } = matchBoundaries(gaps, t.clauses);
+  // complete = 这份实测**整份可信**(exact/nearest 两条路都是全命中才返回); sparse 有缺口,
+  // 只展示不写回 —— 半份实测混着估算写进 timings.json 会让下游拿到自相矛盾的时刻。
+  const complete = (method === 'exact' || method === 'nearest') && meas.length > 0 && meas.every(v => v != null);
+  if (complete) calibratable++;
+
+  const rowsForSlide = t.clauses.slice(1).map((c, k) => ({
+    clause: k + 2, est: +c.start.toFixed(2),
+    meas: meas[k] == null ? null : +meas[k].toFixed(2),
+    diff: meas[k] == null ? null : +(meas[k] - c.start).toFixed(2),
+  }));
+  report.push({ id: t.id, method, gaps: gaps.length, need: t.clauses.length - 1, rows: rowsForSlide });
+
+  if (CALIBRATE && complete) {
+    // 实测边界写回: clauses[].start 与 stage 入场时刻(stage s = 最早属于 s 的句的实测开口 − LEAD)
+    t.clauses.forEach((c, i) => { if (i > 0 && meas[i - 1] != null) c.start = Math.round(meas[i - 1] * 1000) / 1000; });
+    t.clauses.forEach((c, i) => { c.dur = Math.round(((t.clauses[i + 1]?.start ?? t.tts) - c.start) * 1000) / 1000; });
+    const stageTime = {};
+    for (const c of t.clauses) {
+      if (c.stage == null) continue;
+      const v = Math.max(0, c.start - LEAD);
+      stageTime[c.stage] = c.stage in stageTime ? Math.min(stageTime[c.stage], v) : v;
+    }
+    const slide = script.slides.find(s => s.id === t.id);
+    Object.assign(stageTime, slide?.stageTimes ?? {}); // 作者显式覆盖仍然优先
+    t.stages = stageTime;
+  }
+}
+
+if (!report.length) { console.log('所有 slide 都只有单句, 无句间边界可校验。'); process.exit(0); }
+
+let drift = 0, driftN = 0, worst = null;
+const methodLabel = { exact: '精确(标点对齐)', nearest: '最近邻(全命中)', sparse: '静音段不足(未写回)', none: '-' };
+console.log('\n对时对比(估算 vs 静音检测实测):');
+for (const r of report) {
+  console.log(`  ${r.id}: 静音段 ${r.gaps} / 句边界 ${r.need} · 匹配 ${methodLabel[r.method]}`);
+  for (const row of r.rows) {
+    if (row.diff != null) { drift += Math.abs(row.diff); driftN++; if (!worst || Math.abs(row.diff) > Math.abs(worst.diff)) worst = { ...row, id: r.id }; }
+    console.log(`    第${row.clause}句开口 估算 ${row.est}s | 实测 ${row.meas ?? '(未测)'}s | 差 ${row.diff ?? '-'}s`);
+  }
+}
+if (driftN) {
+  const avg = (drift / driftN).toFixed(2);
+  console.log(`\n平均偏差 ${avg}s, 最大 ${worst ? worst.id + ' 第' + worst.clause + '句 ' + worst.diff + 's' : '-'}`);
+  if (!CALIBRATE) console.log('偏差普遍 >0.3s 时, 用 --calibrate 按实测校准后重跑 capture/build-video。');
+  else console.log('已按实测校准(仅精确/最近邻全命中的 slide; 有缺口的只展示不写回)。');
+}
+if (CALIBRATE) {
+  // 一张都没有全命中时不要动文件: 内容没变也该保持字节不变(重排 + 补行尾会让 diff 与
+  // "文件有没有被改"这件事说谎; 2026-09-18 复查 D1)
+  if (!calibratable) {
+    console.log(`没有一张 slide 的句边界被全部实测到, 未改动 ${timingsPath}(见上面的匹配方法与"(未测)"项)`);
+    process.exit(0);
+  }
+  fs.writeFileSync(timingsPath, JSON.stringify(timings, null, 2) + '\n');
+  console.log(`✓ 已写回 ${timingsPath} (校准 ${calibratable}/${report.length} 张, 仅全部句边界都有实测的 slide; 方法 exact/nearest 才算全命中)。下一步: 删 build/frames/ 后重跑 capture --mode motion 与 build-video。`);
+}
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/css-kit.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/css-kit.mjs
new file mode 100644
index 00000000..cfcfddd5
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/css-kit.mjs
@@ -0,0 +1,473 @@
+// html2video-for-mcode · CSS 工具箱的受管块机制(单一来源 → 项目 tokens.css 的可检测传播)。
+//
+// 演进史(每次都是实测踩出来的):
+// 2026-09-18 一轮: 旧 --upgrade-css 用"存在性探针"判定, 项目补过一次就永远报"无需升级"; 追加式升级
+//   压掉项目覆写并留重复段 → 改成"定界块 + 内容 rev + 原位替换"。
+// 2026-09-18 二轮(外部评审 + 复查实测): 机制仍有三条"报 OK 却仍用旧 CSS"的静默路径 ——
+//   ① 同一 kit 出现两个块、后一个陈旧: 旧实现只用 exec(非 g)看**第一个**块, 替换也换第一个,
+//      块后的陈旧副本按层叠继续生效;
+//   ② 块**之外**残留一份陈旧副本(旧版追加升级/手粘留下的): 全文件 never 被扫, 直接报"无需升级";
+//   ③ 收尾定界符被改坏: BLOCK_RE 认不出块, 退到 `src.includes(整段 CSS)` 判定 → 报 ok,
+//      升级还会再吐一个开标记。
+//   三条都源自同一个设计缺口: **判定只看"我认得的那块", 不看"文件里还有没有别的副本"**。
+// 本版据此重写:
+//   · 定位改为 findAllBlocks(全局), 同 id 多块 → 一律判 duplicate 并在升级时归并为一;
+//   · 判定加"受管块之外"扫描: 块外命中工具箱原文/承重规则 → legacy-outside(留档并清理原文副本);
+//   · 开标记没有配对收尾 → broken(不再因为内容字符串命中就算 ok);
+//   · 整个生成物(模板主体 + 三个工具箱子块)包进**一个 tokens 受管区**: 主体改动同样可检测、可原地
+//     替换 —— 项目自己的规则写在 tokens 区之后, 升级一概不动(2026-09-18 用户决定: 旧产物不复用,
+//     所以整段受管化优于"只检测不替换")。
+//
+// 边界: 受管区之内的一切由技能负责, 区之外一律不碰。项目若在区内改过内容, 升级会覆盖(会在 actions 里说明)。
+
+import { createHash } from 'node:crypto';
+import { NOFX_CSS, hasNofxRules } from './nofx-css.mjs';
+import { CHART_CSS, hasChartKit } from './chart-css.mjs';
+import { TABLE_CSS, hasTableKit } from './table-css.mjs';
+
+// 扫描上限的单一来源在 limits.mjs(它对 tokens.css 与 slide HTML 都生效, 不该住在这个 CSS 模块里);
+// 这里再导出一次只为保持既有调用点的 import 路径不变。
+export { MAX_SCAN_BYTES } from './limits.mjs';
+
+export const KITS = [
+  { id: 'nofx', label: 'no-fx 规则', css: NOFX_CSS, probe: hasNofxRules, sig: /\.no-fx\b/ },
+  { id: 'chart', label: '图表工具箱', css: CHART_CSS, probe: hasChartKit, sig: /\.(chart-|chart\b)/ },
+  { id: 'table', label: '表格原语', css: TABLE_CSS, probe: hasTableKit, sig: /\.(tbl|matrix|rank|kv)\b/ },
+];
+
+// 整段受管区的 id: 包住模板主体(主题/字号/fx 工具类)+ 内部三个工具箱子块
+export const TOKENS_ID = 'tokens';
+
+/** 所有受管块 id(单一来源): 判定/升级/测试的正则与循环都从这里取, 加一个工具箱不用改五处 */
+export const BLOCK_IDS = [TOKENS_ID, ...KITS.map(k => k.id)];
+
+
+// rev 对"块里实际会写进项目的 CSS"求哈希: 三个模块的文本逐字决定 rev。
+export const KIT_REV = createHash('sha256')
+  .update(KITS.map(k => k.css).join('\n').replace(/\r\n/g, '\n'))
+  .digest('hex');
+
+export const openTag = (id, rev = KIT_REV) => `/* >>> html2video:${id} rev=${rev} >>> */`;
+export const closeTag = id => `/* <<< html2video:${id} <<< */`;
+export function wrapKit(id, css, rev = KIT_REV) {
+  return `${openTag(id, rev)}\n${String(css).trim()}\n${closeTag(id)}`;
+}
+/** 把整段生成的 tokens.css 包进受管区(内部照旧含三个工具箱子块) */
+export function wrapTokens(text, rev) {
+  return `${openTag(TOKENS_ID, rev)}\n${String(text).trim()}\n${closeTag(TOKENS_ID)}`;
+}
+
+const BLOCK_RE_G = id => new RegExp(
+  `/\\* >>> html2video:${id} rev=([0-9a-f]{8,64}) >>> \\*/([\\s\\S]*?)/\\* <<< html2video:${id} <<< \\*/`, 'g');
+const OPEN_RE_G = id => new RegExp(`/\\* >>> html2video:${id} rev=([0-9a-f]{8,64}) >>> \\*/`, 'g');
+
+/** 所有配对完整的受管块: [{ start, end, rev, body }] —— 用全局匹配, 不再只看第一个 */
+export function findAllBlocks(css, id) {
+  const src = String(css ?? '');
+  const out = [];
+  for (const m of src.matchAll(BLOCK_RE_G(id))) {
+    out.push({ start: m.index, end: m.index + m[0].length, rev: m[1], body: m[2] });
+  }
+  return out;
+}
+
+/** 第一个受管块(兼容旧调用) */
+export function findBlock(css, id) {
+  return findAllBlocks(css, id)[0] ?? null;
+}
+
+/** 开标记总数(含没有配对收尾的孤儿): 用来发现"定界符被改坏" */
+export function countOpenTags(css, id) {
+  return [...String(css ?? '').matchAll(OPEN_RE_G(id))].length;
+}
+
+// ── 受管区间的取法 ───────────────────────────────────────────────
+// 老格式的主体里**嵌着**三个工具箱子块(新版是 tokens 区套住它们), 所以区间天然会嵌套/相交。
+// 第一版拿"按 start 降序逐个裁掉"的写法处理, 嵌套时外层区间的 end 已经被内层删除顶偏了 ——
+// 实测 `outsideRegions` 会把受管区之后的项目规则一起吃掉(out.len 只剩 1 个字符)。
+// 所以统一走"区间并集 + 等长掩码": 一次成型的空格掩码, 偏移照旧对得回原文。
+function regionRanges(src) {
+  const raw = [
+    ...findAllBlocks(src, TOKENS_ID).map(b => [b.start, b.end]),
+    ...KITS.flatMap(k => findAllBlocks(src, k.id)).map(b => [b.start, b.end]),
+  ].sort((a, b) => a[0] - b[0]);
+  const union = [];
+  for (const [s, e] of raw) {
+    const last = union[union.length - 1];
+    if (last && s <= last[1]) last[1] = Math.max(last[1], e);
+    else union.push([s, e]);
+  }
+  return union;
+}
+function maskedAll(src, ranges) {
+  let masked = '', cursor = 0;
+  for (const [s, e] of ranges) { masked += src.slice(cursor, s) + ' '.repeat(e - s); cursor = e; }
+  return masked + src.slice(cursor);
+}
+
+/** 受管区之外的文本(等长掩码, 区内一律变空格; 判定"块外是否还有残留副本"用) */
+export function outsideRegions(css) {
+  const src = String(css ?? '').replace(/\r\n/g, '\n');
+  return maskedAll(src, regionRanges(src));
+}
+
+// ── 区外残留副本的定位 ───────────────────────────────────────────
+// 光比"整段原文"不够: 旧版追加升级/手粘的副本常被改过一两处(比如行高 72→60), 原文比对失配,
+// 但它在层叠上**后出现就赢** —— 于是判定必须按选择器找"区外还有属于我们的规则":
+//   · 工具箱规则: 按选择器签名(.no-fx / .chart-* / .tbl·.matrix·.rank·.kv)归到对应 kit;
+//   · **主体规则**(2026-09-18 复查 A7 补): 选择器出现在当前生成物里的规则(如 :root / .fx-* /
+//     [data-stage] / .stage)归到 TOKENS_ID —— 否则"受管区后面粘着一段旧主体"这种布局会被判 ok,
+//     而那段旧规则按层叠压过受管区(实测: --check-css 报"已是最新", 画面用的是 --fs-h1: 60px)。
+//     A7 这类片段必然"与当前版不同"(否则它就是当前版), 所以主体段不能像工具箱那样要求逐字命中,
+//     改按体量判: ≥400 字符或 ≥3 条 → 作为残留报出来(报而不删, 由人决定它是旧残留还是有意覆写)。
+export function legacyOutsideBlocks(css, { bodyText = null } = {}) {
+  const src = String(css ?? '').replace(/\r\n/g, '\n');
+  // 掩码必须与 src **等长**, 否则算出的偏移对不回原文(第一版就错在这里: 拿 outsideRegions 的
+  // 结果当底, 长度已经变短, 于是位置全偏、一条也扫不出来)。
+  const masked = maskedAll(src, regionRanges(src));
+  const body = bodyText ? String(bodyText).replace(/\r\n/g, '\n') : null;
+  const bodyNorm = body ? body.replace(/\s+/g, ' ') : null;
+  const inBody = sel => {
+    if (!bodyNorm) return false;
+    const s = sel.replace(/\s+/g, ' ').trim();
+    return s.length > 0 && s.length < 400 && bodyNorm.includes(s);
+  };
+  // 区外**所有**规则体(选择器 + 花括号配对的那一段), 再按归属筛
+  const findRuleEnd = from => {
+    let depth = 0;
+    for (let i = from; i < masked.length; i++) {
+      if (masked[i] === '{') depth++;
+      else if (masked[i] === '}') { depth--; if (depth === 0) return i + 1; }
+    }
+    return -1;
+  };
+  const out = [];
+  let acceptedEnd = -1;                       // 已收下的规则覆盖到哪 —— 跳过它内部的嵌套 `{`
+  for (const m of masked.matchAll(/[^\n{};]*\{/g)) {
+    if (m.index < acceptedEnd) continue;      // @keyframes 里的 from{}/to{} 不再算一条: 两条片段重叠时
+                                              // 按偏移删除会错位, 连带把后面的受管区删掉(实测踩到过)
+    const end = findRuleEnd(m.index + m[0].length - 1);
+    if (end < 0) continue;
+    const sel = m[0].slice(0, -1);
+    const kit = KITS.find(k => k.sig.test(sel));
+    const id = kit ? kit.id : inBody(sel) ? TOKENS_ID : null;
+    if (!id) continue;                       // 不是我们的规则(用户自己的选择器) → 不归我们管
+    const label = kit ? kit.label : 'tokens 主体(主题/字号/fx/工具箱)';
+    out.push({ id, label, start: m.index, end, text: src.slice(m.index, end) });
+    acceptedEnd = end;
+  }
+  // 同一归属的相邻规则合并成一段 —— "相邻"要允许**注释**夹在中间(模块原文的规则之间就有
+  // 段落注释; 只认空白的话, 一份整旧副本会被拆成七八段, 单条的那些反而被判成"用户覆写"留着,
+  // 继续按层叠压过受管区 —— 实测踩到过)。合并后按整段长度/条数才判得准; 每段带上 rules,
+  // 清理时按条判定(逐字来自模块原文的才删)。
+  const COMMENT_GAP = /^(?:\s|\/\*[\s\S]*?\*\/)*$/;
+  out.sort((a, b) => a.start - b.start);
+  const merged = [];
+  for (const b of out) {
+    const last = merged[merged.length - 1];
+    if (last && last.id === b.id && COMMENT_GAP.test(src.slice(last.end, b.start))) {
+      last.end = b.end; last.text = src.slice(last.start, b.end);
+      last.rules.push({ start: b.start, end: b.end, text: b.text });
+    } else merged.push({ ...b, rules: [{ start: b.start, end: b.end, text: b.text }] });
+  }
+  // 分类: 大段/多规则 **且含原文的规则** = 旧版整份副本(升级清理);
+  // 单条小规则、或一条原文都不含的多条规则(自己写的样式段) = 覆写(保留, 只提示)。
+  // 加"含原文"这一条是因为: 光看长度会把"用户自己写的三条长规则"当副本删掉 ——
+  // 而一份真副本(哪怕被改过一两处)必然含有多条逐字相同的原文规则。
+  // 主体段(TOKENS_ID)例外: 按体量判 copy(见上面的说明), 但删除照旧只删逐字命中当前版的规则。
+  const norm = s => s.replace(/\s+/g, ' ');
+  const refOf = b => norm(b.id === TOKENS_ID ? String(bodyText ?? '') : KITS.find(k => k.id === b.id).css);
+  return merged.map(b => {
+    const mod = refOf(b);
+    const verbatim = b.rules.filter(r => mod.includes(norm(r.text).trim())).length;
+    const big = b.text.length >= 400 || b.rules.length >= 3;
+    const looksLikeCopy = verbatim >= 2 || verbatim / b.rules.length >= 0.5;
+    const kind = b.id === TOKENS_ID ? (big ? 'copy' : 'override') : (big && looksLikeCopy ? 'copy' : 'override');
+    return { ...b, verbatim, kind };
+  });
+}
+
+// ── 状态判定 ─────────────────────────────────────────────────────
+// 每个工具箱(以及整段 tokens 区)的状态:
+//   ok              受管块存在、rev 与内容都与技能当前一致
+//   stale           受管块存在但 rev 落后或内容被改过 → 升级时原地替换
+//   duplicate       同一 id 有多个受管块 → 升级时归并为一(块外的旧副本按层叠会赢, 必须报出来)
+//   broken          开标记没有配对收尾(定界符被改坏) → 升级时清掉孤儿标记并重建
+//   legacy-outside  受管块之外还留着同一份内容的副本 → 升级时删掉副本
+//   missing         既无受管块也无当前内容
+export function kitStatuses(css, { rev = KIT_REV, expected = null } = {}) {
+  const src = String(css ?? '').replace(/\r\n/g, '\n');   // 行尾归一: rev 是按 LF 算的(git autocrlf 检出会带 CRLF)
+  const legacyBlocks = legacyOutsideBlocks(src, { bodyText: expected?.text ?? null });   // 区外的同类规则(含被改过的副本)
+  // 覆写在受管区之前还是之后, 决定了它到底压不压得住受管区 —— 复查 A6 实测: 老格式文件升级时
+  // 生成物被追加到**文件末尾**, 原来在文件尾的项目覆写就变成"在受管区之前", 于是反被覆盖。
+  // 位置算得出来, 就别再笼统写"后出现者优先"(那句话在 A6 布局下是假的)。
+  const regionStart = findAllBlocks(src, TOKENS_ID)[0]?.start ?? Number.POSITIVE_INFINITY;
+  const overrideNote = (list, label) => {
+    const before = list.filter(b => b.start < regionStart).reduce((a, b) => a + b.rules.length, 0);
+    const total = list.reduce((a, b) => a + b.rules.length, 0);
+    const after = total - before;
+    const parts = [];
+    if (before) parts.push(`${before} 条在受管区**之前**(会被受管区覆盖 —— 想要它们生效, 挪到受管区之后)`);
+    if (after) parts.push(`${after} 条在受管区之后(后出现者优先, 继续生效)`);
+    return `受管区之外还有${label} ${total} 条: ${parts.join('; ')}`;
+  };
+  const statuses = KITS.map(k => {
+    const blocks = findAllBlocks(src, k.id);
+    const openCount = countOpenTags(src, k.id);
+    if (openCount > blocks.length) {
+      return { id: k.id, label: k.label, status: 'broken', detail: `有 ${openCount} 个开标记但只有 ${blocks.length} 个配对收尾(定界符被改坏)` };
+    }
+    if (blocks.length > 1) {
+      return { id: k.id, label: k.label, status: 'duplicate', detail: `同一工具箱出现 ${blocks.length} 个受管块(块外的旧副本按层叠会覆盖新块)` };
+    }
+    const block = blocks[0];
+    const clean = block && block.rev === rev && block.body.trim() === k.css.trim();
+    const copies = legacyBlocks.filter(b => b.id === k.id && b.kind === 'copy');
+    const overrides = legacyBlocks.filter(b => b.id === k.id && b.kind === 'override');
+    if (copies.length) {
+      return {
+        id: k.id, label: k.label, status: 'legacy-outside',
+        detail: `受管区之外还有 ${copies.length} 段旧版副本(整份/多条规则) —— 它们后出现, 按层叠会覆盖受管区; 升级会清掉这几段(原件在 tokens.css.bak)`,
+      };
+    }
+    if (block && clean) {
+      return overrides.length
+        ? { id: k.id, label: k.label, status: 'ok', detail: overrideNote(overrides, '你自己的覆写规则') }
+        : { id: k.id, label: k.label, status: 'ok' };
+    }
+    if (block) {
+      return {
+        id: k.id, label: k.label, status: 'stale',
+        detail: block.rev === rev
+          ? '受管块内容与 rev 不符(被手工改过)'
+          : `受管块 rev ${block.rev.slice(0, 8)} 落后于技能当前 ${rev.slice(0, 8)}`,
+      };
+    }
+    // 老格式(整段裸文本, 没有受管块): 内容与当前版逐字一致时渲染无差异 → 报 ok(不骚扰),
+    // 只提示下次升级会补上受管块。注意这个分支必须排在 legacy-outside 之后 —— 否则
+    // "区外还留着旧副本"会被这里的字符串命中掩盖掉(那正是重写前的旧缺陷)。
+    if (src.includes(k.css.trim())) {
+      return { id: k.id, label: k.label, status: 'ok', detail: '老格式(裸文本, 无受管块): 内容已是当前版, 渲染无差异; 下次 --upgrade-css 会补上受管块' };
+    }
+    return {
+      id: k.id, label: k.label, status: 'missing',
+      detail: k.probe(src) ? '文件里有旧版残留(与当前不同)' : '文件里没有这段',
+    };
+  });
+  if (expected) statuses.unshift(tokensStatus(src, expected));
+  return statuses;
+}
+
+/** 整段 tokens 受管区的状态(单独一个函数, 因为它比工具箱多一层"内容文本"比较) */
+export function tokensStatus(css, { text, rev }) {
+  const src = String(css ?? '').replace(/\r\n/g, '\n');
+  const blocks = findAllBlocks(src, TOKENS_ID);
+  const openCount = countOpenTags(src, TOKENS_ID);
+  const want = String(text).replace(/\r\n/g, '\n').trim();
+  const base = { id: TOKENS_ID, label: 'tokens 主体(主题/字号/fx/工具箱)' };
+  if (openCount > blocks.length) return { ...base, status: 'broken', detail: `有 ${openCount} 个开标记但只有 ${blocks.length} 个配对收尾` };
+  if (blocks.length > 1) return { ...base, status: 'duplicate', detail: `出现 ${blocks.length} 个 tokens 受管区` };
+  if (!blocks.length) {
+    // 老格式(无受管区): 文件里已经有**逐字相同**的当前版生成物时, 渲染结果与升级后完全一致,
+    // 所以仍然报 ok(不骚扰), 只在 detail 里提示下次升级会就地包上受管区。
+    // 但若出现了**两份**(旧追加式升级留下的), 必须报 duplicate: 后出现的那份按层叠会赢,
+    // 不报就会变成"check 绿、成片却是旧值" —— 正是这套机制要堵的静默失效。
+    const n = want ? src.split(want).length - 1 : 0;
+    if (n === 1) {
+      return { ...base, status: 'ok', detail: '老格式(裸文本, 无受管区): 内容已是当前版, 渲染无差异; 下次 --upgrade-css 会就地包上受管区(内容与位置都不变)' };
+    }
+    if (n > 1) {
+      return { ...base, status: 'duplicate', detail: `文件里出现了 ${n} 份当前版生成物(旧追加式升级留下的) —— 后出现的那份按层叠会赢; 升级会归并为一` };
+    }
+    return { ...base, status: 'missing', detail: '没有 tokens 受管区(老项目: 升级会整段重建, 区外的项目自定义规则保留)' };
+  }
+  const b = blocks[0];
+  // 受管区在、内容也对, 也要看**区外**有没有主体残留(A7 实测: 受管区后面粘着一段旧主体,
+  // 旧值按层叠压过新版, 而判定只看区内, 报 ok)。主体段按体量判: ≥400 字符或 ≥3 条 → 报出来。
+  const outsideCopies = legacyOutsideBlocks(src, { bodyText: want }).filter(x => x.id === TOKENS_ID && x.kind === 'copy');
+  if (outsideCopies.length) {
+    const n = outsideCopies.reduce((a, x) => a + x.rules.length, 0);
+    // 说清位置: 在受管区**之前**的会被受管区覆盖(老格式升级就长这样: 生成物被追加到文件末尾,
+    // 原来在文件尾的项目规则反而落到前面), 在之后的才会压过受管区。含糊其辞会让人按错误的
+    // 方向去修(复查 A6 实测: 旧文案写"后出现者按层叠优先", 而实际布局正好相反)。
+    const beforeRules = outsideCopies.reduce((a, x) => a + x.rules.filter(r => r.start < b.start).length, 0);
+    const afterRules = n - beforeRules;
+    const where = [];
+    if (beforeRules) where.push(`${beforeRules} 条在受管区**之前**(会被受管区覆盖 —— 想要它们生效就挪到受管区之后)`);
+    if (afterRules) where.push(`${afterRules} 条在受管区之后(后出现, 按层叠会覆盖受管区 —— 是旧残留就删掉)`);
+    return {
+      ...base, status: 'legacy-outside',
+      detail: `受管区之外还有 ${outsideCopies.length} 段主体规则(共 ${n} 条): ${where.join('; ')}; 升级只清掉与当前版逐字相同的那几条`,
+    };
+  }
+  if (b.rev === rev && b.body.trim() === want) {
+    const overrides = legacyOutsideBlocks(src, { bodyText: want }).filter(x => x.id === TOKENS_ID && x.kind === 'override');
+    const before = overrides.filter(x => x.start < b.start).reduce((a, x) => a + x.rules.length, 0);
+    return overrides.length
+      ? { ...base, status: 'ok', detail: `受管区之外还有你自己的规则 ${overrides.reduce((a, x) => a + x.rules.length, 0)} 条${before ? `(其中 ${before} 条在受管区**之前**, 会被受管区覆盖 —— 要生效就挪到之后)` : '(都在受管区之后, 后出现者优先)'}` }
+      : { ...base, status: 'ok' };
+  }
+  return {
+    ...base, status: 'stale',
+    detail: b.rev === rev ? '受管区内容与 rev 不符(被手工改过)' : `受管区 rev ${b.rev.slice(0, 8)} 落后于技能当前 ${rev.slice(0, 8)}`,
+  };
+}
+
+// ── 升级(原地) ───────────────────────────────────────────────────
+// 只动受管区: tokens 区整体替换(内部含三个工具箱子块), 区外的项目自定义规则一字不动。
+// 老项目(没有 tokens 区)会整段重建, 并且清掉散落在外面的旧 copies。
+export function applyKitUpgrade(css, { rev = KIT_REV, tokensText = null, tokensRev = null } = {}) {
+  const raw = String(css ?? '');
+  let out = raw.replace(/\r\n/g, '\n');   // 与 rev 计算同标准; 写回即统一为 LF
+  const actions = [];
+
+  // 1) 孤儿开标记(定界符被改坏)先清掉 —— 否则重建时会叠加出第二个开标记
+  for (const id of BLOCK_IDS) {
+    const blocks = findAllBlocks(out, id);
+    const opens = countOpenTags(out, id);
+    if (opens > blocks.length) {
+      // 先删掉所有配对块, 再删掉所有开标记(顺序不能反, 否则删标记会破坏配对)
+      for (const b of blocks.reverse()) out = out.slice(0, b.start) + '\n' + out.slice(b.end);
+      out = out.replace(OPEN_RE_G(id), '');
+      actions.push({ id, label: id, action: 'cleaned-broken', reason: `${opens} 个开标记只有 ${blocks.length} 个配对收尾, 已清掉残留标记` });
+    }
+  }
+
+  // 2) 同 id 多块 → 归并为一(保留第一块的位置, 删掉其余)
+  for (const id of BLOCK_IDS) {
+    const blocks = findAllBlocks(out, id);
+    if (blocks.length > 1) {
+      for (const b of blocks.slice(1).reverse()) out = out.slice(0, b.start) + '\n' + out.slice(b.end);
+      actions.push({ id, label: id, action: 'merged-duplicates', reason: `${blocks.length} 个重复受管块已归并(否则块外那份按层叠会赢)` });
+    }
+  }
+
+  // 3) tokens 受管区: 整段替换 / 就地包裹 / 追加到文件尾。
+  //    这一步必须排在"清区外旧副本"**之前**: 老 init 直出的裸文本主体里本来就含三个工具箱的原文,
+  //    而那时它们都还"不在任何受管区里" —— 先清就会把裸文本里的工具箱段删掉, 于是下面的
+  //    indexOf 失配 → 既多插一份新生成物, 又把残缺的旧主体留在区内之外, 按层叠反而压过新版。
+  //    (2026-09-18 实测: 就是这条顺序让"老项目升级"在 C 场景下产出重复内容 + 旧规则压新版。)
+  if (tokensText != null) {
+    const blocks = findAllBlocks(out, TOKENS_ID);
+    const wrapped = `${openTag(TOKENS_ID, tokensRev)}\n${String(tokensText).replace(/\r\n/g, '\n').trim()}\n${closeTag(TOKENS_ID)}`;
+    if (blocks.length === 1) {
+      if (blocks[0].rev !== tokensRev || blocks[0].body.trim() !== String(tokensText).replace(/\r\n/g, '\n').trim()) {
+        out = out.slice(0, blocks[0].start) + wrapped + out.slice(blocks[0].end);
+        actions.push({ id: TOKENS_ID, label: 'tokens 主体', action: 'tokens-replaced', reason: '整段生成物已就地替换为技能当前版本(区之外的项目规则未动)' });
+      }
+    } else {
+      // 没有 tokens 区(老项目/裸文本): 老 init 直出的主体里**嵌着**三个工具箱子块, 所以
+      // "文件里是否已有当前版生成物"必须在**剥掉子块之前**判 —— 先剥就会让 indexOf 失配,
+      // 于是多插一份、旧内容还留在后面按层叠压过新版(实测踩到)。
+      const want = String(tokensText).replace(/\r\n/g, '\n').trim();
+      const at = out.indexOf(want);
+      if (at > -1) {
+        out = out.slice(0, at) + wrapped + out.slice(at + want.length);
+        actions.push({ id: TOKENS_ID, label: 'tokens 主体', action: 'tokens-wrapped', reason: '文件里已有当前版生成物(裸文本), 已就地包进受管区(内容不变, 位置不变)' });
+      } else {
+        // 别的内容: 先清掉散落的工具箱子块与旧副本, 再整段追加到**文件尾** ——
+        // 生成物在层叠上后出现才压得住文件里残留的旧版内容(置于开头会被旧内容反压)
+        for (const k of KITS) {
+          for (const b of findAllBlocks(out, k.id).reverse()) out = out.slice(0, b.start) + '\n' + out.slice(b.end);
+        }
+        out = out.replace(/\s*$/, '\n') + '\n' + wrapped + '\n';
+        actions.push({ id: TOKENS_ID, label: 'tokens 主体', action: 'tokens-appended', reason: '老项目: 整段生成物(含三个工具箱子块)已追加到文件末尾 —— 在你原有规则之后, 所以新版规则在层叠上生效; 你的项目规则原样保留在前(受管区内的规则以新版为准)' });
+      }
+    }
+  } else {
+    // 兼容: 只升级工具箱(不传 tokensText 时按旧语义逐块替换)
+    let statuses = kitStatuses(out, { rev });
+    for (const [i, k] of KITS.entries()) {
+      if (statuses[i].status !== 'stale') continue;
+      const b = findAllBlocks(out, k.id)[0];
+      out = out.slice(0, b.start) + wrapKit(k.id, k.css, rev) + out.slice(b.end);
+      actions.push({ id: k.id, label: k.label, action: 'replaced', reason: statuses[i].detail });
+    }
+    statuses = kitStatuses(out, { rev });
+    for (const [i, k] of KITS.entries()) {
+      const st = statuses[i];
+      if (st.status === 'ok' && !st.detail) continue;
+      const fresh = wrapKit(k.id, k.css, rev);
+      const text = k.css.trim();
+      const at = out.lastIndexOf(text);
+      if (st.status === 'ok' && at > -1) {
+        const before = out.slice(0, at).split(text).join('');
+        out = before + fresh + out.slice(at + text.length);
+        actions.push({ id: k.id, label: k.label, action: 'wrapped', reason: '内容已是当前版, 原地补上受管块定界符(位置不变)' });
+        continue;
+      }
+      out = out.replace(/\s*$/, '\n') + '\n' + fresh + '\n';
+      actions.push({ id: k.id, label: k.label, action: 'appended', reason: '缺失, 已追加受管块' });
+    }
+  }
+
+  // 4) 区外残留的同内容副本: 受管区此时已就位, 这一步在"区之外"的视野里清旧副本
+  //    (只删 kind='copy' 里逐字命中当前版的那些; 用户手写的规则与改过的旧规则都保留 ——
+  //     见 removeLegacyCopies 的注释。bodyText 让主体残留(A7)也走同一条路)
+  out = removeLegacyCopies(out, actions, { bodyText: tokensText });
+
+  return { css: out, actions, normalizedLineEndings: /\r\n/.test(raw) };
+}
+
+// 区外的旧版副本: 逐条删除(升级的语义就是"把生成物之外多余的工具箱内容收拾干净";
+// 备份在 .bak, 日志里会打印删了几条)。**只处理 kind='copy' 的段**, 而且按条判:
+// 一段里"逐字来自模块原文"的前缀才是旧副本, 后面紧挨着的短尾巴(≤2 条)是用户自己写的覆写,
+// 原地保留 —— "受管区之外的规则一概不动"这条承诺不能因为清理而破。
+// 若尾巴很长(被改过多处的整份旧副本), 则整段清掉: 它留在那里就会按层叠压过受管区,
+// 属于必须清掉的错误产物(原文在 .bak, 日志会点名)。
+function removeLegacyCopies(css, actions, { bodyText = null } = {}) {
+  const copies = legacyOutsideBlocks(css, { bodyText }).filter(b => b.kind === 'copy');
+  const spans = [];
+  let tokensRemoved = 0;
+  for (const b of copies) {
+    // 逐字来自"技能当前版"的规则才删: 工具箱对模块原文, 主体对当前生成物。主体段里那些改过的规则
+    // (旧版残留 or 用户有意覆写 —— 光看文本分不出来)一律留在原地, 由人按 check 的提示决定。
+    const ref = (b.id === TOKENS_ID ? String(bodyText ?? '') : KITS.find(k => k.id === b.id).css).replace(/\s+/g, ' ');
+    // 主体段: 逐条判 —— 与当前版逐字相同的都删(它们是重复内容, 删了渲染不变), 改过的都留
+    if (b.id === TOKENS_ID) {
+      const same = b.rules.filter(r => ref.includes(r.text.replace(/\s+/g, ' ').trim()));
+      const diff = b.rules.filter(r => !ref.includes(r.text.replace(/\s+/g, ' ').trim()));
+      for (const r of same) spans.push({ start: r.start, end: r.end });
+      tokensRemoved += same.length;
+      if (diff.length) {
+        const first = diff[0].text.replace(/\s+/g, ' ').trim().slice(0, 60);
+        actions.push({
+          id: TOKENS_ID, label: b.label, action: 'kept-legacy-overrides',
+          reason: `受管区之外有 ${diff.length} 条**与当前版不同**的主体规则, 已原样保留(例: ${first}…) —— 它们是旧版残留还是你有意改的, 工具分不出来: 若是旧残留就删掉, 想让它生效就挪到受管区之后`,
+        });
+      }
+      continue;
+    }
+    let i = 0;
+    while (i < b.rules.length && ref.includes(b.rules[i].text.replace(/\s+/g, ' ').trim())) i++;
+    const tail = b.rules.slice(i);
+    // 工具箱: 尾巴太长 = 改过几处的整份旧副本, 整段清(它压在受管区上就是错误产物)
+    const whole = tail.length > 2;
+    for (const r of (whole ? b.rules : b.rules.slice(0, i))) spans.push({ start: r.start, end: r.end });
+    if (!whole && tail.length) {
+      actions.push({
+        id: b.id, label: b.label, action: 'kept-legacy-overrides',
+        reason: `受管区外与旧副本紧邻的 ${tail.length} 条规则未出现在技能原文里, 判为你自己的覆写并原地保留(它们后出现, 按层叠优先)`,
+      });
+    }
+  }
+  let out = css;
+  // 片段先并集再删: 万一还有相交/嵌套的片段, 按偏移逐个删会错位(删掉后面的受管区)—— 2026-09-18 实测
+  const union = [];
+  for (const s of spans.sort((a, c) => a.start - c.start)) {
+    const last = union[union.length - 1];
+    if (last && s.start <= last.end) last.end = Math.max(last.end, s.end);
+    else union.push({ ...s });
+  }
+  for (const s of union.sort((a, c) => c.start - a.start)) out = out.slice(0, s.start) + out.slice(s.end);
+  const byKit = new Map();
+  for (const b of copies) if (b.id !== TOKENS_ID) byKit.set(b.id, (byKit.get(b.id) ?? 0) + 1);
+  for (const k of KITS) {
+    const n = byKit.get(k.id) ?? 0;
+    if (n) actions.push({ id: k.id, label: k.label, action: 'removed-legacy-outside', reason: `受管区外发现 ${n} 段旧版副本并已删除(后出现会按层叠覆盖受管区; 原件在 tokens.css.bak)` });
+  }
+  if (tokensRemoved) {
+    actions.push({ id: TOKENS_ID, label: 'tokens 主体', action: 'removed-legacy-outside', reason: `受管区外发现 ${tokensRemoved} 条与当前版逐字相同的主体规则, 已删除(它们后出现会按层叠覆盖受管区; 原件在 tokens.css.bak)` });
+  }
+  return out;
+}
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/fetch-official-images.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/fetch-official-images.mjs
new file mode 100644
index 00000000..0d06f590
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/fetch-official-images.mjs
@@ -0,0 +1,291 @@
+#!/usr/bin/env node
+// html2video-for-mcode · 官方站点取图: 用 Playwright 打开官网, 列出候选图并下载
+// 这是 image-sources.md 里成功率最高的路径 B(官方 logo / 产品图 / hero 图), 之前只有描述没有工具。
+//
+// 用法:
+//   node fetch-official-images.mjs <网址> [--list]                        列出候选(默认动作)
+//   node fetch-official-images.mjs <网址> --get 1,3,5 [--out-dir assets]  按序号下载
+//   node fetch-official-images.mjs <网址> --min 800                       只看宽度 ≥800px 的
+//   node fetch-official-images.mjs <网址> --min 800x600                   宽度与高度都要够(横扁图不能用)
+//   node fetch-official-images.mjs <网址> --json                          输出 JSON(供程序处理)
+//   node fetch-official-images.mjs --url <图片URL>[,<URL>...] [--out-dir assets]
+//                                                                         直接下载给定图片 URL
+//   (内置浏览器 inspect 官网 DOM 拿到零散图片 URL 时用 --url: 走同一套 host/重定向/大小/文件名
+//    校验, 且**不需要 playwright** —— 站点要登录/滚动加载、脚本打不开时, 这是唯一落盘手段)
+//   本地页(file://)加 --allow-file; 覆盖已存在文件 / out-dir 出项目 加 --force
+//
+// 纪律(与 image-sources.md 一致): 只取官方域名的资源; 列出的每一项都要人眼过一遍
+// (主体是否居中/有无水印/比例是否合适), 选中的落盘 assets/ 并登记 MANIFEST.md。
+//
+// 网络安全(字符串策略在打开浏览器之前校验, 负向路径无需 playwright):
+//   - 只允许 http(s) 公网页面; file:// 本地页需显式 --allow-file(离线场景)
+//   - 拦内网/本地/元数据地址(loopback、169.254.x、私网段、无点主机名等, 见 url-policy.mjs)
+//   - 下载逐跳过同一策略(maxRedirects=0 手动跟, ≤5 跳), 响应有大小上限(默认 30MB, 读取中途掐断)
+//   - **连接期绑定**(2026-09-25 六审 blocker 1): 预查≠建连 —— DNS rebinding 的缝在"预查时
+//     公网、建连时私网"。所有出网走 policyGet(建连那次 lookup 被否决); 浏览器(Chromium 自己
+//     解析, 注入不了 lookup)的全部流量收进本地否决代理 startVetoProxy, 上游建连走同一否决
+//   - 落盘必须在工作目录内(--out-dir 越界需 --force), 不覆盖已存在文件(需 --force)
+import fs from 'node:fs';
+import { assertContained, canonicalPath as canonical, flagValue, loadPackage, inside, positionals } from './tools.mjs';
+import path from 'node:path';
+import { assertFetchableUrl, assertRedirectTarget, assertResolvedHost, policyGet, startVetoProxy, sanitizeFilename, imageNameFromUrl, MAX_REDIRECTS, PolicyError } from './url-policy.mjs';
+
+const argv = process.argv.slice(2);
+// 位置参数与取值型 flag 的清单统一在 tools.mjs: --json 曾在这里被当作取值型, 于是
+// `--json <网址>` 会把网址吃掉、位置参数成空 → 报用法错误(2026-09-18 复查 D3 的成因)。
+const positional = positionals(argv);
+const FORCE = argv.includes('--force');
+const ALLOW_FILE = argv.includes('--allow-file');
+const MAX_BYTES = Math.max(1, parseFloat(flagValue(argv, '--max-mb', '30')) * 1024 * 1024);
+const URLS = (flagValue(argv, '--url', '') || '').split(',').map(s => s.trim()).filter(Boolean);
+const url = positional[0];
+if (!URLS.length && (!url || !/^(https?|file):/.test(url))) {
+  console.error('用法: node fetch-official-images.mjs <网址> [--list|--get 1,3] [--out-dir assets] [--min 800] [--max-mb 30] [--allow-file] [--force] [--json]\n    或: node fetch-official-images.mjs --url <图片URL>[,<URL>...] [--out-dir assets] [--max-mb 30] [--force]');
+  process.exit(1);
+}
+if (!URLS.length) {
+  try {
+    assertFetchableUrl(url, { allowFile: ALLOW_FILE, where: '页面 URL' });
+  } catch (e) {
+    if (e instanceof PolicyError) { console.error(`✗ ${e.message}`); process.exit(1); }
+    throw e;
+  }
+}
+// --min 支持三种写法: 800(宽≥800) / 800x600(宽高都够) / x600(只限高)
+const MIN_RAW = String(flagValue(argv, '--min', '0') ?? '0').trim().toLowerCase();
+const MIN_M = /^(\d*)(?:x(\d+))?$/.exec(MIN_RAW);
+if (!MIN_M || MIN_RAW === '' || (!MIN_M[1] && !MIN_M[2])) {
+  console.error(`✗ --min 写法不对: ${MIN_RAW}(支持 800 / 800x600 / x600; 只写数字=只限宽)`);
+  process.exit(1);
+}
+const MIN = parseInt(MIN_M[1] || '0', 10) || 0;
+const MIN_H = parseInt(MIN_M[2] || '0', 10) || 0;
+// 落盘收监: 默认 assets/ 必须在当前工作目录(项目根)内; 越界要显式 --force。
+// 比较用规范化路径(二审 P2): macOS 上 /var 与 /private/var 是同一目录的两种写法,
+// 纯字符串比较会把"项目内的合法路径"误判成越界(官方 CI 的 macOS 环境实测踩到)。
+const OUT_DIR = path.resolve(flagValue(argv, '--out-dir', 'assets'));
+const CWD_REAL = canonical(process.cwd()), OUT_REAL = canonical(OUT_DIR);
+if (!inside(CWD_REAL, OUT_REAL) && OUT_REAL !== CWD_REAL) {
+  if (!FORCE) {
+    console.error(`✗ --out-dir 越出当前工作目录: ${OUT_DIR}\n  默认只写项目内(README 的"写入都在项目目录"承诺)。确需外部位置请加 --force`);
+    process.exit(1);
+  }
+  console.warn(`⚠ --force: 下载将写到工作目录之外 ${OUT_DIR}`);
+}
+
+// ── 单张图片 URL 直下模式(内置浏览器 inspect 出零散 URL 时用; 不需要 playwright) ──
+// 与 --get 同一套纪律: 逐跳复核 host 策略、响应大小上限、文件名清洗、已存在不覆盖。
+// 顺序: 跟完重定向拿到响应头 → 按 content-type 定文件名 → **查重** → 才读 body 落盘
+// (否则"已存在"会白下一遍, 网络抖动时还会把"已存在"报成"下载失败")。
+// ── 逐跳策略 + 重定向循环的单一实现(第十三轮 review 去重: fetch 直下与 --get(浏览器请求)曾各写一份,
+// 第二份缺环守卫与 0 字节守卫)── get(url) 由调用方注入, 返回归一化的
+// { status, location, header(name), arrayBuffer() }, 两个客户端共用同一套跳数/环/上限纪律。
+async function followRedirects(src, get, { where = '下载' } = {}) {
+  let current = src, res = null;
+  const seen = new Set([src]);   // 见过就说明在绕圈(A→B→A 这种 ping-pong 靠单步自比较抓不到)
+  for (let hop = 0; hop <= MAX_REDIRECTS; hop++) {
+    assertFetchableUrl(current, { allowFile: false, where: `${where}(第 ${hop + 1} 跳)` });
+    await assertResolvedHost(current, { where: `${where}(第 ${hop + 1} 跳)` });
+    res = await get(current);
+    if (res.status >= 300 && res.status < 400) {
+      const loc = res.location;
+      if (!loc) throw new Error(`HTTP ${res.status} 无 Location`);
+      const next = new URL(loc, current).href;
+      if (next === current) throw new Error(`重定向指向自身(死循环): ${current}`);
+      if (seen.has(next)) throw new Error(`重定向绕圈(来回跳): ${[...seen].join(' → ')} → ${next}\n     多半是站点把不存在的图片转到了错误页`);
+      seen.add(next);
+      current = next;
+      continue;
+    }
+    break;
+  }
+  if (!res || res.status >= 300) throw new Error(`重定向超过 ${MAX_REDIRECTS} 跳, 最后停在: ${current}(多半是站点把不存在的图片转到了错误页)`);
+  if (res.status >= 400) throw new Error(`HTTP ${res.status} ${current}`);
+  const declared = parseInt(res.header('content-length') || '0', 10);
+  if (declared > MAX_BYTES) throw new Error(`响应 ${(declared / 1048576).toFixed(0)}MB 超过上限 ${Math.round(MAX_BYTES / 1048576)}MB(--max-mb 可调)`);
+  return res;
+}
+
+async function openUrl(src) {
+  // 出网走 policyGet: 建连那次 DNS 解析被 checkedLookup 否决(连接期绑定, 不再是"预查后
+  // 任由 fetch 自己再解析一遍" —— 那道缝就是 DNS rebinding 的攻击面)
+  const res = await followRedirects(src, u => policyGet(u, {
+    headers: { 'user-agent': 'html2video-for-mcode/image-fetch' },
+    timeoutMs: 30000,
+    maxBytes: MAX_BYTES,
+  }), { where: '下载' });
+  const type = String(res.header('content-type') || '').split(';')[0].trim();
+  if (!/^image\//.test(type) && !/octet-stream/.test(type)) console.warn(`  ⚠ content-type 不是图片(${type || '未知类型'})—— 人眼确认一下再登记`);
+  return { res, type };
+}
+
+if (URLS.length) {
+  // 建目录这一步也要收监: 叶子文件已经过了 assertContained, 但"先 mkdir 再写叶子"之间隔着一次
+  // 目录创建 —— OUT_DIR 的某个目录段是指向项目外的链接时, mkdir 本身就已经在项目外动手了
+  fs.mkdirSync(assertContained(CWD_REAL, OUT_DIR, { where: '输出目录' }), { recursive: true });
+  let ok = 0;
+  for (const src of URLS) {
+    if (!/^https?:/.test(src)) { console.error(`  ✗ 跳过(只收 http(s)): ${src}`); continue; }
+    try {
+      const { res, type } = await openUrl(src);
+      // 逐文件收监: OUT_DIR 本身过了 canonical 比较, 但它里面的目录段可能是符号链接(复查 B4)
+      const out = assertContained(CWD_REAL, path.join(OUT_DIR, imageNameFromUrl(src, type)), { where: '下载落盘' });
+      if (fs.existsSync(out) && !FORCE) { console.error(`  ✗ 已存在, 不覆盖: ${path.relative(process.cwd(), out)}(要覆盖加 --force)`); continue; }
+      const body = Buffer.from(await res.arrayBuffer());
+      if (!body.length) throw new Error('响应为空(0 字节)—— 多半是 404 错误页或防盗链, 别写空文件');
+      if (body.length > MAX_BYTES) throw new Error(`响应 ${(body.length / 1048576).toFixed(0)}MB 超过上限 ${Math.round(MAX_BYTES / 1048576)}MB(--max-mb 可调)`);
+      fs.writeFileSync(out, body);
+      console.log(`  ✓ ${path.relative(process.cwd(), out)}  (${(body.length / 1024).toFixed(0)}KB, ${type || '未知类型'})`);
+      ok++;
+    } catch (e) {
+      if (e instanceof PolicyError) console.error(`  ✗ ${src}\n     ${e.message}`);
+      else console.error(`  ✗ ${src}\n     下载失败: ${e.message}`);
+    }
+  }
+  console.log(`\n已下载 ${ok}/${URLS.length} 张 → ${path.relative(process.cwd(), OUT_DIR)}`);
+  console.log('下一步(必做): ① node scripts/prep-image.mjs --check <图...>  ② 人眼确认主体居中无水印  ③ 登记 assets/MANIFEST.md(来源 = 图片所在官方页 URL)');
+  process.exit(ok ? 0 : 1);
+}
+
+const playwright = await loadPackage('playwright', { projectDir: process.cwd() });
+if (!playwright) { console.error('✗ 未找到 playwright(已按 项目目录 / 调用目录 / npm 全局 逐个找过)。先执行: npm i playwright && npx playwright install chromium'); process.exit(1); }
+
+// 浏览器全部出网收进本地否决代理(2026-09-25 六审 blocker 1): Chromium 自己解析 DNS,
+// 注入不了 lookup —— 只有把页面导航/Chromium 跟随的重定向/全部子资源收进隧道, 上游建连
+// 的那次解析才能过 checkedLookup(file:// 本地页不经过代理, 也无需)。
+const veto = await startVetoProxy({ maxBytes: MAX_BYTES });
+let browser, context;
+try {
+  browser = await playwright.chromium.launch({ headless: true, proxy: { server: `http://127.0.0.1:${veto.port}` } });
+  context = await browser.newContext({ viewport: { width: 1600, height: 1200 } });
+  const page = await context.newPage();
+  // 逐请求拦截(二审 P1: 策略必须落在**真实请求边界**上): 字符串策略 + **无缓存**的逐请求
+  // 解析预查 —— 这里只做早拦与可命名告警; 权威检查点在否决代理的建连期(预查结果不作准,
+  // 按 host 缓存只会把 rebinding 的时间窗越放越大)。
+  await context.route('**/*', async route => {
+    const u = route.request().url();
+    try {
+      assertFetchableUrl(u, { allowFile: ALLOW_FILE, where: '页面请求' });
+      await assertResolvedHost(u, { where: '页面请求' });
+      await route.continue();
+    } catch (e) {
+      if (e instanceof PolicyError) { console.warn(`  ⚠ 已拦截请求: ${String(u).slice(0, 90)} — ${e.message}`); await route.abort(); }
+      else await route.continue();
+    }
+  });
+
+  await assertResolvedHost(url, { where: '页面 URL' });
+  await page.goto(url, { waitUntil: 'domcontentloaded', timeout: 45000 });
+  await page.waitForTimeout(2500); // 让懒加载图片冒出来
+  await page.evaluate(() => window.scrollTo(0, document.body.scrollHeight));
+  await page.waitForTimeout(1500);
+  await page.evaluate(() => window.scrollTo(0, 0));
+  await page.waitForTimeout(500);
+
+  // 候选来源: <img>/<picture><source> + 元素的计算样式 background-image(官网 hero 常用背景图)
+  const items = await page.evaluate(() => {
+    const out = [];
+    const push = (rawSrc, el) => {
+      if (!rawSrc) return;
+      let abs;
+      try { abs = new URL(rawSrc.split(',')[0].trim().split(' ')[0], location.href).href; } catch { return; }
+      if (!abs || abs.startsWith('data:') || abs.startsWith('blob:')) return;
+      const img = el && el.tagName === 'IMG' ? el : el?.closest?.('picture')?.querySelector('img');
+      out.push({
+        src: abs,
+        w: img?.naturalWidth || 0, h: img?.naturalHeight || 0,
+        alt: (el?.getAttribute?.('alt') || '').slice(0, 60),
+        kind: 'img',
+      });
+    };
+    document.querySelectorAll('img, picture > source').forEach(el => push(el.getAttribute('src') || el.getAttribute('srcset'), el));
+    document.querySelectorAll('body *').forEach(el => {
+      const bg = getComputedStyle(el).backgroundImage;
+      if (bg && bg !== 'none' && bg.includes('url(')) push(bg.slice(bg.indexOf('url(') + 4, bg.lastIndexOf(')')).replace(/["']/g, ''), el);
+    });
+    return out;
+  });
+
+  // 去重 + 尺寸过滤 + 官方域名标注 + host 策略过滤(候选 src 由页面控制, 同样只收公网 http(s)/放行的 file:)
+  const seen = new Set();
+  const candidates = [];
+  const pageHost = (() => { try { return new URL(url).host; } catch { return ''; } })();
+  for (const it of items) {
+    if (seen.has(it.src)) continue;
+    seen.add(it.src);
+    let host = '', ext = '';
+    try { const u = new URL(it.src); host = u.host; ext = (u.pathname.match(/\.(png|jpe?g|webp|avif|svg|gif)$/i) || [''])[0].toLowerCase(); } catch { /* ignore */ }
+    let srcPolicy = 'ok';
+    try {
+      assertFetchableUrl(it.src, { allowFile: ALLOW_FILE, where: '候选图' });
+    } catch (e) {
+      if (!(e instanceof PolicyError)) throw e;
+      srcPolicy = e.reason; // 内网/file/坏协议的 src 不进候选(下载时也过同一策略, 双保险)
+    }
+    if (srcPolicy !== 'ok') continue;
+    const offsite = pageHost && host && !host.endsWith(pageHost.replace(/^www\./, '')) && !pageHost.endsWith(host.replace(/^www\./, ''));
+    const vec = it.w === 0 && ext === '.svg';   // svg 无自然宽度, 不按像素筛
+    // --min 800 只限宽; --min 800x600 宽高都要够(横版视频里"够宽但很扁"的图没法用)
+    const bigEnough = vec || ((!MIN || it.w >= MIN) && (!MIN_H || it.h >= MIN_H));
+    if (!bigEnough) continue;
+    candidates.push({ ...it, host, ext: ext || '(无扩展名)', offsite });
+  }
+
+  if (!candidates.length) { console.log('没找到符合条件的图片(试试去掉 --min, 或换页面 / 产品页 / 新闻页)'); await context.close(); await browser.close(); await veto.close(); process.exit(0); }
+
+  // --json 是布尔 flag: 必须用 includes 判存在。flagValue 取的是"下一个参数",
+  // --json 在末尾(文档写的用法)时取到 undefined → 静默退化成人类可读列表(终轮复查抓到)。
+  if (argv.includes('--json')) { console.log(JSON.stringify(candidates, null, 2)); await context.close(); await browser.close(); await veto.close(); process.exit(0); }
+
+  console.log(`候选图 ${candidates.length} 张(页面 ${pageHost || url}):\n`);
+  candidates.forEach((c, i) => {
+    const size = c.w ? `${c.w}×${c.h}` : '矢量/未知';
+    const warn = c.offsite ? '  ⚠ 非本站域名, 确认是否官方 CDN' : '';
+    console.log(`  [${i + 1}] ${size.padEnd(12)} ${c.ext.padEnd(6)} ${c.host}${warn}\n       ${c.alt ? c.alt + ' — ' : ''}${c.src.slice(0, 120)}`);
+  });
+
+  const get = flagValue(argv, '--get');
+  if (get) {
+    const picks = get.split(',').map(x => parseInt(x.trim(), 10)).filter(n => n >= 1 && n <= candidates.length);
+    if (!picks.length) { console.error('✗ --get 的序号不在范围内'); process.exit(1); }
+    fs.mkdirSync(assertContained(CWD_REAL, OUT_DIR, { where: '输出目录' }), { recursive: true });
+    let ok = 0;
+    for (const n of picks) {
+      const c = candidates[n - 1];
+      const name = sanitizeFilename(c.alt, { maxLen: 30, fallback: `official-${String(n).padStart(2, '0')}` }) + (c.ext.startsWith('.') ? c.ext : '.png');
+      const out = assertContained(CWD_REAL, path.join(OUT_DIR, name), { where: '下载落盘' });
+      if (fs.existsSync(out) && !FORCE) { console.error(`  ✗ [${n}] 已存在, 不覆盖: ${out}(要覆盖加 --force)`); continue; }
+      try {
+        if (c.src.startsWith('file:')) {
+          fs.copyFileSync(new URL(c.src), out); // 本地页面(file://, 需 --allow-file)直拷
+        } else {
+          // 与 --url 直下共用同一套逐跳策略 + 环守卫 + 0 字节守卫(followRedirects 单一实现);
+          // 出网同样走 policyGet(连接期否决) —— context.request 走 Playwright 自己的网络栈,
+          // 解析不受我们控制, 六审 blocker 1 之后不再用它出网
+          const res = await followRedirects(c.src, u => policyGet(u, { timeoutMs: 30000, maxBytes: MAX_BYTES }), { where: '下载' });
+          const body = Buffer.from(await res.arrayBuffer());
+          if (!body.length) throw new Error('响应为空(0 字节)—— 多半是 404 错误页或防盗链, 别写空文件');
+          if (body.length > MAX_BYTES) { throw new Error(`响应 ${(body.length / 1048576).toFixed(0)}MB 超过上限 ${Math.round(MAX_BYTES / 1048576)}MB(--max-mb 可调)`); }
+          fs.writeFileSync(out, body);
+        }
+        console.log(`  ✓ ${path.relative(process.cwd(), out)}  (${(fs.statSync(out).size / 1024).toFixed(0)}KB)`);
+        ok++;
+      } catch (e) { console.error(`  ✗ [${n}] 下载失败: ${e.message}`); }
+    }
+    console.log(`\n已下载 ${ok}/${picks.length} 张 → ${path.relative(process.cwd(), OUT_DIR)}`);
+    // 退出码按成功率给: 全失败还退 0 的话, 上游脚本会把"一张都没落盘"当成功继续往下跑
+    if (!ok) { console.error(`✗ ${picks.length} 张全部下载失败 —— 检查上面的原因(host 策略 / 网络 / --max-mb)`); process.exit(1); }
+    if (ok < picks.length) console.warn(`⚠ 部分失败: ${picks.length - ok} 张没下来(见上面的原因); 已落盘的可以继续用`);
+    console.log('下一步(必做): ① node scripts/prep-image.mjs --check <图...>  ② 人眼确认主体居中无水印  ③ 登记 assets/MANIFEST.md(来源就是本 URL)');
+  } else {
+    console.log('\n选好后: node scripts/fetch-official-images.mjs <网址> --get 1,3 --out-dir assets');
+    console.log('下载前先人眼过一遍: 主体居中 / 无水印无无关 logo / 比例接近 / 分辨率够(image-sources.md §2)。');
+  }
+} catch (e) {
+  console.error(`✗ 取图失败: ${e.message}\n  → 检查网址是否可达; 若站点需要交互(登录/滚动加载), 用内置浏览器手动取。`);
+  process.exitCode = 1;
+} finally {
+  await context?.close();
+  await browser?.close();
+  await veto.close();   // 先关浏览器再关代理: 关代理会掐掉仍在途的隧道连接
+}
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/grab-frames.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/grab-frames.mjs
new file mode 100644
index 00000000..016629bc
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/grab-frames.mjs
@@ -0,0 +1,77 @@
+#!/usr/bin/env node
+// html2video-for-mcode · 成片抽帧核对: 按 build/timings.json 算每张绝对起点, 从 out/final.mp4 抽帧
+// 这是"still 预览看不出、只有成片里才见"那类问题的最终对账手段 —— 图注落进字幕带被字幕压住、
+// 字幕压正文、末级元素没进画面、收尾静止不够, 都只有成片实帧能证明(2026-09-18 实测: 8 张图注
+// 全被字幕盖住, still 预览全绿)。上片前逐张过一遍, 比"整片重看"快得多。
+//
+// 用法:
+//   node grab-frames.mjs <项目目录>                # 每张抽 1 帧(默认取末段: 元素已入场 + 字幕在屏)
+//   node grab-frames.mjs <项目目录> --ids 05,11    # 只抽指定张
+//   node grab-frames.mjs <项目目录> --at 0.5       # 取每张 50% 处(默认 0.93)
+//   node grab-frames.mjs <项目目录> --both         # 每张抽 2 帧(中段 + 末段)
+import fs from 'node:fs';
+import path from 'node:path';
+import { spawnSync } from 'node:child_process';
+import { findTool, flagValue, positionalDir, requireTool, safeId, safeOut, validateTimingsIds } from './tools.mjs';
+
+const argv = process.argv.slice(2);
+const dir = positionalDir(argv);
+// --at 必须在 (0,1]: 静默 clamp/NaN 会让抽帧点变成 NaN 或 0s, 抽出来的帧看着'有图'却不是该看的那一刻
+const AT_RAW = flagValue(argv, '--at', '0.93');
+const AT_NUM = Number(AT_RAW);
+if (!Number.isFinite(AT_NUM) || AT_NUM <= 0 || AT_NUM > 1) {
+  console.error(`✗ --at ${JSON.stringify(AT_RAW)} 非法: 取 (0,1] 的小数(0.93 = 该张 93% 处; 0 或 NaN 会抽到起始帧, 看不到入场结果)`);
+  process.exit(1);
+}
+const at = AT_NUM;
+const both = argv.includes('--both');
+
+const timingsPath = path.join(dir, 'build', 'timings.json');
+if (!fs.existsSync(timingsPath)) { console.error(`✗ 缺 ${timingsPath} — 先跑 plan-timings.mjs 与 build-video.mjs`); process.exit(1); }
+const timings = JSON.parse(fs.readFileSync(timingsPath, 'utf8'));
+validateTimingsIds(timings);
+const final = path.join(dir, 'out', 'final.mp4');
+if (!fs.existsSync(final)) { console.error(`✗ 还没有 ${final} — 先跑 build-video.mjs`); process.exit(1); }
+const FFMPEG = requireTool('ffmpeg', dir);   // 环境错误统一退 2(与 tools.mjs 的约定一致)
+
+const onlyIds = flagValue(argv, '--ids', '') ? new Set(flagValue(argv, '--ids', '').split(',').map(s => safeId(s.trim()))) : null;
+
+const outDir = safeOut(dir, 'build', 'introspect');
+fs.mkdirSync(outDir, { recursive: true });
+
+let start = 0;
+const shots = [];
+// --ids 未命中的 id 点名告警(与 capture 同款: 部分匹配不再静默少做, 2026-09-28 Mavis 交接)
+if (onlyIds) {
+  const have = new Set(timings.slides.map(s => safeId(s.id)));
+  const miss = [...onlyIds].filter(id => !have.has(id));
+  if (miss.length) console.warn(`⚠ --ids 里没有对应 slide 的 id, 已忽略: ${miss.join(', ')}(PowerShell 裸写 03,05 会吃掉前导零 —— 加引号 --ids "03,05" 或逐个传)`);
+}
+for (const s of timings.slides) {
+  const sid = safeId(s.id);
+  const dur = Number(s.duration);
+  if (!Number.isFinite(dur) || dur <= 0) { console.error(`✗ timings 里 ${sid} 的 duration 非法: ${JSON.stringify(s.duration)}`); process.exit(1); }
+  if (!onlyIds || onlyIds.has(sid)) {
+    if (both) {
+      shots.push({ id: sid, t: start + dur * 0.55, tag: 'mid' });
+      shots.push({ id: sid, t: Math.max(0, start + dur - 0.3), tag: 'end' });
+    } else {
+      shots.push({ id: sid, t: start + dur * at, tag: 'end' });
+    }
+  }
+  start += dur;
+}
+
+let bad = 0;
+for (const sh of shots) {
+  // 叶子也要过收监(1.7.5 复查同类收尾, 九轮 review 抓到): id 过了 safeId 只防"路径注入",
+  // 防不了预置在 build/introspect/ 里的**文件符号链接**把 ffmpeg -y 的输出引到项目外。
+  const out = safeOut(dir, 'build', 'introspect', `frame-${sh.id}-${sh.tag}.png`);
+  const r = spawnSync(FFMPEG, ['-y', '-ss', sh.t.toFixed(2), '-i', final, '-frames:v', '1', out], { encoding: 'utf8', windowsHide: true });
+  const ok = fs.existsSync(out) && fs.statSync(out).size > 1000;
+  console.log(`${ok ? '✓' : '✗'} ${sh.id} @${sh.t.toFixed(2)}s (${sh.tag}) → ${path.relative(dir, out)}`);
+  if (!ok) { bad++; console.log('    ', (r.stderr || '').split('\n').slice(-3).join(' ')); }
+}
+if (!shots.length) { console.error('✗ 一帧都没抽到(检查 --ids 是否写错、或 timings.json 的 slides 是不是空的)'); process.exit(1); }
+console.log(`\n共 ${shots.length} 帧 → ${path.relative(dir, outDir)}(逐张看: 图注/字幕有没有打架、末级元素是否都在、收尾静止够不够)`);
+process.exit(bad ? 1 : 0);
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/init-project.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/init-project.mjs
new file mode 100644
index 00000000..a78d06f2
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/init-project.mjs
@@ -0,0 +1,177 @@
+#!/usr/bin/env node
+// html2video-for-mcode · 生成项目骨架
+// 用法: node init-project.mjs <项目目录> [--topic "主题名"] [--force] [--upgrade-css] [--check-css]
+// 目标目录已存在且非空时拒绝执行(会重置 script.json 等 5 个生成文件), 需显式 --force。
+// --upgrade-css: 不生成任何文件, 只把已有项目 tokens.css 里的 tokens 受管区(技能生成的整段主体:
+//   主题令牌 / --fs-* / fx 关键帧与工具类 / .fx-stagger, 三个工具箱子块嵌在其中)更新到技能当前版本
+//   —— 按内容 rev 原地替换, 幂等可重复跑, 写前备份 tokens.css.bak。区外规则**按条判定**: 逐字来自
+//   模块原文的旧副本会被清掉(它们按层叠会压过受管区), 你自己的规则(含文件尾的覆写)保留不动。
+// --check-css: 只检查不修改 —— 受管区落后/缺失/重复/定界符被改坏/区外残留副本时逐项报出并以退出码 1 结束。
+import fs from 'node:fs';
+import path from 'node:path';
+import { KIT_REV, TOKENS_ID, wrapTokens, applyKitUpgrade, kitStatuses, MAX_SCAN_BYTES } from './css-kit.mjs';
+import { generateTokensCss, TOKENS_REV } from './tokens-template.mjs';
+import { VALUE_FLAGS, flagValue, positionalDir, safeOut } from './tools.mjs';
+
+const argv = process.argv.slice(2);
+const hasPositional = argv.some((a, i) => !a.startsWith('--') && !(i > 0 && VALUE_FLAGS.has(argv[i - 1])));
+const dirArg = hasPositional ? positionalDir(argv, { dflt: '' }) : '';
+if (!dirArg) {
+  console.error('用法: node init-project.mjs <项目目录> [--topic "主题名"] [--force] [--upgrade-css] [--check-css]');
+  process.exit(1);
+}
+// --topic 是取值型 flag: 取它的值时不能把值本身当位置参数(否则会在 cwd 下静默建出以主题命名的骨架)
+const topic = flagValue(argv, '--topic', '');
+// topic 会插进模板 HTML 的 .brand 角标, 插入点单独转义(script.json / notes.md 用原文, 不转义)
+const escapeHtml = s => String(s).replace(/[&<>"']/g, ch => ({ '&': '&amp;', '<': '&lt;', '>': '&gt;', '"': '&quot;', "'": '&#39;' }[ch]));
+const topicHtml = escapeHtml(topic);
+const dir = path.resolve(dirArg);
+
+
+const TEMPLATE_HTML = `<!doctype html>
+<!-- 版式示例: statement · 复制改内容, 其他版式片段见 references/authoring.md -->
+<html lang="zh-CN" data-theme="a">
+<head>
+<meta charset="utf-8">
+<link rel="stylesheet" href="tokens.css">
+<style>
+  .layout { position: absolute; inset: 0; display: flex; flex-direction: column;
+    justify-content: safe center; padding: 120px 160px 190px; gap: var(--sp-6); }
+  .layout h1 { font-family: var(--font-display); font-size: var(--fs-display);
+    font-weight: 700; line-height: 1.18; max-width: 1400px; }
+  .layout .sub { font-size: var(--fs-h3); color: var(--muted); max-width: 1200px; }
+  .layout figure { margin-top: var(--sp-4); }
+</style>
+</head>
+<body>
+<div class="stage">
+  <span class="brand">A · ${topicHtml || '主题名'}</span>
+  <main class="layout">
+    <p class="kicker fx-fade" data-stage="1">SECTION · 小节名</p>
+    <h1 class="fx-up" data-stage="1">一句话断言,<span class="accent">关键词</span>点亮。</h1>
+    <p class="fx-fade" data-stage="2">一两句展开: 补充语境或给出解释, 别只有标题。</p>
+    <figure class="fx-grow" data-stage="3">
+      <!-- 视觉锚点: 数据卡 / 图片(.img-frame) / 引语块, 见 authoring.md -->
+      <div class="card" style="display:inline-block;">
+        <div class="gradient-text" style="font-family:var(--font-display); font-size:var(--fs-h1);">9亿</div>
+        <div style="font-size:var(--fs-caption); color:var(--muted);">周活跃用户</div>
+      </div>
+    </figure>
+  </main>
+  <span class="slide-num">02 / 08</span>
+</div>
+</body>
+</html>
+`;
+
+const scriptJson = {
+  _readme: 'html2video-for-mcode 脚本契约。clauses 每个元素=一句口播; stage=这句开口时该入场的视觉层(1/2/3); 可选 text2=双语字幕第二行。tail=收尾留白秒数(默认0.8)。改口播必须过 Gate 1, 之后从 Phase 2 重跑。可选顶层 bgm: "assets/bgm.mp3" 或 {file,volume:0.12,fadeIn:1.5,fadeOut:2.5}。width/height 决定画布: 1920×1080 横屏 / 1080×1920 竖屏。lang 决定口播语种与音色: zh(默认)/en/yue/其他 BCP-47 — 语种必须与音色匹配, 且影响语速基准与字幕行宽校验。speed=试听时与用户定的语速(Gate 4 可复核); transition=切页方式: cut 硬切(默认, 段间不经过黑场) 或 {type:"xfade",duration:0.4} 交叉溶解。',
+  topic,
+  lang: "zh",
+  voice: 'Chinese (Mandarin)_Gentleman',
+  speed: { default: 1.1, first: 1.05, last: 1.05 },
+  transition: { type: 'cut' },
+  fps: 30, width: 1920, height: 1080,
+  slides: [
+    { id: '01', layout: 'title-hero', html: '01-title.html', audio: '01.mp3', title: '', clauses: [{ stage: 1, text: '' }] },
+    { id: '02', layout: 'statement', html: '02-statement.html', audio: '02.mp3', title: '', clauses: [{ stage: 1, text: '' }, { stage: 2, text: '' }] },
+    { id: '03', layout: 'bullets', html: '03-bullets.html', audio: '03.mp3', title: '', clauses: [{ stage: 1, text: '' }, { stage: 2, text: '' }] },
+    { id: '04', layout: 'compare', html: '04-compare.html', audio: '04.mp3', title: '', clauses: [{ stage: 1, text: '' }, { stage: 2, text: '' }] },
+    { id: '05', layout: 'data-viz', html: '05-data.html', audio: '05.mp3', title: '', clauses: [{ stage: 1, text: '' }, { stage: 2, text: '' }] },
+    { id: '06', layout: 'code', html: '06-code.html', audio: '06.mp3', title: '', clauses: [{ stage: 1, text: '' }, { stage: 2, text: '' }] },
+    { id: '07', layout: 'quote', html: '07-quote.html', audio: '07.mp3', title: '', clauses: [{ stage: 1, text: '' }, { stage: 2, text: '' }] },
+    { id: '08', layout: 'closing', html: '08-closing.html', audio: '08.mp3', title: '', clauses: [{ stage: 1, text: '' }] },
+  ],
+};
+
+// 覆盖保护(必须在建目录之前检查, 否则会把自建的空子目录当成"非空"): 目标目录已存在且
+// 非空 → 拒绝。init 会重置 script.json / notes.md 等 5 个文件, 误跑到已开工的项目上会毁掉
+// 全部进度, 所以必须显式 --force。
+const FORCE = argv.includes('--force');
+const GENERATED = ['script.json', 'slides/tokens.css', 'slides/_template.html', 'assets/MANIFEST.md', 'research/notes.md'];
+
+// --upgrade-css / --check-css: 只管 tokens.css 里三个工具箱受管块的新旧, 不生成/不重置任何文件。
+// 老项目缺 no-fx 规则时 <html class="no-fx"> 会静默失效(只关动画不把 opacity 抬回来, 页面反而空白);
+// 缺图表/表格工具箱时新配方会静默半死(条形不生长 / 环形不扫出 / 数字不滚动)。
+// 判定按内容 rev(见 css-kit.mjs), 不再是"文件里出现过标记字符串就算有" —— 那种判定在项目补过
+// 一次后就永远报"无需升级", 源头后续改动再也传不下去(2026-09-18 排查定案的根因)。
+if (argv.includes('--upgrade-css') || argv.includes('--check-css')) {
+  // 落盘收监(与 capture/build-video 同一道): slides/ 这一段**本身**可能是指向项目外的符号链接,
+  // 那样 tokens.css(以及 .bak)就会写到项目外, 而命令照报成功(2026-09-18 复查 B4)
+  const cssPath = safeOut(dir, 'slides', 'tokens.css');
+  if (!fs.existsSync(cssPath)) { console.error(`✗ 找不到 ${cssPath}`); process.exit(1); }
+  const css = fs.readFileSync(cssPath, 'utf8');
+  if (css.length > MAX_SCAN_BYTES) { console.error(`✗ tokens.css 有 ${Math.round(css.length / 1e4) / 100}MB, 超过 ${MAX_SCAN_BYTES / 1e6}MB 上限拒绝扫描(正常项目 ≈15KB; 构造的超大输入会让受管块定位二次方变慢)`); process.exit(1); }
+
+  if (argv.includes('--check-css')) {
+    const st = kitStatuses(css, { expected: { text: generateTokensCss(), rev: TOKENS_REV } });
+    const bad = st.filter(s => s.status !== 'ok');
+    if (!bad.length) {
+      console.log(`✓ tokens.css 已是最新(tokens rev ${TOKENS_REV.slice(0, 8)} / kit rev ${KIT_REV.slice(0, 8)})`);
+      // ok 但带说明的项也打出来: "老格式(裸文本)"、"区外有你自己的覆写"这类信息不影响渲染,
+      // 但决定了下次升级会不会动你的文件; 藏起来只会让人以为工具没看这块(2026-09-18 复查)。
+      for (const s of st.filter(x => x.detail)) console.log(`  提示 ${s.label}: ${s.detail}`);
+      process.exit(0);
+    }
+    for (const s of bad) console.error(`✗ ${s.label}: [${s.status}] ${s.detail}`);
+    console.error('  → node scripts/init-project.mjs <项目目录> --upgrade-css(原地替换受管区, 不动区外的规则)');
+    process.exit(1);
+  }
+
+  const { css: next, actions, normalizedLineEndings } = applyKitUpgrade(css, { tokensText: generateTokensCss(), tokensRev: TOKENS_REV });
+  if (!actions.length) { console.log(`无需升级: ${cssPath} 已是最新(tokens rev ${TOKENS_REV.slice(0, 8)})`); process.exit(0); }
+  const bakPath = safeOut(dir, 'slides', 'tokens.css.bak');   // .bak 也要收监: 它是同一个目录段下的叶子文件
+  if (fs.existsSync(bakPath)) console.warn('⚠ 已存在 tokens.css.bak, 将被本次升级前的备份覆盖(旧备份会丢, 需要留就先改名)');
+  fs.writeFileSync(bakPath, css);   // 覆盖前备份: 升级只应动受管块, 万一不对可整文件回退
+  fs.writeFileSync(cssPath, next);
+  for (const a of actions) console.log(`✓ ${a.label}: ${a.reason}`);
+  if (normalizedLineEndings) console.log('  行尾已统一为 LF(与 rev 哈希同一标准; CSS 语义不变)');
+  console.log(`  备份 → ${cssPath}.bak;受管块之外的内容(含你的覆写)未动`);
+  // 升级只改 tokens.css —— 已渲染产物里还是旧 CSS 的画面, 必须点名要重跑什么(流程审计 P3)
+  console.log('  ⚠ 已生成的 preview/*.png 与 build/frames 仍是旧 CSS 画面: 受影响张重跑 capture(--mode still + --mode motion), 放映页重跑 preview-page.mjs, 再 build-video');
+  process.exit(0);
+}
+
+if (fs.existsSync(dir)) {
+  const existing = fs.readdirSync(dir).filter(e => !GENERATED.includes(e));
+  const wouldOverwrite = GENERATED.filter(f => fs.existsSync(path.join(dir, f)));
+  if ((existing.length || wouldOverwrite.length) && !FORCE) {
+    console.error(`✗ 目标目录已存在且非空: ${dir}\n  init 会重置这些文件(其余不动): ${GENERATED.join(' · ')}`);
+    if (wouldOverwrite.length) console.error(`  其中已存在、将被覆盖的: ${wouldOverwrite.join(' · ')}`);
+    console.error('  确认要重新初始化请加 --force');
+    process.exit(1);
+  }
+  if (FORCE && wouldOverwrite.length) console.warn(`⚠ --force: 将覆盖 ${wouldOverwrite.length} 个生成文件(其余内容不动): ${wouldOverwrite.join(' · ')}`);
+}
+
+// 先把项目根建出来(它就是本次要创建的东西), 之后的每个派生写点才有一个"真实存在"的根可比对:
+// safeOut 比对的是"最深已存在祖先的 realpath 是否仍在根内", 根不存在时会拿父目录去比 → 全部误判越界。
+// 有了根之后, 骨架写点与 --upgrade-css 走同一道收监: --force 重建一个 slides 是指向项目外的
+// junction 的目录时, 不会再把生成物写到项目外(2026-09-18 复查: 这条此前只有升级路径收监了)。
+fs.mkdirSync(dir, { recursive: true });
+for (const d of ['research', 'assets', 'slides', 'audio', 'preview', 'out', 'build', 'asr']) {
+  fs.mkdirSync(safeOut(dir, d), { recursive: true });
+}
+fs.writeFileSync(safeOut(dir, 'slides', 'tokens.css'),
+  '/* 生成物: 下面的 tokens 受管区由 init-project.mjs 生成, 升级走 --upgrade-css 原地替换;\n' +
+  '   你自己的规则写到受管区之外(文件末尾), 升级不会碰 */\n' +
+  wrapTokens(generateTokensCss(), TOKENS_REV) + '\n');
+fs.writeFileSync(safeOut(dir, 'slides', '_template.html'), TEMPLATE_HTML);
+fs.writeFileSync(safeOut(dir, 'script.json'), JSON.stringify(scriptJson, null, 2) + '\n');
+fs.writeFileSync(safeOut(dir, 'assets', 'MANIFEST.md'),
+  `# 素材清单\n\n| 文件 | 内容 | 来源 | 许可 |\n|---|---|---|---|\n\n<!-- 每个素材一行; 来源必须可核查。禁止使用凭空生成的 logo/截图/头像。选图 SOP 见 references/image-sources.md -->\n`);
+fs.writeFileSync(safeOut(dir, 'research', 'notes.md'),
+  `# 调研笔记${topic ? ` · ${topic}` : ''}\n\n## 核心事实\n<!-- 硬规则(见 references/research.md): 关键数字≥2 个独立来源; 一手优先; 标注口径与日期; 查不到出处的不进脚本 -->\n| 事实 | 数值/表述 | 来源(URL/文档) | 口径日期 | 等级(一手/二手/弱) | 第二来源 |\n|---|---|---|---|---|---|\n\n## 不确定项(不进脚本)\n\n## 不该进脚本的内容(传闻/争议/无法核实)\n\n`);
+
+console.log(`已生成项目骨架: ${dir}`);
+console.log(`
+下一步:
+1. 填 research/notes.md(事实性题材必须先搜集, 过 Gate 0)
+2. 逐张填 script.json 的 clauses(内容量规则见 references/authoring.md, 过 Gate 1)
+3. 做 TTS 到 audio/<id>.mp3, 然后:
+   node scripts/plan-timings.mjs "${dir.replace(/\\/g, '/')}"
+4. 每张 HTML 参照 slides/_template.html 写到 slides/(主题见 tokens.css 顶部注释), 然后:
+   node scripts/check-theme.mjs "${dir.replace(/\\/g, '/')}"            # 主题对比度校验
+   node scripts/capture.mjs "${dir.replace(/\\/g, '/')}" --mode still    # Gate 4 终态预览
+   node scripts/capture.mjs "${dir.replace(/\\/g, '/')}" --mode motion   # 动画帧
+   node scripts/build-video.mjs "${dir.replace(/\\/g, '/')}" --asr`);
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/limits.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/limits.mjs
new file mode 100644
index 00000000..e4275788
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/limits.mjs
@@ -0,0 +1,11 @@
+// html2video-for-mcode · 扫描上限(单一来源)
+//
+// 为什么单独一个文件: 这个上限对 **tokens.css 与 slide HTML 都生效**(check-slides / preview-page /
+// capture 两侧都过同一道门), 但它原先住在 css-kit.mjs 里 —— 那是"受管块扫描"引入的, 于是 HTML 侧的
+// 调用点看起来像在依赖 CSS 模块。搬到 tools.mjs 会成环(tools 已经 import css-kit 取 kitStatuses),
+// 所以放在这个谁都能 import 的叶子模块里。
+//
+// 数值的理由: 受管块定位的正则对"无闭合定界符 × N"的构造输入是二次方复杂度 —— 2026-09-18 审计实测
+// 4MB 恶意 tokens.css 会让 --check-css 跑 39s。正常 tokens.css ≈15KB、单张 slide ≈10KB,
+// 2MB 只可能是构造出来的, 所以入口直接拒绝扫描并说明原因。
+export const MAX_SCAN_BYTES = 2_000_000;
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/nofx-css.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/nofx-css.mjs
new file mode 100644
index 00000000..b6808a47
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/nofx-css.mjs
@@ -0,0 +1,19 @@
+// html2video-for-mcode · no-fx 规则的唯一来源。
+// init-project 把它写进 tokens.css; preview-page 在"老项目 tokens.css 里没有这段"时兜底注入,
+// 否则放映页的"关动效对照"会把画面锁在入场前的透明态(看起来像整片空白)。
+// 两处共用一份文本, 避免 CSS 漂移。
+
+export const NOFX_CSS = `/* 一键关全部动效: <html class="no-fx"> 或 .stage.no-fx。
+   注意必须同时把 [data-stage] 的基础态 opacity:0 拉回来 —— 入场效果靠 animation 的 both
+   填充从 0 拉到 1, 只关动画不管基础态 = 元素全部隐形(2026-09-18 实测踩过)。
+   关掉后 motion 捕获自动退化为静态帧, 成片照常出; 字幕(.kit-sub)不受影响。 */
+.no-fx [class*="fx-"], .no-fx .fx-stagger > *, .no-fx .fx-shimmer::after { animation: none !important; }
+.no-fx [data-stage], .no-fx .fx-stagger > * {
+  opacity: 1 !important; transform: none !important; filter: none !important; clip-path: none !important;
+}
+.no-fx .fx-draw { stroke-dasharray: none !important; stroke-dashoffset: 0 !important; }`;
+
+// 判定标准落在"承重"的那条规则上: 只关动画、没把 opacity 抬回来的旧版一律算没有。
+export function hasNofxRules(css) {
+  return /\.no-fx\s+\[data-stage\][^{]*\{[^}]*opacity\s*:\s*1\s*!important/.test(String(css ?? ''));
+}
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/plan-timings.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/plan-timings.mjs
new file mode 100644
index 00000000..7de1b88d
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/plan-timings.mjs
@@ -0,0 +1,136 @@
+#!/usr/bin/env node
+// html2video-for-mcode · 实测对时: script.json + audio/*.mp3 → build/timings.json
+// 这是全流水线时长的唯一事实来源。用法: node plan-timings.mjs <项目目录> [--pacing=<属性值>]
+// 语种由 script.json 的 lang 决定(zh 默认 / en / yue / 其他 BCP-47), 影响语速基准与字幕行宽阈值。
+// timings.json 每个 slide 含 clauses[]: 每句口播的估算开口时刻/时长, 供字幕、ASR 按句切分、对时校准共用。
+import fs from 'node:fs';
+import path from 'node:path';
+import { spawnSync } from 'node:child_process';
+import { positionalDir, probeDuration, requireTool, safeId, safeOut, safeRel, subLineCap, validateScriptPaths } from './tools.mjs';
+
+// 语种相关的计量基准。中文按"字", 英文按"字符"(含词间节奏, 与音节时长大致成正比)。
+// pacing = 每单位每秒的常见语速; emPer = 单字宽按 em 折算(行宽估算用 —— 行宽公式在
+// tools.subLineCap, 字幕胶囊几何的唯一来源也在 tools.SUB_GEOMETRY); pace 区间用于语速异常预警。
+// yearGate = 年份逐位读的闸门只对中/粤语面开(第 24 轮: 用 emPer===1 充当语族判别是排版
+// 属性代理语义, 未来 emPer:1 的语种会误触); 未知语种 fallback 不开闸。
+const ZH_BASE = { unit: '字', pacing: 4.8, paceMin: 3, paceMax: 6.5, emPer: 1, word: null, yearGate: true };
+const LANG_CFG = {
+  zh: ZH_BASE,
+  yue: ZH_BASE,
+  en: { unit: '字符', pacing: 14, paceMin: 9, paceMax: 18, emPer: 0.44, word: 'words', yearGate: false },
+};
+const langCfg = code => LANG_CFG[code] ?? { unit: '字符', pacing: 14, paceMin: 8, paceMax: 20, emPer: 0.44, word: null, yearGate: false };
+
+const argv = process.argv.slice(2);
+const dir = positionalDir(argv);
+const LEAD = 0.2; // 视觉比语音提前出现秒数(广播惯例, 观感同步)
+
+process.env.KIT_PROJECT_DIR = dir;
+const FFPROBE = requireTool('ffprobe', dir);
+
+const scriptPath = path.join(dir, 'script.json');
+if (!fs.existsSync(scriptPath)) { console.error(`✗ 找不到 ${scriptPath}`); process.exit(1); }
+const script = JSON.parse(fs.readFileSync(scriptPath, 'utf8'));
+validateScriptPaths(script, dir); // script.json 是 agent 可编辑文件: id/audio 派生路径先收监
+const fps = script.fps ?? 30;
+const LANG = script.lang ?? 'zh';
+const CFG = langCfg(LANG);
+const pacingArg = argv.find(a => a.startsWith('--pacing='));
+const pacing = pacingArg ? parseFloat(pacingArg.slice(9)) : CFG.pacing;
+// 试听时与用户定的语速(1.6.0): script.speed 此前是纯声明、没有任何脚本读它 → "语速偏慢"只能靠人耳发现。
+// 现在拿它当期望值: 实测语速(字数÷实测音频时长)偏离 期望基准×speed 超过 20% 就告警, 越界也告警。
+const SPEED = (() => {
+  const v = script.speed;
+  const n = Number(typeof v === 'object' && v !== null ? v.default : v);
+  return Number.isFinite(n) && n > 0 ? n : 1.0;
+})();
+
+const probeDur = file => probeDuration(FFPROBE, file);   // 探测收进 tools.mjs 单一实现(第十三轮 review 去重)
+const charCount = s => String(s ?? '').replace(/\s+/g, '').length;
+const r3 = x => Math.round(x * 1000) / 1000;
+
+const rows = [];
+const warns = [];
+for (const s of script.slides) {
+  const audioPath = safeRel(path.join(dir, 'audio'), s.audio ?? `${safeId(s.id)}.mp3`, { where: `slides[${s.id}].audio` });
+  if (!fs.existsSync(audioPath)) { console.error(`✗ 缺音频 ${audioPath} — 先完成 Phase 2 TTS`); process.exit(1); }
+  const tts = probeDur(audioPath);
+  if (tts == null) { console.error(`✗ ffprobe 读不出时长: ${audioPath}`); process.exit(1); }
+
+  // 时间权重按**发音形态**(say ?? text)算 —— 音频念的是 say(第 24 轮: "82.3%"显示 5 字、
+  // 发音"百分之八十二点三"8 字, 按显示形态分摊会把长 say 句的开口系统性估早, 字幕窗与
+  // ASR 切分跟着偏)。字幕行宽(下方 wrap 检查)仍按显示形态 text —— 胶囊几何是屏幕上的事。
+  const spokenLen = c => charCount(c.say ?? c.text);
+  const total = s.clauses.reduce((n, c) => n + spokenLen(c), 0);
+  if (total === 0) warns.push(`${s.id}: clauses 为空, 将整张静止`);
+
+  // 第 k 句开口时刻 ≈ 实测时长 × (前 k-1 句字数占比); stage 取该层最早一句, 再提前 LEAD
+  const clauses = [];
+  const stageTime = {};
+  let cum = 0;
+  for (const c of s.clauses) {
+    const start = total > 0 ? (tts * cum) / total : 0;
+    // say = 发音形态(只喂 TTS/ASR), text = 显示形态(字幕/画面) —— 原样传递给下游; chars = 发音字数(计时簿记)
+    clauses.push({ stage: c.stage ?? null, start: r3(start), chars: spokenLen(c), text: c.text, ...(c.text2 ? { text2: c.text2 } : {}), ...(c.say ? { say: c.say } : {}) });
+    const t = Math.max(0, start - LEAD);
+    if (c.stage != null) stageTime[c.stage] = c.stage in stageTime ? Math.min(stageTime[c.stage], t) : t;
+    cum += spokenLen(c);
+  }
+  clauses.forEach((c, i) => { c.dur = r3((clauses[i + 1]?.start ?? tts) - c.start); });
+  Object.assign(stageTime, s.stageTimes ?? {}); // 显式 stageTimes 覆盖优先
+
+  const tail = s.tail ?? 0.8;
+  const duration = Math.ceil((tts + tail) * fps) / fps; // 对齐帧网格
+  const wps = tts > 0 ? total / tts : 0;
+
+  if (wps > 0 && (wps < CFG.paceMin || wps > CFG.paceMax)) warns.push(`${s.id}: 语速 ${wps.toFixed(1)} ${CFG.unit}/s (${LANG} 常见 ${CFG.paceMin}–${CFG.paceMax}) — 检查 speed 或字数, 或用 --pacing 重估`);
+  // 与 script.speed 对账(1.6.0): 期望 = 基准 × speed; 实测偏离 >20% 说明该段 TTS 没用这个 speed
+  const expect = CFG.pacing * SPEED;
+  if (wps > 0 && Math.abs(wps - expect) > expect * 0.2) {
+    warns.push(`${s.id}: 实测语速 ${wps.toFixed(1)} ${CFG.unit}/s 与 script.speed=${SPEED} 的期望 ${expect.toFixed(1)} 差 ${(Math.abs(wps - expect) / expect * 100).toFixed(0)}% — 该段 TTS 可能没用这个 speed(试听定的是 ${SPEED}), 或字数估算错了; 复核后重做该段 TTS 或改 script.speed`);
+  }
+  if (duration > 15) warns.push(`${s.id}: ${duration.toFixed(1)}s 超过 15s — 建议拆成两张`);
+  const stages = Object.keys(stageTime).map(Number);
+  const last = stages.length ? Math.max(...stages) : 0;
+  if (last > 0 && duration - (stageTime[last] ?? 0) < 1.2) warns.push(`${s.id}: 最后一个 stage 在 ${stageTime[last].toFixed(1)}s, 距收尾不足 1.2s — 观众看不清, 建议 tail 加大或精简口播`);
+  // 字幕行宽来自 tools.subLineCap(胶囊几何唯一来源; 缩放带钳位, 按宽度线性外推是错的 ——
+  // 2026-09-23 复核: 中文 1080≈24 字/行、1920≈33、2560≈35)。折 2 行可读, 折 3 行起才警告。
+  const subCap = subLineCap(script.width ?? 1920, CFG.emPer);
+  for (const c of clauses) {
+    const n = charCount(c.text);
+    const lines = Math.ceil(n / subCap);
+    if (lines >= 3) warns.push(`${s.id} 第 ${clauses.indexOf(c) + 1} 句 ${n} ${CFG.unit}(行宽 ≈${subCap} ${CFG.unit}/行) 会折 ${lines} 行 — 3 行起可读性差, 建议拆句(≤2 行可接受)`);
+    if (c.text2 && c.text2.length > 60) warns.push(`${s.id} 第 ${clauses.indexOf(c) + 1} 句双语第二行 ${c.text2.length} 字符 > 60 — 建议精简译文`);
+    // 年份读法闸门(2026-09-28 用户实测: TTS 把 2026年 念成"两千零二十六年"): 显示与发音
+    // 分离 —— text 保持 2026年 上字幕, 加 say:'二零二六年' 控制发音。判别只钉无歧义的
+    // "四位数字+年"(任意年份都逐位读, 不止 19xx/20xx —— 第 24 轮: 旧正则漏 1897年 这类;
+    // 整读/逐位读是意图问题, 2026点 就该整读 —— 其余数字写法走文档纪律)。
+    if (CFG.yearGate) {
+      const spoken = c.say ?? c.text;
+      const ym = /(?<!\d)\d{4}\s*年(?!\d)/.exec(spoken);   // 数字边界: 12026年 这类长串截出的"2026年"不是年份
+      if (ym) {
+        const digitsRead = ym[0].replace(/\D/g, '').split('').map(d => '零一二三四五六七八九'[+d]).join('');
+        warns.push(`${s.id} 第 ${clauses.indexOf(c) + 1} 句 "${ym[0]}" 会被 TTS 按整数读(2026年→"两千零二十六年") — 年份按播报惯例逐位读: text 原样上字幕, 加 say:"${digitsRead}年…"`);
+      }
+    }
+  }
+
+  rows.push({
+    id: s.id, tts: r3(tts), duration: r3(duration),
+    chars: total, wps: +wps.toFixed(2), stages: stageTime, clauses,
+    script: s.clauses.map(c => c.text).join(''),
+  });
+}
+
+const totalDur = rows.reduce((n, r) => n + r.duration, 0);
+fs.mkdirSync(safeOut(dir, 'build'), { recursive: true });
+fs.writeFileSync(safeOut(dir, 'build', 'timings.json'),
+  JSON.stringify({ fps, lang: LANG, pacing, lead: LEAD, total: r3(totalDur), slides: rows }, null, 2) + '\n');
+
+const unitLabel = `量(${CFG.unit})`;
+console.table(rows.map(({ id, tts, duration, chars, wps, stages }) =>
+  ({ id, 'TTS(s)': tts, '成片(s)': duration, [unitLabel]: chars, [`${CFG.unit}/s`]: wps, 'stage时刻': JSON.stringify(stages) })));
+console.log(`总时长: ${totalDur.toFixed(1)}s · 语言 ${LANG}(${CFG.unit}基准 ${pacing}/${CFG.unit}·s⁻¹ · script.speed=${SPEED}) → build/timings.json`);
+if (SPEED < 0.8 || SPEED > 1.4) warns.unshift(`script.speed = ${SPEED} 超出常规区间 0.8–1.4 — 确认是不是写错(或 TTS 调用与它不一致)`);
+if (warns.length) { console.warn('\n⚠ 警告:'); for (const w of warns) console.warn('  - ' + w); }
+console.log('\n下一步可选: node scripts/check-timing.mjs <项目目录>  用静音检测实测每句开口时刻, 对比/校准估算。');
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/prep-image.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/prep-image.mjs
new file mode 100644
index 00000000..467f60d6
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/prep-image.mjs
@@ -0,0 +1,92 @@
+#!/usr/bin/env node
+// html2video-for-mcode · 配图准备工具(选图 SOP 的执行辅助) — 只用 ffprobe/ffmpeg, 不依赖 Python
+// 用法:
+//   node prep-image.mjs --check <图1> [图2 ...]            看尺寸/比例/裁切风险, 给人做取舍
+//   node prep-image.mjs --crop <in> <out> [--ratio 16:9] [--anchor bottom|top|center]
+// 裁切硬限制(与 image-sources.md 的 SOP 一致): 主体必须完整可见、裁掉面积 ≤20%、输出严格目标比例。
+// 优先靠换图/换版式解决, 本工具只是最后的兜底手段。
+// 边界契约(1.7.5 复查④): --crop 的 in/out 都是**命令行显式给出**的路径(与你直接跑 ffmpeg 同一
+// 信任级), 不在"项目 + 派生路径"的收监范围内; 覆盖已存在 out 仍需 --force。此为 README 声明过的唯一例外。
+import fs from 'node:fs';
+import path from 'node:path';
+import { spawnSync } from 'node:child_process';
+import { flagValue, positionals, probeSize, requireTool } from './tools.mjs';
+
+const argv = process.argv.slice(2);
+// --check / --crop 是布尔开关, 它们后面那串文件名本来就是位置参数(所以不在 tools.VALUE_FLAGS 里)
+const positional = positionals(argv);
+
+// 按当前工作目录找项目内的 ffmpeg-static / ffprobe-static(二审 P2: README 明说支持装在视频项目里)
+const FFMPEG = requireTool('ffmpeg', process.cwd());
+const FFPROBE = requireTool('ffprobe', process.cwd());
+const win = { encoding: 'utf8', windowsHide: true };
+
+function probe(file) {
+  const s = probeSize(FFPROBE, file);   // 探测收进 tools.mjs 单一实现(第十三轮 review 去重)
+  return s ? { w: s[0], h: s[1] } : null;
+}
+const gcd = (a, b) => (b ? gcd(b, a % b) : a);
+const ratioLabel = (w, h) => { const g = gcd(w, h); return `${w / g}:${h / g} (${(w / h).toFixed(2)})`; };
+
+if (argv.includes('--check') || (!argv.includes('--crop') && positional.length)) {
+  const files = positional;
+  if (!files.length) { console.error('用法: node prep-image.mjs --check <图1> [图2 ...]'); process.exit(1); }
+  let risk = 0;
+  for (const f of files) {
+    if (!fs.existsSync(f)) { console.error(`✗ 不存在: ${f}`); process.exitCode = 1; continue; }
+    const s = probe(f);
+    if (!s) { console.error(`✗ 读不出尺寸: ${f}`); process.exitCode = 1; continue; }
+    const r = s.w / s.h;
+    const ratio = 16 / 9;
+    const notes = [];
+    if (Math.abs(r - ratio) < 0.06) notes.push('比例≈16:9, 可直接套 .img-frame 默认框');
+    else if (r > ratio) {
+      // 比 16:9 更宽 → 裁宽度
+      notes.push(`偏宽: 保持高度裁宽度 ${(100 * (1 - (s.h * ratio) / s.w)).toFixed(0)}% 可到 16:9`);
+    } else {
+      // 比 16:9 更高(窄) → 裁高度
+      notes.push(`偏高(窄): 保持宽度裁高度 ${(100 * (1 - (s.w / ratio) / s.h)).toFixed(0)}% 可到 16:9`);
+    }
+    if (s.w < 1200) notes.push('分辨率 <1200px, 上屏会软 — 建议换更大的图');
+    if (Math.abs(r - ratio) >= 0.06) { notes.push('⚠ 主体若在边缘, 先换图(搜「全景/全貌/正面/远景」), 不要硬裁'); risk++; }
+    console.log(`${path.basename(f)}\n  尺寸 ${s.w}×${s.h} · 比例 ${ratioLabel(s.w, s.h)}\n  ${notes.join('\n  ')}`);
+    console.log('  下一步: 肉眼确认主体位置 → 主体在边缘就换图; 居中可小裁, 或改用 .img-frame.contain 留白\n');
+  }
+  if (risk) console.log(`提示: ${risk} 张存在裁切风险 —— 按 references/image-sources.md 的 SOP 优先重搜素材。`);
+  process.exit(process.exitCode ?? 0);
+}
+
+if (argv.includes('--crop')) {
+  const [src, dst] = positional;
+  if (!src || !dst) { console.error('用法: node prep-image.mjs --crop <in> <out> [--ratio 16:9] [--anchor bottom|top|center] [--force]'); process.exit(1); }
+  const FORCE = argv.includes('--force');
+  if (fs.existsSync(dst) && !FORCE) { console.error(`✗ 输出已存在, 不覆盖: ${dst}(要覆盖加 --force)`); process.exit(1); }
+  const [rw, rh] = (flagValue(argv, '--ratio', '16:9')).split(':').map(Number);
+  const anchor = flagValue(argv, '--anchor', 'center');
+  if (!rw || !rh) { console.error('✗ --ratio 形如 16:9'); process.exit(1); }
+  const s = probe(src);
+  if (!s) { console.error(`✗ 读不出尺寸: ${src}`); process.exit(1); }
+  const target = rw / rh, cur = s.w / s.h;
+
+  // 目标裁切框: 比目标更宽 → 裁宽; 比目标更高(窄) → 裁高。绝不放大补边。
+  let cw, ch, x, y;
+  if (cur > target) { ch = s.h; cw = Math.round(s.h * target); }
+  else { cw = s.w; ch = Math.round(s.w / target); }
+  if (cw > s.w || ch > s.h) { console.error('✗ 该比例需要放大补边, 裁不出来 —— 改用 .img-frame.contain 留白'); process.exit(1); }
+  const cutFrac = 1 - (cw * ch) / (s.w * s.h);
+  if (cutFrac > 0.2) {
+    console.error(`✗ 需裁掉 ${(cutFrac * 100).toFixed(0)}% 面积(硬限制 ≤20%)——按 SOP 先换图或用 .img-frame.contain 留白, 不要硬裁`);
+    process.exit(1);
+  }
+  x = Math.round((s.w - cw) / 2);
+  y = anchor === 'top' ? 0 : anchor === 'bottom' ? s.h - ch : Math.round((s.h - ch) / 2);
+  const r = spawnSync(FFMPEG, ['-y', '-v', 'error', '-i', src, '-vf', `crop=${cw}:${ch}:${x}:${y}`, dst], win);
+  if (r.status !== 0) { console.error('✗ 裁切失败:\n' + (r.stderr || '')); process.exit(1); }
+  const out = probe(dst);
+  console.log(`✓ ${path.basename(dst)}  ${s.w}×${s.h} → ${out.w}×${out.h} (${ratioLabel(out.w, out.h)}), 裁掉 ${(cutFrac * 100).toFixed(1)}% 面积, 锚点 ${anchor}`);
+  console.log('  裁完必须肉眼核对主体完整可见 —— 宁可改用 .img-frame + --img-pos, 也不要裁到主体。');
+  process.exit(0);
+}
+
+console.error('用法:\n  node prep-image.mjs --check <图...>\n  node prep-image.mjs --crop <in> <out> [--ratio 16:9] [--anchor bottom|top|center]');
+process.exit(1);
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/preview-page.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/preview-page.mjs
new file mode 100644
index 00000000..4928c96b
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/preview-page.mjs
@@ -0,0 +1,728 @@
+#!/usr/bin/env node
+// html2video-for-mcode · 放映页: preview/play/index.html + 逐张"真实时序"快照。
+// 用法: node preview-page.mjs <项目目录> [--open] [--no-script]
+//
+// 定位: **用浏览器把 HTML 画面放一遍**。只做四件事 —— 翻页(动效开 = 逐级入场)、动效开/关对照、口播开/关、总览;
+// 口播文案是锦上添花(可关), 不做计时器/进度条/播放器那套 UI(要看时间就直接看成片)。
+// 换帧用双缓冲 iframe: 新帧在隐藏帧里加载完才对调显示, 不闪白; UI 文案随 script.lang 中英双语。
+//
+// 为什么不是"直接双击 slides/*.html": 动画延迟(--t1/--t2/--t3)与画布尺寸是渲染管线按
+// timings.json 注入的(tokens.css 里只有占位默认值 --t2:800ms), 直接打开看到的是"所有动画
+// 挤在 2 秒内"的假象。本脚本把注入值原样写进副本的 <html style>, 于是浏览器里的播放时序
+// = 成片时序, 且副本带 <base href> 指回 slides/, 主题与素材照常解析。
+// 另出一份加了 no-fx 的副本, 页面按 X 即可对照"关掉动效后画面是否还完整" —— 这是
+// "元素永久不可见 / 关动效反而更空"那类静默故障的 5 秒自检, 不必等 3–6 分钟的 motion 编码。
+//
+// 口播 UI 按数据决定加不加载: 没有 clauses → 完全不出; 有 clauses没对时 → 只列文案;
+// --no-script 强制不出。布局随窗口自适应(窄窗口/手机口播面板收成底部抽屉, 触屏可左右滑动翻页)。
+import fs from 'node:fs';
+import path from 'node:path';
+import { spawn } from 'node:child_process';
+import { expectedTokens, isMainModule, positionalDir, safeId, safeOut, safeRel, validateScriptPaths } from './tools.mjs';
+import { NOFX_CSS } from './nofx-css.mjs';
+import { KITS, TOKENS_ID, findAllBlocks, kitStatuses, MAX_SCAN_BYTES } from './css-kit.mjs';
+import { generateTokensCss } from './tokens-template.mjs';
+
+const esc = s => String(s ?? '').replace(/[&<>"']/g, c => ({ '&': '&amp;', '<': '&lt;', '>': '&gt;', '"': '&quot;', "'": '&#39;' }[c]));
+
+// ─────────────────────────── 纯函数(供 tests/ 单测) ───────────────────────────
+
+// 与 capture.mjs 的注入保持同一语义: 秒 → 整数毫秒, 负数归零。属性名 --t<stage>。
+export function stageVars(stages = {}, { w, h } = {}) {
+  const out = [];
+  for (const k of Object.keys(stages).sort()) {
+    const v = Number(stages[k]);
+    if (!Number.isFinite(v)) continue;
+    out.push(`--t${Number(k)}:${Math.max(0, Math.round(v * 1000))}ms`);
+  }
+  if (Number.isFinite(w)) out.push(`--stage-w:${w}px`);
+  if (Number.isFinite(h)) out.push(`--stage-h:${h}px`);
+  return out.join(';');
+}
+
+// 从 slide HTML 里读出用到哪些 stage —— 用于"还没对时"时的等间隔兜底
+export function stagesFromHtml(html) {
+  const ids = new Set();
+  for (const m of String(html ?? '').matchAll(/\bdata-stage\s*=\s*["'](\d+)["']/g)) ids.add(Number(m[1]));
+  return [...ids].sort((a, b) => a - b);
+}
+
+// 没有 timings.json(口播还没做)时, 按 0.3s 起步、每层 +1s 错开, 让人至少看清入场顺序;
+// 页面会明确标注"不是成片时序"。tokens.css 的占位值会把所有动画挤在 2 秒内, 更看不懂。
+export function fallbackStages(html, { start = 0.3, step = 1 } = {}) {
+  const out = {};
+  stagesFromHtml(html).forEach((n, idx) => { out[n] = start + idx * step; });
+  return out;
+}
+
+// 属性匹配要认双引号/单引号/无引号三种写法(二审 P2): 旧实现只认双引号, 遇到
+// <html class='theme' style='--t2:800ms'> 会再插一组重复属性, 浏览器保留**先出现的那组** →
+// 实测延迟/画布尺寸全部失效、no-fx 也加不上(实测: 原延迟仍在、没有画布宽、动效关不掉)
+// 前置不能紧跟 `-`/字母数字: 否则 data-style / data-class 这类属性名里含 style/class 的,
+// 会被误认成真属性(于是真 style 没被改、画布尺寸与动画延迟静默丢失)。
+const ATTR_RE = name => new RegExp(`(?<![-\\w])${name}\\s*=\\s*("([^"]*)"|'([^']*)'|([^\\s"'>]+))`, 'i');
+const attrValue = (tag, name) => {
+  const m = ATTR_RE(name).exec(tag);
+  return m ? (m[2] ?? m[3] ?? m[4]) : null;
+};
+
+// 属性值重新序列化进双引号前必须转义(2026-10-10 独立安全审计 H1): 值可能是从**单引号**
+// 属性里取出来的原文, 原样塞进双引号就能提前闭合属性并追加一个活的 on* 处理器 ——
+// 实测 <html class='x" onmouseover=window.FIRED=1 data-x='> 经 addNoFx 后变成
+// class="x" onmouseover=window.FIRED=1 data-x= no-fx", 打开放映页即任意 JS 执行。
+// & 先转, 免得把作者的 &amp; 再编一次(& 本身是实体前缀)。
+const escapeAttrValue = v => String(v)
+  .replace(/&/g, '&amp;').replace(/"/g, '&quot;').replace(/</g, '&lt;').replace(/>/g, '&gt;');
+
+function firstTag(html, name) {
+  // 属性值里可以合法出现 >, 所以标签体不能简单用 [^>]* 一刀切(否则 style='--x: calc(1>0)'
+  // 之后的属性全部读不到); 双引号/单引号区段放行。
+  // 注释里的 <html> 是诱饵: 命中它就会去改注释, 真标签反而没被改(no-fx 加不上 → X 键对照
+  // 静默失效, 画布尺寸/动画延迟也丢)。所以先算出注释区段, 落在里面的命中一律跳过。
+  const comments = [
+    ...[...html.matchAll(/<!--[\s\S]*?-->/g)].map(m => [m.index, m.index + m[0].length]),
+    ...[...html.matchAll(/<!--[\s\S]*$/g)].map(m => [m.index, m.index + m[0].length]),
+  ];
+  const inComment = i => comments.some(([a, b]) => i >= a && i < b);
+  const re = new RegExp(`<${name}\\b(?:[^>"']|"[^"]*"|'[^']*')*>`, 'gi');
+  for (const m of html.matchAll(re)) {
+    if (!inComment(m.index)) return { tag: m[0], index: m.index };
+  }
+  return null;
+}
+
+function setAttr(tag, name, value) {
+  const v = escapeAttrValue(value);
+  if (attrValue(tag, name) !== null) return tag.replace(ATTR_RE(name), () => `${name}="${v}"`);
+  return tag.replace(/\s*\/?>$/, m => ` ${name}="${v}"${m.endsWith('/>') ? '/>' : '>'}`);
+}
+
+function addStyleDecl(tag, decl) {
+  if (!decl) return tag;
+  const cur = attrValue(tag, 'style');
+  if (cur === null) return setAttr(tag, 'style', decl);
+  return setAttr(tag, 'style', `${cur.trim().replace(/;?$/, ';')}${decl}`);
+}
+
+// 把注入值挂在 <html> 上(等价于 capture 的 documentElement.style.setProperty, 优先级最高)
+export function injectHtmlVars(html, decl) {
+  const t = firstTag(html, 'html');
+  if (!t) return html;
+  return html.slice(0, t.index) + addStyleDecl(t.tag, decl) + html.slice(t.index + t.tag.length);
+}
+
+export function addNoFx(html) {
+  const t = firstTag(html, 'html');
+  if (!t) return html;
+  const cls = attrValue(t.tag, 'class');
+  if (cls !== null && /\bno-fx\b/.test(cls)) return html;
+  const tag = setAttr(t.tag, 'class', cls === null ? 'no-fx' : `${cls.trim()} no-fx`);
+  return html.slice(0, t.index) + tag + html.slice(t.index + t.tag.length);
+}
+
+// <base> 必须排在所有相对 URL 之前(slides/ 内含 <link href="tokens.css">, 若排在后面会失效)
+export function addBaseHref(html, href) {
+  const existing = html.match(/<base\b[^>]*>/i);
+  if (existing) return html.replace(existing[0], `<base href="${href}">`);
+  const h = firstTag(html, 'head');
+  if (!h) return html;
+  const at = h.index + h.tag.length;
+  return html.slice(0, at) + `\n<base href="${href}">` + html.slice(at);
+}
+
+// 兜底: 项目 tokens.css 落后于技能当前版本时, 把缺的那段注入副本(排在 tokens.css 之后, 后写覆盖)
+export function injectStyle(html, css, note = '本项目 tokens.css 缺这段规则') {
+  const block = `<style>/* preview-page 兜底注入: ${note} */\n${css}\n</style>\n`;
+  const i = html.search(/<\/head>/i);
+  return i < 0 ? block + html : html.slice(0, i) + block + html.slice(i);
+}
+
+// 动效开的**逐级步进**: 往副本 <head> 开头注入一段在 CSS 动画启动前执行的脚本 ——
+// ?s=k 表示已揭示到第 k 个用到的 stage: 第 k 级带 &anim=1 时当场入场, 更早的级直接终态
+// (-60s 延迟 = 动画早已完成), 更晚的级保持隐藏(+60s 延迟 = 还在等, 等用户按"下一步")。
+// 父页只换 iframe src(与快照同一套机制), 不做任何跨文档访问 —— file:// 下 iframe 是独立源。
+export function addStepScript(html, stageIds = []) {
+  // 形参叫 stageIds: 传进来的是 stagesFromHtml(html) 的 **stage 编号数组**;
+  // 同文件 main() 里另有局部量 stages = timings 的 {1:0.3} 字典 —— 同名不同义, 维护时极易改错(第 25+ 轮外部报告 P3)
+  if (!stageIds.length) return html;
+  const script = `<script>
+/* 放映页逐级步进(动效开): 在动画启动前按 ?s=k 重设各级入场延迟 */
+(function(){try{
+  var q=new URLSearchParams(location.search);var k=parseInt(q.get('s')||'0',10);
+  if(!k)return;var anim=q.get('anim')==='1';var used=${JSON.stringify(stageIds)};
+  used.forEach(function(n,idx){var pos=idx+1,v;
+    if(pos<k)v='-60000ms';else if(pos===k)v=anim?'0ms':'-60000ms';else v='60000ms';
+    document.documentElement.style.setProperty('--t'+n,v);});
+}catch(e){}})();
+</script>`;
+  const m = /<head[^>]*>/i.exec(html);
+  if (!m) return script + html;
+  return html.slice(0, m.index + m[0].length) + '\n' + script + html.slice(m.index + m[0].length);
+}
+
+// ─────────────────────────── 放映页 ───────────────────────────
+
+export function buildPlayPage({
+  topic = '', lang = 'zh', slides = [], cssNote = '', generatedAt = '',
+  narration = true,    // 是否加载口播文案 UI(没有 clauses 或 --no-script 时为 false)
+  timing = false,      // 是否有对时数据(只影响标题上的"(未对时)"标注)
+  fallbackNote = '',   // "还没对时/等间隔预览"的如实说明(与口播 UI 无关, 画面上也要说清)
+  canvas = { w: 1920, h: 1080 },   // 画布尺寸(竖版 1080×1920): 舞台缩放与缩略图比例都按它算
+} = {}) {
+  // UI 双语(实测反馈: 英文项目的放映页整套中文按钮)。lang 来自 script.json(en* → 英文, 其余中文)
+  const T = String(lang).toLowerCase().startsWith('en') ? {
+    title: 'Play', fxOn: 'Motion on', fxOff: 'Motion off', narrOn: 'Narration on', narrOff: 'Narration off',
+    overview: 'Overview', panelH: 'Narration for this slide', untimed: '(untimed)',
+    fxTitle: 'Motion on / off (key X)', narrTitle: 'Narration on / off (key P)',
+    prevTitle: 'Previous level / slide', nextTitle: 'Next level / slide',
+    hintStep: 'reveal level / slide', hintFx: 'motion on / off', hintNarr: 'narration on / off',
+    hintOv: 'overview', hintFs: 'fullscreen', level: 'step',
+    hintKeys: 'Space / PgDn next · PgUp prev · Home / End first / last', hintEsc: 'close overview',
+    edgeFirst: 'Already the first slide', edgeLast: 'Already the last slide',
+    cut: 'Cut', xfade: 'Dissolve', transTitle: 'Slide transition: cut / dissolve (compare, then set script.json transition)', hintTrans: 'slide transition',
+    gen: 'Snapshots generated', genNote: 'same source as the final video (no audio, no subtitles)',
+    noThumbA: 'No thumbnail (run capture first — preview/', ovH: 'Overview · click any slide to jump (thumbnails bundled in play/thumbs/)',
+  } : {
+    title: '放映页', fxOn: '动效开', fxOff: '动效关', narrOn: '口播开', narrOff: '口播关',
+    overview: '总览', panelH: '本张口播文案', untimed: '(未对时)',
+    fxTitle: '动效开 / 动效关(快捷键 X)', narrTitle: '口播开 / 口播关(快捷键 P)',
+    prevTitle: '上一级 / 上一张', nextTitle: '下一级 / 下一张',
+    hintStep: '逐级入场 / 翻页', hintFx: '动效开 / 动效关', hintNarr: '口播开 / 口播关',
+    hintOv: '总览', hintFs: '全屏', level: '级',
+    hintKeys: '空格/PgDn 前进 · PgUp 后退 · Home/End 首末张', hintEsc: '关总览',
+    edgeFirst: '已是第一张', edgeLast: '已是最后一张',
+    cut: '硬切', xfade: '溶解', transTitle: '切页方式:硬切 / 溶解(现场对比后写进 script.json 的 transition)', hintTrans: '切页方式',
+    gen: '快照生成于', genNote: '画面与成片同源(无声、无字幕)',
+    noThumbA: '无缩略图(没跑 capture, 缺 preview/', ovH: '总览 · 点任意一张跳转(缩略图已随 play/thumbs/ 携带)',
+  };
+  const model = slides.map(s => ({
+    id: s.id, name: s.name, title: s.title ?? '',
+    src: s.copy, nofx: s.copyNofx,
+    steps: (s.steps && s.steps.length) ? s.steps : [1],   // 动效开时逐级揭示的 stage 序列
+    clauses: (s.clauses ?? []).map(c => ({ stage: c.stage ?? null, text: c.text ?? '', text2: c.text2 ?? '' })),
+  }));
+  const json = JSON.stringify(model).replace(/</g, '\\u003c');
+  const bodyCls = narration ? '' : 'narr-off';
+  const cssWarn = cssNote ? `<p class="warn">⚠ ${esc(cssNote)}</p>` : '';
+  const notice = fallbackNote ? `<div id="notice">${esc(fallbackNote)}</div>` : '';
+
+  return `<!doctype html>
+<html lang="${esc(lang)}">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width,initial-scale=1,viewport-fit=cover">
+<meta name="color-scheme" content="dark">
+<title>${T.title} · ${esc(topic) || 'html2video'} · html2video-for-mcode</title>
+<style>
+  :root{--c-bg:#0b0c10;--c-panel:#14161c;--c-fg:#e8ebf4;--c-dim:#8b90a0;--c-line:#23252e;--c-acc:#6ea8fe;--c-warn:#ffb454}
+  *{box-sizing:border-box}
+  html{height:100%}
+  body{margin:0;height:100vh;height:100dvh;overflow:hidden;background:var(--c-bg);color:var(--c-fg);
+    display:grid;grid-template-rows:auto minmax(0,1fr) auto;
+    font:14px/1.5 -apple-system,"Segoe UI",system-ui,"Noto Sans CJK SC","Microsoft YaHei",sans-serif}
+  /* ── 顶栏 ── */
+  .top{min-width:0}
+  header{display:flex;align-items:center;gap:12px;padding:9px 16px;border-bottom:1px solid var(--c-line);min-width:0}
+  header .topic{font-weight:600;flex:0 1 auto;min-width:0;overflow:hidden;text-overflow:ellipsis;white-space:nowrap}
+  header .grow{flex:1 1 0;min-width:0}
+  .chip{flex:0 0 auto;color:var(--c-dim);font-variant-numeric:tabular-nums;white-space:nowrap;font:inherit;
+    background:none;border:0;padding:2px 0;margin:0}
+  button.chip{cursor:pointer;border-radius:6px;padding:2px 8px}
+  button.chip:hover{background:#1d2130}
+  button.chip.on{color:var(--c-warn)}
+  #notice{padding:6px 16px;background:#241f14;color:var(--c-warn);font-size:12px;border-bottom:1px solid var(--c-line)}
+  /* ── 舞台 + 口播文案 ── */
+  main{display:flex;min-width:0;min-height:0}
+  #stage{position:relative;flex:1 1 auto;min-width:0;min-height:0;overflow:hidden;background:#000}
+  #fit{position:absolute;left:50%;top:50%;width:${canvas.w}px;height:${canvas.h}px;transform-origin:center center}
+  /* 双缓冲: 两张 iframe 叠放, 新帧在隐藏帧里加载完成才对调显示 —— 换页/逐级不再闪白 */
+  #fit iframe{position:absolute;left:0;top:0;width:${canvas.w}px;height:${canvas.h}px;border:0;background:#fff;opacity:0}
+  #fit iframe.show{opacity:1}
+  /* 溶解对比: 新帧淡入盖住旧帧(硬切时无过渡, 与成片一致) */
+  body.xfade #fit iframe{transition:opacity .4s linear}
+  #hit{position:absolute;inset:0;z-index:2;touch-action:none}   /* 触屏左右滑翻页; 画面是纯 CSS 无需交互 */
+  #panel{flex:0 0 clamp(240px,22vw,340px);min-width:0;overflow:auto;overscroll-behavior:contain;
+    border-left:1px solid var(--c-line);background:var(--c-panel);padding:14px 16px}
+  body.narr-off #panel{display:none}
+  #panel h2{margin:0 0 10px;font-size:13px;color:var(--c-dim);font-weight:600;letter-spacing:.04em}
+  #clauses{list-style:none;margin:0;padding:0}
+  #clauses li{padding:8px 10px;border-radius:8px;border:1px solid transparent;color:var(--c-fg);margin-bottom:6px}
+  #clauses .st{display:inline-block;min-width:22px;margin-right:6px;color:var(--c-acc);font-variant-numeric:tabular-nums;font-size:12px}
+  #clauses .t2{display:block;color:var(--c-dim);font-size:12px;margin-top:2px}
+  .warn{color:var(--c-warn);font-size:12px;margin:8px 0 0}
+  /* ── 底栏 ── */
+  footer{display:flex;flex-wrap:wrap;align-items:center;gap:6px 16px;padding:6px 16px;min-height:30px;
+    border-top:1px solid var(--c-line);color:var(--c-dim);font-size:12px}
+  footer kbd{background:#1d2130;border:1px solid var(--c-line);border-radius:4px;padding:1px 5px;color:var(--c-fg);font-family:inherit}
+  footer .gen{margin-left:auto;opacity:.8}
+  footer .kbd-hints .more{opacity:.72}
+  #edge{position:fixed;left:50%;bottom:76px;transform:translateX(-50%) translateY(6px);background:rgba(12,12,16,.8);color:#fff;padding:8px 16px;border-radius:10px;font-size:15px;opacity:0;pointer-events:none;transition:opacity .18s,transform .18s;z-index:60}
+  #edge.show{opacity:1;transform:translateX(-50%) translateY(0)}
+  .kbd-hints{display:flex;flex-wrap:wrap;gap:6px 16px}   /* 不用内联样式: 否则窄屏媒体查询盖不住它 */
+  #touchbar{display:flex;gap:6px;flex-wrap:wrap}
+  #touchbar button{flex:1 1 auto;min-width:0;font:inherit;font-size:12px;color:var(--c-fg);background:#1a1d26;
+    border:1px solid var(--c-line);border-radius:8px;padding:8px 4px;cursor:pointer;white-space:nowrap;overflow:hidden;text-overflow:ellipsis}
+  #touchbar button:active{background:#252a38}
+  #touchbar button.on{color:var(--c-warn)}   /* "关"态亮警示色: 一眼看出动效/口播被关了 */
+  /* 前后箭头是高频操作, 加大加粗(实测反馈太小) */
+  #touchbar button[data-act="prev"],#touchbar button[data-act="next"]{flex:0 0 76px;font-size:24px;font-weight:600}
+  /* ── 总览 ── */
+  #overlay{position:fixed;inset:0;background:rgba(6,7,10,.975);padding:28px;overflow:auto;display:none;z-index:9}
+  #overlay.show{display:block}
+  #overlay h2{margin:0 0 16px;font-size:15px;color:var(--c-dim);font-weight:600}
+  #grid{display:grid;grid-template-columns:repeat(auto-fill,minmax(min(240px,100%),1fr));gap:14px}
+  #grid figure{margin:0;border:1px solid var(--c-line);border-radius:10px;overflow:hidden;background:#000;cursor:pointer}
+  #grid figure:hover{border-color:var(--c-acc)}
+  #grid img{width:100%;aspect-ratio:${canvas.w}/${canvas.h};object-fit:cover;display:block}
+  #grid .ph{width:100%;aspect-ratio:${canvas.w}/${canvas.h};display:flex;align-items:center;justify-content:center;color:var(--c-dim);font-size:12px}
+  #grid figcaption{padding:8px 10px;font-size:12px;color:var(--c-dim)}
+  #grid figcaption b{color:var(--c-fg);font-weight:600;margin-right:6px}
+  /* ── 窄窗口 / 手机: 口播文案收到底部当抽屉, 键盘提示换成触摸按钮 ── */
+  @media (max-width:860px), (pointer:coarse){
+    footer .kbd-hints{display:none}
+    #touchbar{width:100%}
+  }
+  @media (max-width:860px){
+    header{padding:8px 12px;gap:8px}
+    header .topic{font-size:13px}
+    #notice{padding:6px 12px}
+    main{flex-direction:column}
+    #panel{flex:0 0 auto;max-height:38dvh;border-left:0;border-top:1px solid var(--c-line);padding:10px 12px}
+    footer{padding:6px 12px}
+    #overlay{padding:16px}
+  }
+  @media (max-width:560px){
+    header .topic{display:none}          /* 标题在浏览器标签上还有, 这里让位给页码与按钮 */
+    footer .gen{display:none}
+    #touchbar button{font-size:11px;padding:8px 2px}
+    #overlay h2{font-size:13px}
+  }
+</style>
+</head>
+<body class="${bodyCls}">
+<div class="top">
+  <header>
+    <span class="topic" title="${esc(topic)}">${esc(topic) || T.title}</span>
+    <span class="chip" id="pageno">1 / ${slides.length}</span>
+    <span class="grow"></span>
+  </header>
+  ${notice}
+</div>
+<main>
+  <div id="stage"><div id="fit"><iframe id="frame" title="slide" scrolling="no" frameborder="0"></iframe><iframe id="frameB" title="slide" scrolling="no" frameborder="0" aria-hidden="true"></iframe></div><div id="hit"></div></div>
+  ${narration ? `<aside id="panel">
+    <h2>${T.panelH}${timing ? '' : T.untimed}</h2>
+    <ol id="clauses"></ol>
+    ${cssWarn}
+  </aside>` : ''}
+</main>
+<footer>
+  <span class="kbd-hints">
+    <span><kbd>←</kbd><kbd>→</kbd> ${T.hintStep}</span>
+    <span><kbd>X</kbd> ${T.hintFx}</span>
+    ${narration ? `<span><kbd>P</kbd> ${T.hintNarr}</span>` : ''}
+    <span><kbd>O</kbd> ${T.hintOv}</span>
+    <span><kbd>F</kbd> ${T.hintFs}</span>
+    <span><kbd>T</kbd> ${T.hintTrans}</span>
+    <span class="more">${T.hintKeys} · <kbd>Esc</kbd> ${T.hintEsc}</span>
+  </span>
+  <span id="touchbar">
+    <button type="button" data-act="prev" title="${T.prevTitle}">‹</button>
+    <button type="button" data-act="nofx" id="fxbtn" title="${T.fxTitle}">${T.fxOn}</button>
+    ${narration ? `<button type="button" data-act="narr" id="narrbtn" title="${T.narrTitle}">${T.narrOn}</button>` : ''}
+    <button type="button" data-act="trans" id="transbtn" title="${T.transTitle}">${T.cut}</button>
+    <button type="button" data-act="overview" title="${T.overview}">${T.overview}</button>
+    <button type="button" data-act="next" title="${T.nextTitle}">›</button>
+  </span>
+  <span class="gen">${T.gen} ${esc(generatedAt)} · ${T.genNote}</span>
+</footer>
+<div id="overlay"><h2>${T.ovH}</h2><div id="grid"></div></div>
+<div id="edge"></div>
+<script>
+var S = ${json};
+var T = ${JSON.stringify({ fxOn: T.fxOn, fxOff: T.fxOff, narrOn: T.narrOn, narrOff: T.narrOff, level: T.level, noThumbA: T.noThumbA, cut: T.cut, xfade: T.xfade, hintKeys: T.hintKeys, hintEsc: T.hintEsc, edgeFirst: T.edgeFirst, edgeLast: T.edgeLast })};
+var i = 0, nofx = false, step = 1, animNext = true, trans = 'cut';   // trans: 'cut' 硬切(默认) / 'xfade' 溶解 —— 现场对比用
+// step = 动效开时已揭示到第几个 stage; animNext = 揭示该级时是否当场播入场动画
+
+function el(id){ return document.getElementById(id); }
+function cur(){ return S[i]; }
+
+// 双缓冲换帧: 新地址先喂给隐藏帧, load 后再等两帧 rAF(新文档首帧已画)才对调显示 ——
+// 单 iframe 直接改 src 会先卸载旧文档(白屏一瞬)再加载新文档, 换页/逐级都会闪白(实测反馈)
+var FR = [el('frame'), el('frameB')], front = 0, gen = 0;
+function show(src){
+  gen++; var g = gen, back = FR[1 - front], prev = FR[front];
+  // 入口归一: 后备帧必须从隐藏态出发。xfade 的 400ms 延迟撤场被 gen 守卫作废时,
+  // 旧帧会带着 .show 留到下次换帧 —— 两个 iframe 同 opacity:1, 视觉上被盖住无感,
+  // 但外部脚本按 offsetParent/可见性判"当前帧"必然判错(第 25+ 轮外部报告 P3 实踩)
+  back.classList.remove('show');
+  var raf = window.requestAnimationFrame || function(f){ setTimeout(f, 16); };
+  back.addEventListener('load', function onl(){
+    back.removeEventListener('load', onl);
+    if (g !== gen) return;                 // 等待期间又按了下一步: 这次换帧作废
+    raf(function(){ raf(function(){ if (g === gen) {
+      back.classList.add('show');
+      if (trans === 'xfade') {             // 溶解: 新帧淡入盖住旧帧, 旧帧等过渡结束再撤(绝不经过黑场)
+        setTimeout(function(){ if (g === gen) prev.classList.remove('show'); }, 400);
+      } else { prev.classList.remove('show'); }   // 硬切: 直接换
+      front = 1 - front;
+    } }); });
+  });
+  back.src = './' + src;   // './' 前缀: 强制按相对 URL 解析, 防同名 javascript: 文件名在放映页同源执行
+}
+
+function layout(){
+  var st = el('stage'); if (!st) return;
+  var r = st.getBoundingClientRect();
+  if (!r.width || !r.height) return;
+  var k = Math.min(r.width / ${canvas.w}, r.height / ${canvas.h}) * 0.98;
+  el('fit').style.transform = 'translate(-50%,-50%) scale(' + k + ')';
+}
+window.addEventListener('resize', layout);
+window.addEventListener('orientationchange', function(){ setTimeout(layout, 150); });
+if (window.ResizeObserver) { try { new ResizeObserver(layout).observe(el('stage')); } catch (e) {} }
+document.addEventListener('fullscreenchange', function(){ setTimeout(layout, 60); });
+
+function paintPanel(){
+  var ol = el('clauses'); if (!ol) return;
+  ol.innerHTML = '';
+  (cur().clauses || []).forEach(function(c){
+    var li = document.createElement('li');
+    var st = document.createElement('span'); st.className = 'st';
+    st.textContent = c.stage != null ? 'L' + c.stage : '--';
+    li.appendChild(st);
+    li.appendChild(document.createTextNode(c.text || ''));
+    if (c.text2) { var d = document.createElement('span'); d.className = 't2'; d.textContent = c.text2; li.appendChild(d); }
+    ol.appendChild(li);
+  });
+}
+
+// 底栏开关按当前状态写明"开/关"(实测反馈: 只写"动效/口播"看不出现在是开还是关)
+function paintToggles(){
+  var b = el('fxbtn'); if (b) { b.textContent = nofx ? T.fxOff : T.fxOn; b.className = nofx ? 'on' : ''; }
+  var n = el('narrbtn'); if (n) { var off = document.body.classList.contains('narr-off'); n.textContent = off ? T.narrOff : T.narrOn; n.className = off ? 'on' : ''; }
+  var t = el('transbtn'); if (t) t.textContent = trans === 'cut' ? T.cut : T.xfade;
+}
+
+// multi = 本张当前可逐级(动效开 + 多级 + 文件名没带 ?/# —— 带 ?/# 的副本拼 ?s= 查询会失效, 只能整张放)
+function multi(){ var s = cur(); return !nofx && s.steps && s.steps.length > 1 && !/[?#]/.test(s.src); }
+
+function render(){
+  var s = cur();
+  document.body.classList.toggle('xfade', trans === 'xfade');
+  var lvl = multi() ? (' · ' + T.level + ' ' + step + '/' + s.steps.length) : '';
+  el('pageno').textContent = (i + 1) + ' / ' + S.length + lvl;
+  var src;
+  if (nofx) src = s.nofx;
+  else if (multi()) src = s.src + '?s=' + step + (animNext ? '&anim=1' : '');
+  else src = s.src;
+  show(src);
+  paintToggles(); paintPanel(); layout();
+}
+
+// 动效开: →/← 先逐级揭示(下一级当场入场、上一级直接终态), 到头/回头才翻页;动效关: 直接翻页。
+// 片尾(末张末级)再按 → / 片头(首张第 1 级)再按 ← 都是 no-op —— 旧实现会把末张重置回第 1 级重播、首张跳到末级终态
+// no-op 也要有反馈(toast): 不然用户分不清"到头了"还是"坏了"(外部报告 P2 实测按了 80 次)
+var edgeTimer = 0;
+function edge(msg){
+  var t = el('edge'); if (!t) return;
+  t.textContent = msg; t.classList.add('show');
+  clearTimeout(edgeTimer); edgeTimer = setTimeout(function(){ t.classList.remove('show'); }, 1200);
+}
+function next(){
+  if (i >= S.length - 1 && (!multi() || step >= cur().steps.length)) { edge(T.edgeLast); return; }
+  if (multi() && step < cur().steps.length) { step++; animNext = true; }
+  else { i = Math.min(S.length - 1, i + 1); enterSlide(true); }
+  render();
+}
+function prev(){
+  if (i <= 0 && (!multi() || step <= 1)) { edge(T.edgeFirst); return; }
+  if (multi() && step > 1) { step--; animNext = false; }
+  else { i = Math.max(0, i - 1); enterSlide(false); }
+  render();
+}
+// 进入一张: 向前进 = 只出第 1 级(当场入场);往回跳/总览选张 = 直接整张终态
+function enterSlide(fwd){
+  var s = cur();
+  if (fwd) { step = 1; animNext = true; }
+  else { step = (s.steps && s.steps.length) ? s.steps.length : 1; animNext = false; }
+}
+function go(n){ i = Math.max(0, Math.min(S.length - 1, n)); enterSlide(false); render(); }
+
+function toggleNarr(){
+  if (!el('panel')) return;                    // 本片就没有口播 UI(没 clauses 或 --no-script)
+  document.body.classList.toggle('narr-off');  // 只是 display:none, 所以还能切回来
+  paintToggles();
+  setTimeout(layout, 30);
+}
+
+function buildGrid(){
+  var g = el('grid'); g.innerHTML = '';
+  S.forEach(function(s, n){
+    var f = document.createElement('figure');
+    var img = document.createElement('img');
+    img.src = 'thumbs/' + s.id + '.png';   // 随 play/thumbs/ 携带(生成时拷入): 只拷 play/ 给别人不再全挂
+    img.alt = s.id;
+    img.onerror = function(){ var p = document.createElement('div'); p.className = 'ph'; p.textContent = T.noThumbA + s.id + '.png)'; img.replaceWith(p); };
+    f.appendChild(img);
+    var cap = document.createElement('figcaption');
+    var b = document.createElement('b'); b.textContent = s.id;
+    cap.appendChild(b);
+    cap.appendChild(document.createTextNode(s.title || s.name));
+    f.appendChild(cap);
+    f.onclick = function(){ go(n); el('overlay').classList.remove('show'); };
+    g.appendChild(f);
+  });
+}
+
+var ACT = {
+  prev: prev,
+  next: next,
+  nofx: function(){ nofx = !nofx; render(); },
+  trans: function(){ trans = trans === 'cut' ? 'xfade' : 'cut'; render(); },
+  narr: toggleNarr,
+  overview: function(){ buildGrid(); el('overlay').classList.add('show'); }
+};
+Array.prototype.forEach.call(document.querySelectorAll('#touchbar button'), function(b){
+  b.onclick = function(){ (ACT[b.getAttribute('data-act')] || function(){})(); };
+});
+
+// 编程接口(外部报告 P1: 页面状态全是闭包变量, 自动化只能模拟键盘再拿 innerText 正则抠
+// "级 1/2"、猜哪个 iframe 带 .show —— 实测必错)。外部脚本请用 window.playPage, 不再抠 DOM。
+window.playPage = {
+  get state(){ return {
+    index: i, total: S.length, step: step, steps: (cur().steps || []).length,
+    fx: !nofx, narr: !(el('panel') && document.body.classList.contains('narr-off')),
+    trans: trans, overview: el('overlay').classList.contains('show'),
+  }; },
+  go: go, next: next, prev: prev,
+  setFx: function(b){ nofx = !b; render(); },
+  setNarr: function(b){ if (!el('panel')) return; document.body.classList.toggle('narr-off', !b); paintToggles(); setTimeout(layout, 30); },
+  setTrans: function(s){ trans = s === 'xfade' ? 'xfade' : 'cut'; render(); },
+  overview: function(b){ if (b) ACT.overview(); else el('overlay').classList.remove('show'); },
+};
+
+// 触屏: 左右滑 = 逐级入场/翻页(与按钮、键盘同一套 next/prev;画面本身是纯 CSS, 盖一层透明接收层)
+(function(){
+  var hit = el('hit'); if (!hit) return;
+  var x0 = 0, y0 = 0;
+  hit.addEventListener('touchstart', function(e){ var t = e.changedTouches[0]; x0 = t.clientX; y0 = t.clientY; }, { passive: true });
+  hit.addEventListener('touchend', function(e){
+    var t = e.changedTouches[0], dx = t.clientX - x0, dy = t.clientY - y0;
+    if (Math.abs(dx) >= 45 && Math.abs(dx) > Math.abs(dy)) { if (dx < 0) next(); else prev(); }
+  }, { passive: true });
+})();
+
+document.addEventListener('keydown', function(e){
+  if (e.key === 'ArrowRight' || e.key === 'PageDown' || e.key === ' ') { next(); e.preventDefault(); }
+  else if (e.key === 'ArrowLeft' || e.key === 'PageUp') { prev(); e.preventDefault(); }
+  else if (e.key === 'Home') go(0);
+  else if (e.key === 'End') go(S.length - 1);
+  else if (e.key === 'x' || e.key === 'X') ACT.nofx();
+  else if (e.key === 't' || e.key === 'T') ACT.trans();
+  else if (e.key === 'p' || e.key === 'P') toggleNarr();
+  else if (e.key === 'o' || e.key === 'O') ACT.overview();
+  else if (e.key === 'f' || e.key === 'F') { if (document.fullscreenElement) document.exitFullscreen(); else el('stage').requestFullscreen(); }
+  else if (e.key === 'Escape') el('overlay').classList.remove('show');
+});
+
+// 窄窗口/手机上画面优先: 口播面板默认收起(点"口播文案"或按 P 展开)
+try { if (el('panel') && window.matchMedia('(max-width:860px)').matches) document.body.classList.add('narr-off'); } catch (e) {}
+
+render();
+</script>
+</body>
+</html>
+`;
+}
+
+// ─────────────────────────── 主流程 ───────────────────────────
+
+function main() {
+  const argv = process.argv.slice(2);
+  const dir = positionalDir(argv);
+  const OPEN = argv.includes('--open');
+  const NO_SCRIPT = argv.includes('--no-script');   // 不加载口播文案 UI(只想看 HTML 画面时)
+  process.env.KIT_PROJECT_DIR = dir;
+
+  const scriptPath = path.join(dir, 'script.json');
+  if (!fs.existsSync(scriptPath)) { console.error(`✗ 找不到 ${scriptPath}`); process.exit(1); }
+  const script = JSON.parse(fs.readFileSync(scriptPath, 'utf8'));
+  validateScriptPaths(script, dir); // id/html 派生路径收监(与 capture 同一道门)
+
+  const timingsPath = path.join(dir, 'build', 'timings.json');
+  const hasTimings = fs.existsSync(timingsPath);
+  const timings = hasTimings ? JSON.parse(fs.readFileSync(timingsPath, 'utf8')) : null;
+
+  const slidesDir = path.join(dir, 'slides');
+  const outDir = safeOut(dir, 'preview', 'play');
+  fs.mkdirSync(outDir, { recursive: true });
+
+  // 项目 tokens.css 落后于技能当前版本(缺 no-fx / 图表 / 表格任一段)时给副本兜底注入当前版,
+  // 并在页面与终端如实说明 —— 否则"关动效对照"会把画面锁在入场前的透明态, 旧图表规则也会
+  // 把新版画法(如 .chart-ticks)渲染坏。注入的 <style> 排在 tokens.css 之后, 后写覆盖。
+  const tokensPath = path.join(slidesDir, 'tokens.css');
+  const tokensCss = fs.existsSync(tokensPath) ? fs.readFileSync(tokensPath, 'utf8') : '';
+  // 与 check-slides / init-project 同一道门: 恶意/误写超大 tokens.css 会让受管块扫描二次方变慢(实测 6MB ≈ 129s)
+  if (tokensCss.length > MAX_SCAN_BYTES) {
+    console.error(`✗ tokens.css 有 ${(tokensCss.length / 1e6).toFixed(1)}MB, 超过 ${MAX_SCAN_BYTES / 1e6}MB 上限, 拒绝扫描(正常 ≈15KB; 与 check-slides / init-project --check-css 同一道门)`);
+    process.exit(1);
+  }
+  const bodyNow = generateTokensCss();
+  const kitIssues = kitStatuses(tokensCss, { expected: expectedTokens() })
+    .map(st => ({ ...st, css: st.id === TOKENS_ID ? bodyNow : KITS.find(k => k.id === st.id).css }))
+    .filter(st => st.status !== 'ok');
+  // 整段主体落后(或整段缺失) → 兜底注入**当前整段生成物**: 它本身含三个工具箱子块, 一次就把
+  // 主题令牌/字号/fx/工具箱全补齐。关动效副本连 no-fx 一起要; 动效副本按既有约定不含 no-fx
+  // (它只在关动效那侧有意义), 所以给动效副本注入时把 nofx 子块摘掉。
+  const tokensBad = kitIssues.find(k => k.id === TOKENS_ID);
+  const nofxBlock = findAllBlocks(bodyNow, 'nofx')[0];
+  const bodyNoNofx = nofxBlock ? bodyNow.slice(0, nofxBlock.start) + bodyNow.slice(nofxBlock.end) : bodyNow;
+  const nofxIssue = tokensBad ? { label: tokensBad.label } : kitIssues.find(k => k.id === 'nofx');
+  const chartTable = tokensBad ? [] : kitIssues.filter(k => k.id !== 'nofx');
+  const ctCss = tokensBad ? bodyNoNofx : chartTable.map(k => k.css).join('\n\n');
+  const ctLabel = tokensBad
+    ? [tokensBad.label, ...KITS.filter(k => k.id !== 'nofx').map(k => k.label)].join(' · ')
+    : chartTable.map(k => k.label).join(' · ');
+  const nofxInjected = tokensBad ? bodyNow : nofxIssue ? NOFX_CSS : '';
+  // 页面上用户可见的文字按 script.lang 双语; 终端输出是给 agent 看的, 保持中文
+  const isEn = String(script.lang ?? 'zh').toLowerCase().startsWith('en');
+  const L = (zh, en) => (isEn ? en : zh);
+  // 兜底注入只救放映页 —— 成片读的是项目文件, 所以这句话必须写明"成片仍是旧 CSS"(否则
+  // 会让人以为放映页对了成片就对了, 这正是"改了 CSS 没生效"的错觉来源)。
+  const cssNote = kitIssues.length
+    ? L(`本项目 tokens.css 落后于技能当前版(缺: ${kitIssues.map(k => k.label).join(' · ')}) —— 本页快照已兜底注入当前版, 但**成片仍用项目里的旧 CSS**: 先跑 init-project --upgrade-css, 再 capture / build-video`,
+        `This project's tokens.css is behind the skill's current version (missing: ${kitIssues.map(k => k.label).join(' / ')}) — these snapshots inject the current one as a fallback, but the RENDERED VIDEO still uses the project's old CSS: run init-project --upgrade-css before capture / build-video`)
+    : '';
+
+  const rows = [];
+  let fallbackUsed = false;
+  // 画布值会插进放映页的 CSS 与页面 JS(与 build-video 拼 ffmpeg filter 同一信任级), 先验后用
+  const cvW = Number(script.width ?? 1920), cvH = Number(script.height ?? 1080);
+  if (!Number.isInteger(cvW) || cvW < 16 || cvW > 16384 || !Number.isInteger(cvH) || cvH < 16 || cvH > 16384) {
+    console.error(`✗ script.width/height 非法: ${JSON.stringify(script.width)} × ${JSON.stringify(script.height)} — 需要 16–16384 的整数(误写字符串会直接杀死放映页脚本, 与 build-video 的 clampDim 同一道门)`);
+    process.exit(1);
+  }
+  for (const s of script.slides) {
+    const sid = safeId(s.id);
+    // 安全性照旧收监(越界仍退出 1); 但"这张还没写"按 capture 的惯例跳过并警告 ——
+    // 写到第 3 张就想先放映看看是常态, 不该因为后面几张还没写就整个出不来。
+    const relHtml = s.html ?? `${sid}.html`;
+    const srcPath = safeRel(slidesDir, relHtml, { where: `slides[${sid}].html` });
+    if (!fs.existsSync(srcPath)) { console.warn(`- 跳过 ${sid}: 缺 slides/${relHtml}`); continue; }
+    if (path.dirname(relHtml) !== '.') { console.warn(`- 跳过 ${sid}: slides/${relHtml} 在子目录里 — 放映页快照按平铺 slides/ 设计, 子目录副本与 <base> 会错位; 把 HTML 挪到 slides/ 根再跑`); continue; }
+    const html = fs.readFileSync(srcPath, 'utf8');
+    if (html.length > MAX_SCAN_BYTES) { console.warn(`- 跳过 ${sid}: HTML 超过 ${MAX_SCAN_BYTES / 1e6}MB 上限, 拒绝扫描(正常 ≈10KB)`); continue; }
+    const base = path.basename(srcPath);
+    const t = timings?.slides?.find(x => x.id === s.id);
+
+    // 有对时 → 用实测延迟; 没有(口播还没做) → 按 HTML 里实际用到的 stage 等间隔错开,
+    // 至少能看清入场顺序(占位值会把动画全挤在 2 秒内, 那才是看不懂的)。
+    let stages = t?.stages ?? {};
+    let fallback = false;
+    if (!Object.keys(stages).length) { stages = fallbackStages(html); fallback = Object.keys(stages).length > 0; }
+    if (fallback) fallbackUsed = true;
+    const decl = stageVars(stages, { w: cvW, h: cvH });
+
+    let copy = addBaseHref(injectHtmlVars(html, decl), '../../slides/');
+    copy = `<!-- html2video-for-mcode 放映页快照 · 由 scripts/preview-page.mjs 生成, 请勿编辑;\n     真实文件: slides/${base}(改完 slides 请重跑 preview-page.mjs) -->\n` + copy;
+    copy = addStepScript(copy, stagesFromHtml(html));   // 动效开的逐级步进(?s=k 重设各级延迟)
+    if (ctCss) copy = injectStyle(copy, ctCss, `图表/表格工具箱(${ctLabel})`);
+    // 三个叶子写点全部过 safeOut(1.7.5 复查①): outDir 本身收监了, 但目录里的**叶子**若是
+    // 预置的文件符号链接, writeFileSync 会顺着写出去 —— 叶子也要过收监, 与 capture 等一致。
+    fs.writeFileSync(safeOut(dir, 'preview', 'play', base), copy);
+    const nofxName = base.replace(/\.html?$/i, '') + '.nofx.html';
+    let nofx = addNoFx(copy);
+    if (nofxInjected) nofx = injectStyle(nofx, nofxInjected, 'no-fx 规则');
+    fs.writeFileSync(safeOut(dir, 'preview', 'play', nofxName), nofx);
+
+    const stageVals = Object.values(stages).filter(Number.isFinite);
+    rows.push({
+      id: sid, name: base, title: s.title ?? s.layout ?? '',
+      copy: base, copyNofx: nofxName,
+      steps: stagesFromHtml(html),
+      duration: t?.duration,
+      lastStage: stageVals.length ? Math.max(...stageVals) : null,
+      clauses: t?.clauses ?? (s.clauses ?? []).map(c => ({ stage: c.stage ?? null, start: null, text: c.text, text2: c.text2 })),
+    });
+  }
+
+  const p2 = n => String(n).padStart(2, '0');
+  const now = new Date();
+  if (!rows.length) { console.error('✗ 一张可放映的 slide 都没有(检查 script.json 的 html 字段与 slides/ 目录)'); process.exit(1); }
+
+  // 总览缩略图随 play/thumbs/ 携带(外部报告 P1: 总览曾写死 '../'+id+'.png' 指到父目录的
+  // preview/, 只拷 play/ 或只发 index.html 就全挂)。preview/*.png 是 capture 的产物 —— 缺的
+  // 用占位并如实计数, 没跑过 capture 的项目不再静默全占位。先清空重建, 删掉的旧张不留死图。
+  const thumbsDir = safeOut(dir, 'preview', 'play', 'thumbs');
+  fs.rmSync(thumbsDir, { recursive: true, force: true });
+  fs.mkdirSync(thumbsDir, { recursive: true });
+  let thumbN = 0;
+  for (const r of rows) {
+    const srcPng = path.join(dir, 'preview', `${r.id}.png`);
+    if (fs.existsSync(srcPng)) { fs.copyFileSync(srcPng, safeOut(dir, 'preview', 'play', 'thumbs', `${r.id}.png`)); thumbN++; }
+  }
+
+  // 口播 UI 三种状态: 有文案+有对时 → 标注正常;有文案没对时 → 标注"未对时";没文案或 --no-script → 不出
+  const anyClauses = rows.some(r => r.clauses.length > 0);
+  const narration = anyClauses && !NO_SCRIPT;
+  const timing = narration && !!timings && rows.some(r => r.clauses.some(c => Number.isFinite(c.start)));
+  // 提示与口播无关: 动画时序不是成片的, 这件事看画面的人也必须知道
+  const fallbackNote = [
+    (!hasTimings || fallbackUsed) ? L('口播还没对时: 入场顺序按等间隔 0.3/1.3/2.3s 估算, 不是成片时序(缺 build/timings.json)',
+        'Narration not timed yet: entrance order estimated at 0.3/1.3/2.3s intervals — not the final timing (missing build/timings.json)') : '',
+    L('动画为手动逐级步进: 按 → 先出下一级, 出完再翻页(动效关则直接翻页)',
+        'Animations step manually: press → to reveal the next level first, then the next slide (motion off = switch slides directly)'),
+    narration && !timing ? L('口播还没对时: 只列文案', 'Narration not timed yet: script text only') : '',
+  ].filter(Boolean).join(' · ');
+
+  const page = buildPlayPage({
+    topic: script.topic ?? '', lang: script.lang ?? 'zh', slides: rows, cssNote,
+    narration, timing, fallbackNote,
+    canvas: { w: cvW, h: cvH },
+    generatedAt: `${now.getFullYear()}-${p2(now.getMonth() + 1)}-${p2(now.getDate())} ${p2(now.getHours())}:${p2(now.getMinutes())}`,
+  });
+  fs.writeFileSync(safeOut(dir, 'preview', 'play', 'index.html'), page);
+
+  const modes = !narration ? (NO_SCRIPT ? '关(--no-script)' : '关(本片没有 clauses)') : timing ? '口播文案(有对时)' : '口播文案(未对时)';
+  console.log(`✓ 放映页 → ${path.relative(process.cwd(), path.join(outDir, 'index.html'))}`);
+  console.log(`  快照 ${rows.length} 张 ×2(动效/关动效) → ${path.relative(process.cwd(), outDir)}/${rows[0]?.name ?? '<id>.html'}`);
+  console.log(`  口播 UI: ${modes}`);
+  if (thumbN === rows.length) console.log(`  总览缩略图: ${thumbN}/${rows.length} 张已随 play/thumbs/ 携带`);
+  else console.warn(`⚠ 总览缩略图: 仅 ${thumbN}/${rows.length} 张(其余没跑 capture, 缺 preview/<id>.png —— 总览用占位; 跑 capture 后重新 preview-page 会补齐)`);
+  if (!hasTimings) console.warn('⚠ 缺 build/timings.json: 副本按等间隔预览入场顺序, 不是成片时序 —— 先跑 plan-timings.mjs');
+  else if (fallbackUsed) console.warn('⚠ 部分张在 timings.json 里没有 stage 数据: 那几张按等间隔预览');
+  else console.log(`  已注入实测 stage 延迟: ${rows.map(r => `${r.id}(${r.lastStage != null ? '末层 ' + r.lastStage.toFixed(1) + 's' : '无 stage'})`).join(' ')}`);
+  if (kitIssues.length) console.warn(`⚠ slides/tokens.css 落后于技能当前版(缺: ${kitIssues.map(k => k.label).join(' · ')}) —— 副本已兜底注入;想让成片与项目文件也用上, 跑 node scripts/init-project.mjs <项目目录> --upgrade-css`);
+  console.log(`  操作: ← → 逐级入场/翻页(动效开)或翻页(动效关) · Space/PgDn 前进 · PgUp 后退 · Home/End 首末张${narration ? ' · P 口播开/关' : ''} · X 动效开/关 · T 切页方式 · O 总览(Esc 关) · F 全屏;触屏左右滑同 ← →;外部自动化用 window.playPage(state/go/next/prev/setFx/setNarr/setTrans/overview)`);
+
+  if (OPEN) {
+    const target = path.join(outDir, 'index.html');
+    const { cmd, args, verbatim } = openCmdArgv(target);
+    spawn(cmd, args, { detached: true, stdio: 'ignore', windowsVerbatimArguments: verbatim }).unref();
+    console.log('  已在浏览器打开(若没弹出, 手动双击上面的 index.html)');
+  }
+}
+
+// 打开浏览器的子进程参数。**导出来是为了可测**(单一来源, 同 buildSrt / assertConcatPathSafe 的做法)。
+//
+// 空标题 `""` 不是可选的, 两个原因各自独立:
+//   ① start 的**首个带引号 token 是窗口标题, 不是要打开的东西**。实测(2026-10-10):
+//      `cmd /c start "T" cmd /c <命令>`  → T 成了标题, 后面的命令真的执行了;
+//      `cmd /c start T  cmd /c <命令>`  → T 成了命令(找不到该程序), 后面的命令**根本没跑**。
+//      所以 `start "C:\...\index.html"`(只有一个带引号 token) = 路径被当标题吃掉, 一个都不打开。
+//   ② windowsVerbatimArguments 下 Node 不再替我们给参数加引号, `''` 就真的是空串、`""` 会消失。
+//      空标题必须写成**字面量** '""'。这一条是审计 M1 修复里的返工面: 当时保留了 `''` 作为空标题,
+//      以为 verbatim 只影响引用 —— 结果浏览器打不开了。
+// 引号由我们自己补(q), 不依赖 Node 的"含空格才加引号": Windows 路径允许含 & 而不含空格,
+// 未被引用的 & 会让 cmd 起第二条命令(实测 `&rem` 真的执行了)。
+export function openCmdArgv(target, platform = process.platform) {
+  const q = t => `"${String(t).replace(/"/g, '""')}"`;
+  if (platform === 'win32') return { cmd: 'cmd', args: ['/c', 'start', '""', q(target)], verbatim: true };
+  if (platform === 'darwin') return { cmd: 'open', args: [target], verbatim: false };
+  return { cmd: 'xdg-open', args: [target], verbatim: false };
+}
+
+// 被 tests/ import 时不要跑主流程(只有当脚本直接执行才跑)。判定走 tools.isMainModule:
+// realpath 两侧 + 同名兜底 —— 经符号链接/异形路径调用曾恒假静默 no-op(2026-09-22 云沙箱实测)
+if (isMainModule(import.meta.url)) main();
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/table-css.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/table-css.mjs
new file mode 100644
index 00000000..0dfb1689
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/table-css.mjs
@@ -0,0 +1,50 @@
+// html2video-for-mcode · 表格原语的唯一来源。
+// init-project 把它写进 tokens.css 的受管块; `--upgrade-css` 按内容 rev 原地更新(见 css-kit.mjs)。
+// 为什么要有原语: 以前每种表都靠内联样式各写一套, 行高/对齐/涨跌色各页不一致 ——
+// 而表格最容易出的问题正是"行高压不住、数字不右对齐、涨跌色用错"(见 authoring.md 表格纪律)。
+
+export const TABLE_CSS = `/* ── 表格原语(数据表/规格表/对比矩阵/排名表统一用它) ─────────────
+   1080p 下远处要读得清: 行高 ≥72px、数值右对齐且等宽、表头弱化、只留横线不画竖线。
+   涨跌色一律 var(--up)/var(--down)(财经按受众翻转, 见 compliance.md)。 */
+.tbl { width: 100%; border-collapse: collapse; font-size: var(--fs-body); }   /* 主数据用正文号(≥30px): 视频在手机上也要读得清 */
+.tbl th { text-align: left; font-weight: 500; color: var(--fg-3); font-size: var(--fs-caption);
+  letter-spacing: .04em; padding: 0 0 var(--sp-2); border-bottom: 1px solid var(--line-strong); }
+.tbl td { padding: var(--sp-3) var(--sp-4) var(--sp-3) 0; border-bottom: 1px solid var(--line); }
+.tbl tr:last-child td { border-bottom: 0; }
+.tbl tbody tr { height: 72px; }                       /* 行高保底: 再挤就远处读不出 */
+.tbl .num { text-align: right; padding-right: 0; font-family: var(--font-mono);
+  font-variant-numeric: tabular-nums; }
+.tbl .up { color: var(--up); }
+.tbl .down { color: var(--down); }
+.tbl .muted { color: var(--muted); }
+/* 高亮行(通常是"我们"或当前阶段)与合计行 */
+.tbl tr.key td { background: color-mix(in srgb, var(--accent) 8%, transparent); }
+.tbl tr.key td:first-child { box-shadow: inset 3px 0 0 var(--accent); padding-left: var(--sp-4); }
+.tbl tr.sum td { border-top: 1px solid var(--line-strong); font-weight: 600; color: var(--fg); }
+/* 无表头规格表: 左标签右值, 只有细线 */
+.kv { width: 100%; border-collapse: collapse; font-size: var(--fs-body); }
+.kv td { padding: var(--sp-3) 0; border-bottom: 1px solid var(--line); vertical-align: top; }
+.kv td:first-child { color: var(--muted); width: 32%; }
+.kv td:last-child { text-align: right; font-family: var(--font-mono); font-variant-numeric: tabular-nums; }
+/* 对比矩阵: 行是维度、列是方案, 单元格只放勾叉/等级 */
+.matrix { width: 100%; border-collapse: collapse; font-size: var(--fs-body); text-align: center; }
+.matrix th { color: var(--fg-3); font-size: var(--fs-caption); font-weight: 500; padding: 0 0 var(--sp-2);
+  border-bottom: 1px solid var(--line-strong); }
+.matrix th:first-child, .matrix td:first-child { text-align: left; }
+.matrix td { padding: var(--sp-3) 0; border-bottom: 1px solid var(--line); height: 72px; }
+.matrix .yes { color: var(--accent); font-size: var(--fs-h3); }
+.matrix .no { color: var(--fg-3); }
+/* 高亮列: td 与 th 都要覆盖 —— 只写 td.hi 时表头那格拿不到背景, "我们"那一列看着只亮了一半(2026-09-18 排查抓到) */
+.matrix td.hi, .matrix th.hi { background: color-mix(in srgb, var(--accent) 8%, transparent); }
+/* 排名表: 名次 + 微缩条 + 数值(表格里的条形, 复用 --w 定长) */
+.rank { width: 100%; border-collapse: collapse; font-size: var(--fs-body); }
+.rank td { padding: var(--sp-2) var(--sp-3) var(--sp-2) 0; border-bottom: 1px solid var(--line); height: 72px; }
+.rank .no { color: var(--fg-3); font-family: var(--font-mono); width: 3em; }
+.rank .bar { width: 46%; }
+.rank .bar span { display: block; height: 12px; border-radius: 999px; background: var(--accent); }
+.rank .bar span.dim { background: color-mix(in srgb, var(--accent) 30%, var(--bg)); }
+.rank .val { text-align: right; font-family: var(--font-mono); font-variant-numeric: tabular-nums; }`;
+
+export function hasTableKit(css) {
+  return /\.tbl\s+tbody\s+tr\s*\{[^}]*height:\s*72px/.test(String(css ?? '')) && /\.kv\s*\{/.test(String(css ?? ''));
+}
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/tokens-template.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/tokens-template.mjs
new file mode 100644
index 00000000..ac386890
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/tokens-template.mjs
@@ -0,0 +1,432 @@
+// html2video-for-mcode · tokens.css 的生成模板(整段受管: 由 tokens 块包住, 升级时整体替换)
+// 从 init-project.mjs 抽出, 让 init-project / check-slides / preview-page 共用同一个"当前版本"。
+// 模板里的三处 wrapKit 由 css-kit 提供(内部带 KIT_REV), 所以 TOKENS_REV 天然覆盖三个工具箱的每次改动。
+import { createHash } from 'node:crypto';
+import { wrapKit } from './css-kit.mjs';
+import { NOFX_CSS } from './nofx-css.mjs';
+import { CHART_CSS } from './chart-css.mjs';
+import { TABLE_CSS } from './table-css.mjs';
+
+export const TOKENS_BODY = `/* html2video-for-mcode 设计令牌 · 由 init-project 生成
+   主题切换: <html data-theme="a|b|c|minimal-white|swiss-grid|corporate-clean|editorial-serif|magazine-bold|tokyo-night|catppuccin-mocha|nord|xiaohongshu-white|soft-pastel">
+
+   a / b / c 是初版三主题(向后兼容, 老项目继续有效); 其余 10 套于 2026-09-17 移植自
+   html-ppt-skill (MIT, Copyright (c) 2026 lewis) —— 详见技能内 THIRD-PARTY-NOTICES.md。
+
+   约定: 新增或覆写主题必须覆盖全部颜色令牌 + --accent-ink, 并跑
+   \`node <技能>/scripts/check-theme.mjs <项目目录>\` 校验对比度(深色主题另需覆写 --sub-bg/--sub-ring)。
+
+   受管块: 下面 >>> html2video:nofx / chart / table <<< 三段(带 rev 定界注释)由技能单源生成,
+   不要手工编辑 —— 改了也会在下次 --upgrade-css 时被按源重写; 自己的规则加在文件尾即可(后写覆盖)。
+*/
+:root {
+  /* ── 画布(渲染管线按 script.json 的 width/height 注入真值; 默认横屏 1920×1080) ── */
+  --stage-w: 1920px; --stage-h: 1080px;
+
+  /* ── 颜色 · 表面 ── */
+  --bg: #FAFAF7; --bg-soft: #F4F3ED;
+  --panel: #F1F0EA; --panel-2: #E9E8E0;
+  --line: #E3E1D8; --line-strong: #CFCDC0;
+
+  /* ── 颜色 · 文本(三级) ── */
+  --fg: #1A1A1A; --muted: #6B6B66; --fg-3: #9A9A92;
+
+  /* ── 颜色 · 强调(accent-ink = 压在 accent 上的文字色, 每条主题必须自带对比度注释) ── */
+  --accent: #FF5B2E; --accent-ink: #FFFFFF;  /* on --accent 3.1:1 — 仅可用于大字/图形 */
+  --accent-2: #FF8A5C; --accent-3: #D94A20;
+
+  /* ── 颜色 · 语义(数据页涨跌/正负) ── */
+  --good: #1AAF6C; --warn: #C98500; --bad: #C13A3A;
+
+  /* ── 涨跌专用(指标卡写 var(--up)/var(--down), 别直接写 good/bad):
+     默认 = 欧美读法(绿涨红跌); A股/港股受众在项目 tokens.css 末尾覆写成红涨绿跌:
+     :root { --up: #D92B2B; --down: #12A150; }   见 references/compliance.md ── */
+  --up: var(--good); --down: var(--bad);
+
+  /* ── 渐变 ── */
+  --grad: linear-gradient(135deg, #FF5B2E, #FF8A5C 55%, #D94A20);
+  --grad-soft: linear-gradient(135deg, #FBF0E8, #F7E7DA);
+
+  /* ── 字幕(capture 烧录的字幕读这些钩子; 深色主题必须覆写 --sub-bg 与 --sub-ring) ── */
+  --sub-bg: rgba(12,12,16,.62); --sub-fg: #FFFFFF; --sub-ring: 0 solid transparent;
+
+  /* ── 圆角 / 阴影 ── */
+  --radius: 16px; --radius-sm: 10px; --radius-lg: 24px;
+  --shadow: 0 8px 24px rgba(18,24,40,.07), 0 2px 6px rgba(18,24,40,.04);
+  --shadow-lg: 0 22px 56px rgba(18,24,40,.13), 0 6px 16px rgba(18,24,40,.06);
+
+  /* ── 字体(全部走本地系统栈, 禁外链; 外链字体在离线沙箱会退化成 FOUT/方框) ── */
+  --font-display: "Source Han Serif SC", "Noto Serif CJK SC", "Noto Serif SC", "SimSun", serif;
+  --font-body: "Inter", "Source Han Sans SC", "Noto Sans CJK SC", "Microsoft YaHei", sans-serif;
+  --font-mono: "JetBrains Mono", "Cascadia Code", Consolas, monospace;
+
+  /* ── 字号(7 档 modular scale) ── */
+  --fs-display: 92px; --fs-h1: 64px; --fs-h2: 44px; --fs-h3: 34px;
+  --fs-body: 30px; --fs-caption: 24px; --fs-tiny: 20px;
+
+  /* ── 间距(8 的倍数) ── */
+  --sp-1: 8px; --sp-2: 16px; --sp-3: 24px; --sp-4: 32px;
+  --sp-5: 48px; --sp-6: 64px; --sp-7: 96px; --sp-8: 128px;
+
+  /* ── 缓动 / 时长 ── */
+  --ease-out: cubic-bezier(.16, 1, .3, 1);
+  --ease-in-out: cubic-bezier(.65, 0, .35, 1);
+  --dur-fast: .4s; --dur-mid: .6s; --dur-slow: .9s;
+
+  /* ── stage 延迟: 浏览器里直接打开时的占位值; 渲染管线会按 TTS 实测时长注入真值 ── */
+  --t1: 0ms; --t2: 800ms; --t3: 2000ms;
+}
+
+/* ══ 主题 b · 深色科技绿(初版) ══ */
+[data-theme="b"] {
+  --bg: #0E0F12; --bg-soft: #12141A; --panel: #171A1F; --panel-2: #1E2229;
+  --line: #262A31; --line-strong: #363B45;
+  --fg: #F4F5F3; --muted: #9A9C9F; --fg-3: #6E7276;
+  --accent: #10A37F; --accent-ink: #04110C;  /* on --accent 7.4:1 */
+  --accent-2: #3DD9AC; --accent-3: #0B7A5F;
+  --good: #3DD9AC; --warn: #E0AF68; --bad: #F7768E;
+  --grad: linear-gradient(135deg, #10A37F, #3DD9AC 55%, #0B7A5F);
+  --grad-soft: linear-gradient(135deg, #171A1F, #1E2229);
+  --sub-bg: rgba(0,0,0,.58); --sub-ring: 1px solid rgba(255,255,255,.16);
+  --shadow: 0 10px 30px rgba(0,0,0,.45); --shadow-lg: 0 24px 60px rgba(0,0,0,.6);
+}
+
+/* ══ 主题 c · 极简蓝(初版) ══ */
+[data-theme="c"] {
+  --bg: #FFFFFF; --bg-soft: #F7F8FA; --panel: #F5F6F8; --panel-2: #EDEFF3;
+  --line: #E5E7EB; --line-strong: #CFD3DA;
+  --fg: #111111; --muted: #666666; --fg-3: #9AA0A8;
+  --accent: #1F6FEB; --accent-ink: #FFFFFF;  /* on --accent 4.9:1 */
+  --accent-2: #5B9BFF; --accent-3: #1550B8;
+  --good: #0E9F6E; --warn: #D97706; --bad: #DC2626;
+  --grad: linear-gradient(135deg, #1F6FEB, #5B9BFF 55%, #1550B8);
+  --grad-soft: linear-gradient(135deg, #F0F4FB, #E4ECF7);
+  --shadow: 0 1px 3px rgba(10,37,64,.08), 0 4px 12px rgba(10,37,64,.05);
+  --shadow-lg: 0 4px 12px rgba(10,37,64,.1), 0 16px 40px rgba(10,37,64,.08);
+}
+
+/* ════════════════════════════════════════════════════════════════
+   以下 10 套主题移植自 html-ppt-skill(MIT, © 2026 lewis)
+   映射规则: --surface→--panel, --surface-2→--panel-2, --border→--line,
+   --border-strong→--line-strong, --text-1→--fg, --text-2→--muted, --text-3→--fg-3;
+   外链字体(Playfair/Space Grotesk 等)一律改映射到我们的本地字体栈。
+   每套的 --accent-ink 后保留原作者实测的对比度数值。
+   ════════════════════════════════════════════════════════════════ */
+
+/* 商务汇报 · 极简白, 克制高级 */
+[data-theme="minimal-white"] {
+  --bg: #FFFFFF; --bg-soft: #FAFAFA; --panel: #FFFFFF; --panel-2: #F5F5F6;
+  --line: rgba(17,18,22,.08); --line-strong: rgba(17,18,22,.16);
+  --fg: #0C0D10; --muted: #55596A; --fg-3: #9CA1B0;
+  --accent: #111216; --accent-ink: #FFFFFF;  /* on --accent 18.7:1 */
+  --accent-2: #3B3F4A; --accent-3: #6B6F7A;
+  --good: #1AAF6C; --warn: #C98500; --bad: #C13A3A;
+  --grad: linear-gradient(135deg, #111216, #3B3F4A);
+  --grad-soft: linear-gradient(135deg, #F5F5F6, #FFFFFF);
+  --radius: 14px; --radius-sm: 8px; --radius-lg: 22px;
+  --shadow: 0 1px 2px rgba(17,18,22,.04), 0 8px 24px rgba(17,18,22,.06);
+  --shadow-lg: 0 20px 60px rgba(17,18,22,.1);
+  --font-display: var(--font-body);
+}
+
+/* 商务汇报 · 瑞士网格(Helvetica 感; .stage 带 12 栏竖线) */
+[data-theme="swiss-grid"] {
+  --bg: #FFFFFF; --bg-soft: #F4F4F4; --panel: #FFFFFF; --panel-2: #F4F4F4;
+  --line: #111111; --line-strong: #111111;
+  --fg: #111111; --muted: #444444; --fg-3: #888888;
+  --accent: #D6001C; --accent-ink: #FFFFFF;  /* on --accent 5.4:1 */
+  --accent-2: #111111; --accent-3: #888888;
+  --good: #0F8A2F; --warn: #D38A00; --bad: #D6001C;
+  --grad: linear-gradient(135deg, #D6001C, #111111);
+  --grad-soft: linear-gradient(135deg, #F4F4F4, #FFFFFF);
+  --radius: 0; --radius-sm: 0; --radius-lg: 0;
+  --shadow: none; --shadow-lg: none;
+  --font-display: var(--font-body);
+}
+[data-theme="swiss-grid"] .card { border-top: 2px solid var(--fg); border-bottom: 1px solid var(--fg); border-left: none; border-right: none; box-shadow: none; background: var(--bg); }
+[data-theme="swiss-grid"] .stage { background-image: linear-gradient(90deg, rgba(0,0,0,.04) 1px, transparent 1px); background-size: calc(100% / 12) 100%; }
+
+/* 商务汇报 · 企业商务(深蓝) */
+[data-theme="corporate-clean"] {
+  --bg: #FFFFFF; --bg-soft: #F5F7FA; --panel: #FFFFFF; --panel-2: #F0F3F7;
+  --line: rgba(10,37,64,.12); --line-strong: rgba(10,37,64,.28);
+  --fg: #0A2540; --muted: #425466; --fg-3: #8898AA;
+  --accent: #0A2540; --accent-ink: #FFFFFF;  /* on --accent 15.5:1 */
+  --accent-2: #1D4ED8; --accent-3: #64748B;
+  --good: #0E9F6E; --warn: #D97706; --bad: #DC2626;
+  --grad: linear-gradient(135deg, #0A2540, #1D4ED8);
+  --grad-soft: linear-gradient(135deg, #F0F4FB, #E4ECF7);
+  --radius: 6px; --radius-sm: 4px; --radius-lg: 10px;
+  --shadow: 0 1px 3px rgba(10,37,64,.08), 0 4px 12px rgba(10,37,64,.05);
+  --shadow-lg: 0 4px 12px rgba(10,37,64,.1), 0 16px 40px rgba(10,37,64,.08);
+  --font-display: var(--font-body);
+}
+[data-theme="corporate-clean"] .kicker { color: var(--accent-2); }
+
+/* 编辑杂志 · 杂志风衬线(原主题给标题配斜体, 中文伪斜体观感差故略去) */
+[data-theme="editorial-serif"] {
+  --bg: #FAF7F2; --bg-soft: #F3EFE6; --panel: #FFFFFF; --panel-2: #F7F2E8;
+  --line: rgba(40,28,18,.12); --line-strong: rgba(40,28,18,.24);
+  --fg: #1B1410; --muted: #5C4A3E; --fg-3: #8A7868;
+  --accent: #8A2A1C; --accent-ink: #FFFFFF;  /* on --accent 8.6:1 */
+  --accent-2: #C97A4A; --accent-3: #1B1410;
+  --good: #3F7D4F; --warn: #B07A1F; --bad: #8A2A1C;
+  --grad: linear-gradient(135deg, #8A2A1C, #C97A4A);
+  --grad-soft: linear-gradient(135deg, #FAF7F2, #F3EFE6);
+  --radius: 4px; --radius-sm: 2px; --radius-lg: 8px;
+  --shadow: 0 2px 12px rgba(40,28,18,.06); --shadow-lg: 0 20px 50px rgba(40,28,18,.14);
+}
+
+/* 编辑杂志 · 大标题(硬阴影 + 衬线巨字) */
+[data-theme="magazine-bold"] {
+  --bg: #F5EFE2; --bg-soft: #EBE4D2; --panel: #FBF6E8; --panel-2: #EDE5D0;
+  --line: rgba(10,10,10,.16); --line-strong: #0A0A0A;
+  --fg: #0A0A0A; --muted: #2A2A2A; --fg-3: #6A6458;
+  --accent: #EA5A1A; --accent-ink: #0B1024;  /* on --accent 5.4:1 */
+  --accent-2: #0A0A0A; --accent-3: #C42A10;
+  --good: #2A6A2A; --warn: #EA5A1A; --bad: #C42A10;
+  --grad: linear-gradient(135deg, #EA5A1A, #C42A10);
+  --grad-soft: linear-gradient(135deg, #FBE4D0, #F5D6C0);
+  --radius: 0; --radius-sm: 0; --radius-lg: 2px;
+  --shadow: none; --shadow-lg: 6px 6px 0 var(--accent);
+}
+[data-theme="magazine-bold"] .card { border: 1.5px solid var(--fg); }
+[data-theme="magazine-bold"] .divider-accent { height: 6px; width: 90px; }
+
+/* 科技深色 · tokyo-night */
+[data-theme="tokyo-night"] {
+  --bg: #1A1B26; --bg-soft: #16161E; --panel: #24283B; --panel-2: #2F334D;
+  --line: rgba(192,202,245,.12); --line-strong: rgba(192,202,245,.24);
+  --fg: #C0CAF5; --muted: #A9B1D6; --fg-3: #565F89;
+  --accent: #7AA2F7; --accent-ink: #0B1024;  /* on --accent 7.5:1 */
+  --accent-2: #BB9AF7; --accent-3: #7DCFFF;
+  --good: #9ECE6A; --warn: #E0AF68; --bad: #F7768E;
+  --grad: linear-gradient(135deg, #7AA2F7, #BB9AF7 55%, #F7768E);
+  --grad-soft: linear-gradient(135deg, #24283B, #2F334D);
+  --sub-bg: rgba(0,0,0,.58); --sub-ring: 1px solid rgba(255,255,255,.16);
+  --radius: 12px; --radius-sm: 8px; --radius-lg: 20px;
+  --shadow: 0 10px 30px rgba(0,0,0,.45); --shadow-lg: 0 24px 62px rgba(0,0,0,.6);
+}
+
+/* 科技深色 · catppuccin mocha */
+[data-theme="catppuccin-mocha"] {
+  --bg: #1E1E2E; --bg-soft: #181825; --panel: #313244; --panel-2: #45475A;
+  --line: rgba(205,214,244,.12); --line-strong: rgba(205,214,244,.24);
+  --fg: #CDD6F4; --muted: #A6ADC8; --fg-3: #7F849C;
+  --accent: #CBA6F7; --accent-ink: #0B1024;  /* on --accent 9.3:1 */
+  --accent-2: #89B4FA; --accent-3: #F5C2E7;
+  --good: #A6E3A1; --warn: #F9E2AF; --bad: #F38BA8;
+  --grad: linear-gradient(135deg, #CBA6F7, #89B4FA 50%, #94E2D5);
+  --grad-soft: linear-gradient(135deg, #313244, #45475A);
+  --sub-bg: rgba(0,0,0,.58); --sub-ring: 1px solid rgba(255,255,255,.16);
+  --radius: 14px; --radius-sm: 10px; --radius-lg: 22px;
+  --shadow: 0 10px 30px rgba(0,0,0,.35); --shadow-lg: 0 24px 60px rgba(0,0,0,.5);
+}
+
+/* 科技深色 · nord */
+[data-theme="nord"] {
+  --bg: #2E3440; --bg-soft: #272B35; --panel: #3B4252; --panel-2: #434C5E;
+  --line: rgba(236,239,244,.12); --line-strong: rgba(236,239,244,.24);
+  --fg: #ECEFF4; --muted: #D8DEE9; --fg-3: #7B8394;
+  --accent: #88C0D0; --accent-ink: #0B1024;  /* on --accent 9.4:1 */
+  --accent-2: #81A1C1; --accent-3: #B48EAD;
+  --good: #A3BE8C; --warn: #EBCB8B; --bad: #BF616A;
+  --grad: linear-gradient(135deg, #88C0D0, #81A1C1 50%, #B48EAD);
+  --grad-soft: linear-gradient(135deg, #3B4252, #434C5E);
+  --sub-bg: rgba(0,0,0,.58); --sub-ring: 1px solid rgba(255,255,255,.16);
+  --radius: 12px; --radius-sm: 8px; --radius-lg: 20px;
+  --shadow: 0 10px 30px rgba(0,0,0,.35); --shadow-lg: 0 22px 60px rgba(0,0,0,.5);
+}
+
+/* 消费生活 · 小红书白底高级感 */
+[data-theme="xiaohongshu-white"] {
+  --bg: #FFFDFB; --bg-soft: #FFF6F1; --panel: #FFFFFF; --panel-2: #FFF1EA;
+  --line: rgba(60,30,20,.1); --line-strong: rgba(60,30,20,.22);
+  --fg: #1A1210; --muted: #4F3A32; --fg-3: #A08D85;
+  --accent: #FF2742; --accent-ink: #0B1024;  /* on --accent 5.0:1 */
+  --accent-2: #FF7A90; --accent-3: #FFB38A;
+  --good: #3BA55C; --warn: #F5A524; --bad: #FF2742;
+  --grad: linear-gradient(135deg, #FF2742, #FF7A90 55%, #FFB38A);
+  --grad-soft: linear-gradient(135deg, #FFF6F1, #FFEAE0);
+  --radius: 20px; --radius-sm: 14px; --radius-lg: 28px;
+  --shadow: 0 12px 30px rgba(255,39,66,.08); --shadow-lg: 0 24px 60px rgba(255,39,66,.14);
+}
+
+/* 消费生活 · 柔和马卡龙 */
+[data-theme="soft-pastel"] {
+  --bg: #FDF7FB; --bg-soft: #FBEEF3; --panel: #FFFFFF; --panel-2: #FDF0F5;
+  --line: rgba(120,70,110,.12); --line-strong: rgba(120,70,110,.22);
+  --fg: #3A1F33; --muted: #6B4D62; --fg-3: #A28A99;
+  --accent: #F49BB8; --accent-ink: #0B1024;  /* on --accent 9.2:1 */
+  --accent-2: #B5D5F0; --accent-3: #F7D08A;
+  --good: #9DD9A3; --warn: #F7D08A; --bad: #EF9A9A;
+  --grad: linear-gradient(135deg, #F49BB8, #B5D5F0 55%, #C4A0E8);
+  --grad-soft: linear-gradient(135deg, #FBEEF3, #EAF4FC);
+  --radius: 24px; --radius-sm: 16px; --radius-lg: 32px;
+  --shadow: 0 8px 28px rgba(244,155,184,.18); --shadow-lg: 0 24px 70px rgba(181,213,240,.3);
+}
+
+* { margin: 0; padding: 0; box-sizing: border-box; }
+html, body { width: var(--stage-w, 1920px); height: var(--stage-h, 1080px); overflow: hidden; }
+.stage {
+  position: relative; width: var(--stage-w, 1920px); height: var(--stage-h, 1080px); overflow: hidden;
+  background: var(--bg); color: var(--fg); font-family: var(--font-body);
+}
+.brand {
+  position: absolute; top: 44px; left: 64px;
+  font-size: var(--fs-tiny); letter-spacing: .14em; color: var(--muted);
+}
+.slide-num {
+  position: absolute; bottom: 40px; right: 64px;
+  font-size: var(--fs-tiny); letter-spacing: .1em; color: var(--muted);
+}
+.accent { color: var(--accent); }
+.gradient-text { background: var(--grad); -webkit-background-clip: text; background-clip: text; -webkit-text-fill-color: transparent; color: transparent; }
+
+/* ── 排版角色 ───────────────────────────────────────────────── */
+.kicker { font-size: var(--fs-tiny); font-weight: 600; color: var(--accent); letter-spacing: .08em; }
+.eyebrow { font-size: var(--fs-tiny); font-weight: 500; letter-spacing: .16em; color: var(--fg-3); }
+.lede { font-size: var(--fs-h3); line-height: 1.55; color: var(--muted); font-weight: 300; max-width: 1400px; }
+
+/* ── 卡片 / 标签 / 分隔 ─────────────────────────────────────── */
+.card {
+  background: var(--panel); border: 1px solid var(--line); border-radius: var(--radius);
+  padding: var(--sp-4) var(--sp-5); box-shadow: var(--shadow); position: relative; overflow: hidden;
+}
+.card-soft { background: var(--panel-2); border-color: transparent; box-shadow: none; }
+.card-outline { background: transparent; border: 1.5px solid var(--line-strong); box-shadow: none; }
+.card-accent { border-top: 3px solid var(--accent); }
+.pill {
+  display: inline-block; padding: 4px 14px; border-radius: 999px; font-size: var(--fs-tiny);
+  font-weight: 500; background: var(--panel-2); color: var(--muted); border: 1px solid var(--line);
+}
+.pill-accent {
+  background: var(--panel-2);
+  background: color-mix(in srgb, var(--accent) 14%, transparent);
+  color: var(--accent); border-color: var(--line);
+  border-color: color-mix(in srgb, var(--accent) 30%, transparent);
+}
+.divider-accent { height: 4px; width: 72px; background: var(--accent); border-radius: 2px; }
+
+/* ── 图片框(硬规则: 图片必须套框, 禁止裸放 <img>) ─────────────
+   .img-frame        比例与裁切归框所有; 图片只管填满
+   .img-frame.contain 不裁切(截图/图表/logo 必须用这个), 居中留白
+   --img-ratio       框比例(默认 16/10); --img-pos 主体位置(如 top 保住头部)
+   .img-scrim        底部渐变压暗, 让白字压在照片上仍可读 */
+.img-frame {
+  position: relative; margin: 0; overflow: hidden; display: block;
+  border-radius: var(--radius); box-shadow: var(--shadow); background: var(--panel-2);
+  aspect-ratio: var(--img-ratio, 16/10);
+}
+.img-frame > img { width: 100%; height: 100%; object-fit: cover; object-position: var(--img-pos, center); display: block; }
+.img-frame.contain { background: var(--bg-soft); box-shadow: none; border: 1px solid var(--line); }
+.img-frame.contain > img { object-fit: contain; }
+.img-frame.fill { height: 100%; aspect-ratio: auto; }
+.img-scrim { position: absolute; inset: 0; background: linear-gradient(180deg, transparent 42%, rgba(8,10,20,.72)); }
+.img-cap { margin: 10px 2px 0; font-size: var(--fs-tiny); line-height: 1.5; color: var(--fg-3); }
+.img-tag {
+  position: absolute; top: 14px; left: 14px; padding: 5px 14px; border-radius: 999px;
+  background: rgba(10,12,20,.55); color: #fff; font-size: var(--fs-tiny); letter-spacing: .06em;
+}
+
+/* ── 免责声明 / 出处标注行(受监管题材: 财经/医疗/法律/政策, 见 references/compliance.md)
+   小字不抢视觉; 出现在 closing 张并停留 ≥3s; 口播不念、不进字幕; 加了它也不替代核实。
+   2026-09-28 去掉内置 max-width:1240px —— 那是给单行出处标注设计的, 会把多行整宽
+   免责块卡死在 1240px 居中(项目里显式 width:100% 也压不过它); 需要收窄的项目自己覆写 */
+.disclaimer { font-size: var(--fs-tiny); line-height: 1.6; color: var(--fg-3); }
+.disclaimer-box { border-left: 2px solid var(--line-strong); padding-left: var(--sp-3); }
+
+/* ── 分步入场系统 ──────────────────────────────────────────────
+   要入场的块: data-stage="1|2|3" + fx-* 工具类, 例:
+     <h1 data-stage="1" class="fx-up">大标题</h1>
+     <p  data-stage="2" class="fx-fade">展开内容</p>
+   延迟由管线按口播实测时长注入 --t1/--t2/--t3, HTML 里不要写死。
+   同层内错峰: 用容器 .fx-stagger(基准时刻用 style="--stagger-base:var(--t3)" 指定),
+   或内联 style="animation-delay:calc(var(--t2) + 150ms)"
+   ⚠ data-stage 必须与 fx-* 类同时用 —— 只有 data-stage 没有动画类, 元素会永远停在 opacity:0;
+     容器用 .fx-stagger 时容器自身不要再加 data-stage(子元素自己管入场)。
+   氛围动画(无限循环, 不承载信息)不加 data-stage: fx-pulse / fx-shimmer / fx-kenburns */
+[data-stage] { opacity: 0; --fx-delay: 0ms; }
+[data-stage="1"] { --fx-delay: var(--t1); }
+[data-stage="2"] { --fx-delay: var(--t2); }
+[data-stage="3"] { --fx-delay: var(--t3); }
+
+/* 入场(有限时长, 可被逐帧 seek)。延迟写进 shorthand 的变量槽 —— 不能用独立的
+   animation-delay 声明: 它与下面的 .fx-* 简写同优先级且更靠前, 会被简写重置为 0,
+   导致所有 stage 都在 0 秒入场(2026-09-17 实测确认的坑)。inline animation-delay 仍可覆盖(错峰用)。 */
+.fx-up    { animation: fx-up    var(--dur-mid)  var(--ease-out) var(--fx-delay, 0ms) both; }
+.fx-fade  { animation: fx-fade  var(--dur-fast) var(--ease-out) var(--fx-delay, 0ms) both; }
+.fx-grow  { animation: fx-grow  var(--dur-slow) var(--ease-out) var(--fx-delay, 0ms) both; }
+.fx-blur  { animation: fx-blur  .8s  var(--ease-out) var(--fx-delay, 0ms) both; }   /* 模糊聚焦 */
+.fx-rise  { animation: fx-rise  .9s  var(--ease-out) var(--fx-delay, 0ms) both; }   /* 上浮+微缩放+去模糊 */
+.fx-pop   { animation: fx-pop   .7s  cubic-bezier(.22,1.3,.36,1) var(--fx-delay, 0ms) both; } /* 过冲弹入 */
+.fx-spotlight { animation: fx-spotlight 1.1s var(--ease-out) var(--fx-delay, 0ms) both; }      /* 圆形揭示 */
+.fx-ripple    { animation: fx-ripple    1.2s var(--ease-out) var(--fx-delay, 0ms) both; }      /* 斜角冲开 */
+.fx-glitch    { animation: fx-glitch    .8s  steps(5, end) var(--fx-delay, 0ms) both; }        /* 故障切入 */
+.fx-draw  { animation: fx-draw 1.2s var(--ease-in-out) var(--fx-delay, 0ms) both; } /* 需元素自设 stroke-dasharray:800 */
+.fx-grow-x { transform-origin: left center; animation: fx-grow-x-kf .9s var(--ease-out) var(--fx-delay, 0ms) both; }  /* 图表: 条形从左长出 */
+.fx-grow-y { transform-origin: bottom center; animation: fx-grow-y-kf .9s var(--ease-out) var(--fx-delay, 0ms) both; } /* 图表: 柱状从底长出 */
+
+${wrapKit('nofx', NOFX_CSS)}
+
+@keyframes fx-up   { from { opacity: 0; transform: translateY(32px); } to { opacity: 1; transform: none; } }
+@keyframes fx-fade { from { opacity: 0; } to { opacity: 1; } }
+@keyframes fx-grow { from { opacity: 0; transform: scale(.9); } to { opacity: 1; transform: scale(1); } }
+@keyframes fx-blur { from { opacity: 0; filter: blur(18px); } to { opacity: 1; filter: none; } }
+@keyframes fx-rise { from { opacity: 0; transform: translateY(60px) scale(.97); filter: blur(6px); } to { opacity: 1; transform: none; filter: none; } }
+@keyframes fx-pop  { 0% { opacity: 0; transform: scale(.6); } 60% { transform: scale(1.04); } 100% { opacity: 1; transform: scale(1); } }
+@keyframes fx-spotlight { from { clip-path: circle(0% at 50% 50%); opacity: 1; } to { clip-path: circle(140% at 50% 50%); opacity: 1; } }
+@keyframes fx-ripple    { from { clip-path: circle(0% at 20% 80%); opacity: .4; } to { clip-path: circle(160% at 20% 80%); opacity: 1; } }
+@keyframes fx-glitch {
+  0% { opacity: 0; transform: translateX(0); clip-path: inset(0 0 0 0); }
+  20% { opacity: 1; transform: translateX(-6px); clip-path: inset(20% 0 30% 0); }
+  40% { transform: translateX(4px); clip-path: inset(50% 0 10% 0); }
+  60% { transform: translateX(-3px); clip-path: inset(10% 0 60% 0); }
+  80% { transform: translateX(2px); clip-path: inset(0 0 0 0); }
+  100% { opacity: 1; transform: none; }
+}
+@keyframes fx-draw { from { stroke-dashoffset: 800; opacity: 1; } to { stroke-dashoffset: 0; opacity: 1; } }
+/* 注意: 只做 transform 的动画必须显式带上 opacity:1 —— [data-stage] 的基础态是 opacity:0,
+   靠动画抬回 1; 关键帧不碰 opacity 的动画会让元素永远隐形(2026-09-18 实测) */
+@keyframes fx-grow-x-kf { from { transform: scaleX(0); opacity: 1; } to { transform: scaleX(1); opacity: 1; } }
+@keyframes fx-grow-y-kf { from { transform: scaleY(0); opacity: 1; } to { transform: scaleY(1); opacity: 1; } }
+
+/* 氛围(无限循环, 不参与时长计算) */
+.fx-pulse { animation: fx-pulse 2.4s var(--ease-in-out) infinite; }
+@keyframes fx-pulse { 0%, 100% { opacity: .55; } 50% { opacity: 1; } }
+.fx-shimmer { position: relative; overflow: hidden; }
+.fx-shimmer::after {
+  content: ""; position: absolute; inset: 0;
+  background: linear-gradient(110deg, transparent 40%, rgba(255,255,255,.5) 50%, transparent 60%);
+  transform: translateX(-100%); animation: fx-shimmer-kf 2.4s var(--ease-in-out) infinite;
+}
+@keyframes fx-shimmer-kf { to { transform: translateX(100%); } }
+.fx-kenburns { animation: fx-kenburns-kf 14s ease-in-out infinite alternate; }
+@keyframes fx-kenburns-kf { from { transform: scale(1) translate(0, 0); } to { transform: scale(1.15) translate(-2%, -1%); } }
+
+${wrapKit('chart', CHART_CSS)}
+
+${wrapKit('table', TABLE_CSS)}
+
+/* 依次入场容器: 子元素逐个上浮, 基准时刻取 --stagger-base(默认 --t2)。
+   :not([data-stage]) 是必须的(2026-09-18 实测踩坑): nth-child 延迟规则特异度(0,2,0)高于
+   fx-* 类(0,1,0), 不排除的话会覆盖带 data-stage 子元素自己的 --fx-delay, 元素提前几秒冒出来;
+   语义: 要错峰的块当子元素(不带 data-stage), 要自己管入场时刻的块放容器外面 */
+.fx-stagger > *:not([data-stage]) { opacity: 0; animation: fx-rise .65s var(--ease-out) both; }
+.fx-stagger > *:not([data-stage]):nth-child(1) { animation-delay: calc(var(--stagger-base, var(--t2)) + 0ms); }
+.fx-stagger > *:not([data-stage]):nth-child(2) { animation-delay: calc(var(--stagger-base, var(--t2)) + 120ms); }
+.fx-stagger > *:not([data-stage]):nth-child(3) { animation-delay: calc(var(--stagger-base, var(--t2)) + 240ms); }
+.fx-stagger > *:not([data-stage]):nth-child(4) { animation-delay: calc(var(--stagger-base, var(--t2)) + 360ms); }
+.fx-stagger > *:not([data-stage]):nth-child(5) { animation-delay: calc(var(--stagger-base, var(--t2)) + 480ms); }
+.fx-stagger > *:not([data-stage]):nth-child(6) { animation-delay: calc(var(--stagger-base, var(--t2)) + 600ms); }
+.fx-stagger > *:not([data-stage]):nth-child(7) { animation-delay: calc(var(--stagger-base, var(--t2)) + 720ms); }
+.fx-stagger > *:not([data-stage]):nth-child(n+8) { animation-delay: calc(var(--stagger-base, var(--t2)) + 840ms); }
+`;
+
+/** 生成完整 tokens.css 内容(LF 归一): init-project 写入用它, --upgrade-css 原地替换也用它 */
+export function generateTokensCss() {
+  return TOKENS_BODY.replace(/\r\n/g, '\n');
+}
+
+/** 当前版本的指纹: 内容(含三个工具箱)变了它就变 */
+export const TOKENS_REV = createHash('sha256').update(generateTokensCss()).digest('hex');
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/tools.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/tools.mjs
new file mode 100644
index 00000000..b8240598
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/tools.mjs
@@ -0,0 +1,631 @@
+// html2video-for-mcode · ffmpeg/ffprobe 探测: PATH → node_modules → 常见安装位置
+// 找不到时给可诊断的提示, 而不是让下游脚本报莫名其妙的错。
+import fs from 'node:fs';
+import path from 'node:path';
+import { spawnSync } from 'node:child_process';
+import { kitStatuses } from './css-kit.mjs';
+import { generateTokensCss, TOKENS_REV } from './tokens-template.mjs';
+import { MAX_SCAN_BYTES } from './limits.mjs';
+
+const IS_WIN = process.platform === 'win32';
+const EXE = IS_WIN ? '.exe' : '';
+
+function onPath(name) {
+  const r = spawnSync(name + EXE, ['-version'], { encoding: 'utf8', windowsHide: true });
+  return r.status === 0 && r.stdout ? name + EXE : null;
+}
+
+const SELF_DIR = path.dirname(fileURLToPath(import.meta.url)); // Node 20.11 以下没有 import.meta.dirname, 用 fileURLToPath 保兼容
+
+function candidatePaths(name, projectDir) {
+  const dirs = [];
+  // 调用方项目目录(若给了)、当前工作目录、技能自身位置的**全部祖先**(与 Node import 的模块
+  // 解析同型)。只看固定两级时, 技能被深嵌在 monorepo(plugins/<owner>/<name>/skills/<name>/scripts,
+  // 6 层)会漏掉"仓库根明明装了 ffmpeg-static"的布局 —— 测试进程从仓库根找得到, spawn 出去的
+  // asr.mjs(cwd=临时项目)却找不到, 同一套测试两种结果(发布树实测踩到, 2026-09-28)。
+  // cwd 必须查(二审 P2): README 明说支持把 ffmpeg-static 装在**视频项目**里, 而 prep-image / asr
+  // 是从项目目录被调用的 —— 不查 cwd 就会"项目里明明装了却报找不到"
+  const upDirs = start => {
+    const out = []; let d = path.resolve(start);
+    for (let i = 0; i < 12; i++) {
+      out.push(d);
+      const parent = path.dirname(d);
+      if (parent === d) break;
+      d = parent;
+    }
+    return out;
+  };
+  const anchors = [...new Set([process.env.KIT_PROJECT_DIR, projectDir, process.cwd(), ...upDirs(SELF_DIR)]
+    .filter(Boolean))];
+  for (const base of anchors) {
+    dirs.push(path.join(base, 'node_modules', 'ffmpeg-static'));
+    dirs.push(path.join(base, 'node_modules', 'ffprobe-static', 'bin', process.platform, process.arch));
+  }
+  if (IS_WIN) {
+    const la = process.env.LOCALAPPDATA;
+    if (la) {
+      dirs.push(path.join(la, 'Microsoft', 'WinGet', 'Links'));
+      dirs.push(path.join(la, 'scoop', 'shims'));
+    }
+    dirs.push('C:\\ffmpeg\\bin');
+  }
+  return dirs.map(d => path.join(d, name + EXE));
+}
+
+export function findTool(name, projectDir) {
+  try { const p = onPath(name); if (p) return p; } catch { /* not on PATH */ }
+  if (projectDir) {
+    const local = [
+      path.join(projectDir, 'node_modules', 'ffmpeg-static', name + EXE),
+      path.join(projectDir, 'node_modules', 'ffprobe-static', 'bin', process.platform, process.arch, name + EXE),
+    ];
+    for (const p of local) if (fs.existsSync(p)) return p;
+  }
+  for (const p of candidatePaths(name, projectDir)) if (fs.existsSync(p)) return p;
+  return null;
+}
+
+export function requireTool(name, projectDir) {
+  const p = findTool(name, projectDir);
+  if (!p) {
+    console.error([
+      `✗ 找不到 ${name}。按顺序排查:`,
+      `  1. 系统 PATH 上没有 ${name}`,
+      `  2. 项目/仓库 node_modules 里没有 (试试: npm i ffmpeg-static ffprobe-static)`,
+      `  3. 常见安装位置 (winget Links / scoop / C:\\ffmpeg\\bin) 也没有`,
+      `  最快解法: winget install Gyan.FFmpeg (Windows) · brew install ffmpeg (macOS) · apt install ffmpeg (Linux)`,
+    ].join('\n'));
+    process.exit(2); // 2 = 环境错误, 与 1 (业务错误) 区分
+  }
+  return p;
+}
+
+// ── 输入收监(script.json 是 agent 可编辑文件, 其派生路径必须关进项目目录) ────
+// 威胁模型: script.json 里的 slides[].id / html / audio、顶层 bgm.file 若含
+// "../" 或绝对路径, 可让 fs 读写删与 ffmpeg argv 落到项目目录之外。
+// 所有消费者在 JSON.parse 之后立即调用本节的校验, 失败即退出, 不带病运行。
+
+const ID_RE = /^[A-Za-z0-9_-]{1,64}$/;
+
+export function isSafeId(v) {
+  return typeof v === 'string' && ID_RE.test(v);
+}
+
+// 白名单校验 slide id; 返回原值, 不合格直接退出(并指明是哪个字段)
+export function safeId(v, where = 'slides[].id') {
+  if (!isSafeId(v)) {
+    console.error(`✗ ${where} 非法: ${JSON.stringify(v)} — 只允许字母/数字/下划线/连字符, 长度 1–64`);
+    process.exit(1);
+  }
+  return v;
+}
+
+// 纯字符串判断: resolve 之后 target 是否落在 root 内(target 等于 root 时返回 false)。
+// 这里**不做** realpath —— 符号链接复核在 safeRel / assertContained 里做(它们会取最深已存在祖先的
+// realpath 并 fail closed)。谁需要防"root 自身是指向外面的链接"就用那两个, 别指望这个。
+export function inside(root, target) {
+  const rel = path.relative(path.resolve(root), path.resolve(target));
+  return rel !== '' && !rel.startsWith('..') && !path.isAbsolute(rel);
+}
+
+// 校验"root 下的相对路径"并返回绝对路径:
+//   拒绝绝对路径 → inside(root) → 符号链接复核(取最深已存在祖先的 realpath,
+//   必须仍在 root 的 realpath 内, 防 root 内某段是指向外面的 symlink)
+export function safeRel(root, rel, { where = 'path', mustExist = false } = {}) {
+  if (typeof rel !== 'string' || rel === '') {
+    console.error(`✗ ${where} 非法: ${JSON.stringify(rel)} — 需要非空字符串`);
+    process.exit(1);
+  }
+  if (path.isAbsolute(rel)) {
+    console.error(`✗ ${where} 必须是相对路径: ${JSON.stringify(rel)}`);
+    process.exit(1);
+  }
+  const abs = path.resolve(root, rel);
+  if (!inside(root, abs)) {
+    console.error(`✗ ${where} 越出项目目录: ${JSON.stringify(rel)} → ${abs}`);
+    process.exit(1);
+  }
+  // 符号链接复核: 目标(或其最深"存在或是指向外的悬空链接"的祖先)的 realpath,
+  // 必须仍在 root 的 realpath 内(或是 root realpath 的祖先 —— root 尚不存在时属正常)。
+  // 复查修正(1.7.0): ①回溯用 lstat 而不是 existsSync —— 悬空符号链接 existsSync=false,
+  // 旧写法会把它当"不存在"继续往上走, 从而漏检(写进去就落到项目外); ②realpath 取不到时
+  // **fail closed**(以前 return abs 静默放行, 正好给"造一个解不开的链接"留了门)。
+  // 复查修正(1.7.5): "real 是 root 祖先也放行"的回退只留给 **root 尚不存在** 的场景(合法:
+  // 走过 root 停在已存在祖先)。root 已存在时, 该回退会被"项目内链接指向项目父目录"穿透 ——
+  // 链接的 realpath 是 root 的祖先, 回退放行后, 链接下的**新文件**就写到项目外了(实测复现)。
+  {
+    let probe = abs;
+    while (true) {
+      let st = null;
+      try { st = fs.lstatSync(probe); } catch { /* 真不存在 */ }
+      if (st) break;                          // 存在, 或者是悬空链接(此时也要停下来判)
+      const up = path.dirname(probe);
+      if (up === probe) break;
+      probe = up;
+    }
+    // root 侧与 probe 侧必须用**同一套**规范化(见 canonicalRoot: 七审 blocker 2 的修正)。
+    const { exists: rootExists, real: realRoot } = canonicalRoot(root, where);
+    let real = null;
+    try { real = fs.realpathSync(probe); }
+    catch {
+      // 悬空/自环/无权限的链接: 解不开就不能证明它落在项目内 → 拒绝, 不猜
+      const st = (() => { try { return fs.lstatSync(probe); } catch { return null; } })();
+      if (st && st.isSymbolicLink()) {
+        console.error(`✗ ${where} 命中一个解不开的符号链接: ${path.relative(path.resolve(root), probe) || probe}\n   无法证明它指向项目内, 按越界拒绝(悬空链接常被用来把写入引到项目外)`);
+        process.exit(1);
+      }
+      // 非链接却 realpath 失败(权限等): 同样不放过
+      console.error(`✗ ${where} 无法解析真实路径: ${probe}(realpath 失败)`);
+      process.exit(1);
+    }
+    if (real !== realRoot && !inside(realRoot, real) && !(rootExists ? false : inside(real, realRoot))) {
+      console.error(`✗ ${where} 经符号链接越出项目目录: ${JSON.stringify(rel)}`);
+      process.exit(1);
+    }
+  }
+  if (mustExist && !fs.existsSync(abs)) {
+    console.error(`✗ ${where} 不存在: ${rel}`);
+    process.exit(1);
+  }
+  return abs;
+}
+
+// ── 输出路径收监(2026-09-18 二审 P1: 输出侧此前只查了输入侧) ──────────
+// 威胁模型: 派生输出(preview/<id>.png、build/frames/<id>/…)是"项目内固定位置 + 受监 id",
+// 看起来安全, 但**项目内的目录段本身可能是指向项目外的符号链接**(preview → /tmp/canary,
+// 或 build/frames → 项目外)。此时 writeFileSync/rmSync(recursive) 会穿透符号链接,
+// 在项目外写/删, 而命令一路报成功 —— 用一次性 canary 数据可复现。
+// 规则: 从 root 到目标, **每一段已存在祖先**的 realpath 都必须仍在 root 的 realpath 之内;
+// 叶子自身是符号链接也要拦(否则 ffmpeg/fs 会从那个链接写出去)。
+export function assertContained(root, abs, { where = 'output' } = {}) {
+  // root 侧与 probe 侧同一套规范化(canonicalRoot, 见其注释里的七审 blocker 2 修正)。
+  const { exists: rootExists, real: realRoot } = canonicalRoot(root, where);
+  let probe = path.resolve(abs);
+  while (true) {
+    let st = null;
+    try { st = fs.lstatSync(probe); } catch { /* 真不存在 */ }
+    if (st) break;                                  // 存在或悬空链接: 停下来判
+    const up = path.dirname(probe);
+    if (up === probe) break;
+    probe = up;
+  }
+  let real;
+  try { real = fs.realpathSync(probe); }
+  catch {
+    const st = (() => { try { return fs.lstatSync(probe); } catch { return null; } })();
+    if (st && st.isSymbolicLink()) {
+      console.error(`✗ ${where} 命中解不开的符号链接: ${path.relative(path.resolve(root), probe) || probe}\n   无法证明它指向项目内, 按越界拒绝`);
+      process.exit(1);
+    }
+    console.error(`✗ ${where} 无法解析真实路径: ${probe}(realpath 失败)`);
+    process.exit(1);
+  }
+  // "probe 是 root 的祖先"只在 **root 尚不存在** 时合法(root 到目标之间的段都还没建,
+  // 落盘只会在 root 下新建); root 已存在时绝不放行 —— 那正是"项目内链接指向项目父目录"的
+  // 穿透形状(与 safeRel 的同一处收敛, 1.7.5 已把 safeRel 侧钉死)。
+  if (real !== realRoot && !inside(realRoot, real) && !(rootExists ? false : inside(real, realRoot))) {
+    console.error(`✗ ${where} 经符号链接越出项目目录: ${path.relative(path.resolve(root), abs) || abs}\n   ${probe} → ${real}(项目根 ${realRoot}) — 检查项目内是否有指向外部的符号链接`);
+    process.exit(1);
+  }
+  return abs;
+}
+
+// 派生输出路径的安全拼装: 只接受受监 id 之类的安全片段, 并做上述祖先复核
+export function safeOut(root, ...segs) {
+  const abs = path.join(path.resolve(root), ...segs);
+  if (!inside(path.resolve(root), abs)) {
+    console.error(`✗ 输出路径越出项目目录: ${abs}`);
+    process.exit(1);
+  }
+  return assertContained(root, abs, { where: `输出 ${segs.join('/')}` });
+}
+
+// ── ffprobe 探测(第十三轮 review 去重: duration 三处、尺寸两处逐字重复) ──────────
+// 失败统一返回 null —— 各调用方自己决定报错还是回退, 与既有实现逐字等价。
+export function probeDuration(ffprobe, file) {
+  const r = spawnSync(ffprobe, ['-v', 'error', '-show_entries', 'format=duration', '-of', 'csv=p=0', file],
+    { encoding: 'utf8', windowsHide: true });
+  if (r.status !== 0 || !r.stdout) return null;
+  const d = parseFloat(r.stdout.trim().split('\n')[0]);
+  return Number.isFinite(d) ? d : null;
+}
+export function probeSize(ffprobe, file) {
+  const r = spawnSync(ffprobe, ['-v', 'error', '-select_streams', 'v:0', '-show_entries', 'stream=width,height', '-of', 'csv=p=0', file],
+    { encoding: 'utf8', windowsHide: true });
+  const m = (r.stdout || '').trim().match(/(\d+),(\d+)/);
+  return m ? [parseInt(m[1], 10), parseInt(m[2], 10)] : null;
+}
+
+// ── 对时边界匹配(单一来源, check-timing 消费) ────────────────────────
+// 静音段对齐用**发音形态**(say ?? text)的句读标点 —— 音频里停顿跟的是朗读(1.9.12 第 25 轮:
+// 按 text 对齐时, say 与 text 标点不一致 → 精确模式错位, --calibrate 把错的 start 写回)。
+// 精确模式: 句读标点总数-1 == 静音段数 → 按标点累积下标对齐(最可信)。
+// 最近邻模式: 否则在估算时刻 ±1s 窗内取最近的静音段末端, 时序单调; 找不到就报未测(宁缺勿错)。
+const PAUSE_PUNCT = /[，,、;；:：。！？!?…]/g;
+export function matchBoundaries(gaps, clauses) {
+  const nBound = clauses.length - 1;
+  if (nBound <= 0) return { method: 'none', meas: [] };
+  const marks = clauses.map(c => ((c.say ?? c.text).match(PAUSE_PUNCT) || []).length);
+  const totalMarks = marks.reduce((a, b) => a + b, 0);
+  const meas = new Array(nBound).fill(null);   // 精确模式自己一份
+  if (gaps.length === totalMarks - 1 || gaps.length === totalMarks) {
+    let cum = 0, gi = 0;
+    for (let k = 0; k < nBound; k++) {
+      cum += marks[k]; // 第 k 边界 = 前 k 句累积标点数对应的那个停顿
+      const idx = Math.min(cum - 1, gaps.length - 1);
+      if (idx >= gi) { meas[k] = gaps[idx].end; gi = idx + 1; }
+    }
+    if (meas.every(v => v != null)) return { method: 'exact', meas };
+  }
+  // 最近邻: **重新分配**, 绝不从上面那份里继承已填值 —— 继承的话写回的就是"精确 + 最近邻"的
+  // 混合值, 而 method 只报 nearest, 属于谎报方法(2026-09-18 复查 D1: 精确模式失败即走到此处)。
+  const near = new Array(nBound).fill(null);
+  const est = clauses.slice(1).map(c => c.start);
+  let lastUsed = -1;
+  for (let k = 0; k < nBound; k++) {
+    let best = -1, bestD = Infinity;
+    for (let g = lastUsed + 1; g < gaps.length; g++) {
+      const d = Math.abs(gaps[g].end - est[k]);
+      if (d < bestD && d <= 1.0) { bestD = d; best = g; }
+    }
+    if (best > -1) { near[k] = gaps[best].end; lastUsed = best; }
+  }
+  // 只有**每条边界都有实测**才叫 nearest(可写回); 有缺口的报 sparse, 只展示不写回
+  return { method: near.every(v => v != null) ? 'nearest' : 'sparse', meas: near };
+}
+
+// ── 字幕关键帧(单一来源) ─────────────────────────────────────────────
+// capture 的页面注入(生成 CSS)与窗口自检(算重叠)都从这里取数, 防两处算式漂移。
+// 偏移是"占整张时长的百分比"(0–100)。b 夹在 [a+0.5, 100]: 下一句开口太近时窗口至少留 0.5%,
+// 最后一句起点太晚时不得越过 100%(CSS 关键帧选择器只认 [0%,100%], 越界整条被丢弃)。
+export function subtitleWindows(clauses, duration) {
+  const pct = x => Math.max(0, Math.min(100, (x / duration) * 100));
+  return clauses.map((c, i) => {
+    const a = pct(c.start);
+    const b = Math.min(100, Math.max(pct(clauses[i + 1]?.start ?? duration), a + 0.5));
+    return { a, b, isLast: i === clauses.length - 1 };
+  });
+}
+
+// 每句一对关键帧: 0%→a 藏, a→p1 淡入, [p1,p2] 满亮, p2→b 淡出后藏到片尾。
+// 两个形状约束(2026-09-22 双回归的教训, unit 测试背书):
+// ①同偏移不得声明两次 —— 同偏移多次声明时后者生效, 平台会被尾帧整个吃掉
+//   (灰字幕根因: 末句 b=100 时 "100.000%,100%{opacity:0}" 撞掉 "100.000%{opacity:1}", 整句缓慢淡出);
+// ②两端必须隐藏 —— fill both 下 finish() 后回到 opacity:0, 封面/静帧基底图靠它保持无字幕
+//   (1.9.4 曾改成"末句无尾帧"保住①, 却让终值停在 1: 最后一句被烙进封面与基底)。
+// 于是末句平台收到 99.999%、尾帧单占 100%: 可见部分零淡出, 终值仍归 0。
+export function subtitleKeyframes(clauses, duration) {
+  return subtitleWindows(clauses, duration).map(({ a, b, isLast }, i) => {
+    const win = b - a;
+    const fin = Math.min(Math.max(0.15, win * 0.06), 0.8);
+    const fout = isLast ? 0 : Math.min(Math.max(0.15, win * 0.06), 0.8);
+    const p1 = isLast ? Math.min(a + fin, 99.999) : Math.min(a + fin, b);
+    const p2 = isLast ? 99.999 : Math.max(p1, b - fout);
+    return `@keyframes kit-sub-${i}{0%,${a.toFixed(3)}%{opacity:0}` +
+      `${p1.toFixed(3)}%,${p2.toFixed(3)}%{opacity:1}` +
+      `${b.toFixed(3)}%,100%{opacity:0}}`;
+  }).join('');
+}
+
+// 控制字符与双向覆盖字符剔除(2026-10-10 审计 L4): 字幕/转写文本来自不可信输入,
+// OSC/ANSI 控制序列能把终端染色改标题, bidi 覆盖能让"IMDb"之类显示成别的词。
+// 注意: **不**动普通空格与制表, 也不做大小写或符号改写 —— 只剔掉不可见控制面。
+export function stripControlChars(s) {
+  return String(s ?? '')
+    .replace(/[\u202a-\u202e\u2066-\u2069\u200e\u200f\u202c\u202d\u202e]/g, '')
+    .replace(/[\u0000-\u0008\u000b\u000c\u000e-\u001f\u007f-\u009f]/g, '')
+    .replace(/\r\n?/g, '\n');
+}
+
+// ffmpeg concat 列表的路径守卫(2026-10-10 审计 L6): concat 解复用器用 av_get_token 读条目,
+// 单引号里**放不下**一个单引号, shell 风格的 '\'' 在这里不是转义 —— 路径被悄悄改坏比报错更糟。
+// 路径 = 用户自选项目目录 + 受监 id, 撞上就明确失败, 不产出被改坏的列表文件。
+export function assertConcatPathSafe(p, { where = 'concat 列表' } = {}) {
+  const s = String(p);
+  if (/['\r\n]/.test(s)) {
+    throw new Error(`${where}: 路径里不能有单引号或换行(ffmpeg concat 列表无法安全表示): ${s}\n   请把项目目录或文件名里的 ' 去掉, 或在 script.json 里改用 transition:"xfade"`);
+  }
+  return s;
+}
+
+// ── 字幕文本净化 + SRT 生成(单一来源, 2026-10-10 独立安全审计 H2/L4) ──
+// clause 文案是**不可信输入**, 而 out/subs.srt 是要上传给平台的成品: 文本里带换行就能
+// 伪造成额外的 cue/时间码(实测 2 条 clause 产出 3 个时间码, 伪造块随视频上屏), 控制字符
+// 与 bidi 覆盖字符还能污染播放器/清单的显示。
+// 规则(**承重的一步是最后那条空白折叠** —— JS 的 \s 含 \r \n \u2028 \u2029, 所以换行/CR/
+// 段分隔符都在这一步被折成空格, cue 必然单行): bidi 与控制字符剔除、行内 "-->" 中和、限长。
+// 烧录字幕那条路不需要同一套处理: 那里的文本经 DOM textContent 写入, 没有任何标记语义。
+export function sanitizeCaptionText(s, { maxLen = 400 } = {}) {
+  return String(s ?? '')
+    .replace(/[\u202a-\u202e\u2066-\u2069\u200e\u200f]/g, '')
+    // 行内残留的 "-->" 也要中和: 结构已经安全, 但宽松的 SRT 解析器(ffmpeg 的 srt
+    // demuxer 会**扫行**找时间码)可能把正文里的时间码样式当新 cue 起点。
+    // 字幕几乎不需要字面 "-->", 换成箭头语义不变、歧义消失。
+    .replace(/-->/g, '→')
+    .replace(/[\u0000-\u0008\u000b\u000c\u000e-\u001f\u007f-\u009f]/g, '')
+    .replace(/\s+/g, ' ')
+    .trim()
+    .slice(0, maxLen);
+}
+
+const srtTime = s => {
+  const ms = Math.round((s % 1) * 1000), h = Math.floor(s / 3600), m = Math.floor((s % 3600) / 60), sec = Math.floor(s % 60);
+  return `${String(h).padStart(2, '0')}:${String(m).padStart(2, '0')}:${String(sec).padStart(2, '0')},${String(ms).padStart(3, '0')}`;
+};
+
+// timings.json → SRT 正文。测试直接调它, 所以**别**在这里做 CLI 副作用。
+export function buildSrt(timings) {
+  const cues = [];
+  let idx = 1, cum = 0;
+  for (const t of timings?.slides ?? []) {
+    if (Array.isArray(t.clauses)) {
+      for (let i = 0; i < t.clauses.length; i++) {
+        const c = t.clauses[i];
+        const start = cum + c.start;
+        const end = cum + (t.clauses[i + 1]?.start ?? t.duration);
+        const main = sanitizeCaptionText(c.text);
+        const second = sanitizeCaptionText(c.text2);
+        if (!main && !second) continue;
+        if (end - start < 0.05) continue;
+        cues.push(`${idx++}\n${srtTime(start)} --> ${srtTime(end)}\n${main}${second ? '\n' + second : ''}\n`);
+      }
+    }
+    cum += t.duration;
+  }
+  return { text: cues.join('\n'), count: cues.length };
+}
+
+// ── 路径规范化(比较用) ─────────────────────────────────────────────────
+// 取最深已存在祖先的 realpath 再拼回来: macOS 上 /var/... 与 /private/var/... 是同一目录的两种
+// 写法, 纯字符串比较会把"项目内的合法路径"误判成越界(二审 P2 在官方 CI 的 macOS 上实测踩到)。
+// 取不到 realpath 就返回原样(这是**比较**用的助手, 与 safeRel/assertContained 的 fail-closed 不同)。
+// 2026-09-18 复查 E5: fetch-official-images 里曾有一份同样的实现, 统一到这里。
+export function canonicalPath(p) {
+  const abs = path.resolve(p);
+  let probe = abs;
+  while (!fs.existsSync(probe)) {
+    const up = path.dirname(probe);
+    if (up === probe) return abs;
+    probe = up;
+  }
+  try {
+    const real = fs.realpathSync(probe);
+    return probe === abs ? real : path.join(real, path.relative(probe, abs));
+  } catch { return abs; }
+}
+
+// root 侧的规范化(safeRel / assertContained 共用一份 —— 逐字重复两份正是审查点名的地方):
+// 规则**只有一条**: 必须与 probe 侧同规则, 否则同名不同写法的同一目录会被比成"越界"。
+//   · root 已存在 → realpath(root); 解不开就 fail closed(解不开就不能证明它不是指向外面的链接)
+//   · root 尚不存在 → canonicalPath: 最深已存在祖先的 realpath + 未创建的尾部
+//     (macOS 上 /var/… 与 /private/var/… 是同一目录的两种写法, 拿未规范化字符串去比
+//      probe 的 realpath 会把项目内合法路径判成越界 —— 官方 CI 的 macOS 实测过)
+// 返回 { exists, real }; exists 让调用方决定"root 未建时容忍祖先 probe"这条兜底是否生效。
+export function canonicalRoot(root, where = 'path') {
+  if (!fs.existsSync(root)) return { exists: false, real: canonicalPath(root) };
+  try {
+    return { exists: true, real: fs.realpathSync(path.resolve(root)) };
+  } catch {
+    console.error(`✗ ${where} 无法解析项目目录真实路径: ${root}(realpath 失败)`);
+    process.exit(1);
+  }
+}
+
+// ── 位置参数(项目目录)的取法 ───────────────────────────────────────────
+// 各脚本的用法是 `node x.mjs <项目目录> [--flag 值]`, 但 `<项目目录>` 可以放在任意位置,
+// 于是"第一个非 -- 开头的参数"这个取法会把**取值型 flag 的值**当目录 —— 实测
+// `init-project.mjs --topic 我的主题 <项目目录>` 会在 cwd 下静默建出 `我的主题/` 骨架, 而真正的
+// 项目目录一个字都没写。取值型 flag 名字全技能统一(见下表), 所以用一份清单跳过它们的值。
+// **只列真的带值的**: 布尔开关(--json / --allow-stale-css / --no-subs / --force / --both …)绝不能
+// 进来 —— 进来就会把它后面那个位置参数当"值"吃掉, 于是 <项目目录> 解析成 cwd: 实测
+// `capture --allow-stale-css <项目>` 会对着调用目录干活, 真正的项目一个字节都没被碰。
+// 同理 prep-image 的 --check/--crop 也**不列**: 它们的"参数"本来就是位置参数(一串文件名)。
+// tests/review-round3.test.mjs 有一条从 scripts 源码反扫 argv.includes('--x') 的守卫, 防以后再犯。
+// ── 字幕条(.kit-sub)几何 —— 唯一来源 ─────────────────────────────────
+// capture 注入的字幕胶囊: 宽 = 画布宽 × 0.729167、左右 padding 34px、字号 40px, 其中字号与
+// padding 再乘 --sub-scale = clamp(W/1920, 0.75, 1.25)。行宽(全宽字) = (0.729167·W − 2·34·s)/(40·s·emPer):
+// 中文 1080≈24 字、1920≈33 字、2560≈35 字 —— 2026-09-23 复核: 缩放带钳位, 按宽度线性外推是错的
+// (1.9.9 的 18×W/1080 在 1080 过报 27%、2560 漏报真折行)。emPer = 单字宽按 em 折算(中文全宽 1, 拉丁 ≈0.44)。
+// 2026-09-28 实测勘误(Mavis A/B): 这些行宽值是 width:max-content 生效后的**设计值** —— 旧实现
+// left:50% 的绝对定位可用宽度只有画布一半(1920→960px), max-width 永远够不到, 实际 22 字就折行;
+// 修复后设计值与实测一致(见 subPillCss 的 width:max-content)。
+export const SUB_GEOMETRY = { boxW: 0.729167, padPx: 34, padYPx: 12, fontPx: 40, radiusPx: 14, bottomFrac: 0.077778, sub2Em: 0.74, baseW: 1920 };
+export const subScale = width => Math.min(1.25, Math.max(0.75, (width ?? SUB_GEOMETRY.baseW) / SUB_GEOMETRY.baseW));
+export function subLineCap(width, emPer = 1) {
+  const w = width ?? SUB_GEOMETRY.baseW;
+  const s = subScale(w);
+  return Math.max(8, Math.floor((SUB_GEOMETRY.boxW * w - 2 * SUB_GEOMETRY.padPx * s) / (SUB_GEOMETRY.fontPx * s * emPer)));
+}
+
+// .kit-sub 胶囊样式(单一来源: capture 的页面注入与几何测试共用; 常量全部来自 SUB_GEOMETRY)。
+// width:max-content 是硬约束, 不是优化: left:50% 的绝对定位元素 shrink-to-fit 的可用宽度
+// = 包含块宽 − left 偏移 = 画布的一半(1920→960px), transform 只位移不改布局宽度 ——
+// 于是 max-width: 1400px 永远够不到, 字幕提前一半折行、双行胶囊侵占到距底 230px 压住
+// 图注/免责(2026-09-28 A/B 实测: 22 句 8 句折错; 加 width:max-content 后 0 折行, 恒单行 170px)。
+// 通用教训: left:50%+translateX(-50%) 居中与 max-width 互斥; 要"居中+能撑满"必须显式
+// width:max-content, 或改用 left:0;right:0;margin-inline:auto;max-width:<设计值>。
+export function subPillCss(geom = SUB_GEOMETRY) {
+  return '.kit-sub{position:absolute;left:50%;bottom:calc(var(--stage-h, 1080px) * ' + geom.bottomFrac + ');transform:translateX(-50%);width:max-content;'
+    + 'max-width:calc(var(--stage-w, 1920px) * ' + geom.boxW + ');'
+    + 'background:var(--sub-bg, rgba(12,12,16,.62));color:var(--sub-fg, #fff);'
+    + 'border:var(--sub-ring, 0 solid transparent);'
+    + 'font-size:calc(' + geom.fontPx + 'px * var(--sub-scale, 1));line-height:1.5;'
+    + 'padding:calc(' + geom.padYPx + 'px * var(--sub-scale, 1)) calc(' + geom.padPx + 'px * var(--sub-scale, 1));'
+    + 'border-radius:calc(' + geom.radiusPx + 'px * var(--sub-scale, 1));text-align:center;opacity:0;z-index:9;pointer-events:none;'
+    + 'box-shadow:0 2px 12px rgba(0,0,0,.18)}'
+    + '.kit-sub-2{font-size:' + geom.sub2Em + 'em;opacity:.88;margin-top:6px;letter-spacing:.01em}';
+}
+
+// 字幕带顶部(距画布底的距离): check-slides 的闸门与 authoring 的安全区共用同一公式。
+// 单行 = bottom 偏移(不随 --sub-scale 缩放, CSS 里就是这么写的) + 胶囊高(字号×1.5 行高
+// + 上下 padding, 随缩放) + 2px 渲染余量(实测单行 86px vs 公式 84px)。
+// 双语(text2)再叠一行 .kit-sub-2(0.74em + 6px margin)。
+export function subBandTop(height, width, { bilingual = false } = {}) {
+  const s = subScale(width);
+  const g = SUB_GEOMETRY;
+  const pill = (g.padYPx * 2 + g.fontPx * 1.5) * s + 2;
+  const pill2 = pill + (g.fontPx * g.sub2Em * 1.5 + 6) * s + 2;
+  return Math.round(g.bottomFrac * (height ?? 1080) + (bilingual ? pill2 : pill));
+}
+
+// ── CLI 入口守卫(单一来源) ───────────────────────────────────────────
+// "被 import 不跑主流程, 直接执行才跑" 的判定要稳健: 两侧都取 realpath —— 经符号链接/挂载点
+// 调用时 argv[1] 是调用路径而 import.meta.url 是归一后的路径, 纯字符串比较恒假 → 脚本静默
+// no-op(exit 0 无输出; 2026-09-22 云沙箱实测, 排查花了 10 分钟)。
+// 另注: `node -e "import('...')"` 走的是 import 语义, 不跑 main(设计如此) —— CLI 请 `node <脚本> ...`。
+export function isMainModule(metaUrl, entry = process.argv[1]) {
+  if (!entry) return false;
+  const realpathOf = p => { try { return fs.realpathSync(path.resolve(p)); } catch { return path.resolve(p); } };
+  const meta = fileURLToPath(metaUrl);
+  const eq = process.platform === 'win32'
+    ? (a, b) => a.toLowerCase() === b.toLowerCase()
+    : (a, b) => a === b;
+  if (eq(realpathOf(entry), realpathOf(meta))) return true;
+  // 同名兜底: realpath 归一不了的**未知**路径形态按直接执行处理 —— 这是拿"带提示的多跑一次"
+  // 换"静默不跑"(后者实测烧掉 10 分钟)。命中必须大声宣告, 不许再静默; 误判方向(被同名入口
+  // import)本脚本随后也会响亮报错, 影响面只有 preview-page 一个消费者。
+  if (eq(path.basename(entry), path.basename(meta))) {
+    console.error(`⚠ 入口判定经同名兜底(realpath 未归一: ${entry} vs ${meta}) — 按直接执行处理`);
+    return true;
+  }
+  return false;
+}
+
+export const VALUE_FLAGS = new Set([
+  '--topic', '--mode', '--dsf', '--ids', '--at', '--min', '--max-mb', '--get', '--out', '--out-dir',
+  '--file', '--format', '--from', '--url', '--base-url', '--api-key', '--language',
+  '--ratio', '--timestamp', '--anchor', '--transition',
+]);
+/** 所有位置参数(跳过取值型 flag 的值)。prep-image 这类"一串文件名"的用法用它 */
+export function positionals(argv) {
+  const out = [];
+  for (let i = 0; i < argv.length; i++) {
+    const a = argv[i];
+    if (typeof a !== 'string') continue;
+    if (a.startsWith('--')) { if (VALUE_FLAGS.has(a)) i++; continue; }
+    out.push(a);
+  }
+  return out;
+}
+/** 取位置参数(项目目录): 第一个非 flag 参数; 没有位置参数时用 dflt(默认 cwd) */
+export function positionalDir(argv, { dflt = '.' } = {}) {
+  const first = positionals(argv)[0];
+  return path.resolve(first ?? dflt);
+}
+/** 取值型 flag: flag('--mode', 'still') —— 没有该 flag 时返回 dflt */
+export const flagValue = (argv, name, dflt) => {
+  const i = argv.indexOf(name);
+  return i > -1 ? argv[i + 1] : dflt;
+};
+
+// ── 切页方式的单一解析来源 ─────────────────────────────────────────────
+// script.json 的 transition 允许三种写法: 省略 / "cut" / "xfade" / {type:"xfade",duration:0.4}。
+// 2026-09-18 复查 D6: check-slides 认裸字符串、build-video 只认对象里的 type —— 于是
+// `"xfade"`(字符串)在闸门里显示"交叉溶解", 成片却是硬切, 两边说法不一致还没有任何报错。
+// 现在两边都调这个函数: 同一个值必然得到同一个结论, 非法值都在自己的入口报错。
+export const XFADE_DEFAULT_DUR = 0.4;
+export function readTransition(v, { where = 'script.transition' } = {}) {
+  const raw = v == null ? 'cut' : (typeof v === 'string' ? v : v?.type);
+  const type = typeof v === 'object' && v !== null && raw == null ? 'cut' : raw;
+  if (!['cut', 'xfade'].includes(type)) {
+    throw new Error(`${where} 非法: ${JSON.stringify(typeof v === 'object' ? v?.type : v)} — 只支持 "cut"(硬切, 默认) 或 "xfade"(交叉溶解), 或 {type,duration} 对象`);
+  }
+  const durRaw = typeof v === 'object' && v !== null ? (v.duration ?? XFADE_DEFAULT_DUR) : XFADE_DEFAULT_DUR;
+  const dur = Number(durRaw);
+  if (type === 'xfade' && (!Number.isFinite(dur) || dur <= 0 || dur > 2)) {
+    throw new Error(`${where}.duration 非法: ${JSON.stringify(durRaw)} — 需要 0–2 秒`);
+  }
+  return { type, dur: type === 'xfade' ? dur : 0, label: type === 'cut' ? '硬切 cut' : `交叉溶解 xfade ${dur}s` };
+}
+
+// ── tokens.css 受管块新鲜度闸门(entry point 共用) ──────────────────────
+// 为什么要在 capture / build-video 入口硬拦: 受管块落后时**画面照出, 只是旧样式** ——
+// 新类没有样式、旧规则按层叠压过新版, 全程没有任何报错, 成片出来才发现(audit 2026-09-18)。
+// 所以"陈旧"必须是阻断级, 并把修复命令直接给到手上; 确实要用旧 CSS 出片时显式 --allow-stale-css。
+const EXPECT_TOKENS = { text: generateTokensCss(), rev: TOKENS_REV };
+/** 受管区的新鲜度判据(技能当前版正文 + rev) —— check-slides / preview-page 与本模块共用同一份,
+ *  免得各自再拼一次 generateTokensCss()(散着拼就会出现"某个调用点忘了带 rev"这类漂移) */
+export const expectedTokens = () => EXPECT_TOKENS;
+/** tokens.css 里所有非 ok 的受管块(缺/旧/重复/坏定界/区外残留副本) */
+export function cssStaleIssues(cssText) {
+  return kitStatuses(cssText, { expected: EXPECT_TOKENS }).filter(s => s.status !== 'ok');
+}
+/** 读项目 tokens.css 并做新鲜度闸门: 落后 → 打印清单 + 修复命令并退出 1(除非 allowStale) */
+export function requireFreshCss(dir, { who, allowStale = false } = {}) {
+  const cssPath = path.join(dir, 'slides', 'tokens.css');
+  if (!fs.existsSync(cssPath)) return '';
+  // 扫描上限与 check-slides / init-project / preview-page 同一道门(声明在 limits.mjs):
+  // 少了这一步, 这里就是唯一一个会去读 6MB tokens.css 的入口。
+  const size = fs.statSync(cssPath).size;
+  if (size > MAX_SCAN_BYTES) {
+    console.error(`✗ ${who}: slides/tokens.css 有 ${(size / 1e6).toFixed(1)}MB, 超过 ${MAX_SCAN_BYTES / 1e6}MB 上限, 拒绝扫描(正常项目 ≈15KB)`);
+    process.exit(1);
+  }
+  const css = fs.readFileSync(cssPath, 'utf8');
+  const bad = cssStaleIssues(css);
+  if (!bad.length) return css;
+  const lines = bad.map(s => `   - ${s.label}: [${s.status}] ${s.detail}`);
+  if (allowStale) {
+    console.warn(`⚠ ${who}: slides/tokens.css 的受管块与技能当前版不一致(--allow-stale-css 已显式跳过闸门):\n${lines.join('\n')}`);
+    return css;
+  }
+  console.error(`✗ ${who}: slides/tokens.css 的受管块与技能当前版不一致 —— 画面会照出, 但用的是旧 CSS(新类静默无样式, 全程不报错):\n${lines.join('\n')}`);
+  console.error(`  → 先修: node scripts/init-project.mjs ${dir} --upgrade-css(原地替换受管区, 不动区外的规则; 原件备份在 tokens.css.bak)`);
+  console.error('  → 确实要用项目里的旧 CSS 出片: 加 --allow-stale-css 显式跳过');
+  process.exit(1);
+}
+
+// 一次性走查 script.json 的全部路径派生字段(两个调用点: 每个 consumer 读入后)
+export function validateScriptPaths(script, dir) {
+  const list = Array.isArray(script?.slides) ? script.slides : [];
+  for (const s of list) {
+    safeId(s.id, `slides[].id (${JSON.stringify(s.title ?? '')})`);
+    if (s.html !== undefined) safeRel(path.join(dir, 'slides'), s.html, { where: `slides[${s.id}].html` });
+    if (s.audio !== undefined) safeRel(path.join(dir, 'audio'), s.audio, { where: `slides[${s.id}].audio` });
+  }
+  const bgm = script?.bgm;
+  const bgmFile = typeof bgm === 'string' ? bgm : bgm?.file;
+  if (bgmFile !== undefined) safeRel(dir, bgmFile, { where: 'bgm.file' });
+}
+
+// timings.json 的 slides[].id 是 script.json 的二次传播, 消费侧同样校验
+export function validateTimingsIds(timings) {
+  const list = Array.isArray(timings?.slides) ? timings.slides : [];
+  for (const t of list) safeId(t.id, 'timings.slides[].id');
+}
+
+// ── Node 包解析(playwright) ──────────────────────────────────────
+// 技能通常装在 ~/.claude/skills/ 或 ~/.openclaw/skills/ —— 不在项目树里,
+// 而 Node 的 import 是按"脚本所在位置"向上找 node_modules 的, 于是会出现
+// "项目里明明装了 playwright 却报找不到"。这里按多个锚点依次尝试。
+import { createRequire } from 'node:module';
+import { pathToFileURL, fileURLToPath } from 'node:url';
+
+const PKG_ANCHORS = () => {
+  const list = [process.env.KIT_PROJECT_DIR, process.cwd()];
+  // 全局 npm root(全局装的 playwright 也能用)
+  for (const npm of IS_WIN ? ['npm.cmd', 'npm'] : ['npm']) {
+    try {
+      const r = spawnSync(npm, ['root', '-g'], { encoding: 'utf8', windowsHide: true });
+      if (r.status === 0 && r.stdout) { list.push(r.stdout.trim()); break; }
+    } catch { /* ignore */ }
+  }
+  return [...new Set(list.filter(Boolean))];
+};
+
+export async function loadPackage(name, { projectDir } = {}) {
+  // CJS 包(如 playwright)动态 import 后命名导出可能拿不到, 统一解包出真实模块对象
+  const unwrap = mod => (mod && (mod.chromium || mod.default?.chromium)) ? (mod.chromium ? mod : mod.default) : mod;
+  // 1) 脚本自身位置(技能装在项目内, 或全局 node_modules 可被解析时)
+  try { const m = unwrap(await import(name)); if (m?.default !== undefined || m) return m; } catch { /* 继续找 */ }
+  // 2) 项目目录 / 调用目录 / 全局 npm root, 逐个用 createRequire 解析真实入口
+  const anchors = [...new Set([projectDir, ...PKG_ANCHORS()].filter(Boolean))];
+  for (const base of anchors) {
+    try {
+      const req = createRequire(path.join(base, 'package.json'));
+      const entry = req.resolve(name);
+      return unwrap(await import(pathToFileURL(entry).href));
+    } catch { /* 试下一个 */ }
+  }
+  return null;
+}
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/url-policy.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/url-policy.mjs
new file mode 100644
index 00000000..e0225440
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/scripts/url-policy.mjs
@@ -0,0 +1,355 @@
+// html2video-for-mcode · 网络 URL 策略(纯函数, 供 asr.mjs / fetch-official-images.mjs 与测试复用)
+// 两个职责:
+//   1. 出网端点白名单 —— asr.mjs 只把 API Key 发往官方 MiniMax 域; 换端点必须显式
+//      --allow-any-endpoint(危险项), 防止被环境变量/提示词偷换后凭证外发。
+//   2. 抓图目标校验 —— fetch-official-images.mjs 的页面 URL、每个候选图 src、每一跳
+//      重定向都过同一套 host 策略: 拦 loopback / 链路本地(含云元数据)/私网 / 无点主机名,
+//      只允许 http(s), file:// 需显式 --allow-file, 禁 userinfo。
+// host 判定是纯函数; DNS 相关(assertResolvedHost/checkedLookup/policyGet/startVetoProxy)
+// 做 IO —— 那是它们的职责。抛 PolicyError(带 reason code)由调用方决定退出码与文案。
+
+import dns from 'node:dns';
+import http from 'node:http';
+import https from 'node:https';
+import net from 'node:net';
+
+export const OFFICIAL_ASR_BASES = ['https://api.minimaxi.com', 'https://api.minimax.io'];
+
+export class PolicyError extends Error {
+  constructor(reason, detail) {
+    super(`${reason}: ${detail ?? ''}`);
+    this.reason = reason;
+    this.detail = detail;
+  }
+}
+
+// ── host 判定 ────────────────────────────────────────────────────
+
+function ipv4ToLong(ip) {
+  const parts = ip.split('.').map(Number);
+  if (parts.length !== 4 || parts.some(p => !Number.isInteger(p) || p < 0 || p > 255)) return null;
+  return parts.reduce((a, p) => a * 256 + p, 0);
+}
+
+// [网络基址, 前缀位数]; 计算一律走无符号, 避开 int32 符号位比较的坑
+// 2026-09-18 复查补全: 除私网/回环/链路本地外, 还要拦保留段与文档段 ——
+// 前者常被内网设备使用, 后者虽不可路由, 但放行等于把"地址分类"这层判断交给下游。
+const BLOCKED_V4 = [
+  [0b00000000, 8],        // 0.0.0.0/8        "this host"
+  [10 << 24, 8],          // 10.0.0.0/8       私网
+  [100 << 24 | 64 << 16, 10], // 100.64.0.0/10 CGNAT(常被漏掉的内网段)
+  [127 << 24, 8],         // 127.0.0.0/8      loopback
+  [169 << 24 | 254 << 16, 16], // 169.254.0.0/16 链路本地(含云元数据 169.254.169.254)
+  [172 << 24 | 16 << 16, 12], // 172.16.0.0/12  私网
+  [192 << 24 | 0 << 16, 24], // 192.0.0.0/24   IETF 协议保留
+  [192 << 24 | 0 << 16 | 2 << 8, 24], // 192.0.2.0/24 TEST-NET-1
+  [192 << 24 | 168 << 16, 16],// 192.168.0.0/16 私网
+  [198 << 24 | 18 << 16, 15], // 198.18.0.0/15  基准测试段
+  [198 << 24 | 51 << 16 | 100 << 8, 24], // 198.51.100.0/24 TEST-NET-2
+  [203 << 24 | 0 << 16 | 113 << 8, 24],  // 203.0.113.0/24 TEST-NET-3
+  [224 << 24, 4],         // 224.0.0.0/4     组播
+  [240 << 24, 4],         // 240.0.0.0/4     保留(含 255.255.255.255)
+].map(([base, bits]) => [base >>> 0, bits]);
+
+// host 是否属于禁止出网/抓取的地址(或形态)
+export function isBlockedHost(hostname) {
+  // 尾点 FQDN 归一: `localhost.` 与 `localhost` 是同一个名字(RFC 1034 绝对名),
+  // 只判裸名会让 `http://localhost./` 直接穿过去(2026-09-18 实测放行)
+  const h = String(hostname ?? '').toLowerCase().replace(/^\[|\]$/g, '').replace(/\.+$/, '');
+  if (!h) return true;
+  if (h.includes(':')) { // IPv6
+    if (h === '::' || h === '::1') return true;                    // unspecified / loopback
+    if (/^f[cd][0-9a-f]{2}:/.test(h)) return true;                 // fc00::/7 私网(ULA)
+    if (/^fe[89ab][0-9a-f]:/.test(h)) return true;                 // fe80::/10 链路本地
+    if (/^fec[0-9a-f]:/.test(h)) return true;                      // fec0::/10 站点本地(已废弃)
+    if (/^2001:0?db8:/.test(h)) return true;                       // 2001:db8::/32 文档段
+    // 内嵌 IPv4 的三种载体都要换算回点分再判, 否则可借壳穿透:
+    //   ::ffff:x:y   IPv4-mapped(WHATWG URL 会把点分规范成 hex 组)
+    //   64:ff9b::x:y NAT64   6to4 2002:AABB:CCDD::/48(前 32 位就是 IPv4)
+    const embedded = (() => {
+      let m = /^::ffff:(.+)$/.exec(h);
+      if (m) return m[1];
+      // 展开写法 0:0:0:0:0:ffff:7f00:1 —— 同一台机器的 net.isIP() 认它是 IPv6(6),
+      // 而它不以 ::ffff: 开头, 旧写法整条漏判(2026-10-10 审计 L5; 目前经 new URL() 规范化
+      // 后不可达, 属于潜伏缺陷, 但这是安全函数, 补上更便宜)。
+      m = /^(?:0*:)*ffff:([0-9a-f]{1,4}):([0-9a-f]{1,4})$/.exec(h);
+      if (m) return `${m[1]}:${m[2]}`;
+      m = /^64:ff9b::(.+)$/.exec(h);
+      if (m) return m[1];
+      m = /^2002:([0-9a-f]{1,4}):([0-9a-f]{1,4})/.exec(h);
+      if (m) {
+        const hi = parseInt(m[1], 16) >>> 0, lo = parseInt(m[2], 16) >>> 0;
+        return `${(hi >> 8) & 255}.${hi & 255}.${(lo >> 8) & 255}.${lo & 255}`;
+      }
+      return null;
+    })();
+    if (embedded) {
+      const m = /^([0-9a-f]{1,4}):([0-9a-f]{1,4})$/.exec(embedded);   // 两个 hex 组 = 32 位 IPv4
+      if (m) {
+        const hi = parseInt(m[1], 16) >>> 0, lo = parseInt(m[2], 16) >>> 0;
+        return isBlockedHost(`${(hi >> 8) & 255}.${hi & 255}.${(lo >> 8) & 255}.${lo & 255}`);
+      }
+      return isBlockedHost(embedded);      // 已经是点分或还带别的形态, 继续判
+    }
+    return false;
+  }
+  const ipv4 = ipv4ToLong(h);
+  if (ipv4 !== null) {
+    for (const [base, prefixBits] of BLOCKED_V4) {
+      const mask = prefixBits === 0 ? 0 : (0xffffffff << (32 - prefixBits)) >>> 0;
+      if (((ipv4 & mask) >>> 0) === ((base & mask) >>> 0)) return true;
+    }
+    return false;
+  }
+  // 非字面量: 无点主机名(localhost / 内网裸名 / NetBIOS)拦下; 内网风格后缀拦下
+  if (!h.includes('.')) return true;
+  if (['.localhost', '.local', '.internal', '.localdomain', '.home.arpa'].some(sfx => h.endsWith(sfx))) return true;
+  return false;
+}
+
+// ── 出网端点(asr) ────────────────────────────────────────────────
+
+// base 必须是官方端点; 允许用 --base-url/env 在两个官方域之间切换(region 对齐),
+// 任何其他值都拒 —— 除非 allowAny(显式危险项)。http 一律拒绝(官方端点全是 https)。
+export function assertAsrEndpoint(base, { allowAny = false } = {}) {
+  const normalized = String(base ?? '').replace(/\/+$/, '');
+  if (OFFICIAL_ASR_BASES.includes(normalized)) return normalized;
+  if (allowAny) {
+    let u;
+    try { u = new URL(normalized); } catch { throw new PolicyError('bad-url', normalized); }
+    if (u.protocol !== 'https:' && u.protocol !== 'http:') throw new PolicyError('bad-scheme', normalized);
+    console.warn(`⚠ --allow-any-endpoint: API Key 将发往非官方端点 ${normalized}(自担风险, 仅用于自建网关/测试)`);
+    return normalized;
+  }
+  throw new PolicyError('endpoint-not-allowed',
+    `${normalized} 不在 ASR 端点白名单(${OFFICIAL_ASR_BASES.join(' / ')})。确需自定义请显式加 --allow-any-endpoint`);
+}
+
+// ── 抓图目标(fetch-official-images) ─────────────────────────────
+
+// 校验一个要访问的 URL; allowFile 时放行 file: 协议(本地离线页是文档化场景)
+export function assertFetchableUrl(raw, { allowFile = false, where = 'url' } = {}) {
+  let u;
+  try { u = new URL(raw); } catch { throw new PolicyError('bad-url', `${where}: ${raw}`); }
+  if (u.protocol === 'file:') {
+    if (!allowFile) throw new PolicyError('file-not-allowed', `${where}: file:// 需要显式 --allow-file`);
+    if (u.username || u.password) throw new PolicyError('userinfo', where);
+    return u;
+  }
+  if (u.protocol !== 'https:' && u.protocol !== 'http:') throw new PolicyError('bad-scheme', `${where}: ${u.protocol}`);
+  if (u.username || u.password) throw new PolicyError('userinfo', `${where}: 带凭据的 URL`); // 凭据会进日志/请求头
+  if (isBlockedHost(u.hostname)) throw new PolicyError('blocked-host', `${where}: ${u.hostname} 是内网/本地/元数据地址`);
+  return u;
+}
+
+// 重定向跳转是否允许(逐跳调用; 每一跳都要过同一套校验)
+export function assertRedirectTarget(location, { where = 'redirect' } = {}) {
+  return assertFetchableUrl(location, { allowFile: false, where });
+}
+
+export const MAX_REDIRECTS = 5;
+
+// 解析后再验一层(2026-09-18 二审 P1): 字符串层拦得住字面 IP, 拦不住"公网域名解析回内网"
+// (DNS rebinding / 内网域名)。**每个真实请求**都要过这里 —— 页面导航、它的重定向、浏览器子资源、
+// 逐跳下载 —— 否则策略只覆盖了"最初输入的那个 URL"。lookup 可注入(测试用假解析器, 无需真 DNS)。
+// 2026-09-25 六审 blocker 1: 解析失败/空答案不再放行 —— 旧实现 catch 后 return 等于
+// "解析不了就放行"(fail open), 钉死的语义应该是 fail closed。
+export async function assertResolvedHost(urlStr, { lookup, where = 'url' } = {}) {
+  let h;
+  try { h = new URL(urlStr).hostname; } catch { return; }
+  if (!h || !h.includes('.')) return;              // 裸主机名/非法 URL 已被字符串层拦掉
+  const doLookup = lookup || (async name => dns.promises.lookup(name, { all: true }));
+  let addr;
+  try {
+    addr = await doLookup(h);
+  } catch (e) {
+    throw new PolicyError('dns-error', `${where}: ${h} 解析失败(${e?.message ?? e}) — fail closed, 拒绝继续`);
+  }
+  const list = (Array.isArray(addr) ? addr : [addr]).filter(a => a && typeof a.address === 'string');
+  if (!list.length) throw new PolicyError('dns-error', `${where}: ${h} 解析结果为空 — fail closed`);
+  const bad = list.find(a => isBlockedHost(a.address));
+  if (bad) throw new PolicyError('dns-rebinding', `${where}: ${h} 解析到被拦地址 ${bad.address}(DNS rebinding 或内网域名? 换官方 CDN 直链)`);
+}
+
+// ── 连接期绑定(2026-09-25 六审 blocker 1) ─────────────────────────
+// 预查(assertResolvedHost)与真正建连之间隔着一次**独立解析** —— 攻击面就是这道缝:
+// 预查时公网、建连时私网(DNS rebinding), 或两台解析器答案不一致。唯一权威的检查点是
+// socket 建连用的那次 lookup。本节三个件:
+//   checkedLookup   把否决权挂到 http(s)/net 的 lookup 选项上(Node lookup 兼容签名)
+//   policyGet       配套出网 GET(不跟重定向, 由调用方逐跳复核), 建连带否决
+//   startVetoProxy  浏览器(Chromium 自己解析, 注入不了 lookup)的全部出网收进本地隧道,
+//                   上游建连仍走同一否决
+
+// Node lookup 签名 (hostname, options, callback) 的策略包装: 解析失败按失败处理(建连即败,
+// fail closed); 结果里**任何一个**地址被拦就整体拒绝 —— Happy-Eyeballs 会在地址列表里挑,
+// "部分放行"等于没拦。resolve 可注入(与 assertResolvedHost 的 lookup 同一形态, 测试免真 DNS)。
+export function checkedLookup(resolve) {
+  const doResolve = resolve || ((name, opts) => dns.promises.lookup(name, { family: opts?.family ?? 0, hints: opts?.hints, all: true }));
+  return (hostname, options, callback) => {
+    doResolve(hostname, options).then(raw => {
+      const all = (Array.isArray(raw) ? raw : [raw]).filter(a => a && typeof a.address === 'string');
+      const fam = options?.family;
+      const pool = (fam === 4 || fam === 6) ? all.filter(a => a.family === fam) : all;
+      if (!pool.length) return callback(new Error(`dns: ${hostname} 无可用解析结果${fam ? `(family ${fam})` : ''}`));
+      const bad = pool.find(a => isBlockedHost(a.address));
+      if (bad) return callback(new PolicyError('dns-rebinding', `建连: ${hostname} 解析到被拦地址 ${bad.address}(连接期否决)`));
+      // 只行使否决权, 不改变解析语义: 原样回传让 net 自己挑地址
+      if (options?.all) return callback(null, pool);
+      callback(null, pool[0].address, pool[0].family);
+    }, callback);
+  };
+}
+
+// 出网 GET, 建连那次解析经 checkedLookup 否决。不跟重定向 —— 重定向由调用方逐跳复核
+// (fetch-official-images 的 followRedirects), 每一跳的建连都被单独否决。
+// 返回与 followRedirects 的 get 契约一致的归一化结果。maxBytes 在**流式读取时**掐断,
+// 不信用 content-length(恶意服务器会谎报)。createConnection 仅测试注入假上游用。
+export function policyGet(rawUrl, { headers = {}, timeoutMs = 30000, maxBytes = Infinity, lookup, createConnection } = {}) {
+  return new Promise((resolve, reject) => {
+    let u;
+    try { u = new URL(rawUrl); } catch { return reject(new PolicyError('bad-url', String(rawUrl))); }
+    // 建连**之前**的字符串层 host 策略(七审 blocker 1): IP 字面量目标根本不走 lookup ——
+    // Node 直接连 IP, checkedLookup 对它是空转, 于是 "http://127.0.0.1:8080/" 这类目标
+    // 没有任何否决点(生产 startVetoProxy 的绝对形式 HTTP 分支经本地 origin 实测拿到 200
+    // 且 origin 实收请求)。这一层是所有 policyGet 调用方的**唯一**字面量入口。
+    try { assertFetchableUrl(u.href, { where: 'policyGet' }); }
+    catch (e) { return reject(e); }
+    const mod = u.protocol === 'https:' ? https : u.protocol === 'http:' ? http : null;
+    if (!mod) return reject(new PolicyError('bad-scheme', String(u.protocol)));
+    const req = mod.request(u, {
+      headers,
+      lookup: checkedLookup(lookup),
+      ...(createConnection ? { createConnection } : { agent: false }),   // agent:false = 不留池化连接(工具进程要能自然退出)
+    }, res => {
+      const chunks = []; let total = 0;
+      res.on('data', c => {
+        total += c.length;
+        if (total > maxBytes) { res.destroy(new Error(`响应超限: 已收 ${(total / 1048576).toFixed(0)}MB 超过上限(读取中途掐断, 不信用 content-length)`)); return; }
+        chunks.push(c);
+      });
+      res.on('end', () => resolve({
+        status: res.statusCode,
+        location: res.headers.location,
+        headers: { ...res.headers },
+        header: n => res.headers[String(n).toLowerCase()] || '',
+        arrayBuffer: async () => Buffer.concat(chunks),
+      }));
+      res.on('error', reject);
+    });
+    req.setTimeout(timeoutMs, () => req.destroy(new Error(`请求超时(${timeoutMs}ms): ${u.host}`)));
+    req.on('error', reject);
+    req.end();
+  });
+}
+
+// 本地否决代理: Chromium 的页面导航/子资源自己解析 DNS, 没有 lookup 钩子可注入 ——
+// 把它的全部出网收进这条隧道才能把策略绑到"实际建连的那次解析"上:
+//   CONNECT(https) → 上游建连带 checkedLookup, 建连前还做一次字符串层+解析预否决(可命名告警)
+//   绝对形式 http  → 经 policyGet 出网(连接期否决), 原样回贴
+// 只监听 127.0.0.1(它本身就是策略咽喉, 不是出网目标)。connect 可注入(测试假上游)。
+const HOP_BY_HOP = ['connection', 'keep-alive', 'proxy-connection', 'te', 'trailer', 'transfer-encoding', 'upgrade'];
+export async function startVetoProxy({ lookup, connect = net.connect, maxBytes = 30 * 1024 * 1024 } = {}) {
+  const sockets = new Set();
+  const track = s => { sockets.add(s); s.on('close', () => sockets.delete(s)); return s; };
+  const server = http.createServer((req, res) => {
+    if (!req.url || !/^https?:\/\//i.test(req.url)) { res.writeHead(400); return res.end('veto-proxy: 只收绝对形式 URL'); }
+    // 每个代理 HTTP 请求都先过 host 策略(与下面 CONNECT 分支同形的可命名预否决): 字面量
+    // 目标不触发 checkedLookup, 缺这道预否决时 policyGet 之前的绝对形式分支等于没有闸门。
+    try { assertFetchableUrl(req.url, { where: '代理 HTTP' }); }
+    catch (e) { res.writeHead(502, { 'content-type': 'text/plain; charset=utf-8' }); return res.end(`veto-proxy: ${e.message}`); }
+    const fwd = {};
+    for (const [k, v] of Object.entries(req.headers)) {
+      if (k === 'host' || HOP_BY_HOP.includes(k)) continue;   // Host 由目标 URL 决定, 逐跳头不转发
+      fwd[k] = v;
+    }
+    policyGet(req.url, { headers: fwd, lookup, createConnection: connect, maxBytes, timeoutMs: 45000 })
+      .then(async r => {
+        const out = { ...r.headers };
+        delete out['transfer-encoding']; delete out['connection'];   // 已整包缓冲, 传输头自算
+        res.writeHead(r.status, out);
+        res.end(await r.arrayBuffer());
+      })
+      .catch(e => {
+        try { res.writeHead(502, { 'content-type': 'text/plain; charset=utf-8' }); res.end(`veto-proxy: ${e.message}`); }
+        catch { /* 客户端已断 */ }
+      });
+  });
+  server.on('connect', (req, clientSocket, head) => {
+    // 可命名的拒绝: 隧道没建起来时回一个 502 + 原因, 而不是静默掐断 —— 客户端拿到的是
+    // "代理拒绝了并说明理由", 排查不必去猜。绝对形式 HTTP 分支同样带名字, 两个分支对齐。
+    const fail = (why) => {
+      const body = `veto-proxy: ${why ?? 'refused'}`;
+      try {
+        clientSocket.write('HTTP/1.1 502 Bad Gateway\r\nContent-Type: text/plain; charset=utf-8\r\n' +
+          `Content-Length: ${Buffer.byteLength(body)}\r\n\r\n${body}`);
+      } catch { /* 客户端已断 */ }
+      clientSocket.destroy();
+    };
+    const m = /^(\[[^\]]+\]|[^:]+):(\d+)$/.exec(req.url || '');
+    if (!m) return fail();
+    const host = m[1], port = parseInt(m[2], 10);
+    const bare = host.replace(/^\[|\]$/g, '');
+    const literal = /^[\d.]+$/.test(bare) || (bare.includes(':') && /^[0-9a-f:]+$/i.test(bare));
+    (async () => {
+      if (isBlockedHost(bare)) throw new PolicyError('blocked-host', `代理 CONNECT: ${bare}`);
+      // 名字(非字面 IP)先做一次可命名的预否决; 权威检查点仍是下面建连里的 checkedLookup
+      if (!literal && (bare.includes('.') || bare.includes(':'))) {
+        await assertResolvedHost(`https://${bare}/`, { lookup, where: '代理 CONNECT' });
+      }
+      const up = track(connect({ host, port, lookup: checkedLookup(lookup), noDelay: true }, () => {
+        track(clientSocket);
+        clientSocket.write('HTTP/1.1 200 Connection Established\r\n\r\n');
+        if (head && head.length) up.write(head);
+        up.pipe(clientSocket);
+        clientSocket.pipe(up);
+      }));
+      up.on('error', e => fail(`上游连接失败: ${e?.message ?? e}`));
+      clientSocket.on('error', () => up.destroy());
+      up.on('close', () => clientSocket.destroy());
+      clientSocket.on('close', () => up.destroy());
+    })().catch(e => fail(e?.message ?? e));
+  });
+  server.on('connection', track);
+  await new Promise(r => server.listen(0, '127.0.0.1', r));
+  return {
+    port: server.address().port,
+    close: async () => { for (const s of sockets) s.destroy(); await new Promise(r => server.close(r)); },
+  };
+}
+
+// ── 落盘文件名(抓图下载) ─────────────────────────────────────────
+
+const WIN_RESERVED = /^(con|prn|aux|nul|com[1-9]|lpt[1-9])$/i;
+
+// alt/标题转成安全文件名: 去路径分隔符与控制字符、压长度、挡 Windows 保留名与开头点/空格
+export function sanitizeFilename(name, { maxLen = 30, fallback = 'official' } = {}) {
+  let s = String(name ?? '')
+    .replace(/[\u0000-\u001f\u007f]/g, '')
+    .replace(/[\\/:*?"<>|]+/g, '-')
+    .replace(/\s+/g, '-')
+    .replace(/^[.\s-]+/, '')            // 不允许开头点/空格/连字符(点开头=隐藏文件)
+    .replace(/[.\s-]+$/, '')
+    .slice(0, maxLen)
+    .replace(/^[.\s-]+|[.\s-]+$/g, ''); // 截断后再清一次边缘
+  if (!s || WIN_RESERVED.test(s)) s = fallback;
+  return s;
+}
+
+// 图片 URL → 落盘文件名(下载零散 URL 时用): 取路径末段, 扩展名优先用 URL 自己的,
+// 没有/不像扩展名就按 content-type 推, 都不行给 .bin(交给人眼确认后再登记)。
+const EXT_BY_TYPE = {
+  'image/png': '.png', 'image/jpeg': '.jpg', 'image/webp': '.webp',
+  'image/gif': '.gif', 'image/svg+xml': '.svg', 'image/avif': '.avif',
+};
+
+export function imageNameFromUrl(raw, contentType = '') {
+  let pathname = '', base = '';
+  try {
+    pathname = new URL(raw).pathname;
+    base = decodeURIComponent(pathname.split('/').filter(Boolean).pop() || '');
+  } catch { /* 非法 URL: 走 fallback */ }
+  const extInPath = /^\.[a-z0-9]{1,5}$/i.test(pathname.slice(pathname.lastIndexOf('.'))) ? pathname.slice(pathname.lastIndexOf('.')).toLowerCase() : '';
+  const ext = extInPath || EXT_BY_TYPE[String(contentType).split(';')[0].trim().toLowerCase()] || '.bin';
+  const stem = base.replace(/\.[a-z0-9]{1,5}$/i, '');
+  return sanitizeFilename(stem, { maxLen: 30, fallback: 'official' }) + ext;
+}
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/chart-kit.test.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/chart-kit.test.mjs
new file mode 100644
index 00000000..30eb76f8
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/chart-kit.test.mjs
@@ -0,0 +1,77 @@
+// 图表工具箱: 模板里必须真带着这些规则, 且"关动效 = 终态"这条不变量成立。
+// 覆盖过三类静默失败: 缺 keyframes → 条形不生长; 静态默认 ≠ 终值 → 关动效后是空的;
+// 柱高相对整列算 → 被 flex-shrink 压回, 柱高不再等于数值。
+import { test, describe } from 'node:test';
+import assert from 'node:assert/strict';
+import fs from 'node:fs';
+import path from 'node:path';
+import { runSkill, tmpdir } from './helpers.mjs';
+import { CHART_CSS, hasChartKit } from '../scripts/chart-css.mjs';
+
+const proj = tmpdir();
+const r = runSkill('init-project.mjs', [proj, '--topic', 'T']);
+assert.equal(r.status, 0, r.stderr);
+const css = fs.readFileSync(path.join(proj, 'slides', 'tokens.css'), 'utf8');
+
+describe('图表工具箱 · 模板内容', () => {
+  test('五个新动效类都在, 且关键帧都声明了 opacity(否则入场后永远隐形)', () => {
+    for (const cls of ['fx-grow-w', 'fx-grow-h', 'fx-sweep', 'fx-count', 'fx-dot']) {
+      assert.ok(css.includes('.' + cls + ' {') || css.includes('.' + cls + '{'), `缺 .${cls}`);
+    }
+    const names = ['fx-grow-w-kf', 'fx-grow-h-kf', 'fx-sweep-kf', 'fx-count-kf', 'fx-dot-kf'];
+    for (const n of names) {
+      const i = css.indexOf('@keyframes ' + n);
+      assert.ok(i > -1, `缺关键帧 ${n}`);
+      const body = css.slice(i, css.indexOf('\n}', i));
+      assert.match(body, /opacity\s*:/, `${n} 缺 opacity(该元素会永远隐形)`);
+    }
+  });
+
+  test('注册属性存在(没有它, 扫出/数字滚动无法被逐帧 seek)', () => {
+    assert.match(css, /@property\s+--pv\s*\{\s*syntax:\s*'<number>'/);
+    assert.match(css, /@property\s+--cnt\s*\{\s*syntax:\s*'<integer>'/);
+  });
+
+  test('静态默认 = 终值(关掉动效后画面停在完成态, 不是空的)', () => {
+    assert.match(css, /\.fx-sweep\s*\{[^}]*--pv:\s*var\(--p-to/);
+    assert.match(css, /\.fx-count\s*\{[^}]*--cnt:\s*var\(--n-to/);
+    assert.match(css, /\.fx-grow-w\s*\{[^}]*width:\s*var\(--w/);
+  });
+
+  test('柱状三层结构齐全(绘图区 / 高度基准 / 标签), 网格在绘图区里', () => {
+    assert.match(css, /\.chart-plot-cell\s*\{/, '缺 .chart-plot-cell');
+    assert.match(css, /\.chart-plot-cell\.grid\s*\{[^}]*repeating-linear-gradient/, '网格必须画在绘图区上(否则与柱高两把尺子)');
+    assert.match(css, /\.chart-bar-wrap\s*\{/, '缺 .chart-bar-wrap(柱高的百分比基准)');
+    assert.match(css, /\.chart-bar-wrap\s+\.chart-val\s*\{[^}]*position:\s*absolute/, '数值要贴各自柱顶');
+    assert.match(css, /\.chart-col\s*\{[^}]*grid-template-rows/, '列必须是 [绘图区 1fr] + [轴标签]');
+    assert.match(css, /\.chart-plot-cell\s+\.chart-target\s*\{/, '目标虚线要落在绘图区里(才有 0 与轴上限的参照)');
+  });
+
+  test('横向条形的可比性: 行是三列定宽 grid, 数值外置不影响轨道长度', () => {
+    assert.match(css, /\.chart-row\s*\{[^}]*grid-template-columns:\s*var\(--lbl-w[^;]*var\(--val-w/, '行必须是定宽三列');
+    assert.match(css, /\.chart-val\.out\s*\{/, '缺 .chart-val.out(条外数值)');
+    assert.match(css, /\.chart-bar\.dim\s+\.chart-val\s*\{/, '弱化条上的数值要自动改墨色(白字读不出来)');
+  });
+});
+
+describe('图表工具箱 · 老项目升级', () => {
+  test('--upgrade-css 会把图表工具箱补进老 tokens.css, 且幂等', () => {
+    const old = tmpdir();
+    fs.mkdirSync(path.join(old, 'slides'), { recursive: true });
+    fs.writeFileSync(path.join(old, 'slides', 'tokens.css'), ':root { --accent: #111; }\n');
+    assert.equal(hasChartKit(':root { --accent: #111; }'), false);
+    const r1 = runSkill('init-project.mjs', [old, '--upgrade-css']);
+    assert.equal(r1.status, 0, r1.stderr);
+    const after = fs.readFileSync(path.join(old, 'slides', 'tokens.css'), 'utf8');
+    assert.match(after, /@property\s+--pv/);
+    assert.match(after, /fx-sweep-kf/);
+    assert.ok(after.startsWith(':root { --accent: #111; }'), '原内容必须在前');
+    const r2 = runSkill('init-project.mjs', [old, '--upgrade-css']);
+    assert.equal(fs.readFileSync(path.join(old, 'slides', 'tokens.css'), 'utf8'), after, '第二次跑不得再追加');
+    assert.ok(r2.stdout.includes('无需升级'));
+  });
+
+  test('CHART_CSS 自身通过 hasChartKit 判定(防止判定与内容脱节)', () => {
+    assert.equal(hasChartKit(CHART_CSS), true);
+  });
+});
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/connect-binding.test.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/connect-binding.test.mjs
new file mode 100644
index 00000000..0f9b9e41
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/connect-binding.test.mjs
@@ -0,0 +1,432 @@
+// 连接期绑定(2026-09-25 六审 blocker 1): 两件事 —— ① 解析失败从"放行"改为 fail closed;
+// ② 策略从"请求前预查"绑到"实际建连的那次解析"上(预查时公网、建连时私网 = DNS rebinding 的缝)。
+// 判据锚在"被拒目标收到 0 个连接"这类几何事实上, 不看文案; 假解析器注入走 url-policy 的
+// lookup/resolve 接缝, 不需要真 DNS。
+import { test, describe } from 'node:test';
+import assert from 'node:assert/strict';
+import net from 'node:net';
+import http from 'node:http';
+import fs from 'node:fs';
+import path from 'node:path';
+import { runSkill, tmpdir, SCRIPTS } from './helpers.mjs';
+import { assertResolvedHost, checkedLookup, policyGet, startVetoProxy, PolicyError } from '../scripts/url-policy.mjs';
+
+const PUB = [{ address: '93.184.216.34', family: 4 }];   // example.com 的公网地址, 只当"公网样子"用
+const LOOPBACK = [{ address: '127.0.0.1', family: 4 }];
+
+const listen = (srv, host = '127.0.0.1') => new Promise(r => srv.listen(0, host, r));
+
+describe('fail closed: 解析失败/空答案不再是放行(六审 blocker 1 前半)', () => {
+  test('assertResolvedHost: 解析器抛错 → 拒绝(旧实现静默放行)', async () => {
+    await assert.rejects(
+      () => assertResolvedHost('https://gone.example.com/x.png', { lookup: async () => { throw new Error('ENOTFOUND gone.example.com'); } }),
+      e => e instanceof PolicyError,
+    );
+  });
+  test('assertResolvedHost: 空答案 → 拒绝', async () => {
+    await assert.rejects(
+      () => assertResolvedHost('https://empty.example.com/x', { lookup: async () => [] }),
+      e => e instanceof PolicyError,
+    );
+  });
+});
+
+describe('checkedLookup(连接期否决, Node lookup 兼容签名)', () => {
+  // 断言不能写进 lookup 的 callback 里直接抛 —— 抛错会落进 .then 吞成 unhandled rejection,
+  // 测试假绿(否决被临时拿掉时当场抓到); 一律先 await 捕获结果再断言。
+  const call = (lk, host, opts) => new Promise(res => lk(host, opts, (...a) => res(a)));
+
+  test('公网地址原样放行(单地址与 all 两种回传形态)', async () => {
+    const lk = checkedLookup(async () => PUB);
+    const [e1, list] = await call(lk, 'ok.example.com', { all: true });
+    assert.equal(e1, null); assert.deepEqual(list, PUB);
+    const [e2, addr, fam] = await call(lk, 'ok.example.com', {});
+    assert.equal(e2, null); assert.equal(addr, '93.184.216.34'); assert.equal(fam, 4);
+  });
+  test('解析到被拦地址 → callback 收到 dns-rebinding, 建连必败', async () => {
+    const lk = checkedLookup(async () => LOOPBACK);
+    const [e] = await call(lk, 'rebind.example.com', { all: true });
+    assert.ok(e instanceof PolicyError, '应抛 PolicyError, 实得: ' + e);
+    assert.equal(e.reason, 'dns-rebinding');
+  });
+  test('解析器报错 → 错误透传(建连失败, 不降级放行)', async () => {
+    const lk = checkedLookup(async () => { throw new Error('ESERVFAIL'); });
+    const [e] = await call(lk, 'x.example.com', { all: true });
+    assert.match(String(e?.message), /ESERVFAIL/);
+  });
+});
+
+describe('policyGet: 建连那次解析被否决(rebinding 的权威检查点)', () => {
+  test('连接期解析回环 → 拒绝, 目标服务器收到 0 个请求(真走 net.lookup 路径, 无注入建连)', async () => {
+    let hits = 0;
+    const srv = http.createServer((q, s) => { hits++; s.writeHead(200); s.end('x'); });
+    await listen(srv);
+    try {
+      await assert.rejects(
+        () => policyGet(`http://rebind.example.test:${srv.address().port}/a.png`, { lookup: async () => LOOPBACK }),
+        e => /被拦地址|dns-rebinding/.test(e.message),
+      );
+      assert.equal(hits, 0, '否决必须发生在连接之前 —— 服务器一个请求都不能收到');
+    } finally { srv.close(); }
+  });
+  test('管道贯通: 状态/头/体回传正确, lookup 确实接线到建连层', async () => {
+    const srv = http.createServer((q, s) => { s.writeHead(200, { 'content-type': 'image/png' }); s.end('PNGDATA'); });
+    await listen(srv);
+    let sawLookup = false;
+    try {
+      const res = await policyGet(`http://ok.example.test:${srv.address().port}/a.png`, {
+        lookup: async () => PUB,
+        // createConnection 仅替换传输(测试假上游): lookup 必须作为建连参数传进来, 否则否决没接线
+        createConnection: opts => { sawLookup = typeof opts.lookup === 'function'; return net.connect({ host: '127.0.0.1', port: srv.address().port }); },
+      });
+      assert.equal(sawLookup, true, 'lookup 必须传进建连层');
+      assert.equal(res.status, 200);
+      assert.equal(res.header('content-type'), 'image/png');
+      assert.equal(String(await res.arrayBuffer()), 'PNGDATA');
+    } finally { srv.close(); }
+  });
+  test('响应超限在读取中途掐断(不信用 content-length)', async () => {
+    const srv = http.createServer((q, s) => { s.writeHead(200); s.end('A'.repeat(64)); });  // 不写长度=分块, 全量送达
+    await listen(srv);
+    try {
+      await assert.rejects(
+        () => policyGet(`http://ok.example.test:${srv.address().port}/big`, {
+          lookup: async () => PUB,
+          createConnection: () => net.connect({ host: '127.0.0.1', port: srv.address().port }),
+          maxBytes: 16,
+        }),
+        e => /超限/.test(e.message),
+      );
+    } finally { srv.close(); }
+  });
+});
+
+describe('startVetoProxy: 浏览器全部出网的否决隧道', () => {
+  const lookupByName = map => async name => map[name] ?? PUB;
+  // 注入的 connect 必须保留第二参 connectListener —— 生产路径 net.connect(options, cb) 靠它回调
+  const connectTo = port => (opts, cb) => net.connect({ host: '127.0.0.1', port }, cb);
+
+  function connectHandshake(port) {
+    // 模拟浏览器侧: 裸 socket 发 CONNECT / 绝对形式请求由各用例自己拼
+    return net.connect({ host: '127.0.0.1', port });
+  }
+
+  test('CONNECT 公网名 → 隧道建立且双向通(回显上游)', async () => {
+    const echo = net.createServer(sock => sock.on('data', d => sock.write(d)));
+    await listen(echo);
+    const proxy = await startVetoProxy({ lookup: lookupByName({}), connect: connectTo(echo.address().port) });
+    try {
+      const verdict = await new Promise((resolve, reject) => {
+        const c = connectHandshake(proxy.port);
+        let phase = 'head', buf = '';
+        const timer = setTimeout(() => { c.destroy(); reject(new Error('隧道握手/回显超时')); }, 5000);
+        c.on('connect', () => c.write('CONNECT ok.example.com:443 HTTP/1.1\r\n\r\n'));
+        c.on('data', d => {
+          buf += d.toString('latin1');
+          if (phase === 'head') {
+            if (!buf.includes('\r\n\r\n')) return;
+            assert.ok(/^HTTP\/1\.1 200/.test(buf), `握手应 200: ${buf.slice(0, 40)}`);
+            phase = 'echo'; c.write('PING');
+          } else if (buf.endsWith('PING')) { clearTimeout(timer); c.destroy(); resolve('tunnel-ok'); }
+        });
+        c.on('error', e => { clearTimeout(timer); reject(e); });
+      });
+      assert.equal(verdict, 'tunnel-ok');
+    } finally { proxy.close(); echo.close(); }
+  });
+
+  test('CONNECT 解析回环(rebinding) → 建连前拒绝, 上游收到 0 个连接', async () => {
+    const echo = net.createServer(sock => sock.on('data', d => sock.write(d)));
+    await listen(echo);
+    let upstreamConnects = 0;
+    const proxy = await startVetoProxy({
+      lookup: lookupByName({ 'rebind.example.com': LOOPBACK }),
+      connect: opts => { upstreamConnects++; return connectTo(echo.address().port)(opts); },
+    });
+    try {
+      const seen = await new Promise(resolve => {
+        let buf = '';
+        const c = connectHandshake(proxy.port);
+        c.on('connect', () => c.write('CONNECT rebind.example.com:443 HTTP/1.1\r\n\r\n'));
+        c.on('data', d => { buf += d.toString(); c.destroy(); resolve(buf); });
+        c.on('error', () => resolve(buf));   // destroy 可能以 ECONNRESET 形态到达
+        c.on('close', () => resolve(buf));
+        setTimeout(() => { c.destroy(); resolve(buf); }, 5000);
+      });
+      // 1.9.17: CONNECT 的拒绝从"静默掐断"改为可命名的 502 + 原因(不建隧道, 原因写明)
+      assert.match(seen, /^HTTP\/1\.\d 502/, `拒绝应是 502, 实得: ${seen.slice(0, 120)}`);
+      assert.match(seen, /dns-rebinding|连接期否决/, `拒绝要点名连接期否决: ${seen.slice(0, 200)}`);
+      assert.ok(!/200 Connection Established/.test(seen), '不许建立隧道');
+      assert.equal(upstreamConnects, 0, '否决必须发生在上游建连之前');
+    } finally { proxy.close(); echo.close(); }
+  });
+
+test('CONNECT 字面内网地址(字符串层) → 拒绝可命名(1.9.17: 从静默掐断改为 502 + 原因)', async () => {
+      let upstreamConnects = 0;
+      const proxy = await startVetoProxy({ lookup: lookupByName({}), connect: opts => { upstreamConnects++; return net.connect(opts); } });
+      try {
+        const seen = await new Promise(resolve => {
+          let buf = '';
+          const c = connectHandshake(proxy.port);
+          c.on('connect', () => c.write('CONNECT 169.254.169.254:80 HTTP/1.1\r\n\r\n'));
+          c.on('data', d => { buf += d.toString(); c.destroy(); resolve(buf); });
+          c.on('error', () => resolve(buf));
+          c.on('close', () => resolve(buf));
+          setTimeout(() => { c.destroy(); resolve(buf); }, 5000);
+        });
+        assert.match(seen, /^HTTP\/1\.\d 502/, `拒绝应是可读的 502, 实得: ${seen.slice(0, 120)}`);
+        assert.match(seen, /blocked-host/, `拒绝要点名原因: ${seen.slice(0, 200)}`);
+        assert.ok(!/200 Connection Established/.test(seen), '不许建立隧道');
+        assert.equal(upstreamConnects, 0, '否决必须发生在上游建连之前');
+      } finally { proxy.close(); }
+    });
+
+  test('绝对形式 http: 连接期解析回环 → 502 且源站 0 请求(生产路径, 未注入建连)', async () => {
+    let hits = 0;
+    const srv = http.createServer((q, s) => { hits++; s.writeHead(200); s.end('page'); });
+    await listen(srv);
+    const proxy = await startVetoProxy({ lookup: lookupByName({ 'rebind.example.test': LOOPBACK }) });
+    try {
+      const res = await new Promise((resolve, reject) => {
+        const timer = setTimeout(() => { req.destroy(); reject(new Error('代理绝对形式请求超时')); }, 5000);
+        const req = http.get({ host: '127.0.0.1', port: proxy.port, path: `http://rebind.example.test:${srv.address().port}/page` }, r => {
+          const chunks = []; r.on('data', c => chunks.push(c)); r.on('end', () => { clearTimeout(timer); resolve({ status: r.statusCode, body: Buffer.concat(chunks).toString() }); });
+        });
+        req.on('error', e => { clearTimeout(timer); reject(e); });
+      });
+      assert.equal(res.status, 502, '回源被否决应回 502');
+      assert.equal(hits, 0, '源站一个请求都不能收到');
+    } finally { proxy.close(); srv.close(); }
+  });
+
+  test('绝对形式 http: 公网名 → 经代理透传, 状态/头/体一致(传输注入假上游)', async () => {
+    let hits = 0;
+    const srv = http.createServer((q, s) => { hits++; s.writeHead(200, { 'content-type': 'text/html' }); s.end('<html>ok</html>'); });
+    await listen(srv);
+    const proxy = await startVetoProxy({ lookup: lookupByName({}), connect: connectTo(srv.address().port) });
+    try {
+      const res = await new Promise((resolve, reject) => {
+        const timer = setTimeout(() => { req.destroy(); reject(new Error('代理绝对形式请求超时')); }, 5000);
+        const req = http.get({ host: '127.0.0.1', port: proxy.port, path: `http://ok.example.test:${srv.address().port}/page` }, r => {
+          const chunks = []; r.on('data', c => chunks.push(c)); r.on('end', () => { clearTimeout(timer); resolve({ status: r.statusCode, type: r.headers['content-type'], body: Buffer.concat(chunks).toString() }); });
+        });
+        req.on('error', e => { clearTimeout(timer); reject(e); });
+      });
+      assert.equal(res.status, 200);
+      assert.equal(res.type, 'text/html');
+      assert.equal(res.body, '<html>ok</html>');
+      assert.equal(hits, 1);
+    } finally { proxy.close(); srv.close(); }
+  });
+});
+
+// CLI 的 --url 负向接线(127.0.0.1 字面量 + localtest.me 解析回环)由 review-round2 的既有用例
+// 覆盖 —— 它们走的就是换装后的 openUrl→policyGet 路径; 本文件不重复造环境依赖的判据
+// (.invalid 之类"必然解析失败"的域名在不同 DNS 环境行为不一, 不作测试依据)。
+
+describe('渲染 Chromium 也必须进否决隧道(第 23 轮收口, 结构钉)', () => {
+  test('capture.mjs 的浏览器启动带 proxy(去隧道 = slide 子资源可直连私网, 第 23 轮半做项)', () => {
+    const src = fs.readFileSync(path.join(SCRIPTS, 'capture.mjs'), 'utf8');
+    assert.match(src, /startVetoProxy/, 'capture 必须启动否决代理');
+    assert.match(src, /proxy:\s*\{\s*server:\s*`http:\/\/127\.0\.0\.1:\$\{veto\.port\}`/, 'launch 必须把浏览器流量收进隧道');
+  });
+});
+
+describe('浏览器模式接线(带否决代理启动; 需 playwright/chromium, 缺则按能力 skip)', () => {
+  test('file:// 页面 --allow-file --list --json 正常列出候选(代理在链上不碍事)', async t => {
+    const d = tmpdir();
+    // 1×1 PNG: naturalWidth=1, 过滤器(默认 --min 0)必放行
+    fs.writeFileSync(path.join(d, 'shot.png'), Buffer.from('iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mP8z8BQDwAEhQGAhKmMIQAAAABJRU5ErkJggg==', 'base64'));
+    fs.writeFileSync(path.join(d, 'page.html'), '<!doctype html><html><body><img src="shot.png" alt="hero"></body></html>');
+    const pageUrl = 'file:///' + path.join(d, 'page.html').replace(/\\/g, '/');
+    const r = await runSkill('fetch-official-images.mjs', [pageUrl, '--allow-file', '--json'], { cwd: d });
+    const out = String(r.stdout + r.stderr);
+    if (/未找到 playwright|chromium 未安装|Executable doesn't exist|Looks like Playwright/.test(out)) return t.skip('无 playwright/chromium');
+    assert.equal(r.status, 0, out.slice(0, 400));
+    assert.match(out, /shot\.png/, '候选里应列出页面图片: ' + out.slice(0, 300));
+  });
+});
+
+// 七审 blocker 1: 生产 startVetoProxy 的**绝对形式 HTTP** 分支此前不进 host 策略 —— 而
+// IP 字面量目标 Node 不调用 lookup, checkedLookup 对它是空转。本地 origin 实测拿到 200 且
+// 实收请求。判据钉两件事: ①被拒目标收到 **0 个请求**; ②502 正文点名 blocked-host ——
+// 否则"目标不可达"同样表现为 502, 会让断言假绿。
+describe('七审 blocker 1 · 生产否决代理: IP 字面量目标 0 请求', () => {
+  const viaProxy = (proxyPort, target) => new Promise(resolve => {
+    const req = http.request({
+      host: '127.0.0.1', port: proxyPort, path: target,      // 绝对形式 URL = 代理请求行
+      headers: { host: new URL(target).host },
+    }, r => {
+      const chunks = [];
+      r.on('data', c => chunks.push(c));
+      r.on('end', () => resolve({ status: r.statusCode, body: Buffer.concat(chunks).toString('utf8') }));
+    });
+    req.on('error', e => resolve({ status: 'ERR', body: e.message }));
+    req.end();
+  });
+  const countingOrigin = async () => {
+    const hits = { n: 0 };
+    const srv = http.createServer((req, res) => { hits.n++; res.writeHead(200); res.end('SECRET'); });
+    await listen(srv);
+    return { srv, hits, port: srv.address().port };
+  };
+
+  test('loopback 字面量 → 0 请求, 且拒绝来自策略(正文点名 blocked-host)', async () => {
+    const { srv, hits, port } = await countingOrigin();
+    const proxy = await startVetoProxy({});                  // 生产参数: 不注入 connect/lookup
+    try {
+      const r = await viaProxy(proxy.port, `http://127.0.0.1:${port}/secret`);
+      assert.equal(r.status, 502, r.body);
+      assert.match(r.body, /blocked-host/, `拒绝必须来自 host 策略, 实得: ${r.body.slice(0, 200)}`);
+      assert.equal(hits.n, 0, '被拒目标必须收到 0 个请求');
+    } finally { await proxy.close(); srv.close(); }
+  });
+
+  test('私网/元数据/IPv6/零/裸名 字面量 → 全部点名 blocked-host(不可达目标也要证明是策略拦的)', async () => {
+    const proxy = await startVetoProxy({});
+    try {
+      for (const u of ['http://192.168.1.1/admin', 'http://169.254.169.254/latest/meta-data/',
+        'http://[::1]:9/x', 'http://0.0.0.0:9/x', 'http://10.0.0.5/internal', 'http://localhost:9/x',
+        'http://100.64.0.1/x', 'http://198.18.0.1/x']) {
+        const r = await viaProxy(proxy.port, u);
+        assert.equal(r.status, 502, `${u} → ${r.status} ${r.body.slice(0, 120)}`);
+        assert.match(r.body, /blocked-host/, `${u} 的 502 必须点名 host 策略, 实得: ${r.body.slice(0, 160)}`);
+      }
+    } finally { await proxy.close(); }
+  });
+
+  test('主机名解析回环 → 仍由 checkedLookup 在建连期否决(0 请求, 正文点名 dns-rebinding)', async () => {
+    const { srv, hits, port } = await countingOrigin();
+    const proxy = await startVetoProxy({ lookup: async () => LOOPBACK });   // 真建连路径
+    try {
+      const r = await viaProxy(proxy.port, `http://rebind.example.test:${port}/secret`);
+      assert.equal(r.status, 502, r.body);
+      assert.match(r.body, /dns-rebinding|连接期否决/, r.body.slice(0, 200));
+      assert.equal(hits.n, 0, '被拒目标必须收到 0 个请求');
+    } finally { await proxy.close(); srv.close(); }
+  });
+
+  test('正向对照: 放行的主机名仍能走通绝对形式 HTTP 分支(别把管道焊死)', async () => {
+    const upstream = http.createServer((req, res) => { res.writeHead(200, { 'content-type': 'text/plain' }); res.end('upstream-ok'); });
+    await listen(upstream);
+    const upPort = upstream.address().port;
+    const proxy = await startVetoProxy({
+      lookup: async () => PUB,
+      connect: () => net.connect({ host: '127.0.0.1', port: upPort }),   // 假上游: 只换传输
+    });
+    try {
+      const r = await viaProxy(proxy.port, 'http://cdn.example.com/a.png');
+      assert.equal(r.status, 200, r.body.slice(0, 200));
+      assert.match(r.body, /upstream-ok/);
+    } finally { await proxy.close(); upstream.close(); }
+  });
+});
+
+// ── 自审轮(1.9.17): 把两层否决各自钉住 ──────────────────────────────────
+// policyGet 的 assertFetchableUrl(where:'policyGet') 与代理 HTTP 分支的预否决(where:'代理 HTTP')
+// 此前各自撤掉都测不出来(只撤一处, 另一处照样拦)。两处 where 标签不同 → 断言各自的名字,
+// 两层就都成了载荷: 少一层, 对应断言立刻红。
+describe('自审 1.9.17 · 两层否决各自有名字、各自被钉住', () => {
+  test('policyGet 直连字面量 → 拒, 理由带 policyGet 这一层的名字', async () => {
+    let hits = 0;
+    const srv = http.createServer((req, res) => { hits++; res.writeHead(200); res.end('SECRET'); });
+    await listen(srv);
+    const port = srv.address().port;
+    try {
+      await assert.rejects(
+        () => policyGet(`http://127.0.0.1:${port}/secret`),
+        e => {
+          assert.equal(e.reason, 'blocked-host', `应点名 blocked-host, 实得 ${e.reason}`);
+          assert.match(e.message, /policyGet/, `拒绝应带本层名字, 实得: ${e.message}`);
+          return true;
+        },
+      );
+      assert.equal(hits, 0, '被拒目标必须收到 0 个请求');
+    } finally { srv.close(); }
+  });
+
+  test('代理明文 HTTP 分支的拒绝带"代理 HTTP"这一层的名字(撤掉这层预否决则断言红)', async () => {
+    const proxy = await startVetoProxy({});
+    try {
+      const r = await new Promise(resolve => {
+        const req = http.request({
+          host: '127.0.0.1', port: proxy.port, path: 'http://127.0.0.1:9/x',
+          headers: { host: '127.0.0.1:9' },
+        }, res => {
+          const c = [];
+          res.on('data', d => c.push(d));
+          res.on('end', () => resolve({ status: res.statusCode, body: Buffer.concat(c).toString() }));
+        });
+        req.on('error', e => resolve({ status: 'ERR', body: e.message }));
+        req.end();
+      });
+      assert.equal(r.status, 502, r.body);
+      assert.match(r.body, /代理 HTTP/, `应带代理层名字, 实得: ${r.body.slice(0, 160)}`);
+      assert.match(r.body, /blocked-host/);
+    } finally { await proxy.close(); }
+  });
+
+  test('CONNECT 分支的拒绝同样可命名(不是静默掐断)', async () => {
+    const proxy = await startVetoProxy({});
+    try {
+      // 用裸 socket: Node 的 http.request 对 method:'CONNECT' 走 connect 事件而非 response,
+      // 挂在 response 上会永远等不到(挂死过一次)。
+      const r = await new Promise(resolve => {
+        const sock = net.connect(proxy.port, '127.0.0.1', () => {
+          sock.write('CONNECT 169.254.169.254:80 HTTP/1.1\r\nHost: 169.254.169.254:80\r\n\r\n');
+        });
+        let buf = '';
+        const done = () => {
+          const head = buf.split('\r\n\r\n')[0] || '';
+          const status = Number(/HTTP\/1\.\d (\d{3})/.exec(head)?.[1] ?? 0);
+          const body = buf.slice(head.length + 4);
+          sock.destroy();
+          resolve({ status, body });
+        };
+        sock.setTimeout(4000, done);
+        sock.on('data', d => { buf += d.toString(); if (buf.includes('\r\n\r\n')) done(); });
+        sock.on('error', e => resolve({ status: 'ERR', body: e.message }));
+      });
+      assert.equal(r.status, 502, `CONNECT 拒绝应是 502, 实得 ${r.status} ${r.body.slice(0, 120)}`);
+      assert.match(r.body, /blocked-host/, `CONNECT 的拒绝也要点名原因: ${r.body.slice(0, 160)}`);
+    } finally { await proxy.close(); }
+  });
+});
+
+// 浏览器侧不变量: 幻灯片里的 http://127.0.0.1 子资源**必须**经过否决代理(Chromium 对 loopback
+// 有隐式代理绕过; 目前 Playwright + 显式 proxy 实测是走代理的, 但那是**依赖默认值**的行为 ——
+// 这里把它钉成契约, 默认值一变 CI 立刻报)。
+describe('自审 1.9.17 · 浏览器的 loopback 子资源必须经过否决代理(需 chromium)', () => {
+  test('capture 形态的浏览器: 页面里的 loopback 图片 → origin 收到 0 个请求', async t => {
+    const pw = await import('node:module').then(m => m.createRequire(import.meta.url));
+    let chromium;
+    for (const p of ['playwright', 'C:/Users/mjc39/.minimax/skills/node_modules/playwright']) {
+      try { ({ chromium } = pw(p)); break; } catch { /* 下一处 */ }
+    }
+    if (!chromium) return t.skip('无 playwright');
+    const dir = tmpdir();
+    let hits = 0;
+    const origin = http.createServer((req, res) => { hits++; res.writeHead(200, { 'content-type': 'image/png' }); res.end('x'); });
+    await listen(origin);
+    const port = origin.address().port;
+    const slide = path.join(dir, 'slide.html');
+    fs.writeFileSync(slide, `<!doctype html><html><body><img id="i" src="http://127.0.0.1:${port}/pixel.png" width="8" height="8"></body></html>`);
+    const veto = await startVetoProxy({});
+    let browser;
+    try {
+      browser = await chromium.launch({ headless: true, proxy: { server: `http://127.0.0.1:${veto.port}` } });
+      const ctx = await browser.newContext({ viewport: { width: 320, height: 240 } });
+      const p2 = await ctx.newPage();
+      await p2.goto('file:///' + slide.replace(/\\/g, '/'));
+      await p2.waitForTimeout(1200);
+      assert.equal(hits, 0, '浏览器必须把 loopback 子资源交给否决代理(而不是绕过它直连)');
+    } catch (e) {
+      if (/Executable doesn't exist|browserType.launch|Looks like Playwright/i.test(String(e?.message ?? e)))
+        return t.skip('chromium 未安装');
+      throw e;
+    } finally {
+      if (browser) await browser.close().catch(() => {});
+      await veto.close(); origin.close();
+    }
+  });
+});
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/cover-transition.test.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/cover-transition.test.mjs
new file mode 100644
index 00000000..fd5571f2
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/cover-transition.test.mjs
@@ -0,0 +1,120 @@
+// 1.6.0 回归测试: 封面内嵌与首帧非黑、切页无黑帧(硬切/溶解)、语速闸门、单级/多级都能出片
+import { test, describe } from 'node:test';
+import assert from 'node:assert/strict';
+import fs from 'node:fs';
+import path from 'node:path';
+import { spawnSync } from 'node:child_process';
+import { findTool, loadPackage, safeId } from '../scripts/tools.mjs';
+import { runSkill, tmpdir } from './helpers.mjs';
+
+const FFMPEG = findTool('ffmpeg');
+const FFPROBE = findTool('ffprobe');
+const ffprobe = args => (spawnSync(FFPROBE, ['-v', 'error', ...args], { encoding: 'utf8', windowsHide: true }).stdout || '').trim();
+
+// 造一个 n 张的项目(每张 1 段静音 + 多级入场), 返回项目目录
+function mkVideoProject({ n = 3, stagesPerSlide = 2, speed = 1.1, transition = null, audioSec = 3 } = {}) {
+  const proj = tmpdir();
+  assert.equal(runSkill('init-project.mjs', [proj, '--topic', 'CoverTest']).status, 0);
+  for (let i = 1; i <= n; i++) {
+    const id = String(i).padStart(2, '0');
+    const g = spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'anullsrc=r=32000:cl=mono', '-t', String(audioSec),
+      '-c:a', 'libmp3lame', '-b:a', '64k', '-y', path.join(proj, 'audio', `${id}.mp3`)], { windowsHide: true });
+    assert.equal(g.status, 0, g.stderr ?? '');
+  }
+  const slides = [];
+  for (let i = 1; i <= n; i++) {
+    const id = String(i).padStart(2, '0');
+    const cls = [];
+    for (let k = 1; k <= stagesPerSlide; k++) cls.push({ stage: k, text: `${'\u5b57'.repeat(6)}${i}-${k}\u3002` });
+    slides.push({ id, layout: 'statement', html: `${id}-s.html`, audio: `${id}.mp3`, title: `\u7b2c${i}\u5f20`, clauses: cls });
+    const layers = Array.from({ length: stagesPerSlide }, (_, k) =>
+      `<p class="fx-fade" data-stage="${k + 1}">\u5c42 ${k + 1}</p>`).join('');
+    fs.writeFileSync(path.join(proj, 'slides', `${id}-s.html`),
+      `<!doctype html><html data-theme="a"><head><meta charset="utf-8"><link rel="stylesheet" href="tokens.css"></head>
+<body><div class="stage"><h1 class="fx-rise" data-stage="1">\u7b2c ${i} \u5f20</h1>${layers}</div></body></html>`);
+  }
+  const scriptPath = path.join(proj, 'script.json');
+  const script = JSON.parse(fs.readFileSync(scriptPath, 'utf8'));
+  script.slides = slides;
+  script.speed = { default: speed, first: speed, last: speed };
+  if (transition) script.transition = transition;
+  else delete script.transition;
+  fs.writeFileSync(scriptPath, JSON.stringify(script, null, 2));
+  return proj;
+}
+
+const meanLuma = (file, t) => {
+  const o = spawnSync(FFMPEG, ['-hide_banner', '-ss', String(t), '-i', file, '-frames:v', '1',
+    '-vf', 'signalstats,metadata=print:file=-', '-f', 'null', '-'], { encoding: 'utf8', windowsHide: true });
+  const m = /YAVG=([0-9.]+)/.exec((o.stdout || '') + (o.stderr || ''));
+  return m ? parseFloat(m[1]) : null;
+};
+
+describe('1.6.0 · 封面与切页(需 ffmpeg+playwright)', () => {
+  const needTools = async t => {
+    if (!FFMPEG || !FFPROBE) { t.skip('无 ffmpeg/ffprobe'); return false; }
+    const pw = await loadPackage('playwright');
+    if (!pw) { t.skip('无 playwright'); return false; }
+    try { const b = await pw.chromium.launch({ headless: true }); await b.close(); }
+    catch { t.skip('chromium 未安装'); return false; }
+    return true;
+  };
+
+  test('封面: capture 出 preview/cover.png, 成片内嵌 attached_pic + 导出 out/cover.png, 首帧非黑', async t => {
+    if (!await needTools(t)) return;
+    const proj = mkVideoProject({ n: 3 });
+    assert.equal(runSkill('plan-timings.mjs', [proj]).status, 0);
+    assert.equal(runSkill('capture.mjs', [proj, '--mode', 'motion']).status, 0);
+    const cover = path.join(proj, 'preview', 'cover.png');
+    assert.ok(fs.existsSync(cover), 'capture 要为第 1 张出 preview/cover.png');
+    assert.equal(runSkill('build-video.mjs', [proj]).status, 0);
+    const final = path.join(proj, 'out', 'final.mp4');
+    assert.ok(fs.existsSync(path.join(proj, 'out', 'cover.png')), '要导出 out/cover.png 供平台上传');
+    // 内嵌封面流
+    const disp = ffprobe(['-select_streams', 'v', '-show_entries', 'stream_disposition=attached_pic', '-of', 'csv=p=0', final]);
+    assert.match(disp, /1/, `成片里要有 attached_pic 流, 实际: ${JSON.stringify(disp)}`);
+    // 首帧不是黑的(封面溶解后 ≈ 封面本身; 阈值取远高于黑电平 16)
+    const l0 = meanLuma(final, 0.05);
+    assert.ok(l0 != null && l0 > 60, `首帧亮度 ${l0} 太低 —— 分享出去的缩略图会很难看`);
+    // 封面尺寸 = 画布尺寸
+    const sz = ffprobe(['-select_streams', 'v', '-show_entries', 'stream=width,height', '-of', 'csv=p=0', path.join(proj, 'out', 'cover.png')]);
+    assert.equal(sz.split(',').slice(0, 2).join('x'), '1920x1080', `封面尺寸应为画布尺寸, 实际 ${sz}`);
+  });
+
+  test('切页: 硬切与溶解都不经过黑场, 总时长都等于 timings.total', async t => {
+    if (!await needTools(t)) return;
+    const proj = mkVideoProject({ n: 3 });
+    assert.equal(runSkill('plan-timings.mjs', [proj]).status, 0);
+    assert.equal(runSkill('capture.mjs', [proj, '--mode', 'motion']).status, 0);
+    const timings = JSON.parse(fs.readFileSync(path.join(proj, 'build', 'timings.json'), 'utf8'));
+    const final = path.join(proj, 'out', 'final.mp4');
+    const cuts = [];
+    let cum = 0;
+    for (let i = 0; i < timings.slides.length - 1; i++) { cum += timings.slides[i].duration; cuts.push(cum); }
+    for (const mode of ['cut', 'xfade']) {
+      const r = runSkill('build-video.mjs', [proj, ...(mode === 'xfade' ? ['--transition', 'xfade'] : [])]);
+      assert.equal(r.status, 0, `${mode}: ${r.stdout}${r.stderr}`);
+      const dur = parseFloat(ffprobe(['-show_entries', 'format=duration', '-of', 'csv=p=0', final]));
+      assert.ok(Math.abs(dur - timings.total) <= 0.25, `${mode} 总时长 ${dur} 应≈${timings.total}`);
+      for (const c of cuts) {
+        const l = meanLuma(final, Math.max(0, c - 0.1));
+        assert.ok(l != null && l > 60, `${mode} 在切页 ${c.toFixed(2)}s 检出黑帧(亮度 ${l})`);
+      }
+    }
+  });
+
+  test('语速闸门: 实测语速与 script.speed 对不上要告警; 单级入场也能出片', async t => {
+    if (!await needTools(t)) return;
+    // 语速 1.1 → 期望 ≈5.3 字/s; 这里给 3s 静音 + 13 字 → ≈4.3 字/s, 偏离 ~18% 时不报,
+    // 把 speed 写成 1.4(期望 6.7)就必然报警 —— 用它验证闸门真的接上了 script.speed
+    const proj = mkVideoProject({ n: 1, stagesPerSlide: 1, speed: 1.4, audioSec: 3 });
+    const r = runSkill('plan-timings.mjs', [proj]);
+    assert.equal(r.status, 0, r.stdout + r.stderr);
+    assert.match(r.stdout + r.stderr, /script\.speed=1\.4/, '应在汇总行回显 script.speed');
+    assert.match(r.stdout + r.stderr, /实测语速|超出常规区间/, '语速与 speed 对不上时必须告警(而不是等用户听出来)');
+    assert.equal(runSkill('capture.mjs', [proj, '--mode', 'motion']).status, 0);
+    const b = runSkill('build-video.mjs', [proj]);
+    assert.equal(b.status, 0, b.stdout + b.stderr);   // 单级(一次性出现)的 slide 也要能正常出片
+    assert.ok(fs.existsSync(path.join(proj, 'out', 'final.mp4')));
+  });
+});
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/css-kit.test.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/css-kit.test.mjs
new file mode 100644
index 00000000..056efa1c
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/css-kit.test.mjs
@@ -0,0 +1,468 @@
+// CSS 工具箱受管块: "块在但内容旧"必须被发现并原地修复。
+// 这是"改了 CSS 但项目里没生效"的根因回归 —— 2026-09-18 双代理审计定案: 旧判定只探
+// 1–2 个标记字符串, 项目补过一次后就永远报"无需升级"(源模块后续改动永不传播);
+// 追加式升级还会压掉项目端覆写并留下整份重复块。本文件逐条锁死这些行为。
+//
+// 2026-09-18 第二轮(整段受管化 + 三条静默路径): 生成物主体也进了受管区(`tokens`), 于是
+// 判定多了 duplicate(同 id 多块 / 整份重复) / broken(开标记无配对收尾) / legacy-outside
+// (受管区之外残留旧副本) 三个状态; 升级顺序也定死: 先落 tokens 区, 再清区外副本 ——
+// 反过来的话, 裸文本主体里的工具箱原文会被当区外副本删掉, 于是 indexOf 失配 → 又多插一份、
+// 旧主体还留在后面按层叠压过新版(本项目实测踩到, 见"整段受管化 · 顺序与层叠"用例)。
+import { test, describe } from 'node:test';
+import assert from 'node:assert/strict';
+import fs from 'node:fs';
+import path from 'node:path';
+import { runSkill, mkproj, tmpdir, SCRIPTS, LEGACY_TOKENS } from './helpers.mjs';
+
+const { KITS, KIT_REV, TOKENS_ID, wrapKit, wrapTokens, findAllBlocks, findBlock, outsideRegions,
+  kitStatuses, tokensStatus, applyKitUpgrade, MAX_SCAN_BYTES } =
+  await import('file://' + path.join(SCRIPTS, 'css-kit.mjs').replace(/\\/g, '/'));
+const { TOKENS_REV, generateTokensCss } =
+  await import('file://' + path.join(SCRIPTS, 'tokens-template.mjs').replace(/\\/g, '/'));
+const { TABLE_CSS } =
+  await import('file://' + path.join(SCRIPTS, 'table-css.mjs').replace(/\\/g, '/'));
+const { NOFX_CSS } =
+  await import('file://' + path.join(SCRIPTS, 'nofx-css.mjs').replace(/\\/g, '/'));
+const { CHART_CSS } =
+  await import('file://' + path.join(SCRIPTS, 'chart-css.mjs').replace(/\\/g, '/'));
+
+const EXPECT = { text: generateTokensCss(), rev: TOKENS_REV };
+const allStatus = css => kitStatuses(css, { expected: EXPECT });
+const initProj = () => {
+  const p = tmpdir();
+  const r = runSkill('init-project.mjs', [p, '--topic', 'T']);
+  assert.equal(r.status, 0, r.stderr);
+  return p;
+};
+const tokensOf = p => path.join(p, 'slides', 'tokens.css');
+const read = p => fs.readFileSync(tokensOf(p), 'utf8');
+const write = (p, css) => fs.writeFileSync(tokensOf(p), css);
+// 把三个受管块的 rev 全部换成假值 = 模拟"源模块后来改了, 项目里还是旧 rev"
+const withStaleRev = css => {
+  let out = css;
+  for (const k of KITS) {
+    const b = findBlock(out, k.id);
+    out = out.slice(0, b.start) + wrapKit(k.id, k.css, '0000deadbeef0000') + out.slice(b.end);
+  }
+  return out;
+};
+// 老格式(老版 init 直出): 只有整段生成物外面没有 tokens 定界注释, **内嵌的 kit 子块照样带定界**
+// (老模板把三个工具箱块放在主体里)。所以这里只抹掉 tokens 那一对, 不动 kit 的。
+const stripTokensTag = css => css
+  .replace(/\/\* >>> html2video:tokens rev=[0-9a-f]{8,64} >>> \*\/\n?/, '')
+  .replace(/\/\* <<< html2video:tokens <<< \*\/\n?/, '');
+
+describe('受管块 · 单元', () => {
+  test('KIT_REV 是内容哈希(64 位 hex), wrapKit/findBlock 可往返', () => {
+    assert.match(KIT_REV, /^[0-9a-f]{64}$/);
+    assert.match(TOKENS_REV, /^[0-9a-f]{64}$/);
+    for (const k of KITS) {
+      const wrapped = wrapKit(k.id, k.css);
+      const b = findBlock(wrapped, k.id);
+      assert.ok(b, `${k.id} 应能找回受管块`);
+      assert.equal(b.rev, KIT_REV);
+      assert.equal(b.body.trim(), k.css.trim());
+      assert.equal(findBlock(`/* 无关 */\n.foo{}`, k.id), null);
+    }
+    const t = findAllBlocks(wrapTokens(generateTokensCss(), TOKENS_REV), TOKENS_ID);
+    assert.equal(t.length, 1);
+    assert.equal(t[0].rev, TOKENS_REV);
+    assert.equal(t[0].body.trim(), EXPECT.text.trim());
+  });
+
+  test('内容变了 rev 必须变(判定依据是内容, 不是"出现过没有")', () => {
+    const r1 = applyKitUpgrade(`:root{}\n${KITS.map(k => wrapKit(k.id, k.css, 'a'.repeat(64))).join('\n')}\n`);
+    assert.equal(r1.actions.filter(a => a.action === 'replaced').length, 3, '假 rev 的块全部判旧');
+    const mutated = KITS[2].css.replace('72px', '60px');
+    const st = kitStatuses(`${wrapKit(KITS[2].id, mutated)}\n`);
+    assert.equal(st[2].status, 'stale', '块内容与 rev 不符应判 stale(被手工改过)');
+  });
+});
+
+describe('受管块 · 新项目直出', () => {
+  test('init 生成的 tokens.css 带 tokens 受管区(内含三个工具箱), --check-css 绿', () => {
+    const p = initProj();
+    const css = read(p);
+    for (const k of KITS) {
+      const b = findBlock(css, k.id);
+      assert.ok(b, `缺 ${k.id} 受管块`);
+      assert.equal(b.rev, KIT_REV, `${k.id} rev 应为当前`);
+    }
+    const t = findAllBlocks(css, TOKENS_ID);
+    assert.equal(t.length, 1, 'tokens 受管区恰好一个');
+    assert.equal(t[0].rev, TOKENS_REV, 'tokens 区 rev 应为当前');
+    assert.equal(t[0].body.trim(), EXPECT.text.trim(), 'tokens 区内容应与当前模板逐字节一致');
+    assert.equal((css.match(/rev=[0-9a-f]{64}/g) ?? []).length, 4, 'rev 标记恰好四处(tokens + 三个工具箱; 头部注释不携带 rev)');
+    assert.ok(allStatus(css).every(s => s.status === 'ok'), JSON.stringify(allStatus(css).map(s => s.id + ':' + s.status)));
+    const r = runSkill('init-project.mjs', [p, '--check-css']);
+    assert.equal(r.status, 0, r.stderr);
+  });
+});
+
+describe('受管块 · 块在但旧(旧判定的漏检区)', () => {
+  test('块 rev 落后 → 原地替换: 行数不变、无重复、.bak 先落盘', () => {
+    const p = initProj();
+    const fresh = read(p);
+    write(p, withStaleRev(fresh));
+    const before = read(p);
+    assert.notEqual(before, fresh);
+    const r = runSkill('init-project.mjs', [p, '--upgrade-css']);
+    assert.equal(r.status, 0, r.stderr);
+    const after = read(p);
+    assert.equal(after, fresh, '按 rev 原地替换后应与最新模板逐字节一致');
+    assert.equal(fs.readFileSync(tokensOf(p) + '.bak', 'utf8'), before, '写前必须先备份');
+    assert.equal((after.match(/@property --pv/g) ?? []).length, 1, '不得出现重复块');
+  });
+
+  test('块被手工改坏(rev 没变) → --check-css 报 stale, 升级按源重写', () => {
+    const p = initProj();
+    const tampered = read(p).replace('.tbl tbody tr { height: 72px; }', '.tbl tbody tr { height: 60px; }');
+    assert.notEqual(tampered, read(p), '前置: 确实改到了受管块内的规则');
+    write(p, tampered);
+    const chk = runSkill('init-project.mjs', [p, '--check-css']);
+    assert.equal(chk.status, 1, '块内容与 rev 不符必须被发现');
+    assert.ok(chk.stderr.includes('手工改过'), chk.stderr);
+    runSkill('init-project.mjs', [p, '--upgrade-css']);
+    assert.ok(read(p).includes('.tbl tbody tr { height: 72px; }'), '升级后按技能当前源恢复');
+  });
+
+  test('用户覆写在受管块之后 → 升级不碰它(旧追加式会把它压掉)', () => {
+    const p = initProj();
+    const fresh = read(p);
+    const override = '\n/* 项目端覆写: 行高按这台投影调过 */\n.tbl tbody tr { height: 96px; }\n';
+    write(p, withStaleRev(fresh) + override);
+    runSkill('init-project.mjs', [p, '--upgrade-css']);
+    const after = read(p);
+    assert.ok(after.trimEnd().endsWith('.tbl tbody tr { height: 96px; }'), '覆写必须原样留在文件尾');
+    assert.equal((after.match(/height: 96px/g) ?? []).length, 1, '覆写不得被复制或挪动');
+    const b = findBlock(after, 'table');
+    assert.ok(b.start < after.indexOf('height: 96px'), '受管块仍在覆写之前(覆写继续生效)');
+  });
+
+  test('旧判定核心缺陷不复存在: 升级过的项目在源模块再改后仍能继续升级(审计实验 D)', () => {
+    const p = initProj();
+    runSkill('init-project.mjs', [p, '--upgrade-css']);           // 第一次: 无需升级
+    // 模拟源模块又改了: 把项目里的块回滚成"旧 rev + 旧内容"(少一条规则)
+    let css = read(p);
+    const k = KITS.find(x => x.id === 'chart');
+    const b = findBlock(css, 'chart');
+    const older = wrapKit('chart', k.css.replace(/\/\* 左侧刻度列[\s\S]*?padding-right: var\(--sp-2\); \}/, ''), '1111222233334444');
+    css = css.slice(0, b.start) + older + css.slice(b.end);
+    write(p, css);
+    assert.ok(!read(p).includes('.chart-ticks'), '前置: .chart-ticks 已不在项目里');
+    const chk = runSkill('init-project.mjs', [p, '--check-css']);
+    assert.equal(chk.status, 1, '旧 rev 必须报落后(旧实现此处报"无需升级")');
+    runSkill('init-project.mjs', [p, '--upgrade-css']);
+    assert.ok(read(p).includes('.chart-ticks'), '源里的新规则必须传播进项目');
+  });
+});
+
+describe('受管块 · 无定界的老文件迁移', () => {
+  test('内容已是当前版(老 init 直出的裸文本) → check 绿, 升级原地包裹且与新模板逐字节一致', () => {
+    const p = initProj();
+    const fresh = read(p);
+    const legacy = stripTokensTag(fresh);
+    assert.ok(!/>>> html2video:tokens rev=/.test(legacy), '前置: tokens 定界已去掉(头部注释里的示例文字无 rev, 不算)');
+    assert.ok(/>>> html2video:nofx rev=/.test(legacy), '前置: 内嵌的 kit 子块照旧留着(老格式就是这样)');
+    write(p, legacy);
+    const chk = runSkill('init-project.mjs', [p, '--check-css']);
+    assert.equal(chk.status, 0, '内容当前就该是 ok, 不该骚扰');
+    assert.ok((chk.stdout + chk.stderr).includes('老格式'), '要说明是靠裸文本内容判定的');
+    const st = tokensStatus(legacy, EXPECT);
+    assert.equal(st.status, 'ok', JSON.stringify(st));
+    assert.ok(st.detail.includes('无受管区'), st.detail);
+    runSkill('init-project.mjs', [p, '--upgrade-css']);
+    assert.equal(read(p), fresh, '原地包裹后应回到与新模板完全一致');
+  });
+
+  test('裸文本重复两份(旧追加时代留的) → check 报错并归并为一', () => {
+    const p = initProj();
+    const fresh = read(p);
+    const legacy = stripTokensTag(fresh);
+    // (a) 整份生成物出现两次 → tokens 判 duplicate
+    const dup = legacy + '\n' + EXPECT.text.trim() + '\n';
+    write(p, dup);
+    assert.equal(tokensStatus(dup, EXPECT).status, 'duplicate', '整份生成物出现两次必须报出来');
+    assert.equal(runSkill('init-project.mjs', [p, '--check-css']).status, 1, '重复的生成物不许静默通过');
+    runSkill('init-project.mjs', [p, '--upgrade-css']);
+    assert.equal(findAllBlocks(read(p), TOKENS_ID).length, 1, '归并后只剩一个 tokens 区');
+    assert.equal((read(p).match(/>>> html2video:table/g) ?? []).length, 1);
+    assert.equal(runSkill('init-project.mjs', [p, '--check-css']).status, 0);
+    // (b) 只有某个工具箱被追加过第二份(无定界的裸副本) → 该 kit 判 legacy-outside
+    write(p, legacy + '\n' + KITS[2].css + '\n');
+    assert.equal(kitStatuses(read(p), { expected: EXPECT }).find(s => s.id === 'table').status, 'legacy-outside');
+    assert.equal(runSkill('init-project.mjs', [p, '--check-css']).status, 1, '区外多一份表格原语同样不许通过');
+    runSkill('init-project.mjs', [p, '--upgrade-css']);
+    const after = read(p);
+    assert.equal((after.match(/\.tbl \{[^}]*border-collapse/g) ?? []).length, 1, '重复的表格原语必须只剩一份');
+    assert.equal((after.match(/>>> html2video:table/g) ?? []).length, 1);
+  });
+
+  test('完全缺失 → 文件尾追加受管区; 原有内容在前; 幂等', () => {
+    const p = mkproj(tmpdir());
+    const r1 = runSkill('init-project.mjs', [p, '--upgrade-css']);
+    assert.equal(r1.status, 0, r1.stderr);
+    const once = read(p);
+    assert.ok(once.startsWith(':root { --accent: #111; }'), '原有内容必须在前');
+    for (const k of KITS) assert.ok(findBlock(once, k.id), `缺 ${k.id} 受管块`);
+    assert.ok(findBlock(once, TOKENS_ID), '缺 tokens 受管区');
+    assert.ok(once.indexOf(TOKENS_ID) > once.indexOf('--accent: #111'), '受管区必须排在原有内容之后(后写才压得住旧值)');
+    const r2 = runSkill('init-project.mjs', [p, '--upgrade-css']);
+    assert.equal(r2.status, 0);
+    assert.equal(read(p), once, '第二次不得再追加');
+    assert.ok(r2.stdout.includes('无需升级'));
+  });
+
+  test('旧版残留(整份旧副本) → 已落入层叠死区: 升级清掉副本、当前版进受管区、可重复升级', () => {
+    // 探针规则(.tbl 行高/.kv)保持原样、其余内容是旧版 —— 探针通过但逐字不含当前版, 即"旧版残留"
+    const oldTable = TABLE_CSS.replace(/\.rank td \{[^}]*\}/, m => m.replace('72px', '60px'));
+    const p = mkproj(tmpdir(), { tokens: oldTable });
+    const r = runSkill('init-project.mjs', [p, '--upgrade-css']);
+    assert.equal(r.status, 0, r.stderr);
+    const css = read(p);
+    // 旧实现是"把当前版追加到旧规则之后靠后写覆盖": 一旦旧副本又被写回文件尾就又落后了。
+    // 现在整份旧副本按 kind='copy' 直接清掉(原件在 .bak), 所以文件里只剩当前版一条。
+    assert.equal((css.match(/\.rank td \{[^}]*\}/g) ?? []).length, 1, '旧副本必须被清掉, 当前版只剩一条');
+    assert.ok(css.includes('72px'), '内容应是当前版(72px)');
+    assert.ok(!/height: 60px/.test(css), '旧值(60px)不得残留');
+    assert.ok(css.trimEnd().endsWith('<<< */'), '受管区必须落在文件末尾(层叠上后写才压得住前面的旧内容)');
+    assert.equal(runSkill('init-project.mjs', [p, '--check-css']).status, 0, '升级后必须绿');
+    assert.ok(r.stdout.includes('清理') || r.stdout.includes('受管区'), r.stdout);
+  });
+});
+
+describe('受管块 · 整段受管化(2026-09-18 补: 四条静默路径的逐个回归)', () => {
+  test('主体陈旧 → tokens 区被整体就地替换, 区外的项目规则一字不动', () => {
+    const body = generateTokensCss().replace('--fs-h1: 64px', '--fs-h1: 60px');   // 模拟旧版主体
+    assert.notEqual(body, generateTokensCss(), '前置: 造出了与当前不同的主体内容');
+    assert.ok(!body.includes('--fs-h1: 64px'), '前置: 旧值确实换掉了');
+    const css = `${wrapTokens(body, 'cafebabecafebabe')}\n.proj { color: blue }\n`;
+    const st = tokensStatus(css, EXPECT);
+    assert.equal(st.status, 'stale', JSON.stringify(st));
+    const up = applyKitUpgrade(css, { tokensText: EXPECT.text, tokensRev: TOKENS_REV }).css;
+    assert.equal(findAllBlocks(up, TOKENS_ID).length, 1);
+    assert.equal(findAllBlocks(up, TOKENS_ID)[0].body.trim(), EXPECT.text.trim(), '整段应换成当前版');
+    assert.ok(!up.includes('--fs-h1: 60px'), '旧主体的值不得残留');
+    assert.ok(up.includes('.proj { color: blue }'), '区外的项目规则必须保留');
+    assert.equal(allStatus(up).find(s => s.id === TOKENS_ID).status, 'ok');
+  });
+
+  test('同一 id 两个块 → 报 duplicate, 升级归并为一(块外那份按层叠会赢)', () => {
+    const one = wrapKit('table', TABLE_CSS, 'a'.repeat(16));
+    const two = wrapKit('table', TABLE_CSS.replace('72px', '60px'), 'b'.repeat(16));
+    const css = `${one}\n${two}\n`;
+    const st = kitStatuses(css);
+    assert.equal(st[2].status, 'duplicate', JSON.stringify(st.map(s => s.id + ':' + s.status)));
+    const up = applyKitUpgrade(css, { tokensText: EXPECT.text, tokensRev: TOKENS_REV }).css;
+    assert.equal(findAllBlocks(up, 'table').length, 1, '归并为一');
+    assert.equal(findAllBlocks(up, 'table')[0].body.trim(), TABLE_CSS.trim(), '留下的是当前版');
+    assert.equal((up.match(/height: 60px/g) ?? []).length, 0, '块外那份旧副本必须清掉');
+  });
+
+  test('开标记无配对收尾 → 报 broken(旧实现因字符串命中反而报 ok), 升级后不得叠加第二个开标记', () => {
+    const css = `${wrapKit('nofx', NOFX_CSS, 'a'.repeat(16))}\n/* >>> html2video:table rev=b4c5d6e7 >>> */\n.tbl { color: red }\n`;
+    const st = kitStatuses(css);
+    assert.equal(st[2].status, 'broken', JSON.stringify(st.map(s => s.id + ':' + s.status)));
+    const up = applyKitUpgrade(css, { tokensText: EXPECT.text, tokensRev: TOKENS_REV }).css;
+    const opens = up.match(/>>> html2video:table rev=/g) ?? [];
+    assert.equal(opens.length, findAllBlocks(up, 'table').length, '开标记数必须与配对块数一致');
+    assert.equal(opens.length, 1);
+    assert.ok(up.includes('.tbl { color: red }'), '用户自己写的规则保留');
+  });
+
+  test('受管区之外残留旧副本 → 报 legacy-outside; 升级清副本但保留用户单条覆写', () => {
+    const override = '.tbl tbody tr { height: 96px; }';
+    // tokens 区里本来就含一份当前的 table 子块, 区外这份是旧副本(后出现 → 按层叠会赢)
+    const css = `${wrapTokens(generateTokensCss(), TOKENS_REV)}\n${TABLE_CSS}\n${override}\n`;
+    const st = kitStatuses(css, { expected: EXPECT });
+    const table = st.find(s => s.id === 'table');
+    assert.equal(table.status, 'legacy-outside', st.map(s => s.id + ':' + s.status).join(' '));
+    assert.ok(table.detail.includes('受管区之外'), table.detail);
+    const up = applyKitUpgrade(css, { tokensText: EXPECT.text, tokensRev: TOKENS_REV });
+    const actions = up.actions.map(a => a.action);
+    assert.ok(actions.includes('removed-legacy-outside'), actions.join(','));
+    assert.ok(actions.includes('kept-legacy-overrides'), '紧邻旧副本的用户覆写必须被单独保留: ' + actions.join(','));
+    assert.equal(findAllBlocks(up.css, 'table').length, 1);
+    assert.equal((up.css.match(/\.tbl tbody tr \{ height: 72px; \}/g) ?? []).length, 1, '受管区里的当前版只剩一条');
+    assert.ok(up.css.includes(override), '用户单条覆写必须原样保留(区外规则一概不动)');
+    assert.ok(allStatus(up.css).every(s => s.status === 'ok'), allStatus(up.css).map(s => s.id + ':' + s.status).join(' '));
+  });
+
+  test('区外整份旧副本被改过几处(不止一处) → 整段清掉: 它是压在受管区上的错误产物', () => {
+    const modCopy = TABLE_CSS.replace('72px', '60px');
+    const override = '.tbl tbody tr { height: 96px; }';
+    const css = `${wrapTokens(generateTokensCss(), TOKENS_REV)}\n${modCopy}\n${override}\n`;
+    assert.equal(kitStatuses(css, { expected: EXPECT }).find(s => s.id === 'table').status, 'legacy-outside');
+    const up = applyKitUpgrade(css, { tokensText: EXPECT.text, tokensRev: TOKENS_REV }).css;
+    assert.ok(!/height: 60px/.test(up), '被改过的整份旧副本必须清掉(否则它会按层叠压过受管区)');
+    assert.ok(up.includes(override), '与之相邻的用户覆写仍然保留');
+    assert.equal(allStatus(up).every(s => s.status === 'ok'), true, allStatus(up).map(s => s.id + ':' + s.status).join(' '));
+  });
+
+  test('老格式裸文本 + 旧主体并存 → 生成物排在旧内容之后(层叠上后写才生效), 且幂等', () => {
+    // 顺序回归: 升级必须"先判是否已有当前版主体、再剥内嵌子块"。反过来的话裸文本主体里的工具箱原文
+    // 会被当区外副本删掉 → indexOf 失配 → 插到文件末尾之外的位置 + 旧主体留在后面压过新版(实测踩到过)。
+    const oldBody = generateTokensCss().replace('--fs-h1: 64px', '--fs-h1: 60px');
+    const css = `/* 老项目 */\n${oldBody}\n.proj { color: blue }\n`;
+    const up = applyKitUpgrade(css, { tokensText: EXPECT.text, tokensRev: TOKENS_REV }).css;
+    assert.equal((up.match(/>>> html2video:tokens/g) ?? []).length, 1, '只能有一个受管区(不得多插一份)');
+    assert.ok(up.indexOf('>>> html2video:tokens') > up.indexOf('--fs-h1: 60px'), '受管区必须在旧主体之后(否则旧值压过新版)');
+    assert.ok(up.includes('.proj { color: blue }'), '项目规则保留');
+    for (const k of KITS) assert.equal(findAllBlocks(up, k.id).length, 1, `${k.id} 恰好一个受管块`);
+    assert.equal(applyKitUpgrade(up, { tokensText: EXPECT.text, tokensRev: TOKENS_REV }).css, up, '幂等');
+    // 逐字与当前版相同的旧主体规则会被清掉; 与当前版**不同**的那条留下并如实报出来
+    // (工具分不出"旧残留"还是"你有意改的", 所以报出来由人决定 —— 复查 A7 定下的语义)
+    const st = allStatus(up);
+    const tokens = st.find(s => s.id === TOKENS_ID);
+    assert.equal(tokens.status, 'legacy-outside', st.map(s => s.id + ':' + s.status).join(' '));
+    assert.match(tokens.detail, /与当前版不同|逐字相同/, tokens.detail);
+    assert.ok(st.filter(s => s.id !== TOKENS_ID).every(s => s.status === 'ok'), st.map(s => s.id + ':' + s.status).join(' '));
+    // 关键的用户可见结论: 与当前版不同的那条旧规则虽然留着(工具分不出来是不是有意改的),
+    // 但它在受管区**之前**, 所以渲染用的是受管区里的当前值 —— 这就是"区域排在最后"的意义
+    assert.equal(up.split('--fs-h1: 60px').length - 1, 1, '与当前版不同的那条只应留一条(其余逐字相同的旧规则都该被清掉)');
+    assert.ok(up.includes('--fs-h1: 64px'), '受管区里应有当前值');
+    assert.ok(up.indexOf('--fs-h1: 64px') > up.indexOf('--fs-h1: 60px'), '当前值必须出现在旧值之后(层叠上生效)');
+  });
+
+  test('区外扫描的可见范围: outsideRegions 把受管区(含内部工具箱子块)整体挖掉', () => {
+    const css = `${wrapTokens(generateTokensCss(), TOKENS_REV)}\n.keepme { color: red }\n`;
+    const out = outsideRegions(css);
+    assert.ok(out.includes('.keepme'), '区外内容应保留在视野里(嵌套区间的掩码不许吃掉它)');
+    assert.equal(out.length, css.length, '掩码与原文等长, 偏移才对得回原文');
+    assert.ok(!out.includes('.no-fx'), 'tokens 区内的工具箱内容不得计入区外');
+    assert.ok(!out.includes('.chart-bar'), 'tokens 区内的图表工具内容不得计入区外');
+  });
+});
+
+describe('--check-css CLI 语义', () => {
+  test('落后 → 退出 1 并给出修复命令; 升级后 → 退出 0', () => {
+    const p = mkproj(tmpdir(), { tokens: LEGACY_TOKENS });
+    const bad = runSkill('init-project.mjs', [p, '--check-css']);
+    assert.equal(bad.status, 1);
+    assert.ok(bad.stderr.includes('--upgrade-css'), '要给出修复动作');
+    runSkill('init-project.mjs', [p, '--upgrade-css']);
+    assert.equal(runSkill('init-project.mjs', [p, '--check-css']).status, 0);
+  });
+});
+
+describe('受管块 · 行尾与病态输入(2026-09-18 三维审计补的回归)', () => {
+  test('CRLF 文件(git autocrlf 检出)不误报: check 绿; 升级后受管块统一为 LF 并如实说明', () => {
+    const p = initProj();
+    const fresh = read(p);
+    write(p, fresh.replace(/\n/g, '\r\n'));
+    assert.equal(runSkill('init-project.mjs', [p, '--check-css']).status, 0, '只换了行尾, 不该判"被手工改过"');
+    write(p, withStaleRev(read(p)).replace(/\n/g, '\r\n'));   // 行尾 + 旧 rev 一起, 走真升级路径
+    const r = runSkill('init-project.mjs', [p, '--upgrade-css']);
+    assert.equal(r.status, 0, r.stderr);
+    assert.ok(r.stdout.includes('行尾已统一为 LF'), r.stdout);
+    assert.ok(!findBlock(read(p), 'chart').body.includes('\r'), '受管块内不得再残留 CRLF');
+    assert.equal(runSkill('init-project.mjs', [p, '--check-css']).status, 0);
+  });
+
+  test('.bak 已存在 → 显式警告后再覆盖(旧备份无声丢是不对的)', () => {
+    const p = initProj();
+    const stale = withStaleRev(read(p));
+    write(p, stale);
+    fs.writeFileSync(tokensOf(p) + '.bak', 'OLD PRECIOUS BACKUP');
+    const r = runSkill('init-project.mjs', [p, '--upgrade-css']);
+    assert.equal(r.status, 0, r.stderr);
+    assert.ok((r.stdout + r.stderr).includes('将被本次升级前的备份覆盖'), '必须警告');
+    assert.ok(fs.readFileSync(tokensOf(p) + '.bak', 'utf8').includes('0000dead'), '.bak 应是升级前的内容');
+  });
+
+  test('病态互踩: nofx 裸文本嵌在 chart 旧受管块内 → 一次升级全部到位(不再"上次成功这次又落后")', () => {
+    const hostile = `:root{--a:1}\n${wrapKit('chart', CHART_CSS + '\n' + NOFX_CSS, 'a'.repeat(64))}\n`;
+    const p = mkproj(tmpdir(), { tokens: hostile });
+    const r = runSkill('init-project.mjs', [p, '--upgrade-css']);
+    assert.equal(r.status, 0, r.stderr);
+    assert.equal(runSkill('init-project.mjs', [p, '--check-css']).status, 0, '一次升级后三段必须全部就位');
+    for (const k of KITS) assert.ok(findBlock(read(p), k.id), `缺 ${k.id} 受管块`);
+  });
+
+  test('超大 tokens.css → 拒绝扫描(构造输入会让受管块定位二次方变慢, 4MB 实测 39s)', () => {
+    const p = mkproj(tmpdir(), { tokens: '/* >>> html2video:nofx rev=aaaaaaaa >>> */\n' + '/* '.repeat(MAX_SCAN_BYTES / 3 + 10) });
+    const r = runSkill('init-project.mjs', [p, '--check-css']);
+    assert.equal(r.status, 1);
+    assert.ok((r.stdout + r.stderr).includes('拒绝扫描'), '要说明是拒绝扫描而不是慢慢算');
+  });
+});
+
+describe('陈旧 CSS 闸门(2026-09-18 复查: 陈旧必须阻断, 不许静默出旧画面)', () => {
+  const slideHtml = `<!doctype html><html data-theme="a"><head><meta charset="utf-8"><link rel="stylesheet" href="tokens.css"></head>
+<body><div class="stage"><h1 class="fx-rise" data-stage="1">标题</h1></div></body></html>`;
+  const staleProj = () => {
+    const p = mkproj(tmpdir(), { tokens: LEGACY_TOKENS, slides: [{ id: '01', html: '01.html', audio: '01.mp3', clauses: [{ stage: 1, text: '一句。' }] }] });
+    fs.writeFileSync(path.join(p, 'slides', '01.html'), slideHtml);
+    return p;
+  };
+
+  test('capture: 受管块落后 → 退出 1 并给修复命令; --allow-stale-css 显式放行', () => {
+    const p = staleProj();
+    const r = runSkill('capture.mjs', [p]);
+    assert.equal(r.status, 1, '旧 CSS 会照常出图且全程不报错 —— 必须在入口停住');
+    assert.ok(r.stderr.includes('受管块与技能当前版不一致'), r.stderr.slice(-300));
+    assert.ok(r.stderr.includes('--upgrade-css'), '要把修复命令给到手上');
+    assert.ok(r.stderr.includes(p), '命令里要有本项目路径(可直接粘)');
+    const r2 = runSkill('capture.mjs', [p, '--allow-stale-css']);
+    assert.ok(r2.stderr.includes('已显式跳过闸门'), '显式跳过要留痕, 实际: ' + r2.stderr.slice(-200));
+  });
+
+  test('build-video: 同一道闸门(成片比预览图更贵, 更不能静默用旧 CSS)', () => {
+    const p = staleProj();
+    const r = runSkill('build-video.mjs', [p]);
+    assert.equal(r.status, 1);
+    assert.ok(r.stderr.includes('build-video') && r.stderr.includes('受管块与技能当前版不一致'), r.stderr.slice(-300));
+    assert.ok(r.stderr.includes('--allow-stale-css'));
+  });
+
+  test('check-slides: 受管块落后从"提示"升为"错误"级(阻断截图), 升级后恢复绿', () => {
+    const p = staleProj();
+    const r = runSkill('check-slides.mjs', [p]);
+    assert.equal(r.status, 1, '提示级不得再放行陈旧 CSS');
+    assert.ok(r.stdout.includes('✗ tokens.css'), r.stdout.slice(-400));
+    assert.ok(r.stdout.includes('--upgrade-css'), '报错里要给修复命令');
+    assert.equal(runSkill('init-project.mjs', [p, '--upgrade-css']).status, 0);
+    assert.equal(runSkill('check-slides.mjs', [p]).status, 0, '升级后必须恢复');
+  });
+
+  test('真项目形状(项目规则 + 当前受管区) → 三个入口都不拦, 且把"区外覆写"作为提示打出来', () => {
+    const p = mkproj(tmpdir(), { slides: [{ id: '01', html: '01.html', audio: '01.mp3', clauses: [{ stage: 1, text: '一句。' }] }] });
+    fs.writeFileSync(path.join(p, 'slides', '01.html'), slideHtml);
+    fs.appendFileSync(path.join(p, 'slides', 'tokens.css'), '\n.tbl tbody tr { height: 96px; }\n');
+    assert.notEqual(runSkill('check-slides.mjs', [p]).status, 1, '区外单条覆写不该让截图停摆');
+    const r = runSkill('init-project.mjs', [p, '--check-css']);
+    assert.equal(r.status, 0, r.stderr);
+    assert.ok(r.stdout.includes('覆写'), '要告诉用户区外有他自己的覆写: ' + r.stdout);
+  });
+});
+
+describe('复查 A6/A7: "check 绿但画面是旧值"的最后两条路径(2026-09-18 第三轮 review 实测)', () => {
+  test('A7: 受管区之后粘着一段旧主体 → 必须报 legacy-outside(旧实现报"已是最新")', () => {
+    // 旧实现只扫"工具箱选择器"与"逐字重复的整份主体", 于是这种布局被判 ok, 而那段旧规则
+    // 按层叠压过受管区 —— 实测画面用的是 --fs-h1: 60px, check 却报 ✓ 已是最新。
+    const oldFragment = generateTokensCss().replace('--fs-h1: 64px', '--fs-h1: 60px').slice(0, 4000);
+    const css = `${wrapTokens(generateTokensCss(), TOKENS_REV)}\n\n/* 旧版残留(手粘) */\n${oldFragment}\n`;
+    const st = tokensStatus(css, EXPECT);
+    assert.equal(st.status, 'legacy-outside', JSON.stringify(st));
+    assert.match(st.detail, /主体规则/, st.detail);
+    assert.match(st.detail, /逐字相同/, '要说清升级会清掉哪一部分');
+    // 升级: 逐字命中当前版的旧规则被清掉; 改过的那条留着并再次如实报出来(工具分不出旧残留与有意覆写)
+    const up = applyKitUpgrade(css, { tokensText: EXPECT.text, tokensRev: TOKENS_REV });
+    assert.equal(up.actions.filter(a => a.action === 'removed-legacy-outside').length, 1, up.actions.map(a => a.action).join(','));
+    assert.equal(tokensStatus(up.css, EXPECT).status, 'legacy-outside', '改过的那条还在, 仍要报(由人决定)');
+    assert.ok(up.actions.some(a => a.action === 'kept-legacy-overrides' && /与当前版不同/.test(a.reason)), up.actions.map(a => a.action).join(','));
+  });
+
+  test('A6: 老格式 + 文件尾的项目覆写 → 升级后必须说清它被受管区覆盖了(旧文案谎称"后出现者优先")', () => {
+    const oldBody = generateTokensCss().replace('--fs-h1: 64px', '--fs-h1: 60px');
+    const css = `/* 老项目 */\n${oldBody}\n:root { --accent: #D92B2B; }\n`;
+    const up = applyKitUpgrade(css, { tokensText: EXPECT.text, tokensRev: TOKENS_REV });
+    const regionAt = up.css.indexOf('>>> html2video:tokens');
+    const overrideAt = up.css.indexOf('--accent: #D92B2B');
+    assert.ok(overrideAt > -1 && overrideAt < regionAt, '前置: 老格式升级会把覆写留在受管区之前');
+    const st = tokensStatus(up.css, EXPECT);
+    assert.notEqual(st.status, 'ok', '这条改过的旧规则会一直被报出来, 由人决定: ' + JSON.stringify(st));
+    assert.match(st.detail, /受管区\*\*之前\*\*/, '必须在状态里点名"会被受管区覆盖"(旧文案笼统写"后出现者优先", 与实际布局相反): ' + st.detail);
+    assert.ok(/挪到受管区之后/.test(st.detail), '并给出可执行的做法: ' + st.detail);
+  });
+});
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/endpoint-allowlist.test.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/endpoint-allowlist.test.mjs
new file mode 100644
index 00000000..742d7788
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/endpoint-allowlist.test.mjs
@@ -0,0 +1,169 @@
+// 安全边界 · ASR 端点白名单(API Key 只发官方域)
+// 对应评审意见 3: credential exfiltration via --base-url / MINIMAX_BASE_URL。
+import { test, describe } from 'node:test';
+import assert from 'node:assert/strict';
+import http from 'node:http';
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+import { spawnSync } from 'node:child_process';
+import { assertAsrEndpoint, PolicyError } from '../scripts/url-policy.mjs';
+import { findTool } from '../scripts/tools.mjs';
+import { runSkill, runSkillAsync, tmpdir } from './helpers.mjs';
+
+describe('assertAsrEndpoint(纯函数)', () => {
+  test('两个官方域放行(含尾斜杠归一)', () => {
+    assert.equal(assertAsrEndpoint('https://api.minimaxi.com'), 'https://api.minimaxi.com');
+    assert.equal(assertAsrEndpoint('https://api.minimaxi.com/'), 'https://api.minimaxi.com');
+    assert.equal(assertAsrEndpoint('https://api.minimax.io'), 'https://api.minimax.io');
+  });
+  test('官方域的 http 变体也拒绝', () => {
+    assert.throws(() => assertAsrEndpoint('http://api.minimaxi.com'), PolicyError);
+  });
+  test('任意第三方域拒绝', () => {
+    for (const b of ['https://evil.example', 'https://api.minimaxi.com.evil.example', 'https://127.0.0.1:9456']) {
+      assert.throws(() => assertAsrEndpoint(b), PolicyError, b);
+    }
+  });
+  test('allowAny 显式放行(自建网关/测试, 自担风险)', () => {
+    assert.equal(assertAsrEndpoint('http://127.0.0.1:9456', { allowAny: true }), 'http://127.0.0.1:9456');
+    assert.throws(() => assertAsrEndpoint('ftp://x', { allowAny: true }), PolicyError);
+  });
+});
+
+describe('asr.mjs 进程级行为', () => {
+  test('MINIMAX_BASE_URL=第三方域 + 假 Key → 退出 1, 不发起请求(与提供方无关: 端点闸门在提供方解析之前)', () => {
+    const r = runSkill('asr.mjs', ['--api-key', 'sk-test-not-real', '--base-url', 'https://evil.example', '--file', 'whatever.mp3']);
+    assert.notEqual(r.status, 0);
+    assert.ok((r.stderr + r.stdout).includes('端点'), '应说明端点被拒');
+  });
+
+  test('--allow-any-endpoint 指向本地服务器 → 请求到达且带着 Bearer 假 Key(证明门在 fetch 之前且可显式放行)', async () => {
+    if (!findTool('ffprobe')) return; // 主 CI 无 ffmpeg: 只留纯函数/负向证据
+    // 造一个 1s 静音 mp3
+    const work = tmpdir();
+    const mp3 = path.join(work, 'clip.mp3');
+    const FFMPEG = findTool('ffmpeg');
+    const g = spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'anullsrc=r=32000:cl=mono', '-t', '1', '-c:a', 'libmp3lame', '-y', mp3], { windowsHide: true });
+    if (g.status !== 0) return; // 造不出音频就跳过
+
+    let sawAuth = null;
+    const server = http.createServer((req, res) => {
+      sawAuth = req.headers.authorization ?? '(none)';
+      // Connection:close 让 undici 不留 keep-alive 连接 —— 否则 server.close() 永远等不到
+      // socket 全关, 测试进程挂死(本次实测踩过: 整个 node --test 卡到超时)
+      res.writeHead(200, { 'content-type': 'application/json', connection: 'close' });
+      res.end(JSON.stringify({ text: '' }));
+    });
+    await new Promise(resolve => server.listen(0, '127.0.0.1', resolve));
+    const port = server.address().port;
+    try {
+      // 必须 runSkillAsync: spawnSync 冻结事件循环, 本地服务器 accept 不了子进程请求 → 互等死锁
+      // --provider api 显式定向: 本机装了 mmx ≥1.0.26 时自动解析会走 mmx, 绕开这台桩服务器
+      const r = await runSkillAsync('asr.mjs', ['--provider', 'api', '--api-key', 'sk-test-not-real', '--allow-any-endpoint',
+        '--base-url', `http://127.0.0.1:${port}`, '--file', mp3]);
+      assert.equal(sawAuth, 'Bearer sk-test-not-real', '本地服务器应收到 Bearer 假 Key');
+      // 脚本本身成功与否不强断言(响应是假 JSON); 关键证据是 sawAuth
+      assert.ok(r.status === 0 || (r.stdout + r.stderr).length > 0);
+    } finally {
+      server.closeAllConnections?.();
+      server.close();
+      server.unref();
+    }
+  });
+});
+
+describe('asr.mjs 提供方选择(mmx cli ≥1.0.26 缺省首选, REST 为显式后备)', () => {
+  test('--provider 非法值 → 退出 1 并点名只认 mmx|api', () => {
+    const r = runSkill('asr.mjs', ['--provider', 'whisper', '--file', 'x.mp3']);
+    assert.notEqual(r.status, 0);
+    assert.ok((r.stderr + r.stdout).includes('--provider'), '应点名 --provider 的合法值');
+  });
+
+  test('--provider api 且无 Key → 退出 2, 引导同时给出两条路(装 mmx / 配 Key)', () => {
+    const r = runSkill('asr.mjs', ['--provider', 'api', '--file', 'x.mp3'], { env: { ...process.env, MINIMAX_API_KEY: '' } });
+    assert.equal(r.status, 2);
+    const out = r.stderr + r.stdout;
+    assert.match(out, /mmx-cli@latest/, '引导应包含升级/安装 mmx');
+    assert.match(out, /MINIMAX_API_KEY/, '引导应包含 REST 后备的 Key 配法');
+  });
+
+  // 假 mmx: 记录收到的参数与"子进程是否看到 MINIMAX_API_KEY", 按 --out 指定的路径写回 JSON 文档。
+  // POSIX 形态是带 shebang 的 node 脚本(chmod +x), Windows 形态是 mmx.cmd 包一层 node ——
+  // 与 asr.mjs 的 mmxRun 封装(Windows 经 cmd.exe /c)对得上。
+  const makeShim = (bin, log, { fail } = {}) => {
+    const body = [
+      "import fs from 'node:fs';",
+      `const a = process.argv.slice(2);`,
+      `fs.appendFileSync(${JSON.stringify(log)}, JSON.stringify(a) + '\\n' + 'ENVKEY=' + (process.env.MINIMAX_API_KEY === undefined ? 'absent' : 'present') + '\\n');`,
+      `if (a.includes('--version')) { console.log('mmx 1.0.26'); process.exit(0); }`,
+      ...(fail ? [`if (a.includes('transcribe')) { console.error('boom-quota'); process.exit(3); }`] : []),
+      "const i = a.indexOf('--out');",
+      "if (i > -1) fs.writeFileSync(a[i + 1], JSON.stringify({ text: '你好世界二零二六', duration: 1.23, segments: [{ id: 0, start: 0.1, end: 0.4, text: '你好' }] }));",
+      "process.exit(0);",
+    ].join('\n');
+    if (process.platform === 'win32') {
+      fs.writeFileSync(path.join(bin, 'mmx-shim.mjs'), body);
+      fs.writeFileSync(path.join(bin, 'mmx.cmd'), '@node "%~dp0mmx-shim.mjs" %*\r\n');
+    } else {
+      const f = path.join(bin, 'mmx');
+      fs.writeFileSync(f, '#!/usr/bin/env node\n' + body);
+      fs.chmodSync(f, 0o755);
+    }
+  };
+  const runShimVersion = (bin) => {
+    const env = { ...process.env, PATH: `${bin}${path.delimiter}${process.env.PATH}` };
+    delete env.MINIMAX_API_KEY; // 与 mmxRun 同纪律: 测试自己的探测也别贡献 ENVKEY=present(否则机器导出过 Key 就假红)
+    return process.platform === 'win32'
+      ? spawnSync(process.env.ComSpec || 'cmd.exe', ['/d', '/s', '/c', 'mmx', '--version'], { encoding: 'utf8', windowsHide: true, env })
+      : spawnSync('mmx', ['--version'], { encoding: 'utf8', windowsHide: true, env });
+  };
+  // 两个 shim 用例的公共前置; 不满足能力时 t.skip 并返回 null(调用方直接 return)
+  const setupShimCase = (t, { fail } = {}) => {
+    if (!findTool('ffprobe')) { t.skip('无 ffprobe: prepareForUpload 走不到提供方分支'); return null; }
+    const bin = tmpdir(); const log = path.join(bin, 'args.log');
+    makeShim(bin, log, { fail });
+    if (runShimVersion(bin).status !== 0) { t.skip('当前环境执行不了 mmx shim(可执行脚本受限)'); return null; }
+    // 项目路径故意带 & 和空格: Windows 上 mmxRun 经 cmd.exe 拼命令行, 引用必须保住元字符(1.9.1 修复的回归)
+    const proj = path.join(tmpdir(), 'a&b c');
+    fs.mkdirSync(proj, { recursive: true });
+    const mp3 = path.join(proj, 'clip.mp3');
+    const g = spawnSync(findTool('ffmpeg'), ['-v', 'error', '-f', 'lavfi', '-i', 'anullsrc=r=32000:cl=mono', '-t', '1', '-c:a', 'libmp3lame', '-y', mp3], { windowsHide: true });
+    if (g.status !== 0) { t.skip('造不出音频'); return null; }
+    const env = { ...process.env, PATH: `${bin}${path.delimiter}${process.env.PATH}`, MINIMAX_API_KEY: 'should-not-reach-child' };
+    return { bin, log, proj, mp3, env };
+  };
+
+  test('mmx shim: --provider mmx 调 speech transcribe(含 --timestamp-level 映射), 元字符路径不破, 输出经 os.tmpdir() 中转且 Key 不流向子进程', async (t) => {
+    const ctx = setupShimCase(t);
+    if (!ctx) return;
+    const before = new Set(fs.readdirSync(os.tmpdir()));
+    const r = await runSkillAsync('asr.mjs', [ctx.proj, '--file', ctx.mp3, '--provider', 'mmx', '--language', 'zh', '--format', 'verbose_json', '--timestamp', 'word'], { env: ctx.env });
+    assert.equal(r.status, 0, r.stdout + r.stderr);
+    assert.ok((r.stdout || '').includes('你好世界二零二六'), '应打印 shim 文档里解析出的 text');
+    const lines = fs.readFileSync(ctx.log, 'utf8').trim().split('\n');
+    const tr = lines.filter(l => l.startsWith('[')).map(l => JSON.parse(l)).find(a => a.includes('transcribe'));
+    assert.ok(tr, '应调用 speech transcribe');
+    for (const want of ['speech', 'transcribe', '--file', '--model', 'asr-1.0', '--response-format', 'verbose_json', '--timestamp-level', 'word', '--language', 'zh']) assert.ok(tr.includes(want), `参数缺 ${want}: ${JSON.stringify(tr)}`);
+    const outIdx = tr.indexOf('--out');
+    assert.ok(tr[outIdx + 1].includes('asr-mmx-'), '--out 必须落在专属临时目录');
+    // 不能只看第一条 ENVKEY(那是 setup 探测写的, 空过): 整份日志里任何一次 mmx 子进程调用
+    // (--version 探测也算)都不许看到 Key —— 剥除回退到任何一处, 这里就红(1.9.2 修正断言空过)。
+    const envLines = lines.filter(l => l.startsWith('ENVKEY='));
+    assert.ok(envLines.length > 0 && envLines.every(l => l === 'ENVKEY=absent'),
+      `mmx 子进程(探测与 transcribe 全算)不得看到 MINIMAX_API_KEY: ${envLines.join(', ')}`);
+    // 临时目录必须清理干净(与转码临时目录同一条纪律)
+    assert.deepEqual(fs.readdirSync(os.tmpdir()).filter(f => f.startsWith('asr-mmx-') && !before.has(f)), [], 'os.tmpdir() 的 mmx 中转目录必须已清理');
+  });
+
+  test('mmx shim 失败(退出 3) → asr 非零退出, 透出 mmx 的 stderr 与修法提示', async (t) => {
+    const ctx = setupShimCase(t, { fail: true });
+    if (!ctx) return;
+    const r = await runSkillAsync('asr.mjs', [ctx.proj, '--file', ctx.mp3, '--provider', 'mmx'], { env: ctx.env });
+    assert.notEqual(r.status, 0, 'mmx 失败不能被当成功');
+    const out = r.stdout + r.stderr;
+    assert.match(out, /mmx speech transcribe 失败/, '应点名提供方与命令');
+    assert.match(out, /boom-quota/, '应透出 mmx 的 stderr');
+    assert.match(out, /mmx-cli@latest/, '应给出升级提示');
+  });
+});
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/fetch-policy.test.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/fetch-policy.test.mjs
new file mode 100644
index 00000000..da705838
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/fetch-policy.test.mjs
@@ -0,0 +1,164 @@
+// 安全边界 · 抓图 URL 策略(SSRF / 本地文件读 / 重定向 / 文件名)
+// 对应评审意见 4: SSRF and local-file-read behavior in fetch-official-images.mjs。
+import { test, describe } from 'node:test';
+import assert from 'node:assert/strict';
+import {
+  isBlockedHost, assertFetchableUrl, assertRedirectTarget, sanitizeFilename, imageNameFromUrl, MAX_REDIRECTS, PolicyError,
+} from '../scripts/url-policy.mjs';
+
+describe('isBlockedHost(内网/本地/元数据/裸主机名)', () => {
+  const blocked = [
+    '127.0.0.1', '127.8.9.10',            // loopback
+    '169.254.169.254', '169.254.0.9',     // 链路本地(含云元数据)
+    '10.0.0.5', '172.16.0.1', '172.31.9.9', '192.168.1.1', // 私网
+    '100.64.0.1',                          // CGNAT
+    '0.0.0.0',
+    '::1', '::', 'fe80::1', 'fc00::1', 'fd12:3456::1', // IPv6 loopback/链路本地/ULA
+    '::ffff:127.0.0.1', '::ffff:10.0.0.1', // IPv4-mapped
+    'localhost', 'intranet-app',           // 无点主机名
+    'box.local', 'corp.internal', 'x.home.arpa',
+    '',                                    // 空
+  ];
+  const allowed = [
+    '8.8.8.8', '1.1.1.1', '172.32.0.1', '172.15.0.1', // 公网(含私网段边界外)
+    'api.minimaxi.com', 'example.com', 'cdn.example.co.uk',
+    '::ffff:8.8.8.8', '2606:4700::1111',
+  ];
+  for (const h of blocked) test(`拦 ${h}`, () => assert.equal(isBlockedHost(h), true, h));
+  for (const h of allowed) test(`放 ${h}`, () => assert.equal(isBlockedHost(h), false, h));
+});
+
+describe('assertFetchableUrl', () => {
+  test('公网 https/http 放行', () => {
+    assertFetchableUrl('https://example.com/a.png');
+    assertFetchableUrl('http://example.com/a.png');
+  });
+  test('内网/元数据/裸名拒绝', () => {
+    for (const u of ['https://127.0.0.1/x', 'http://169.254.169.254/latest', 'https://192.168.0.1/', 'http://localhost/', 'http://intranet/x']) {
+      assert.throws(() => assertFetchableUrl(u), PolicyError, u);
+    }
+  });
+  test('file:// 默认拒绝, --allow-file 放行', () => {
+    assert.throws(() => assertFetchableUrl('file:///C:/page.html'), PolicyError);
+    assertFetchableUrl('file:///C:/page.html', { allowFile: true });
+  });
+  test('带 userinfo 的 URL 拒绝(凭据会泄漏进日志)', () => {
+    assert.throws(() => assertFetchableUrl('https://user:pw@example.com/a.png'), PolicyError);
+  });
+  test('非 http(s)/file 协议拒绝', () => {
+    for (const u of ['ftp://example.com/a', 'javascript:alert(1)', 'data:text/html,x']) {
+      assert.throws(() => assertFetchableUrl(u), PolicyError, u);
+    }
+  });
+});
+
+describe('重定向目标逐跳校验', () => {
+  test('重定向到私网/元数据/file: 拒绝', () => {
+    for (const loc of ['http://10.0.0.5/x.png', 'http://169.254.169.254/a', 'file:///C:/Windows/win.ini', 'https://evil.example']) {
+      if (loc === 'https://evil.example') continue; // 公网第三方在抓图语境是允许的(offsite 仅标注)
+      assert.throws(() => assertRedirectTarget(loc), PolicyError, loc);
+    }
+  });
+  test('重定向到公网放行; MAX_REDIRECTS 有上限', () => {
+    assertRedirectTarget('https://cdn.example.com/a.png');
+    assert.equal(typeof MAX_REDIRECTS, 'number');
+    assert.ok(MAX_REDIRECTS >= 1 && MAX_REDIRECTS <= 10, '重定向上限应在 1–10 跳');
+  });
+});
+
+describe('sanitizeFilename(页面控制的 alt → 落盘名)', () => {
+  test('路径分隔符与控制字符被替换', () => {
+    const n = sanitizeFilename('brand/logo\\main v2');
+    assert.ok(!n.includes('/') && !n.includes('\\'), n);
+  });
+  test('Windows 保留名被替换', () => {
+    for (const bad of ['CON', 'con', 'NUL', 'com1', 'LPT9']) {
+      assert.ok(sanitizeFilename(bad) !== bad, `保留名 ${bad} 必须被换掉`);
+    }
+  });
+  test('开头点/空格被剥(不产生隐藏文件)', () => {
+    assert.ok(!sanitizeFilename('  .hidden').startsWith('.'), sanitizeFilename('  .hidden'));
+  });
+  test('空/undefined → fallback', () => {
+    assert.equal(sanitizeFilename(''), 'official');
+    assert.equal(sanitizeFilename(undefined), 'official');
+  });
+  test('长度有上限', () => {
+    assert.ok(sanitizeFilename('x'.repeat(500)).length <= 30);
+  });
+});
+
+describe('imageNameFromUrl(零散图片 URL → 落盘文件名)', () => {
+  test('URL 自己的扩展名优先', () => {
+    assert.equal(imageNameFromUrl('https://cdn.x/img/hero-2026.png'), 'hero-2026.png');
+    assert.equal(imageNameFromUrl('https://cdn.x/a/CEO%20photo.jpg?v=2'), 'CEO-photo.jpg');
+  });
+  test('没有扩展名 → 按 content-type 推', () => {
+    assert.equal(imageNameFromUrl('https://cdn.x/image', 'image/webp; charset=binary'), 'image.webp');
+    assert.equal(imageNameFromUrl('https://cdn.x/image', 'image/svg+xml'), 'image.svg');
+    assert.equal(imageNameFromUrl('https://cdn.x/image', 'text/html'), 'image.bin');
+  });
+  test('路径穿越 / 保留名 / 空路径都不会逃出 out-dir', () => {
+    assert.equal(imageNameFromUrl('https://cdn.x/../../etc/passwd'), 'passwd.bin');
+    assert.equal(imageNameFromUrl('https://cdn.x/CON.png'), 'official.png');
+    assert.equal(imageNameFromUrl('https://cdn.x/', 'image/png'), 'official.png');
+    assert.ok(!imageNameFromUrl('https://cdn.x/%2e%2e%2fescape.png').includes('/'));
+  });
+  test('长度有上限且不含路径分隔符', () => {
+    const n = imageNameFromUrl('https://cdn.x/' + 'a'.repeat(300) + '.png');
+    assert.ok(n.length <= 34, n);
+    assert.ok(!/[\/:*?"<>|]/.test(n), n);
+  });
+});
+
+describe('IPv4-mapped IPv6 的十六进制形式(经 new URL() 规范化后的真实到达形态)', () => {
+  // WHATWG URL 会把 http://[::ffff:127.0.0.1]/ 的 hostname 规范化成 ::ffff:7f00:1 再送检 ——
+  // 只测点分形式会漏掉这条真实穿透路径(2026-09-18 安全审计端到端打到了 loopback 与云元数据)
+  const blocked = [
+    'http://[::ffff:7f00:1]/x',                    // 127.0.0.1 loopback
+    'http://[::ffff:a9fe:a9fe]/latest/meta-data/', // 169.254.169.254 云元数据
+    'http://[::ffff:c0a8:101]/router',             // 192.168.1.1
+    'http://[::ffff:a00:5]/internal',              // 10.0.0.5
+    'http://[0:0:0:0:0:ffff:a9fe:a9fe]/',          // 长形式(解析器会压缩)
+  ];
+  for (const u of blocked) {
+    test(`拦 ${u}`, () => assert.throws(() => assertFetchableUrl(u), PolicyError));
+  }
+  test('放 公网映射形式 [::ffff:808:808](= 8.8.8.8): 换算回点分后判定, 不是一刀切', () => {
+    assert.doesNotThrow(() => assertFetchableUrl('http://[::ffff:808:808]/pub.png'));
+  });
+});
+
+describe('2026-09-18 复查补全: 尾点 FQDN / NAT64 / 6to4 / 保留段', () => {
+  // 这一组是把"外部评审打穿的两例 + 同类载体"固化成负例: 尾点 FQDN 与 NAT64 当时都是实测放行
+  const blocked = [
+    'localhost.',                       // 尾点绝对名, 与 localhost 同一主机
+    'LOCALHOST.',                       // 大小写
+    '127.0.0.1.',                       // 字面量带尾点
+    '10.0.0.1.',
+    'printer.local.',
+    '240.0.0.1',                        // 240/4 保留
+    '255.255.255.255',
+    '224.0.0.1',                        // 组播
+    '198.18.0.1',                       // 基准测试段
+    '192.0.0.1',                        // IETF 保留
+    '192.0.2.5',                        // TEST-NET-1
+    '198.51.100.7',                     // TEST-NET-2
+    '203.0.113.9',                      // TEST-NET-3
+    '[fec0::1]',                        // 站点本地(废弃)
+    '[2001:db8::1]',                    // 文档段
+    '[64:ff9b::7f00:1]',                // NAT64 内嵌 127.0.0.1(实测曾放行)
+    '[64:ff9b::a9fe:a9fe]',             // NAT64 内嵌 169.254.169.254
+    '[2002:7f00:0001::]',               // 6to4 内嵌 127.0.0.1
+    '[2002:c0a8:0101::]',               // 6to4 内嵌 192.168.1.1
+  ];
+  for (const h of blocked) test(`拦 ${h}`, () => assert.equal(isBlockedHost(h), true, h));
+
+  const allowed = [
+    '93.184.216.34',                    // 公网 IPv4
+    'cdn.example.com.',                 // 公网域名的尾点写法照常放行
+    '[2001:4860:4860::8888]',           // 公网 IPv6 (Google DNS)
+    '[2002:0808:0808::]',               // 6to4 内嵌 8.8.8.8 —— 换算后是公网, 不该一刀切拦
+  ];
+  for (const h of allowed) test(`放 ${h}`, () => assert.equal(isBlockedHost(h), false, h));
+});
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/helpers.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/helpers.mjs
new file mode 100644
index 00000000..a486f6e8
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/helpers.mjs
@@ -0,0 +1,69 @@
+// html2video-for-mcode · 测试共用件(文件名不匹配 *.test.mjs, 不会被 node --test 当作用例执行)
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+import { spawnSync } from 'node:child_process';
+import { fileURLToPath, pathToFileURL } from 'node:url';
+import { wrapTokens } from '../scripts/css-kit.mjs';
+import { generateTokensCss, TOKENS_REV } from '../scripts/tokens-template.mjs';
+
+export const SCRIPTS = path.join(path.dirname(fileURLToPath(import.meta.url)), '..', 'scripts');
+export const skill = name => path.join(SCRIPTS, name);
+
+// 跑一个技能脚本, 返回 { status, stdout, stderr }
+export function run(cmd, args, opts = {}) {
+  const r = spawnSync(cmd, args, { encoding: 'utf8', windowsHide: true, ...opts });
+  return { status: r.status, stdout: r.stdout ?? '', stderr: r.stderr ?? '' };
+}
+export const runSkill = (name, args, opts = {}) => run(process.execPath, [skill(name), ...args], opts);
+
+// 异步版: 子进程与父进程需要并发交互时必须用这个 —— spawnSync 会冻结事件循环,
+// 父进程里的本地服务器 accept 不了子进程的连接, 互相等死(本次实测踩过)
+import { spawn } from 'node:child_process';
+export function runSkillAsync(name, args, opts = {}) {
+  return new Promise(resolve => {
+    const c = spawn(process.execPath, [skill(name), ...args], { windowsHide: true, ...opts });
+    let stdout = '', stderr = '';
+    c.stdout.on('data', d => { stdout += d; });
+    c.stderr.on('data', d => { stderr += d; });
+    c.on('error', err => resolve({ status: -1, stdout, stderr: stderr + String(err) }));
+    c.on('close', status => resolve({ status, stdout, stderr }));
+  });
+}
+
+// 临时目录(进程退出时统一清理 —— 此前上万个 h2v-test-* 堆在 %TEMP%, 第八轮复查 F2)
+const madeDirs = new Set();
+export const tmpdir = () => {
+  const d = fs.mkdtempSync(path.join(os.tmpdir(), 'h2v-test-'));
+  madeDirs.add(d);
+  return d;
+};
+process.once('exit', () => { for (const d of madeDirs) fs.rmSync(d, { recursive: true, force: true }); });
+
+// 在 subprocess 里调用 tools.mjs 的退出型助手(safeId/safeRel 会 process.exit, 只能这样测)
+export function probeHelper(expr, argsJson) {
+  const code = `
+    import { ${expr} } from ${JSON.stringify(pathToFileURL(skill('tools.mjs')).href)};
+    const args = ${argsJson};
+    const v = ${expr}(...args);
+    process.stdout.write('RETURN:' + String(v));
+  `;
+  return run(process.execPath, ['--input-type=module', '-e', code]);
+}
+
+// 最小可跑的项目骨架(check-slides 需要 tokens.css; 其余脚本按需补)
+// tokens.css 默认写成**真项目的样子**: 项目自己的规则 + 技能当前的受管区。2026-09-18 起
+// check-slides / capture / build-video 把"受管区缺失或落后"当阻断级(旧 CSS 会静默出片),
+// 所以默认骨架要是新鲜的 —— 要测"落后/无受管区"的场景, 显式传 tokens 或 LEGACY_TOKENS。
+export const FRESH_TOKENS = ':root { --accent: #111; }\n' + wrapTokens(generateTokensCss(), TOKENS_REV) + '\n';
+export const LEGACY_TOKENS = ':root { --accent: #111; }\n';   // 老项目形状: 没有受管区
+export function mkproj(dir, { slides = [], tokens = FRESH_TOKENS } = {}) {
+  fs.mkdirSync(path.join(dir, 'slides'), { recursive: true });
+  fs.mkdirSync(path.join(dir, 'audio'), { recursive: true });
+  fs.mkdirSync(path.join(dir, 'build'), { recursive: true });
+  fs.writeFileSync(path.join(dir, 'slides', 'tokens.css'), tokens);
+  fs.writeFileSync(path.join(dir, 'script.json'), JSON.stringify({
+    topic: 't', fps: 30, width: 1920, height: 1080, slides,
+  }, null, 2));
+  return dir;
+}
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/no-clobber.test.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/no-clobber.test.mjs
new file mode 100644
index 00000000..c62d3bfc
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/no-clobber.test.mjs
@@ -0,0 +1,121 @@
+// 安全边界 · 覆盖拒绝(不再无条件覆写用户/agent 已有内容)
+// 对应评审意见 2: destructive overwrite behavior。
+import { test, describe } from 'node:test';
+import assert from 'node:assert/strict';
+import fs from 'node:fs';
+import path from 'node:path';
+import { runSkill, mkproj, tmpdir } from './helpers.mjs';
+
+describe('init-project: 非空目录拒绝 / --force 只重置生成文件', () => {
+  test('空目录 → 正常生成', () => {
+    const proj = tmpdir();
+    const r = runSkill('init-project.mjs', [proj, '--topic', 'T']);
+    assert.equal(r.status, 0, r.stderr);
+    assert.ok(fs.existsSync(path.join(proj, 'script.json')));
+  });
+
+  test('非空目录(有无关文件) → 拒绝, 退出 1, 无关文件不动', () => {
+    const proj = tmpdir();
+    fs.writeFileSync(path.join(proj, 'user-notes.txt'), 'mine');
+    const r = runSkill('init-project.mjs', [proj, '--topic', 'T']);
+    assert.notEqual(r.status, 0);
+    assert.ok(r.stderr.includes('--force'), '应提示 --force');
+    assert.equal(fs.readFileSync(path.join(proj, 'user-notes.txt'), 'utf8'), 'mine', '无关文件不得被动');
+  });
+
+  test('已有 script.json 的项目 → 拒绝并列出将被覆写的文件', () => {
+    const proj = mkproj(tmpdir(), { slides: [{ id: '01', clauses: [{ stage: 1, text: '已填的口播' }] }] });
+    const r = runSkill('init-project.mjs', [proj]);
+    assert.notEqual(r.status, 0);
+    assert.ok(r.stderr.includes('script.json'), '应点名 script.json');
+    assert.equal(JSON.parse(fs.readFileSync(path.join(proj, 'script.json'), 'utf8')).slides[0].clauses[0].text, '已填的口播', '已填内容不得被清');
+  });
+
+  test('--force → 重新生成, 无关文件保留', () => {
+    const proj = mkproj(tmpdir(), { slides: [{ id: '01', clauses: [{ stage: 1, text: '旧' }] }] });
+    fs.writeFileSync(path.join(proj, 'user-notes.txt'), 'mine');
+    const r = runSkill('init-project.mjs', [proj, '--force']);
+    assert.equal(r.status, 0, r.stderr);
+    assert.equal(fs.readFileSync(path.join(proj, 'user-notes.txt'), 'utf8'), 'mine', '--force 也不删无关文件');
+    const script = JSON.parse(fs.readFileSync(path.join(proj, 'script.json'), 'utf8'));
+    assert.equal(script.slides[0].clauses[0].text, '', '--force 后 script.json 被重置');
+  });
+
+  test('--topic 的 HTML 转义: <script> 不会原样进模板', () => {
+    const proj = tmpdir();
+    const r = runSkill('init-project.mjs', [proj, '--topic', 'A<B>&"C"']);
+    assert.equal(r.status, 0, r.stderr);
+    const tpl = fs.readFileSync(path.join(proj, 'slides', '_template.html'), 'utf8');
+    assert.ok(!tpl.includes('A<B>'), '模板里不得出现未转义的 <');
+    const script = JSON.parse(fs.readFileSync(path.join(proj, 'script.json'), 'utf8'));
+    assert.equal(script.topic, 'A<B>&"C"', 'script.json 存原文');
+  });
+});
+
+describe('fetch-official-images: 参数校验在打开浏览器之前完成', () => {
+  test('内网地址 → 退出 1(无需 playwright)', () => {
+    for (const u of ['http://127.0.0.1:8080/', 'http://192.168.1.1/', 'http://169.254.169.254/latest/meta-data/', 'http://10.0.0.5/', 'http://localhost/x']) {
+      const r = runSkill('fetch-official-images.mjs', [u]);
+      assert.notEqual(r.status, 0, `${u} 必须拒绝`);
+    }
+  });
+  test('file:// 无 --allow-file → 退出 1', () => {
+    const r = runSkill('fetch-official-images.mjs', ['file:///C:/some/page.html']);
+    assert.notEqual(r.status, 0);
+    assert.ok(r.stderr.includes('--allow-file'), '应提示 --allow-file');
+  });
+  test('--out-dir 越出 cwd 且无 --force → 退出 1', () => {
+    const outside = tmpdir();
+    const r = runSkill('fetch-official-images.mjs',
+      ['https://example.com/', '--out-dir', path.join(outside, 'assets')], { cwd: tmpdir() });
+    assert.notEqual(r.status, 0);
+    assert.ok(r.stderr.includes('--out-dir') || r.stderr.includes('--force'));
+  });
+});
+
+describe('fetch-official-images --url: 零散图片 URL 也走同一套策略(不看页面、不需 playwright)', () => {
+  const proj = () => { const d = tmpdir(); fs.mkdirSync(path.join(d, 'assets'), { recursive: true }); return d; };
+  test('内网 / 元数据 / 裸主机名 / IPv6 → 退出 1, 且 assets/ 里不落任何文件', () => {
+    for (const u of ['http://169.254.169.254/a.png', 'http://127.0.0.1/a.png', 'http://192.168.1.10/logo.svg',
+      'http://10.0.0.5/a.png', 'http://[::1]/a.png', 'http://localhost/x.png', 'http://intranet-app/x.png']) {
+      const d = proj();
+      const r = runSkill('fetch-official-images.mjs', ['--url', u, '--out-dir', path.join(d, 'assets')], { cwd: d });
+      assert.equal(r.status, 1, `${u} 必须拒绝`);
+      assert.deepEqual(fs.readdirSync(path.join(d, 'assets')), [], `${u} 不得落盘`);
+    }
+  });
+  test('只收 http(s): ftp:// 与 file:// 直接跳过', () => {
+    const d = proj();
+    for (const u of ['ftp://example.com/a.png', 'file:///D:/x.png']) {
+      const r = runSkill('fetch-official-images.mjs', ['--url', u, '--out-dir', path.join(d, 'assets')], { cwd: d });
+      assert.equal(r.status, 1);
+      assert.ok((r.stdout + r.stderr).includes('只收 http(s)'), r.stdout + r.stderr);
+    }
+  });
+  test('--out-dir 越界且无 --force → 退出 1(与页面模式同一道门)', () => {
+    const outside = tmpdir();
+    const r = runSkill('fetch-official-images.mjs', ['--url', 'https://example.com/a.png', '--out-dir', path.join(outside, 'assets')], { cwd: tmpdir() });
+    assert.equal(r.status, 1);
+    assert.ok(r.stderr.includes('--out-dir'), r.stderr);
+  });
+  test('没有 --url 也没有页面 URL → 打印用法并退出 1', () => {
+    const r = runSkill('fetch-official-images.mjs', []);
+    assert.equal(r.status, 1);
+    assert.ok(r.stderr.includes('--url'), '用法里要提到 --url');
+  });
+});
+
+describe('prep-image --crop: 输出已存在 → 拒绝(校验在 ffmpeg 之前)', () => {
+  test('dst 已存在且无 --force → 退出 1 且文件内容不变', () => {
+    const proj = tmpdir();
+    const src = path.join(proj, 'in.png');
+    const dst = path.join(proj, 'out.png');
+    fs.writeFileSync(src, 'fake-png');
+    fs.writeFileSync(dst, 'PRECIOUS');
+    const r = runSkill('prep-image.mjs', ['--crop', src, dst]);
+    // 无 ffmpeg 时本检查依然应先于工具解析... 事实是 requireTool 在顶部:
+    // 无 ffmpeg → 退出 2(同样没碰 dst); 有 ffmpeg → 覆盖拒绝退出 1。两种都不得改写 dst。
+    assert.notEqual(r.status, 0);
+    assert.equal(fs.readFileSync(dst, 'utf8'), 'PRECIOUS', '目标文件不得被改写');
+  });
+});
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/preview-page.test.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/preview-page.test.mjs
new file mode 100644
index 00000000..0c391f5c
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/preview-page.test.mjs
@@ -0,0 +1,472 @@
+// 放映页(preview-page): 快照时序注入 + no-fx 兜底 + 自包含 + 路径收监
+// 这是"用户自己在浏览器里先放映一遍"的交付物, 也是"元素永久不可见"那类静默故障的目视闸门。
+import { test, describe } from 'node:test';
+import assert from 'node:assert/strict';
+import fs from 'node:fs';
+import path from 'node:path';
+import { runSkill, mkproj, tmpdir, SCRIPTS, LEGACY_TOKENS } from './helpers.mjs';
+
+const mod = await import('file://' + path.join(SCRIPTS, 'preview-page.mjs').replace(/\\/g, '/'));
+const { stageVars, injectHtmlVars, addNoFx, addBaseHref, injectStyle, buildPlayPage, stagesFromHtml, fallbackStages } = mod;
+const { NOFX_CSS } = await import('file://' + path.join(SCRIPTS, 'nofx-css.mjs').replace(/\\/g, '/'));
+const { CHART_CSS } = await import('file://' + path.join(SCRIPTS, 'chart-css.mjs').replace(/\\/g, '/'));
+const { TABLE_CSS } = await import('file://' + path.join(SCRIPTS, 'table-css.mjs').replace(/\\/g, '/'));
+const { generateTokensCss } = await import('file://' + path.join(SCRIPTS, 'tokens-template.mjs').replace(/\\/g, '/'));
+
+const SLIDE = `<!doctype html>
+<html lang="zh-CN" data-theme="a">
+<head>
+<meta charset="utf-8">
+<link rel="stylesheet" href="tokens.css">
+</head>
+<body><div class="stage"><p class="fx-up" data-stage="1">文案</p><p class="fx-fade" data-stage="2">展开</p></div></body>
+</html>
+`;
+
+describe('预览放映页 · 纯函数', () => {
+  test('stageVars: 秒 → 整数毫秒, 负数归零, 非数值跳过', () => {
+    assert.equal(stageVars({ 1: 0, 2: 4.8655 }, { w: 1920, h: 1080 }),
+      '--t1:0ms;--t2:4866ms;--stage-w:1920px;--stage-h:1080px');
+    assert.equal(stageVars({ 2: -3 }, {}), '--t2:0ms');
+    assert.equal(stageVars({ 2: 'x' }, {}), '');
+  });
+
+  test('injectHtmlVars: 写进 <html> 的 style(等价 capture 的 documentElement 注入)', () => {
+    const out = injectHtmlVars(SLIDE, '--t2:4866ms');
+    assert.match(out, /<html lang="zh-CN" data-theme="a" style="--t2:4866ms">/);
+    assert.ok(out.includes('<link rel="stylesheet" href="tokens.css">'), '正文不得被改动');
+  });
+
+  test('injectHtmlVars: 已有 style 时合并而不是覆盖', () => {
+    const html = '<html style="color:red">';
+    assert.equal(injectHtmlVars(html, '--t1:0ms'), '<html style="color:red;--t1:0ms">');
+  });
+
+  test('addNoFx: 无 class 时新增, 有 class 时追加, 重复调用不叠加', () => {
+    assert.match(addNoFx(SLIDE), /<html lang="zh-CN" data-theme="a" class="no-fx">/);
+    const once = addNoFx('<html class="deck">');
+    assert.equal(once, '<html class="deck no-fx">');
+    assert.equal(addNoFx(once), once);
+  });
+
+  test('addBaseHref: 插在 <head> 之后、link 之前(base 必须早于相对 URL)', () => {
+    const out = addBaseHref(SLIDE, '../../slides/');
+    assert.ok(out.indexOf('<base href="../../slides/">') < out.indexOf('href="tokens.css"'), 'base 必须排在相对链接前面');
+  });
+
+  test('injectStyle: 插到 </head> 前(tokens.css 之后, 后写覆盖)', () => {
+    const out = injectStyle(SLIDE, '.no-fx [data-stage]{opacity:1 !important}');
+    assert.ok(out.indexOf('tokens.css') < out.indexOf('.no-fx [data-stage]'));
+    assert.ok(out.indexOf('.no-fx [data-stage]') < out.indexOf('</head>'));
+  });
+
+  test('buildPlayPage: 自包含(无外链/无外站脚本)且含全部张次', () => {
+    const page = buildPlayPage({
+      topic: 'T', slides: [
+        { id: '01', name: '01.html', copy: '01.html', copyNofx: '01.nofx.html', clauses: [{ stage: 1, text: '甲' }] },
+        { id: '02', name: '02.html', copy: '02.html', copyNofx: '02.nofx.html', clauses: [] },
+      ], cssNote: '注意 X',
+    });
+    assert.ok(!/https?:/.test(page), '不得引用任何外部 URL(离线双击可用)');
+    assert.ok(!/<script[^>]+src=/i.test(page), '不得外链脚本');
+    assert.ok(page.indexOf('01.html') < page.indexOf('02.html'), '张次顺序保持');
+    assert.ok(page.includes('01.nofx.html'), '关动效副本要在模型里');
+    assert.ok(page.includes('注意 X'), 'tokens.css 落后时应把提示带给用户');
+  });
+
+  test('buildPlayPage: 不做播放器那套 UI(没有计时器/进度条/常驻动画循环)', () => {
+    const page = buildPlayPage({ topic: 'T', slides: [{ id: '01', name: 'a.html', copy: 'a.html', copyNofx: 'a.nofx.html', clauses: [] }] });
+    assert.ok(!page.includes('id="clock"'), '不要走秒的计时器');
+    assert.ok(!page.includes('id="bar"'), '不要时间轴进度条');
+    assert.ok(!page.includes('setInterval'), '不要常驻定时循环');
+    assert.ok(!/requestAnimationFrame\((loop|tick|frame)/.test(page), '不要常驻 rAF 动画循环 —— 双缓冲换帧的一次性 rAF 等帧不算');
+    assert.ok(!/配音/.test(page), '页面不得出现"配音"字样 —— 此前把"关动效对照"写成"关配音画对比", 让人以为在管音频');
+  });
+
+  test('buildPlayPage: 随窗口自适应(视口栏 / dvh / 窄屏媒体查询 / 触屏滑动)', () => {
+    const page = buildPlayPage({ topic: 'T', slides: [{ id: '01', name: 'a.html', copy: 'a.html', copyNofx: 'a.nofx.html', clauses: [] }] });
+    assert.match(page, /name="viewport"[^>]*width=device-width/, '手机需要视口声明');
+    assert.ok(page.includes('100dvh'), '用动态视口高度, 免得手机上被地址栏切掉');
+    assert.match(page, /@media \(max-width:860px\)/, '窄窗口要有独立布局(面板收到底部)');
+    assert.ok(page.includes('(pointer:coarse)'), '触屏要有触摸条而不是键盘提示');
+    assert.match(page, /class="kbd-hints"/, '键盘提示必须是类(内联样式会盖住媒体查询, 手机上就藏不掉)');
+    assert.ok(!page.includes('style="display:flex'), '键盘提示不得用内联 display');
+    assert.ok(page.includes('touchstart') && page.includes('touchend'), '触屏要能左右滑翻页');
+  });
+
+  test('buildPlayPage: 口播 UI 三态(有/未对时/不出)', () => {
+    const withNarr = buildPlayPage({ topic: 'T', timing: true, slides: [{ id: '01', name: 'a.html', copy: 'a.html', copyNofx: 'a.nofx.html', clauses: [{ stage: 1, text: '甲' }] }] });
+    assert.ok(withNarr.includes('id="panel"') && withNarr.includes('本张口播文案'));
+    assert.ok(!withNarr.includes('(未对时)'));
+    const noTiming = buildPlayPage({ topic: 'T', timing: false, slides: [{ id: '01', name: 'a.html', copy: 'a.html', copyNofx: 'a.nofx.html', clauses: [{ stage: 1, text: '甲' }] }] });
+    assert.ok(noTiming.includes('本张口播文案(未对时)'), '没对时要说清楚');
+    const off = buildPlayPage({ topic: 'T', narration: false, slides: [{ id: '01', name: 'a.html', copy: 'a.html', copyNofx: 'a.nofx.html', clauses: [] }] });
+    assert.ok(!off.includes('id="panel"'), '没有口播内容就不该有面板');
+    assert.match(off, /<body class="narr-off">/);
+    assert.ok(!off.includes('data-act="narr"'), '触摸条也不该留口播按钮');
+  });
+
+  test('stagesFromHtml / fallbackStages: 没对时时按 HTML 里用到的 stage 等间隔排', () => {
+    assert.deepEqual(stagesFromHtml(SLIDE), [1, 2]);
+    assert.deepEqual(stagesFromHtml('<p data-stage="2"></p><p data-stage="2"></p>'), [2]);
+    assert.deepEqual(fallbackStages(SLIDE), { 1: 0.3, 2: 1.3 });
+    assert.deepEqual(fallbackStages('<p data-stage="3"></p>'), { 3: 0.3 });
+    assert.deepEqual(fallbackStages('<p>无 stage</p>'), {});
+  });
+});
+
+describe('预览放映页 · 落到磁盘', () => {
+  const build = (proj, { tokens, timings } = {}) => {
+    mkproj(proj, {
+      slides: [{ id: '01', layout: 'statement', html: '01-title.html', audio: '01.mp3', title: '开场', clauses: [{ stage: 1, text: '第一句。' }, { stage: 2, text: '第二句。' }] }],
+      ...(tokens ? { tokens } : {}),
+    });
+    fs.writeFileSync(path.join(proj, 'slides', '01-title.html'), SLIDE);
+    if (timings) {
+      fs.writeFileSync(path.join(proj, 'build', 'timings.json'), JSON.stringify({
+        fps: 30, slides: [{ id: '01', duration: 6.3, stages: { 1: 0, 2: 4.8655 }, clauses: [{ stage: 1, start: 0, text: '第一句。' }, { stage: 2, start: 4.87, text: '第二句。' }] }],
+      }));
+    }
+    return proj;
+  };
+
+  test('生成放映页 + 两份快照, 副本带真实时序', () => {
+    const proj = build(tmpdir(), { timings: true });
+    const r = runSkill('preview-page.mjs', [proj]);
+    assert.equal(r.status, 0, r.stderr);
+    const out = path.join(proj, 'preview', 'play');
+    for (const f of ['index.html', '01-title.html', '01-title.nofx.html']) assert.ok(fs.existsSync(path.join(out, f)), `缺 ${f}`);
+    const copy = fs.readFileSync(path.join(out, '01-title.html'), 'utf8');
+    assert.ok(copy.includes('style="--t1:0ms;--t2:4866ms;'), '注入实测延迟(tokens.css 占位值是 800ms)');
+    assert.ok(copy.includes('<base href="../../slides/">'));
+    assert.ok(copy.includes('请勿编辑'), '副本要标明是快照');
+    assert.ok(fs.readFileSync(path.join(proj, 'slides', '01-title.html'), 'utf8').includes('data-theme="a"'), '原文件不得被改写');
+    assert.ok(r.stdout.includes('4866') || r.stdout.includes('末层'), '应打印注入结果');
+  });
+
+  test('老项目 tokens.css 落后(无受管区) → 副本兜底注入当前版并如实警告', () => {
+    const proj = build(tmpdir(), { tokens: LEGACY_TOKENS, timings: true });
+    const r = runSkill('preview-page.mjs', [proj]);
+    assert.equal(r.status, 0, r.stderr);
+    const out = path.join(proj, 'preview', 'play');
+    const nofx = fs.readFileSync(path.join(out, '01-title.nofx.html'), 'utf8');
+    const copy = fs.readFileSync(path.join(out, '01-title.html'), 'utf8');
+    assert.ok(/class="[^"]*no-fx/.test(nofx), '副本根元素要带 no-fx');
+    assert.ok(nofx.includes('opacity: 1 !important'), '旧 tokens.css 下必须兜底注入 opacity 重置, 否则对照是空白');
+    assert.ok(!copy.includes('no-fx 规则'), '动效副本不得注入 no-fx 规则(它只在关动效副本里有意义)');
+    assert.ok(copy.includes('图表工具箱'), '图表工具箱落后时动效副本也要兜底(否则旧规则把新版画法渲染坏)');
+    assert.ok(r.stderr.includes('no-fx'), '应在终端说明 tokens.css 落后');
+  });
+
+  test('新 tokens.css(tokens 主体 + 三段工具箱都当前) → 不注入, 保持纯快照', () => {
+    // "当前"的判据是**整段生成物在场**(受管区包裹, 或老格式的等价裸文本) —— 只贴三段工具箱
+    // 而没有主题令牌/字号/fx 的文件不算当前(渲染结果与技能当前版不同, 该兜底就兜底)。
+    const tokens = generateTokensCss();
+    const proj = build(tmpdir(), { tokens, timings: true });
+    const r = runSkill('preview-page.mjs', [proj]);
+    assert.equal(r.status, 0, r.stderr);
+    const outDir = path.join(proj, 'preview', 'play');
+    assert.ok(!fs.readFileSync(path.join(outDir, '01-title.nofx.html'), 'utf8').includes('兜底注入'));
+    assert.ok(!fs.readFileSync(path.join(outDir, '01-title.html'), 'utf8').includes('兜底注入'));
+    assert.ok(!r.stderr.includes('落后'));
+  });
+
+  test('老格式裸文本(无 tokens 受管区但内容已是当前版) → 同样不注入: 渲染无差异就不该骚扰', () => {
+    // 真实老格式: 只有外层 tokens 定界没有, 内嵌的三个工具箱子块照旧带定界(老模板就是那样)
+    const legacy = generateTokensCss()
+      .replace(/^\/\* >>> html2video:tokens rev=[0-9a-f]{8,64} >>> \*\/\n?/, '')
+      .replace(/\n?\/\* <<< html2video:tokens <<< \*\/\n?$/, '');
+    assert.ok(!/>>> html2video:tokens rev=/.test(legacy), '前置: 外层 tokens 定界已去掉');
+    assert.ok(/>>> html2video:nofx rev=/.test(legacy), '前置: 内嵌 kit 子块照旧留着');
+    const proj = build(tmpdir(), { tokens: legacy, timings: true });
+    const r = runSkill('preview-page.mjs', [proj]);
+    assert.equal(r.status, 0, r.stderr);
+    assert.ok(!fs.readFileSync(path.join(proj, 'preview', 'play', '01-title.html'), 'utf8').includes('兜底注入'));
+    assert.ok(!r.stderr.includes('落后'));
+  });
+
+  test('整段主体缺失(只贴了三段工具箱) → 兜底注入当前整段, 页面写明"成片仍是旧 CSS"', () => {
+    const tokens = `:root{--accent:#111}\n${NOFX_CSS}\n${CHART_CSS}\n${TABLE_CSS}\n`;
+    const proj = build(tmpdir(), { tokens, timings: true });
+    const r = runSkill('preview-page.mjs', [proj]);
+    assert.equal(r.status, 0, r.stderr);
+    const main = fs.readFileSync(path.join(proj, 'preview', 'play', '01-title.html'), 'utf8');
+    const nofx = fs.readFileSync(path.join(proj, 'preview', 'play', '01-title.nofx.html'), 'utf8');
+    assert.ok(main.includes('--fs-body'), '主题令牌/字号要兜底补上(缺了画面直接走样)');
+    assert.ok(!main.includes('no-fx 规则'), '动效副本不得注入 no-fx 规则');
+    assert.ok(nofx.includes('opacity: 1 !important'), '关动效副本要连 no-fx 一起兜底');
+    const page = fs.readFileSync(path.join(proj, 'preview', 'play', 'index.html'), 'utf8');
+    assert.ok(page.includes('成片仍用项目里的旧 CSS'), '放映页对了不等于成片对了 —— 这句话必须在页面上');
+    assert.ok(r.stderr.includes('--upgrade-css'), '终端也要给修复命令');
+  });
+
+  test('没有 timings.json → 照常出页, 但明确警告时序不是成片的', () => {
+    const proj = build(tmpdir());
+    const r = runSkill('preview-page.mjs', [proj]);
+    assert.equal(r.status, 0, r.stderr);
+    assert.ok(r.stderr.includes('timings.json'), '应提示先跑 plan-timings');
+    const page = fs.readFileSync(path.join(proj, 'preview', 'play', 'index.html'), 'utf8');
+    assert.ok(page.includes('等间隔'), '页面上要说清动画是按等间隔预览的');
+    assert.ok(page.includes('未对时'), '口播标题要标未对时');
+  });
+
+  test('--no-script: 不加载口播 UI, 但时序提示照留(它讲的是画面)', () => {
+    const proj = build(tmpdir());
+    const r = runSkill('preview-page.mjs', [proj, '--no-script']);
+    assert.equal(r.status, 0, r.stderr);
+    assert.ok(r.stdout.includes('关(--no-script)'));
+    const page = fs.readFileSync(path.join(proj, 'preview', 'play', 'index.html'), 'utf8');
+    assert.ok(!page.includes('id="panel"'));
+    assert.ok(!page.includes('data-act="narr"'));
+    assert.ok(page.includes('等间隔'), '口播关了, 动画时序的提示仍要有');
+  });
+
+  test('script.json 没有 clauses → 口播 UI 自动不加载', () => {
+    const proj = tmpdir();
+    mkproj(proj, { slides: [{ id: '01', layout: 'statement', html: '01-title.html', audio: '01.mp3', clauses: [] }] });
+    fs.writeFileSync(path.join(proj, 'slides', '01-title.html'), SLIDE);
+    const r = runSkill('preview-page.mjs', [proj]);
+    assert.equal(r.status, 0, r.stderr);
+    assert.ok(r.stdout.includes('本片没有 clauses'));
+    assert.ok(!fs.readFileSync(path.join(proj, 'preview', 'play', 'index.html'), 'utf8').includes('id="panel"'));
+  });
+
+  test('slides[].html 越界 → 退出 1, 项目外不落任何文件', () => {
+    const proj = build(tmpdir(), { timings: true });
+    const outside = tmpdir();
+    const canary = path.join(outside, 'canary.txt');
+    fs.writeFileSync(canary, 'SAFE');
+    const j = JSON.parse(fs.readFileSync(path.join(proj, 'script.json'), 'utf8'));
+    j.slides[0].html = '../../' + path.basename(outside) + '/pwn.html';
+    fs.writeFileSync(path.join(proj, 'script.json'), JSON.stringify(j));
+    const r = runSkill('preview-page.mjs', [proj]);
+    assert.equal(r.status, 1, '越界必须被 safeRel 拦下');
+    assert.equal(fs.readFileSync(canary, 'utf8'), 'SAFE');
+    assert.ok(!fs.existsSync(path.join(outside, 'pwn.html')));
+  });
+
+  test('部分张还没写 → 跳过并警告, 其余照常出页(与 capture 的惯例一致)', () => {
+    const proj = build(tmpdir(), { timings: true });
+    const j = JSON.parse(fs.readFileSync(path.join(proj, 'script.json'), 'utf8'));
+    j.slides.push({ id: '02', layout: 'bullets', html: '02-bullets.html', audio: '02.mp3', clauses: [{ stage: 1, text: '还没写。' }] });
+    fs.writeFileSync(path.join(proj, 'script.json'), JSON.stringify(j));
+    const r = runSkill('preview-page.mjs', [proj]);
+    assert.equal(r.status, 0, r.stderr);
+    assert.ok(r.stderr.includes('跳过 02'), '应点名跳过哪张');
+    assert.ok(fs.existsSync(path.join(proj, 'preview', 'play', '01-title.html')), '已写的张照常出');
+    assert.ok(!fs.existsSync(path.join(proj, 'preview', 'play', '02-bullets.html')));
+  });
+
+  test('一张可放映的都没有 → 退出 1', () => {
+    const proj = build(tmpdir(), { timings: true });
+    fs.rmSync(path.join(proj, 'slides', '01-title.html'));
+    const r = runSkill('preview-page.mjs', [proj]);
+    assert.equal(r.status, 1);
+    assert.ok(r.stderr.includes('可放映'));
+  });
+
+  test('缺 script.json → 退出 1', () => {
+    const r = runSkill('preview-page.mjs', [tmpdir()]);
+    assert.equal(r.status, 1);
+  });
+});
+
+describe('预览放映页 · 竖版画布与 iframe 安全(2026-09-18 流程/安全审计补)', () => {
+  const SLIDES_ARG = [{ id: '01', name: 'a.html', copy: 'a.html', copyNofx: 'a.nofx.html', clauses: [] }];
+
+  test('buildPlayPage: 画布参数化 — 竖版尺寸/缩放除数/缩略图比例都按 1080×1920', () => {
+    const page = buildPlayPage({ topic: 'T', canvas: { w: 1080, h: 1920 }, slides: SLIDES_ARG.map(s => ({ ...s })) });
+    assert.ok(page.includes('width:1080px;height:1920px'), '舞台与 iframe 要用竖版尺寸');
+    assert.ok(page.includes('r.width / 1080, r.height / 1920'), '缩放除数要按竖版算');
+    assert.ok(page.includes('aspect-ratio:1080/1920'), '总览缩略图比例要跟画布');
+    assert.ok(!page.includes('aspect-ratio:16/9'), '不得残留横屏写死比例');
+  });
+
+  test('iframe src 一律加 ./ 前缀: javascript: 文件名不会被当 scheme 在放映页同源执行', () => {
+    const page = buildPlayPage({ topic: 'T', slides: SLIDES_ARG.map(s => ({ ...s })) });
+    assert.match(page, /src\s*=\s*'\.\/'\s*\+\s*src/, '必须强制相对解析(赋值统一走 ./ + src 变量)');
+  });
+
+  test('竖版项目落盘: script.json 1080×1920 → index.html 用竖版画布', () => {
+    const proj = tmpdir();
+    mkproj(proj, { slides: [{ id: '01', layout: 'statement', html: '01-title.html', audio: '01.mp3', title: '竖版', clauses: [{ stage: 1, text: '一句。' }] }] });
+    const j = JSON.parse(fs.readFileSync(path.join(proj, 'script.json'), 'utf8'));
+    j.width = 1080; j.height = 1920;
+    fs.writeFileSync(path.join(proj, 'script.json'), JSON.stringify(j));
+    fs.writeFileSync(path.join(proj, 'slides', '01-title.html'), SLIDE);
+    const r = runSkill('preview-page.mjs', [proj]);
+    assert.equal(r.status, 0, r.stderr);
+    const page = fs.readFileSync(path.join(proj, 'preview', 'play', 'index.html'), 'utf8');
+    assert.ok(page.includes('width:1080px;height:1920px'), '放映页要按竖版渲染, 否则 Gate 4 画面被切掉近半');
+  });
+});
+
+describe('预览放映页 · 逐级步进(动效开)+ 无重播(2026-09-18 实测反馈)', () => {
+  const ARG = [{ id: '01', name: 'a.html', copy: 'a.html', copyNofx: 'a.nofx.html', clauses: [], steps: [1, 2] }];
+
+  test('开关移到底栏且写明状态(动效开/关 · 口播开/关), 顶部不再有动效开关;重播已移除', () => {
+    const page = buildPlayPage({ topic: 'T', slides: ARG.map(s => ({ ...s })) });
+    assert.ok(!page.includes('id="mode"'), '顶部动效开关要移掉(实测反馈: 与底栏开关重复)');
+    assert.match(page, /data-act="nofx"[^>]*id="fxbtn"[^>]*>动效开</, '底栏动效开关, 默认态写"动效开"');
+    assert.match(page, /data-act="narr"[^>]*id="narrbtn"[^>]*>口播开</, '底栏口播开关, 默认态写"口播开"');
+    assert.ok(page.includes('paintToggles'), '开关文案要随状态重绘(动效关/口播关)');
+    assert.ok(!page.includes('data-act="replay"'), '不得再有重播按钮');
+    assert.ok(!/<kbd>R<\/kbd>/.test(page), '不得再有 R 键提示');
+    assert.ok(page.includes('逐级入场 / 翻页'), '键盘提示要说明逐级行为');
+    assert.match(page, /#touchbar button\[data-act="prev"\]/, '前后箭头要有加大样式');
+    const noNarr = buildPlayPage({ topic: 'T', narration: false, slides: ARG.map(s => ({ ...s })) });
+    assert.ok(!noNarr.includes('id="narrbtn"'), '没有口播 UI 就不该有口播开关');
+  });
+
+  test('UI 双语: lang=en → 整套英文文案, 页面不残留中文 UI 字样(2026-09-18 实测反馈)', () => {
+    const en = buildPlayPage({ topic: 'T', lang: 'en', slides: [{ id: '01', name: 'a.html', copy: 'a.html', copyNofx: 'a.nofx.html', clauses: [{ stage: 1, text: 'Hi' }], steps: [1, 2] }] });
+    assert.match(en, /<html lang="en">/);
+    assert.match(en, /id="fxbtn"[^>]*>Motion on</, '动效开关英文');
+    assert.match(en, /id="narrbtn"[^>]*>Narration on</, '口播开关英文');
+    assert.ok(en.includes('Narration for this slide'), '口播面板标题英文');
+    assert.ok(en.includes('>Overview<'), '总览按钮英文');
+    assert.ok(en.includes('No thumbnail (run capture first — preview/'), '缩略图缺失占位英文');
+    assert.ok(en.includes('reveal level / slide'), '键盘提示英文(断言可见文本; 代码注释里的中文不算)');
+    assert.ok(!en.includes('>动效开<') && !en.includes('>口播开<'), '按钮不得残留中文');
+    assert.ok(!en.includes('本张口播文案') && !en.includes('快照生成于') && !en.includes('无缩略图'), '面板/页脚/占位不得残留中文');
+    const zh = buildPlayPage({ topic: 'T', lang: 'zh', slides: [{ id: '01', name: 'a.html', copy: 'a.html', copyNofx: 'a.nofx.html', clauses: [{ stage: 1, text: '甲' }] }] });
+    assert.ok(zh.includes('>动效开</button>') && zh.includes('>口播开</button>'), '中文默认不变');
+  });
+
+  test('双缓冲换帧: 两张叠放 iframe, load 后等两帧 rAF 才对调 —— 单 iframe 改 src 换页会闪白(实测反馈)', () => {
+    const page = buildPlayPage({ topic: 'T', slides: ARG.map(s => ({ ...s })) });
+    assert.ok(page.includes('id="frame"') && page.includes('id="frameB"'), '要有一前一后两张 iframe');
+    assert.match(page, /#fit iframe\{[^}]*position:absolute/, '两张 iframe 必须叠放, 隐藏帧加载完再显示');
+    assert.match(page, /#fit iframe\.show\{opacity:1\}/, '默认 opacity:0, 显示态由 .show 控制');
+    assert.ok(page.includes("back.addEventListener('load'"), '新帧要等 load 事件(文档就绪)才换');
+    assert.match(page, /raf\(function\(\)\{ raf\(function\(\)\{ if \(g === gen\)/, 'load 后还要等两帧 rAF(新文档首帧已画)才对调, 否则仍会闪');
+    assert.match(page, /src\s*=\s*'\.\/'\s*\+\s*src/, "换帧也必须走 './' + src 强制相对解析(javascript: 文件名不得当 scheme 执行)");
+  });
+
+  test('多级张的模型带 steps, 快照副本头部带步进脚本(?s=k 重设各级延迟)', () => {
+    const page = buildPlayPage({ topic: 'T', slides: ARG.map(s => ({ ...s })) });
+    assert.ok(page.includes('"steps":[1,2]'), '模型要带逐级序列');
+    const proj = tmpdir();
+    mkproj(proj, { slides: [{ id: '01', layout: 'statement', html: '01-title.html', audio: '01.mp3', clauses: [{ stage: 1, text: '第一句。' }, { stage: 2, text: '第二句。' }] }] });
+    fs.writeFileSync(path.join(proj, 'slides', '01-title.html'), SLIDE);   // SLIDE 用了 stage 1/2
+    const r = runSkill('preview-page.mjs', [proj]);
+    assert.equal(r.status, 0, r.stderr);
+    const copy = fs.readFileSync(path.join(proj, 'preview', 'play', '01-title.html'), 'utf8');
+    assert.ok(copy.includes('逐级步进'), '副本要带步进脚本说明');
+    assert.ok(/var used=\[1,2\]/.test(copy), '步进脚本要写入该张用到的 stage 序列');
+    assert.ok(copy.indexOf('逐级步进') < copy.indexOf('tokens.css'), '脚本必须在样式表之前(动画启动前执行)');
+    assert.ok(fs.readFileSync(path.join(proj, 'preview', 'play', 'index.html'), 'utf8').includes('"steps":[1,2]'));
+  });
+
+  test('行为级: ?s=1 隐藏二级 · ?s=2&anim=1 当场入场 · 回退不重播(需 Playwright,缺则 skip)', async t => {
+    const { loadPackage } = await import('file://' + path.join(SCRIPTS, 'tools.mjs').replace(/\\/g, '/'));
+    const playwright = await loadPackage('playwright');
+    if (!playwright) return t.skip('无 playwright');
+    let browser;
+    try { browser = await playwright.chromium.launch({ headless: true }); }
+    catch { return t.skip('chromium 未安装'); }
+    try {
+      // 必须用 init-project 的完整 tokens(fx 类/[data-stage] 基础态都在里面);
+      // mkproj 的极简 tokens 没有这些规则, 元素根本不参与动画 —— 首版就错在这里
+      const proj = tmpdir();
+      assert.equal(runSkill('init-project.mjs', [proj, '--topic', 'T']).status, 0);
+      fs.writeFileSync(path.join(proj, 'script.json'), JSON.stringify({ topic: 't', fps: 30, width: 1920, height: 1080, slides: [{ id: '01', layout: 'statement', html: '01-title.html', audio: '01.mp3', clauses: [{ stage: 1, text: '一。' }, { stage: 2, text: '二。' }] }] }, null, 2));
+      fs.writeFileSync(path.join(proj, 'slides', '01-title.html'), SLIDE);
+      assert.equal(runSkill('preview-page.mjs', [proj]).status, 0);
+      const p = await browser.newPage({ viewport: { width: 1280, height: 720 } });
+      const url = s => 'file:///' + path.join(proj, 'preview', 'play', '01-title.html').replace(/\\/g, '/') + s;
+      const op = sel => p.evaluate(x => { const e = document.querySelector(x); return e ? getComputedStyle(e).opacity : null; }, sel);
+      await p.goto(url('?s=1&anim=1')); await p.waitForTimeout(1500);
+      assert.equal(await op('[data-stage="1"]'), '1', '第 1 级可见');
+      assert.equal(await op('[data-stage="2"]'), '0', '第 2 级必须还在等(不能自己出来)');
+      await p.goto(url('?s=2&anim=1')); await p.waitForTimeout(1200);
+      assert.equal(await op('[data-stage="2"]'), '1', '按下一步后第 2 级当场入场');
+      await p.goto(url('?s=2')); await p.waitForTimeout(250);
+      assert.equal(await op('[data-stage="2"]'), '1', '回退/跳转态直接终态, 不用等动画');
+      await p.goto(url('?s=1&anim=1')); await p.waitForTimeout(250);
+      assert.equal(await op('[data-stage="2"]'), '0', '回到 ?s=1 二级重新隐藏');
+    } finally {
+      await browser.close();   // 断言失败也必须关掉浏览器, 否则事件循环挂着 node --test 永不退出(实测踩过)
+    }
+  });
+});
+
+// 2026-09-22 云沙箱实测: 入口守卫经符号链接/异形路径调用恒假 → 脚本静默 no-op(exit 0 无输出,
+// 排查了 10 分钟)。isMainModule 两侧 realpath + 同名兜底, 这里把三种形态钉死。
+test('isMainModule: 入口判定稳健(符号链接/同名兜底/被 import 不跑)', async t => {
+  const { isMainModule } = await import('file://' + path.join(SCRIPTS, 'tools.mjs').split(path.sep).join('/'));
+  const { pathToFileURL } = await import('node:url');
+  const dir = tmpdir();
+  const real = path.join(dir, 'probe.mjs');
+  fs.writeFileSync(real, '// probe\n');
+  const meta = pathToFileURL(real).href;
+  assert.equal(isMainModule(meta, real), true, '同一路径直接执行必须为真');
+  assert.equal(isMainModule(meta, path.join(dir, 'other.mjs')), false, '别的脚本 import 时必须为假');
+  assert.equal(isMainModule(meta, null), false, '无入口(如 node -e "import()")为假 —— 注意传 null: 传 undefined 会命中默认参数 process.argv[1], 测不到 !entry 分支');
+  assert.equal(isMainModule(meta, path.join(dir, 'sub', 'probe.mjs')), true, '同名不同目录 = 直接执行兜底(宁可多跑, 拒绝静默)');
+  const link = path.join(dir, 'link.mjs');
+  try { fs.symlinkSync(real, link, 'file'); } catch { return t.skip('当前环境建不了文件符号链接(Windows 需开发者模式)'); }
+  assert.equal(isMainModule(meta, link), true, '经符号链接调用必须为真(realpath 两侧) — 修复前恒假');
+});
+
+describe('外部报告(放映页六条) · P1a/P1b/P2/P3 修复', () => {
+  const page = buildPlayPage({ topic: 'T', slides: [{ id: '01', name: 'a.html', copy: 'a.html', copyNofx: 'a.nofx.html', clauses: [{ stage: 1, text: '甲' }], steps: [1] }] });
+
+  test('P1a 总览缩略图: 相对路径指向 play/thumbs/, 不再指到父目录', () => {
+    assert.ok(page.includes("img.src = 'thumbs/' + s.id + '.png'"), '总览缩略图必须用随页携带的 thumbs/ 相对路径');
+    assert.ok(!page.includes("'../' + s.id"), '不得再写死 ../ 指到父目录 preview/(只拷 play/ 就全挂)');
+  });
+
+  test('P1b 编程接口: window.playPage 暴露 state/go/next/prev/setFx/setNarr/setTrans/overview', () => {
+    assert.ok(page.includes('window.playPage = {'), '必须暴露 window.playPage(状态全是闭包变量, 外部自动化只能模拟键盘抠 innerText)');
+    for (const k of ['get state()', 'go: go', 'next: next', 'prev: prev', 'setFx:', 'setNarr:', 'setTrans:', 'overview:']) {
+      assert.ok(page.includes(k), `playPage 缺 ${k}`);
+    }
+  });
+
+  test('P2a 键位提示: footer 列全 Space/PgDn/PgUp/Home/End/Esc(键盘实际都支持)', () => {
+    assert.ok(page.includes('hintKeys') && page.includes('hintEsc'), '文案键要进页面');
+    const zh = buildPlayPage({ topic: 'T', lang: 'zh', slides: [{ id: '01', name: 'a.html', copy: 'a.html', copyNofx: 'a.nofx.html' }] });
+    assert.ok(zh.includes('空格/PgDn 前进'), '中文键位提示');
+    assert.ok(zh.includes('Home/End 首末张'), '中文键位提示(首末张)');
+    const en = buildPlayPage({ topic: 'T', lang: 'en', slides: [{ id: '01', name: 'a.html', copy: 'a.html', copyNofx: 'a.nofx.html' }] });
+    assert.ok(en.includes('Home / End first / last'), '英文键位提示');
+  });
+
+  test('P2b 边界反馈: 首/末 no-op 弹 toast, 不是毫无反应', () => {
+    assert.ok(page.includes('id="edge"'), '要有 toast 元素');
+    assert.ok(/next\(\)\{\s*if \(i >= S\.length - 1[^}]*\{ edge\(T\.edgeLast\); return; \}/.test(page), 'next 到末张必须 edge 反馈');
+    assert.ok(/prev\(\)\{\s*if \(i <= 0[^}]*\{ edge\(T\.edgeFirst\); return; \}/.test(page), 'prev 到首张必须 edge 反馈');
+  });
+
+  test('P3a xfade 状态泄漏: 换帧入口归一后备帧隐藏态', () => {
+    // 400ms 延迟撤场被 gen 作废时旧帧带 .show 留存, 两个 iframe 同 opacity:1 —— 外部脚本按
+    // offsetParent 判当前帧必错(报告人实踩)。入口一行归一即封死所有泄漏路径。
+    assert.ok(/function show\(src\)\{[\s\S]{0,400}back\.classList\.remove\('show'\)/.test(page), 'show() 入口必须先把后备帧归一到隐藏态');
+  });
+
+  test('P1a(CLI): preview/*.png 拷进 play/thumbs/, 没跑 capture 时如实计数告警', () => {
+    const proj = tmpdir();
+    const build2 = proj; // 复用上面的 build 夹具太重, 这里最小化: script + slides + (可选)preview png
+    fs.mkdirSync(path.join(build2, 'slides'), { recursive: true });
+    fs.writeFileSync(path.join(build2, 'script.json'), JSON.stringify({
+      topic: 'T', lang: 'zh', speed: 1.1,
+      slides: [{ id: '01', layout: 'title', html: '01.html', audio: '01.mp3', title: 'A', clauses: [{ stage: 1, text: '甲。' }] },
+               { id: '02', layout: 'title', html: '02.html', audio: '02.mp3', title: 'B', clauses: [{ stage: 1, text: '乙。' }] }],
+    }));
+    fs.writeFileSync(path.join(build2, 'slides', '01.html'), SLIDE.replace(/01-title/g, '01'));
+    fs.writeFileSync(path.join(build2, 'slides', '02.html'), SLIDE.replace(/01-title/g, '02'));
+    fs.mkdirSync(path.join(build2, 'preview'), { recursive: true });
+    fs.writeFileSync(path.join(build2, 'preview', '01.png'), 'PNG');   // 只有 01 跑过 capture
+    const r = runSkill('preview-page.mjs', [proj]);
+    assert.equal(r.status, 0, r.stderr);
+    assert.ok(fs.existsSync(path.join(proj, 'preview', 'play', 'thumbs', '01.png')), '存在的缩略图必须拷进 play/thumbs/');
+    assert.ok(!fs.existsSync(path.join(proj, 'preview', 'play', 'thumbs', '02.png')), '没跑 capture 的不得凭空造图');
+    assert.match(r.stdout + r.stderr, /1\/2 张/, '终端必须如实计数(其余没跑 capture)');
+    // 陈旧缩略图清理: 重新生成后老 thumbs 目录里不该留已删张次的死图
+    fs.writeFileSync(path.join(proj, 'preview', 'play', 'thumbs', '99.png'), 'STALE');
+    const r2 = runSkill('preview-page.mjs', [proj]);
+    assert.equal(r2.status, 0, r2.stderr);
+    assert.ok(!fs.existsSync(path.join(proj, 'preview', 'play', 'thumbs', '99.png')), '重建 thumbs/ 时陈旧死图必须清掉');
+  });
+});
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/render-smoke.test.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/render-smoke.test.mjs
new file mode 100644
index 00000000..3802be82
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/render-smoke.test.mjs
@@ -0,0 +1,510 @@
+// 渲染构建冒烟: init → 静音音频 → 最小 slide → plan-timings → check-slides → capture still → build-video
+// 对应评审意见 5: a minimal render/checker/build dry-run。
+// 无 ffmpeg / playwright / chromium 时逐级 skip 并说明原因(主 CI 只跑上面的安全测试;
+// 本文件由 .github/workflows/html2video-for-mcode-smoke.yml 在装好依赖的环境里真实执行)。
+import { test } from 'node:test';
+import assert from 'node:assert/strict';
+import fs from 'node:fs';
+import path from 'node:path';
+import { spawnSync } from 'node:child_process';
+import { findTool, loadPackage } from '../scripts/tools.mjs';
+import { runSkill, tmpdir, SCRIPTS } from './helpers.mjs';
+
+const FFMPEG = findTool('ffmpeg');
+const FFPROBE = findTool('ffprobe');
+let playwright = null;
+
+test('冒烟: 全链路出片', async t => {
+  if (!FFMPEG || !FFPROBE) return t.skip('无 ffmpeg/ffprobe(主 CI 环境; 由 scoped smoke workflow 覆盖)');
+  playwright = await loadPackage('playwright');
+  if (!playwright) return t.skip('无 playwright');
+  try {
+    const b = await playwright.chromium.launch({ headless: true });
+    await b.close();
+  } catch {
+    return t.skip('chromium 未安装(npx playwright install chromium)');
+  }
+
+  const proj = tmpdir();
+  // 1. init
+  let r = runSkill('init-project.mjs', [proj, '--topic', 'Smoke']);
+  assert.equal(r.status, 0, r.stderr);
+  // 2. 静音口播(1 张, 3s)
+  const mp3 = path.join(proj, 'audio', '01.mp3');
+  const g = spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'anullsrc=r=32000:cl=mono', '-t', '3',
+    '-c:a', 'libmp3lame', '-b:a', '64k', '-y', mp3], { windowsHide: true });
+  assert.equal(g.status, 0, g.stderr ?? '');
+  // 3. 只留 1 张 slide
+  const scriptPath = path.join(proj, 'script.json');
+  const script = JSON.parse(fs.readFileSync(scriptPath, 'utf8'));
+  script.slides = script.slides.filter(s => s.id === '01');
+  // 第 2 句带 say(发音分离): 显示形态 82.3% 上字幕, TTS/ASR 吃 百分之八十二点三
+  script.slides[0].clauses = [{ stage: 1, text: '冒烟测试。' }, { stage: 2, text: '第二句展开,营收增长82.3%。', say: '第二句展开,营收增长百分之八十二点三。' }];
+  fs.writeFileSync(scriptPath, JSON.stringify(script, null, 2));
+  // 4. 最小 slide(过得了 check-slides: data-stage 配 fx 类, 变量有定义)
+  fs.writeFileSync(path.join(proj, 'slides', '01-title.html'),
+    `<!doctype html><html data-theme="a"><head><meta charset="utf-8"><link rel="stylesheet" href="tokens.css"></head>
+<body><div class="stage"><span class="brand">冒烟</span>
+<h1 class="fx-rise" data-stage="1">冒烟测试</h1>
+<p class="fx-fade" data-stage="2">第二层展开内容</p>
+<div class="slide-num">01</div></div></body></html>`);
+  // 5. plan-timings
+  r = runSkill('plan-timings.mjs', [proj]);
+  assert.equal(r.status, 0, r.stdout + r.stderr);
+  const timings = JSON.parse(fs.readFileSync(path.join(proj, 'build', 'timings.json'), 'utf8'));
+  assert.ok(timings.total > 2.5 && timings.total < 5, `总时长应≈3s+tai1, 实际 ${timings.total}`);
+  // 6. check-slides(静态闸门, 必须全绿)
+  r = runSkill('check-slides.mjs', [proj]);
+  assert.equal(r.status, 0, r.stdout + r.stderr);
+  // 7. capture still(--ids 带一个不存在的 id: 部分匹配必须点名告警, 不静默少做 —— Mavis 交接 §4)
+  r = runSkill('capture.mjs', [proj, '--mode', 'still', '--ids', '01,99']);
+  assert.equal(r.status, 0, r.stdout + r.stderr);
+  assert.ok((r.stdout + r.stderr).includes('已忽略: 99'), '没命中的 id 必须点名: ' + (r.stdout + r.stderr).slice(-300));
+  assert.ok(fs.existsSync(path.join(proj, 'preview', '01.png')), '应产出 preview/01.png');
+  // 8. build-video(不带 --asr, 不碰网络)
+  r = runSkill('build-video.mjs', [proj]);
+  assert.equal(r.status, 0, r.stdout + r.stderr);
+  const final = path.join(proj, 'out', 'final.mp4');
+  assert.ok(fs.existsSync(final), '应产出 out/final.mp4');
+  const d = spawnSync(FFPROBE, ['-v', 'error', '-show_entries', 'format=duration', '-of', 'csv=p=0', final],
+    { encoding: 'utf8', windowsHide: true });
+  const dur = parseFloat((d.stdout ?? '').trim());
+  assert.ok(Math.abs(dur - timings.total) <= 0.3, `成片时长 ${dur} 应≈timings ${timings.total}`);
+  // 9. build-video --asr: 清单期望文必须取发音形态(say ?? text), 字幕/srt 恒为显示形态
+  r = runSkill('build-video.mjs', [proj, '--asr']);
+  assert.equal(r.status, 0, r.stdout + r.stderr);
+  const ck = fs.readFileSync(path.join(proj, 'asr', 'checklist.md'), 'utf8');
+  assert.match(ck, /part-01-2\.mp3[^|]*\|[^|]*\|[^|]*百分之八十二点三/, 'ASR 清单第 2 句期望文应为 say 形态: ' + ck);
+  assert.ok(!/82\.3%[^|]*\| *\| *\|/.test(ck), '清单期望文不得回退显示形态(那会把正确发音判成数字不符)');
+  const srt = fs.readFileSync(path.join(proj, 'out', 'subs.srt'), 'utf8');
+  assert.match(srt, /82\.3%/, 'srt 恒为显示形态');
+  assert.ok(!srt.includes('百分之八十二点三'), 'say 不得泄漏进 srt/字幕');
+});
+
+test('still 复截后直接 build-video: 必须点名"静态图出片"警告(still 会作废帧目录, 动画无声丢失)', async t => {
+  if (!FFMPEG || !FFPROBE) return t.skip('无 ffmpeg/ffprobe(主 CI 环境; 由 scoped smoke workflow 覆盖)');
+  playwright = await loadPackage('playwright');
+  if (!playwright) return t.skip('无 playwright');
+  try {
+    const b = await playwright.chromium.launch({ headless: true });
+    await b.close();
+  } catch {
+    return t.skip('chromium 未安装(npx playwright install chromium)');
+  }
+  const proj = tmpdir();
+  let r = runSkill('init-project.mjs', [proj, '--topic', 'StillWarn']);
+  assert.equal(r.status, 0, r.stderr);
+  const mp3 = path.join(proj, 'audio', '01.mp3');
+  const g = spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'anullsrc=r=32000:cl=mono', '-t', '2',
+    '-c:a', 'libmp3lame', '-b:a', '64k', '-y', mp3], { windowsHide: true });
+  assert.equal(g.status, 0, g.stderr ?? '');
+  const scriptPath = path.join(proj, 'script.json');
+  const script = JSON.parse(fs.readFileSync(scriptPath, 'utf8'));
+  script.slides = script.slides.filter(s => s.id === '01');
+  script.slides[0].clauses = [{ stage: 1, text: '静音。' }];
+  fs.writeFileSync(scriptPath, JSON.stringify(script, null, 2));
+  fs.writeFileSync(path.join(proj, 'slides', '01-title.html'),
+    `<!doctype html><html data-theme="a"><head><meta charset="utf-8"><link rel="stylesheet" href="tokens.css"></head>
+<body><div class="stage"><h1 class="fx-rise" data-stage="1">静音</h1></div></body></html>`);
+  assert.equal(runSkill('plan-timings.mjs', [proj]).status, 0);
+  assert.equal(runSkill('capture.mjs', [proj, '--mode', 'still']).status, 0);   // 只 still, 不 motion
+  r = runSkill('build-video.mjs', [proj]);                                       // 真编码(2.9s 静音段, 秒级; dry-run 的自检会因无产物误报)
+  assert.equal(r.status, 0, r.stdout + r.stderr);
+  assert.ok((r.stdout + r.stderr).includes('静态图出片'), '回退静态图出片时必须点名警告并给补法');
+  assert.ok((r.stdout + r.stderr).includes('--mode motion --ids 01'), '要给出补跑命令');
+});
+
+test('字幕时间轴: 动画窗之后的字幕变化必须进画面(二审 P1: 冻帧导致后半句字幕永不出现)', async t => {
+  if (!FFMPEG || !FFPROBE) return t.skip('无 ffmpeg/ffprobe(主 CI 环境; 由 scoped smoke workflow 覆盖)');
+  playwright = await loadPackage('playwright');
+  if (!playwright) return t.skip('无 playwright');
+  try { const b = await playwright.chromium.launch({ headless: true }); await b.close(); }
+  catch { return t.skip('chromium 未安装(npx playwright install chromium)'); }
+
+  const proj = tmpdir();
+  assert.equal(runSkill('init-project.mjs', [proj, '--topic', 'SubTimeline']).status, 0);
+  const mp3 = path.join(proj, 'audio', '01.mp3');
+  const g = spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'anullsrc=r=32000:cl=mono', '-t', '6',
+    '-c:a', 'libmp3lame', '-b:a', '64k', '-y', mp3], { windowsHide: true });
+  assert.equal(g.status, 0, g.stderr ?? '');
+  const scriptPath = path.join(proj, 'script.json');
+  const script = JSON.parse(fs.readFileSync(scriptPath, 'utf8'));
+  script.slides = script.slides.filter(s => s.id === '01');
+  script.slides[0].clauses = [{ stage: 1, text: 'AAAAAAAAAAAAAAAA' }, { stage: 1, text: 'BBBB' }];
+  fs.writeFileSync(scriptPath, JSON.stringify(script, null, 2));
+  // 动画 1s 内结束, 第二句开口在 3.5s —— 正落在动画窗之外(修复前这里被 tpad 冻帧吞掉)
+  fs.writeFileSync(path.join(proj, 'slides', '01-title.html'),
+    `<!doctype html><html data-theme="a"><head><meta charset="utf-8"><link rel="stylesheet" href="tokens.css"></head>
+<body><div class="stage"><h1 class="fx-rise" data-stage="1">字幕时间轴</h1></div></body></html>`);
+  assert.equal(runSkill('plan-timings.mjs', [proj]).status, 0);
+  const tj = path.join(proj, 'build', 'timings.json');
+  const timings = JSON.parse(fs.readFileSync(tj, 'utf8'));
+  timings.slides[0].clauses = [
+    { stage: 1, start: 0.3, text: 'AAAAAAAAAAAAAAAA' },
+    { stage: 1, start: 3.5, text: 'BBBB' },
+  ];
+  fs.writeFileSync(tj, JSON.stringify(timings, null, 2));
+
+  let r = runSkill('capture.mjs', [proj, '--mode', 'motion']);
+  assert.equal(r.status, 0, r.stdout + r.stderr);
+  const man = JSON.parse(fs.readFileSync(path.join(proj, 'build', 'substills', '01.json'), 'utf8'));
+  assert.ok(man.framesCover > 0 && man.framesCover < 3.5, `帧序列只该覆盖动画窗, 实际 ${man.framesCover}`);
+  const late = man.stills.filter(s => s.start >= 3.4);
+  assert.equal(late.length, 1, `第二句(3.5s 起)必须有一条"帧覆盖不到"的字幕图, 实际 ${JSON.stringify(man.stills)}`);
+  assert.ok(fs.existsSync(path.join(proj, 'build', 'substills', '01', `s${late[0].k}.png`)), '字幕图文件要真落盘');
+
+  r = runSkill('build-video.mjs', [proj]);
+  assert.equal(r.status, 0, r.stdout + r.stderr);
+
+  // 成片里第 2 句窗内的画面必须贴合第 2 句字幕图、而不是第 1 句(修复前这里冻在第 1 句 = 后半句永不出现)
+  // 判据用"抽帧成 PNG 再逐像素比": 双输入 blend 会因两条流 seek 起点不同产生假差异(实测踩过)
+  const final = path.join(proj, 'out', 'final.mp4');
+  const bd = path.join(proj, 'build', 'bandcheck');
+  fs.mkdirSync(bd, { recursive: true });
+  const crop = 'crop=1920:260:0:820';
+  const grab = (t, out) => { spawnSync(FFMPEG, ['-y', '-hide_banner', '-loglevel', 'error', '-ss', String(t), '-i', final, '-frames:v', '1', '-vf', crop, out], { windowsHide: true }); return out; };
+  const still = (k, out) => { spawnSync(FFMPEG, ['-y', '-hide_banner', '-loglevel', 'error', '-i', path.join(proj, 'build', 'substills', '01', `s${k}.png`), '-vf', crop, out], { windowsHide: true }); return out; };
+  const diff = (a, b) => {
+    const o = spawnSync(FFMPEG, ['-hide_banner', '-loglevel', 'error', '-i', a, '-i', b,
+      '-lavfi', '[0:v][1:v]blend=all_mode=difference,signalstats,metadata=print:file=-', '-frames:v', '1', '-f', 'null', '-'],
+      { encoding: 'utf8', windowsHide: true });
+    const m = /YAVG=([0-9.]+)/.exec((o.stdout || '') + (o.stderr || ''));
+    return m ? parseFloat(m[1]) : null;
+  };
+  const fLate = grab(5.0, path.join(bd, 't50.png'));
+  const sOwn = still(late[0].k, path.join(bd, 'own.png'));
+  const sOther = still(0, path.join(bd, 'other.png'));
+  const dOwn = diff(fLate, sOwn), dOther = diff(fLate, sOther);
+  assert.ok(dOwn != null && dOther != null, '应能算出字幕带差异');
+  assert.ok(dOwn < 2.0, `5.0s 的画面应贴合第 2 句字幕图, 实际差异 ${dOwn}`);
+  // 2026-09-22 重标定: 第 2 句文本改为**不同长度**。原来两句等长(A/B 各 16 字), 修复灰字幕 bug 后
+  // 两张字幕图都满亮, 纯字形差只剩 ~3.7 —— 旧阈值 4.0 其实是靠"末句静帧截在淡变上"的灰度差撑过的;
+  // 长度不同 → 底板宽度也不同, 判据与字形渲染/平台字体解耦(冻帧时该差异 ≈ 0, 依旧必红)。
+  assert.ok(dOther > 4.0, `5.0s 的画面不该还是第 1 句(冻帧), 与第 1 句字幕图的差异只有 ${dOther}`);
+});
+
+test('字幕时间轴(静态/no-fx 路径): 没有帧序列时也要逐句烧字幕', async t => {
+  if (!FFMPEG || !FFPROBE) return t.skip('无 ffmpeg/ffprobe(主 CI 环境; 由 scoped smoke workflow 覆盖)');
+  playwright = await loadPackage('playwright');
+  if (!playwright) return t.skip('无 playwright');
+  try { const b = await playwright.chromium.launch({ headless: true }); await b.close(); }
+  catch { return t.skip('chromium 未安装(npx playwright install chromium)'); }
+
+  const proj = tmpdir();
+  assert.equal(runSkill('init-project.mjs', [proj, '--topic', 'SubStatic']).status, 0);
+  const mp3 = path.join(proj, 'audio', '01.mp3');
+  spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'anullsrc=r=32000:cl=mono', '-t', '6',
+    '-c:a', 'libmp3lame', '-b:a', '64k', '-y', mp3], { windowsHide: true });
+  const scriptPath = path.join(proj, 'script.json');
+  const script = JSON.parse(fs.readFileSync(scriptPath, 'utf8'));
+  script.slides = script.slides.filter(s => s.id === '01');
+  script.slides[0].clauses = [{ stage: 1, text: 'AAAAAAAAAAAAAAAA' }, { stage: 1, text: 'BBBB' }];
+  fs.writeFileSync(scriptPath, JSON.stringify(script, null, 2));
+  // no-fx: 整张没有入场动画 → capture 走静态路径(修复前只截一帧 → 整片没有字幕)
+  fs.writeFileSync(path.join(proj, 'slides', '01-title.html'),
+    `<!doctype html><html data-theme="a" class="no-fx"><head><meta charset="utf-8"><link rel="stylesheet" href="tokens.css"></head>
+<body><div class="stage"><h1 data-stage="1">关动效也要有字幕</h1></div></body></html>`);
+  assert.equal(runSkill('plan-timings.mjs', [proj]).status, 0);
+  const tj = path.join(proj, 'build', 'timings.json');
+  const timings = JSON.parse(fs.readFileSync(tj, 'utf8'));
+  timings.slides[0].clauses = [
+    { stage: 1, start: 0.3, text: 'AAAAAAAAAAAAAAAA' },
+    { stage: 1, start: 3.5, text: 'BBBB' },
+  ];
+  fs.writeFileSync(tj, JSON.stringify(timings, null, 2));
+
+  assert.equal(runSkill('capture.mjs', [proj, '--mode', 'still']).status, 0);
+  const man = JSON.parse(fs.readFileSync(path.join(proj, 'build', 'substills', '01.json'), 'utf8'));
+  assert.equal(man.stills.length, 2, `静态路径要为每句都出字幕图, 实际 ${JSON.stringify(man.stills)}`);
+  assert.equal(man.framesCover, 0, '静态路径没有帧序列');
+
+  const r = runSkill('build-video.mjs', [proj]);
+  assert.equal(r.status, 0, r.stdout + r.stderr);   // 拼接正确性由 build-video 的时长自检背书(段长之和 ≠ 整张时长即退出 1)
+  const final = path.join(proj, 'out', 'final.mp4');
+  const bd = path.join(proj, 'build', 'bandcheck');
+  fs.mkdirSync(bd, { recursive: true });
+  const crop = 'crop=1920:260:0:820';
+  const grab = (t, out) => { spawnSync(FFMPEG, ['-y', '-hide_banner', '-loglevel', 'error', '-ss', String(t), '-i', final, '-frames:v', '1', '-vf', crop, out], { windowsHide: true }); return out; };
+  const still = (k, out) => { spawnSync(FFMPEG, ['-y', '-hide_banner', '-loglevel', 'error', '-i', path.join(proj, 'build', 'substills', '01', `s${k}.png`), '-vf', crop, out], { windowsHide: true }); return out; };
+  const diff = (a, b) => {
+    const o = spawnSync(FFMPEG, ['-hide_banner', '-loglevel', 'error', '-i', a, '-i', b,
+      '-lavfi', '[0:v][1:v]blend=all_mode=difference,signalstats,metadata=print:file=-', '-frames:v', '1', '-f', 'null', '-'],
+      { encoding: 'utf8', windowsHide: true });
+    const m = /YAVG=([0-9.]+)/.exec((o.stdout || '') + (o.stderr || ''));
+    return m ? parseFloat(m[1]) : null;
+  };
+  // 关动效时: 5.0s 的画面必须是第 2 句(修复前整片一帧字幕都没有 → 与两句都不贴合)
+  const dOwn = diff(grab(5.0, path.join(bd, 't50.png')), still(1, path.join(bd, 'own.png')));
+  const dOther = diff(grab(5.0, path.join(bd, 't50b.png')), still(0, path.join(bd, 'other.png')));
+  assert.ok(dOwn != null && dOwn < 2.0, `关动效时 5.0s 应显示第 2 句字幕, 与第 2 句字幕图差异 ${dOwn}(修复前无字幕)`);
+  // 2026-09-22 重标定: 第 2 句改不同长度(几何差), 旧等长 A/B 串的阈值靠灰字幕 bug 的灰度差撑过 —— 同上
+  assert.ok(dOther != null && dOther > 4.0, `不该显示第 1 句, 差异 ${dOther}`);
+});
+
+test('grab-frames: 按 timings 抽成片实帧(图注/字幕带那类 still 看不出的问题, 最终对账手段)', async t => {
+  if (!FFMPEG) return t.skip('无 ffmpeg(主 CI 环境; 由 scoped smoke workflow 覆盖)');
+  const proj = tmpdir();
+  fs.mkdirSync(path.join(proj, 'build'), { recursive: true });
+  fs.mkdirSync(path.join(proj, 'out'), { recursive: true });
+  fs.writeFileSync(path.join(proj, 'build', 'timings.json'), JSON.stringify({
+    fps: 30, total: 2.2,
+    slides: [{ id: '01', duration: 1.1 }, { id: '02', duration: 1.1 }],
+  }));
+  const final = path.join(proj, 'out', 'final.mp4');
+  // 片长要比 timings.total 略长: 末段抽帧点(≈总长 0.93 处)必须落在真实帧上(真实成片也如此)
+  const g = spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'testsrc=size=320x240:rate=10',
+    '-t', '2.6', '-pix_fmt', 'yuv420p', '-y', final], { windowsHide: true });
+  assert.equal(g.status, 0, g.stderr ?? '');
+
+  let r = runSkill('grab-frames.mjs', [proj]);
+  assert.equal(r.status, 0, r.stdout + r.stderr);
+  for (const id of ['01', '02']) {
+    assert.ok(fs.statSync(path.join(proj, 'build', 'introspect', `frame-${id}-end.png`)).size > 1000, `${id} 的末段帧要真抽出来`);
+  }
+  r = runSkill('grab-frames.mjs', [proj, '--ids', '02', '--both']);
+  assert.equal(r.status, 0, r.stdout + r.stderr);
+  assert.ok(fs.existsSync(path.join(proj, 'build', 'introspect', 'frame-02-mid.png')), '--both 要有中段帧');
+  assert.ok(!fs.existsSync(path.join(proj, 'build', 'introspect', 'frame-01-mid.png')), '--ids 限定只抽 02');
+
+  // 负向: 没有成片 → 退出 1 并点名
+  fs.renameSync(final, final + '.bak');
+  r = runSkill('grab-frames.mjs', [proj]);
+  assert.equal(r.status, 1);
+  assert.ok(r.stderr.includes('final.mp4'), '要说清缺什么');
+});
+
+test('受管区升级真的改变渲染: 旧 CSS 出图 → 升级 → 画面按技能当前版重出(像素级)', async t => {
+  if (!FFMPEG || !FFPROBE) return t.skip('无 ffmpeg/ffprobe(主 CI 环境; 由 scoped smoke workflow 覆盖)');
+  playwright = await loadPackage('playwright');
+  if (!playwright) return t.skip('无 playwright');
+  try { const b = await playwright.chromium.launch({ headless: true }); await b.close(); } catch {
+    return t.skip('chromium 未安装(npx playwright install chromium)');
+  }
+  const { generateTokensCss, TOKENS_REV } = await import(
+    'file://' + path.join(SCRIPTS, 'tokens-template.mjs').replace(/\\/g, '/'));
+  const { wrapTokens } = await import(
+    'file://' + path.join(SCRIPTS, 'css-kit.mjs').replace(/\\/g, '/'));
+
+  const proj = tmpdir();
+  let r = runSkill('init-project.mjs', [proj, '--topic', 'CssUpgrade']);
+  assert.equal(r.status, 0, r.stderr);
+  const g = spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'anullsrc=r=32000:cl=mono', '-t', '2',
+    '-c:a', 'libmp3lame', '-b:a', '64k', '-y', path.join(proj, 'audio', '01.mp3')], { windowsHide: true });
+  assert.equal(g.status, 0, g.stderr ?? '');
+  const scriptPath = path.join(proj, 'script.json');
+  const script = JSON.parse(fs.readFileSync(scriptPath, 'utf8'));
+  script.slides = script.slides.filter(s => s.id === '01');
+  script.slides[0].clauses = [{ stage: 1, text: '升级检查。' }];
+  fs.writeFileSync(scriptPath, JSON.stringify(script, null, 2));
+  fs.writeFileSync(path.join(proj, 'slides', '01-title.html'),
+    `<!doctype html><html data-theme="a"><head><meta charset="utf-8"><link rel="stylesheet" href="tokens.css"></head>
+<body><div class="stage"><h1 class="fx-rise" data-stage="1">升级检查</h1></div></body></html>`);
+  assert.equal(runSkill('plan-timings.mjs', [proj]).status, 0);
+  const still = () => {
+    assert.equal(runSkill('capture.mjs', [proj, '--mode', 'still', '--allow-stale-css']).status, 0);
+    const o = spawnSync(FFMPEG, ['-hide_banner', '-loglevel', 'error', '-i', path.join(proj, 'preview', '01.png'),
+      '-vf', 'signalstats,metadata=print:file=-', '-frames:v', '1', '-f', 'null', '-'], { encoding: 'utf8', windowsHide: true });
+    const m = /YAVG=([0-9.]+)/.exec((o.stdout || '') + (o.stderr || ''));
+    return m ? parseFloat(m[1]) : null;
+  };
+
+  // 当前版(init 直出)渲染一次做基准
+  const baseline = still();
+  assert.ok(baseline != null, '应能读到基准画面的平均亮度');
+
+  // 模拟"老项目": 把受管区内容换成改过一处令牌的旧版主体(画面整体变暗一点, 但仍是同构图)
+  const oldBody = generateTokensCss().replace('--bg: #FAFAF7', '--bg: #D8D2C4');
+  assert.notEqual(oldBody, generateTokensCss(), '前置: 主体内容确实改了(否则这个用例证明不了任何事)');
+  fs.writeFileSync(path.join(proj, 'slides', 'tokens.css'),
+    ':root { --accent: #111; }\n' + wrapTokens(oldBody, TOKENS_REV) + '\n');
+  assert.equal(runSkill('check-slides.mjs', [proj]).status, 1, '内容被改过的受管区必须被静态闸门拦住');
+  assert.equal(runSkill('capture.mjs', [proj]).status, 1, 'capture 入口也必须拦住(不许静默出旧画面)');
+  const stale = still();   // 显式放行才出得来
+  assert.ok(Math.abs(stale - baseline) > 0.5,
+    `旧 CSS 与当前版的画面必须真的不同(基准 ${baseline}, 旧版 ${stale}) —— 否则"升级生效"这件事无从谈起`);
+
+  // 升级: 受管区整段替换回技能当前版, 区外项目规则保留
+  const up = runSkill('init-project.mjs', [proj, '--upgrade-css']);
+  assert.equal(up.status, 0, up.stderr);
+  assert.ok(up.stdout.includes('tokens 主体'), '要说明整段生成物被替换: ' + up.stdout.slice(0, 200));
+  assert.equal(runSkill('check-slides.mjs', [proj]).status, 0, '升级后静态闸门必须恢复');
+  const after = still();
+  assert.ok(Math.abs(after - baseline) <= 0.05,
+    `升级后必须回到技能当前版的画面(基准 ${baseline}, 升级后 ${after}) —— 这条把"改了 CSS 项目里真的生效"钉死在像素上`);
+});
+
+// 2026-09-22 实测踩坑(域外反馈): 内容总高溢出可用区时, 普通 flex center 把溢出推到画布上方裁掉
+// (kicker 被顶到 y≈17px、溢出 ≈125px), 截图目检却当成"没渲染"放过。capture 的几何自检要点名数字;
+// 骨架 .layout 的 safe center + 上下 padding 让顶部永不被裁 —— 两个都钉死在这里。
+test('capture 布局几何自检: 内容溢出/顶部被裁要点名(不是"没渲染")', async t => {
+  if (!FFMPEG) return t.skip('无 ffmpeg(主 CI 环境; 由 scoped smoke workflow 覆盖)');
+  const playwright = await loadPackage('playwright');
+  if (!playwright) return t.skip('无 playwright');
+  try { const b = await playwright.chromium.launch({ headless: true }); await b.close(); } catch {
+    return t.skip('chromium 未安装(npx playwright install chromium)');
+  }
+  const proj = tmpdir();
+  assert.equal(runSkill('init-project.mjs', [proj, '--topic', 'GeoOverflow']).status, 0);
+  const g = spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'anullsrc=r=32000:cl=mono', '-t', '3',
+    '-c:a', 'libmp3lame', '-b:a', '64k', '-y', path.join(proj, 'audio', '01.mp3')], { windowsHide: true });
+  assert.equal(g.status, 0, g.stderr ?? '');
+  const sp = path.join(proj, 'script.json');
+  const s = JSON.parse(fs.readFileSync(sp, 'utf8'));
+  s.slides = [{ id: '01', html: '01-title.html', audio: '01.mp3', clauses: [{ stage: 1, text: '一句。' }] }];
+  fs.writeFileSync(sp, JSON.stringify(s, null, 2));
+  const slide = path.join(proj, 'slides', '01-title.html');
+  // 坑形态(复刻实测现场): 整块普通 center + 字幕带 padding-bottom:280px + 950px 高内容块
+  fs.writeFileSync(slide, `<!doctype html><html data-theme="a"><head><meta charset="utf-8"><link rel="stylesheet" href="tokens.css">
+<style>.layout { position:absolute; inset:0; display:flex; flex-direction:column; justify-content:center; padding:0 160px 280px; gap:32px }
+.kicker{font-size:20px}.tall{height:950px;background:var(--panel-2)}</style></head>
+<body><div class="stage"><main class="layout"><div class="kicker">小标题</div><h1>标题两行标题两行</h1><div class="tall"></div></main></div></body></html>`);
+  assert.equal(runSkill('plan-timings.mjs', [proj]).status, 0);
+  const bad = runSkill('capture.mjs', [proj, '--mode', 'still']);
+  assert.equal(bad.status, 0, bad.stdout + bad.stderr);
+  assert.match(bad.stdout + bad.stderr, /布局几何/, '溢出必须被点名: ' + (bad.stdout + bad.stderr).slice(-400));
+  assert.match(bad.stdout + bad.stderr, /溢出|贴顶/, '要给出数字方向的定位');
+  // 矮内容不许误报(03-kpi-grid 那种健康页)
+  fs.writeFileSync(slide, `<!doctype html><html data-theme="a"><head><meta charset="utf-8"><link rel="stylesheet" href="tokens.css">
+<style>.layout { position:absolute; inset:0; display:flex; flex-direction:column; justify-content:center; padding:0 160px 280px; gap:32px }</style></head>
+<body><div class="stage"><main class="layout"><h1>短标题</h1><p>一点内容</p></main></div></body></html>`);
+  const ok = runSkill('capture.mjs', [proj, '--mode', 'still']);
+  assert.equal(ok.status, 0, ok.stdout + ok.stderr);
+  assert.ok(!/布局几何/.test(ok.stdout + ok.stderr), '矮内容不该报: ' + (ok.stdout + ok.stderr).slice(-300));
+});
+
+test('骨架 .layout 必须是 safe center + 上下 padding(溢出退回顶部对齐, 顶部永不被裁)', () => {
+  const proj = tmpdir();
+  assert.equal(runSkill('init-project.mjs', [proj, '--topic', 'TplContract']).status, 0);
+  const tpl = fs.readFileSync(path.join(proj, 'slides', '_template.html'), 'utf8');
+  assert.match(tpl, /justify-content:\s*safe center/, '普通 center 溢出时会把内容推上画布裁掉(实测 kicker y≈17px)');
+  assert.match(tpl, /padding:\s*120px\s+160px\s+190px/, '上 padding 给品牌栏、下 padding 给字幕带');
+});
+
+// 2026-09-22 云沙箱反馈 + 2026-09-23 复核: 字幕"会换行"警告曾按 18 字一刀切 —— 18 是被取代的
+// 竖屏基准, 真实行宽 = tools.subLineCap(缩放带钳位, 非线性): 中文 1080≈24 字/行、1920≈33、2560≈35。
+// 折 2 行可读, 折 3 行起才警告。样例覆盖 2/3/4 行三档。
+test('plan-timings 字幕行宽: 折 2 行不报、3 行起报(样例 2/3/4 行)', async t => {
+  if (!FFMPEG) return t.skip('无 ffmpeg/ffprobe(ffprobe 实测音频)');
+  const proj = tmpdir();
+  assert.equal(runSkill('init-project.mjs', [proj, '--topic', 'SubCap']).status, 0);
+  const g = spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'anullsrc=r=32000:cl=mono', '-t', '3',
+    '-c:a', 'libmp3lame', '-b:a', '64k', '-y', path.join(proj, 'audio', '01.mp3')], { windowsHide: true });
+  assert.equal(g.status, 0, g.stderr ?? '');
+  const sp = path.join(proj, 'script.json');
+  const s = JSON.parse(fs.readFileSync(sp, 'utf8'));
+  s.width = 1920; s.height = 1080;
+  s.slides = [{ id: '01', html: '01-title.html', audio: '01.mp3', clauses: [
+    { stage: 1, text: '一'.repeat(60) },    // 60 ÷ ≈33 字/行 = 2 行 —— 可读, 不报
+    { stage: 2, text: '二'.repeat(70) },    // 70 = 3 行 —— 报
+    { stage: 3, text: '三'.repeat(100) },   // 100 = 4 行 —— 报
+  ] }];
+  fs.writeFileSync(sp, JSON.stringify(s, null, 2));
+  fs.writeFileSync(path.join(proj, 'slides', '01-title.html'),
+    '<!doctype html><html><head><meta charset="utf-8"></head><body><div class="stage"></div></body></html>');
+  const r = runSkill('plan-timings.mjs', [proj]);
+  assert.equal(r.status, 0, r.stdout + r.stderr);
+  const out = r.stdout + r.stderr;
+  assert.ok(!/第 1 句.*会折/.test(out), '折 2 行不该报(1920 宽行宽 ≈33 字/行): ' + out.slice(-300));
+  assert.match(out, /第 2 句.*会折 3 行/, '折 3 行必须报: ' + out.slice(-300));
+  assert.match(out, /第 3 句.*会折 4 行/, '折 4 行必须报: ' + out.slice(-300));
+});
+
+// 2026-09-28 用户实测: TTS 把 2026年 念成"两千零二十六年"。显示与发音分离 —— text(2026年)
+// 上字幕, say(二零二六年)喂 TTS; 闸门只钉无歧义的"四位数字+年", 有 say 即静默, say 原样传递。
+test('plan-timings 年份读法闸门: 无 say 必须点名(给逐位写法), 有 say 静默, en 不误报, say 传递到 timings', async t => {
+  if (!FFMPEG || !FFPROBE) return t.skip('无 ffmpeg/ffprobe(ffprobe 实测音频)');
+  const proj = tmpdir();
+  assert.equal(runSkill('init-project.mjs', [proj, '--topic', 'YearGate']).status, 0);
+  const g = spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'anullsrc=r=32000:cl=mono', '-t', '3',
+    '-c:a', 'libmp3lame', '-b:a', '64k', '-y', path.join(proj, 'audio', '01.mp3')], { windowsHide: true });
+  assert.equal(g.status, 0, g.stderr ?? '');
+  const sp = path.join(proj, 'script.json');
+  const s = JSON.parse(fs.readFileSync(sp, 'utf8'));
+  s.slides = [{ id: '01', html: '01-title.html', audio: '01.mp3', clauses: [
+    { stage: 1, text: '回顾2026年,七条要闻速览。' },                                   // 无 say → 必须点名
+    { stage: 2, text: '展望2027年,增长82.3%。', say: '展望二零二七年,增长百分之八十二点三。' },  // 有 say → 静默; 且 say(19) 比 text(16) 长 → 时间权重的判别样本
+    { stage: 3, text: '远溯1897年,一切开始。' },                                       // 19xx/20xx 之外的四位年份同样要报
+    { stage: 4, text: '档案编号12026年入库。' },                                        // 长数字串内部的"2026年"不是年份, 不得误报
+  ] }];
+  fs.writeFileSync(sp, JSON.stringify(s, null, 2));
+  fs.writeFileSync(path.join(proj, 'slides', '01-title.html'),
+    '<!doctype html><html><head><meta charset="utf-8"></head><body><div class="stage"></div></body></html>');
+  const r = runSkill('plan-timings.mjs', [proj]);
+  assert.equal(r.status, 0, r.stdout + r.stderr);
+  const out = r.stdout + r.stderr;
+  assert.match(out, /第 1 句.*2026年.*逐位读.*say:"二零二六年/, '无 say 的年份必须点名并给出逐位写法: ' + out.slice(-500));
+  assert.ok(!/第 2 句.*(整读|逐位)/.test(out), '有 say 的年份不得再报: ' + out.slice(-500));
+  assert.match(out, /第 3 句.*1897年.*say:"一八九七年/, '闸门不得只认 19xx/20xx —— 任意四位年份都逐位读(第 24 轮): ' + out.slice(-500));
+  assert.ok(!/第 4 句.*(整读|逐位)/.test(out), '长数字串内部的"2026年"不是年份, 数字边界必须挡住(第 25 轮): ' + out.slice(-500));
+  // say 原样传递(build-video 的 ASR 清单要消费它, 丢了清单就回退显示形态)
+  const timings = JSON.parse(fs.readFileSync(path.join(proj, 'build', 'timings.json'), 'utf8'));
+  assert.equal(timings.slides[0].clauses[1].say, '展望二零二七年,增长百分之八十二点三。');
+  assert.ok(!('say' in timings.slides[0].clauses[0]), '无 say 的句子不得凭空造出 say');
+  // 时间权重按**发音形态**分摊(第 24 轮): 用实测 tts 反推两个候选 —— 发音字数 15/19/13/12 vs 显示 15/16/13/12,
+  // 第 2 句开口只有按 say 加权才对得上; 按显示形态分摊会系统性估早(字幕窗/ASR 切分跟着偏)。
+  const tts0 = timings.slides[0].tts;
+  const expSpoken = (tts0 * 15) / (15 + 19 + 13 + 12), expText = (tts0 * 15) / (15 + 16 + 13 + 12);
+  const start2 = timings.slides[0].clauses[1].start;
+  assert.ok(Math.abs(start2 - expSpoken) < 0.02, `开口应按 say 加权≈${expSpoken.toFixed(3)}(实测 ${start2})`);
+  assert.ok(Math.abs(start2 - expText) > 0.04, `与显示形态加权(${expText.toFixed(3)})必须可区分, 否则判据失效`);
+  // 英语面不误报(英语 TTS 自己读 twenty twenty-six, 且正则本身要求"年"字)
+  s.lang = 'en';
+  s.slides[0].clauses = [{ stage: 1, text: 'Looking back at 2026, seven stories.' }];
+  fs.writeFileSync(sp, JSON.stringify(s, null, 2));
+  const r2 = runSkill('plan-timings.mjs', [proj]);
+  assert.equal(r2.status, 0, r2.stdout + r2.stderr);
+  assert.ok(!/逐位读/.test(r2.stdout + r2.stderr), '英语面不得误报年份闸门: ' + (r2.stdout + r2.stderr).slice(-300));
+});
+
+// 2026-09-28 Mavis 交接 P0: left:50% 的绝对定位可用宽度只有画布一半, max-width(设计 1400)永远
+// 够不到 —— 字幕提前一半折行(22 句实测 8 句折错), 双行侵占到距底 230px 压住图注/免责。
+// 修法 = width:max-content(subPillCss 单一来源)。判据: 30 字句必须单行且宽 >1000px。
+test('字幕胶囊: width:max-content 生效 — 30 字句单行且宽 >1000px(旧实现封顶 960 折两行)', async t => {
+  playwright = playwright || await loadPackage('playwright');
+  if (!playwright) return t.skip('无 playwright');
+  let browser;
+  try { browser = await playwright.chromium.launch({ headless: true }); } catch { return t.skip('chromium 未安装(npx playwright install chromium)'); }
+  try {
+    const { subPillCss } = await import('file://' + path.join(SCRIPTS, 'tools.mjs').split(path.sep).join('/'));
+    const page = await browser.newPage({ viewport: { width: 1920, height: 1080 } });
+    await page.setContent('<div class="stage" style="position:relative;width:1920px;height:1080px"></div>');
+    await page.evaluate(() => {
+      const el = document.documentElement;
+      el.style.setProperty('--stage-w', '1920px'); el.style.setProperty('--stage-h', '1080px'); el.style.setProperty('--sub-scale', '1');
+    });
+    await page.addStyleTag({ content: subPillCss() });
+    const g = await page.evaluate(() => {
+      const d = document.createElement('div'); d.className = 'kit-sub'; d.style.opacity = '1'; d.textContent = '一'.repeat(30);
+      document.querySelector('.stage').appendChild(d);
+      const r = d.getBoundingClientRect();
+      return { w: Math.round(r.width), h: Math.round(r.height) };
+    });
+    assert.ok(g.w > 1000, `30 字句应撑到 >1000px(设计上限 1400): 实测 ${g.w}px —— 封顶 960 说明 max-width 仍够不到`);
+    assert.ok(g.h < 100, `必须单行(高≈86px): 实测 ${g.h}px —— 146px = 折成了两行`);
+  } finally { await browser.close(); }
+});
+
+// 2026-09-23 复核钉死: 胶囊几何唯一来源 tools.subLineCap —— 缩放带钳位, 按宽度线性外推是错的
+// (1.9.9 的 18×W/1080 在 1080 过报 27%、2560 漏报真折行)。数字来自 .kit-sub 的 CSS 常量。
+test('subLineCap: 字幕胶囊行宽几何(钳位后非线性, 数字钉死)', async () => {
+  const { subLineCap, subScale } = await import('file://' + path.join(SCRIPTS, 'tools.mjs').split(path.sep).join('/'));
+  assert.equal(subScale(1080), 0.75, '1080 宽钳到下限 0.75');
+  assert.equal(subLineCap(1080), 24);          // (0.729167·1080 − 68·0.75)/(40·0.75) ≈ 24.6
+  assert.equal(subLineCap(1920), 33);          // (0.729167·1920 − 68)/40 ≈ 33.3
+  assert.equal(subLineCap(2560), 35);          // 缩放钳到 1.25 ≈ 35.6 —— 线性外推会给 42(错)
+  assert.equal(subLineCap(1080, 0.44), 55);    // 拉丁 ≈0.44em/字符
+});
+
+// 2026-09-28 第 22 轮 review 判断项收口: 上面只钉了 tools 侧数字, 若 capture 侧改回硬编码
+// (两份几何曾经漂移过、正是 1.9.10 单源化的起因), 257 例照样全绿 —— 把"消费方必须吃 SUB_GEOMETRY"
+// 钉成结构断言: 除 tools.mjs 外任何脚本不得再出现胶囊宽度字面量, capture 必须以 geom.* 注入。
+test('胶囊几何单源: 除 tools.mjs 外无字面量, capture 必须消费 SUB_GEOMETRY', () => {
+  for (const f of fs.readdirSync(SCRIPTS)) {
+    if (!f.endsWith('.mjs') || f === 'tools.mjs') continue;
+    const src = fs.readFileSync(path.join(SCRIPTS, f), 'utf8');
+    assert.ok(!/0\.729167/.test(src), `${f} 不得硬编码胶囊宽度(唯一来源 tools.SUB_GEOMETRY)`);
+  }
+  const cap = fs.readFileSync(path.join(SCRIPTS, 'capture.mjs'), 'utf8');
+  assert.match(cap, /subPillCss/, 'capture 的字幕胶囊样式必须来自 tools.subPillCss(硬编码即两份几何)');
+  assert.match(cap, /css: subPillCss\(\) \+ subtitleKeyframes/, '注入样式 = 胶囊样式 + 关键帧, 同一来源拼接');
+});
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/review-round2.test.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/review-round2.test.mjs
new file mode 100644
index 00000000..83f743e2
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/review-round2.test.mjs
@@ -0,0 +1,290 @@
+// 二审(2026-09-18)修复的回归测试: 输出收监 / 网络边界 / 属性解析 / ASR 错误计入 / ffmpeg 发现 / 路径别名
+import { test, describe } from 'node:test';
+import assert from 'node:assert/strict';
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+import http from 'node:http';
+import { spawnSync } from 'node:child_process';
+import { runSkill, runSkillAsync, mkproj, tmpdir, skill } from './helpers.mjs';
+import { assertResolvedHost, PolicyError } from '../scripts/url-policy.mjs';
+import { SCRIPTS } from './helpers.mjs';
+import { findTool, loadPackage } from '../scripts/tools.mjs';
+
+const HAS_PW = !!(await loadPackage('playwright'));
+const { injectHtmlVars, addNoFx } = await import('file://' + path.join(SCRIPTS, 'preview-page.mjs').replace(/\\/g, '/'));
+
+describe('二审 · 输出收监(项目内目录段是符号链接时, 不得写到/删到项目外)', () => {
+  // Windows 无开发者模式时 symlinkSync 会失败, 但**目录 junction 不需要管理员**(cmd mklink /J);
+  // 两条路都试, 都不行才 skip —— 这个 canary 用例必须在能建的平台上真跑
+  const linkDir = (target, linkPath) => {
+    try { fs.symlinkSync(target, linkPath, 'dir'); return true; } catch { /* 试 junction */ }
+    try {
+      const r = spawnSync('cmd', ['/c', 'mklink', '/J', linkPath, target], { encoding: 'utf8', windowsHide: true });
+      return r.status === 0 && fs.existsSync(linkPath);
+    } catch { return false; }
+  };
+  const canLink = (() => {
+    try {
+      const a = tmpdir(), b = tmpdir();
+      const ok = linkDir(b, path.join(a, 'l'));
+      fs.rmSync(a, { recursive: true, force: true });
+      return ok;
+    } catch { return false; }
+  })();
+
+  // 这两个 canary 都要靠 capture / preview-page 真跑到"收监检查"才观察得到越界拒绝 ——
+  // 而两个脚本启动第一件事就是加载 playwright: 无 playwright 时它们先退出 2,
+  // 用例要么误判"通过"(preview-page 只断言非零), 要么误报"失败"(capture 还断言报错文案)。
+  // 所以无 playwright 必须带原因跳过(与 ASR 用例同理) —— 否则在无 playwright 但可建链接的
+  // 环境(ubuntu 裸 CI)上, 这两条会把宿主主 CI 染红(终轮复查在 Spec 轴实测抓到)。
+  const CANARY_SKIP = [
+    !canLink && '当前环境建不了符号链接/junction',
+    !HAS_PW && '无 playwright(capture/preview-page 启动即需要, 观察不到收监检查; 真跑由 scoped smoke workflow 覆盖)',
+  ].filter(Boolean).join(' · ');
+  test('capture: build/frames 是指向项目外的符号链接 → 拒绝, canary 目录完好', { skip: CANARY_SKIP || undefined }, () => {
+    const proj = tmpdir();
+    const canary = tmpdir();
+    fs.writeFileSync(path.join(canary, 'keep.txt'), 'DO NOT DELETE');
+    mkproj(proj, { slides: [{ id: '01', html: '01.html', audio: '01.mp3', clauses: [{ stage: 1, text: 'x' }] }] });
+    fs.writeFileSync(path.join(proj, 'build', 'timings.json'), JSON.stringify({ fps: 30, total: 1, slides: [{ id: '01', duration: 1, stages: { 1: 0 }, clauses: [{ stage: 1, start: 0, text: 'x' }] }] }));
+    fs.writeFileSync(path.join(proj, 'slides', '01.html'), '<html><head></head><body><p class="fx-fade" data-stage="1">x</p></body></html>');
+    fs.rmSync(path.join(proj, 'build', 'frames'), { recursive: true, force: true });
+    assert.ok(linkDir(canary, path.join(proj, 'build', 'frames')), '测试前置: 链接要建成功');
+    const r = runSkill('capture.mjs', [proj, '--mode', 'motion']);
+    assert.notEqual(r.status, 0, '输出目录经符号链接越界必须拒绝');
+    assert.ok(/符号链接|越出项目目录/.test(r.stdout + r.stderr), `要点明原因, 实际: ${(r.stdout + r.stderr).slice(-200)}`);
+    assert.equal(fs.readFileSync(path.join(canary, 'keep.txt'), 'utf8'), 'DO NOT DELETE', 'canary 必须完好');
+  });
+
+  test('preview-page: preview 是指向项目外的符号链接 → 拒绝写快照', { skip: CANARY_SKIP || undefined }, () => {
+    const proj = tmpdir();
+    const canary = tmpdir();
+    mkproj(proj, { slides: [{ id: '01', html: '01.html', audio: '01.mp3', clauses: [{ stage: 1, text: 'x' }] }] });
+    fs.writeFileSync(path.join(proj, 'slides', '01.html'), '<html><head></head><body><p class="fx-fade" data-stage="1">x</p></body></html>');
+    assert.ok(linkDir(canary, path.join(proj, 'preview')), '测试前置: 链接要建成功');
+    const r = runSkill('preview-page.mjs', [proj]);
+    assert.notEqual(r.status, 0);
+    assert.deepEqual(fs.readdirSync(canary), [], 'canary 目录里不得出现任何文件');
+  });
+
+  test('preview-page: outDir 里的叶子(index.html)是文件符号链接 → 拒绝, canary 文件完好(1.7.5 复查①)', { skip: CANARY_SKIP || undefined }, (t) => {
+    // 目录收监了不等于叶子安全: 预置在 preview/play/ 里的**文件**符号链接会把 writeFileSync 引到
+    // 项目外。文件符号链接不需要管理员, 但需要 Windows 开发者模式 —— 建不了就带因跳过(ubuntu CI 真跑)。
+    const proj = tmpdir();
+    const canaryFile = path.join(tmpdir(), 'victim.html');
+    fs.writeFileSync(canaryFile, 'ORIGINAL');
+    mkproj(proj, { slides: [{ id: '01', html: '01.html', audio: '01.mp3', clauses: [{ stage: 1, text: 'x' }] }] });
+    fs.writeFileSync(path.join(proj, 'slides', '01.html'), '<html><head></head><body><p class="fx-fade" data-stage="1">x</p></body></html>');
+    fs.mkdirSync(path.join(proj, 'preview', 'play'), { recursive: true });
+    let made = false;
+    try { fs.symlinkSync(canaryFile, path.join(proj, 'preview', 'play', 'index.html'), 'file'); made = true; } catch { /* 需要开发者模式 */ }
+    if (!made) return t.skip('当前环境建不了文件符号链接(Windows 需开发者模式; ubuntu CI 真跑)');
+    const r = runSkill('preview-page.mjs', [proj]);
+    assert.notEqual(r.status, 0, `叶子符号链接必须拒绝, 实际: ${(r.stdout + r.stderr).slice(-200)}`);
+    assert.equal(fs.readFileSync(canaryFile, 'utf8'), 'ORIGINAL', 'canary 文件不得被改写');
+
+    // 场景②: 逐张叶子(01.html)同样收监 —— 三个叶子写点中 index.html 之外的路径(十三轮 review 指出
+    // canary 只埋了 index 一个)。独立项目, 埋 01.html, expect 拒绝且 canary 完好。
+    const proj2 = tmpdir();
+    const canary2 = path.join(tmpdir(), 'victim2.html');
+    fs.writeFileSync(canary2, 'ORIGINAL');
+    mkproj(proj2, { slides: [{ id: '01', html: '01.html', audio: '01.mp3', clauses: [{ stage: 1, text: 'x' }] }] });
+    fs.writeFileSync(path.join(proj2, 'slides', '01.html'), '<html><head></head><body><p class="fx-fade" data-stage="1">x</p></body></html>');
+    fs.mkdirSync(path.join(proj2, 'preview', 'play'), { recursive: true });
+    try { fs.symlinkSync(canary2, path.join(proj2, 'preview', 'play', '01.html'), 'file'); } catch {
+      return t.skip('当前环境建不了文件符号链接(Windows 需开发者模式; ubuntu CI 真跑)');
+    }
+    const r2 = runSkill('preview-page.mjs', [proj2]);
+    assert.notEqual(r2.status, 0, `逐张叶子符号链接必须拒绝, 实际: ${(r2.stdout + r2.stderr).slice(-200)}`);
+    assert.equal(fs.readFileSync(canary2, 'utf8'), 'ORIGINAL', 'canary 文件不得被改写');
+  });
+
+  test('grab-frames: introspect 叶子是文件符号链接 → 拒绝, canary 完好(九轮 review 同类收尾)', { skip: !findTool('ffmpeg') && '无 ffmpeg(需从成片抽帧; 真跑由 scoped smoke workflow 覆盖)' }, (t) => {
+    // 与 preview-page 叶子同类: 目录收监了, 但 frame-<id>-<tag>.png 若是预置的文件符号链接,
+    // ffmpeg -y 会顺着写出去。文件符号链接需要 Windows 开发者模式 —— 建不了带因跳过, CI 真跑。
+    const proj = tmpdir();
+    const canaryFile = path.join(tmpdir(), 'victim.png');
+    fs.writeFileSync(canaryFile, 'ORIGINAL');
+    mkproj(proj, { slides: [{ id: '01', html: '01.html', audio: '01.mp3', clauses: [{ stage: 1, text: 'x' }] }] });
+    fs.writeFileSync(path.join(proj, 'slides', '01.html'), '<html><head></head><body><p class="fx-fade" data-stage="1">x</p></body></html>');
+    fs.mkdirSync(path.join(proj, 'build'), { recursive: true });
+    fs.writeFileSync(path.join(proj, 'build', 'timings.json'), JSON.stringify(
+      { fps: 30, total: 1, slides: [{ id: '01', duration: 1, stages: { 1: 0 }, clauses: [{ stage: 1, start: 0, text: 'x' }] }] }));
+    const FFMPEG = findTool('ffmpeg');
+    fs.mkdirSync(path.join(proj, 'out'), { recursive: true });
+    assert.equal(spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'color=c=0x808080:s=640x360:r=30', '-t', '1',
+      '-c:v', 'libx264', '-pix_fmt', 'yuv420p', '-y', path.join(proj, 'out', 'final.mp4')],
+      { windowsHide: true }).status, 0, '应能造出占位成片');
+    fs.mkdirSync(path.join(proj, 'build', 'introspect'), { recursive: true });
+    let made = false;
+    try { fs.symlinkSync(canaryFile, path.join(proj, 'build', 'introspect', 'frame-01-end.png'), 'file'); made = true; } catch { /* 需要开发者模式 */ }
+    if (!made) return t.skip('当前环境建不了文件符号链接(Windows 需开发者模式; ubuntu CI 真跑)');
+    const r = runSkill('grab-frames.mjs', [proj]);
+    assert.notEqual(r.status, 0, `叶子符号链接必须拒绝, 实际: ${(r.stdout + r.stderr).slice(-200)}`);
+    assert.equal(fs.readFileSync(canaryFile, 'utf8'), 'ORIGINAL', 'canary 文件不得被 ffmpeg 改写');
+  });
+
+  test('asr 超限转码的临时文件不落输入目录(十轮 review: 同类最后一处)', { skip: !findTool('ffmpeg') && '无 ffmpeg(需真转码 500s+ 音频)' }, async (t) => {
+    // 旧实现把 .asr-<名>.16k.mp3 写在输入文件旁边 —— 输入在项目内时, 预置的同名文件符号链接
+    // 会被 ffmpeg -y 写穿。现在临时文件进 os.tmpdir() 专属目录: 输入目录里不得出现任何转码产物。
+    const FFMPEG = findTool('ffmpeg');
+    const proj = tmpdir();
+    mkproj(proj, { slides: [] });
+    fs.mkdirSync(path.join(proj, 'audio'), { recursive: true });
+    const big = path.join(proj, 'audio', 'long.mp3');
+    assert.equal(spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'anullsrc=r=32000:cl=mono',
+      '-t', '501', '-c:a', 'libmp3lame', '-b:a', '32k', '-y', big], { windowsHide: true }).status, 0,
+      '应能造出 501s 音频(超过 500s 上限 → 触发转码)');
+    const canaryFile = path.join(tmpdir(), 'victim.mp3');
+    fs.writeFileSync(canaryFile, 'ORIGINAL');
+    const planted = path.join(proj, 'audio', '.asr-long.16k.mp3');
+    let made = false;
+    try { fs.symlinkSync(canaryFile, planted, 'file'); made = true; } catch { /* 需要开发者模式 */ }
+    if (!made) return t.skip('当前环境建不了文件符号链接(Windows 需开发者模式; ubuntu CI 真跑)');
+    // 注意: planted 本身就是个存在的路径(链接), 不能用 existsSync 断言"没出现"(必真) ——
+    // 用目录前后 diff 断言"没有**新**条目"(1.7.7 canary 的这个断言被十一轮 review 抓到不可满足)。
+    const before = new Set(fs.readdirSync(path.join(proj, 'audio')));
+    const r = await runSkillAsync('asr.mjs', [proj, '--file', big, '--provider', 'api',
+      '--allow-any-endpoint', '--base-url', 'http://127.0.0.1:9'],   // 转码后请求死端口 → 快速失败, 不碰真网
+      { env: { ...process.env, MINIMAX_API_KEY: 'fake-key-for-test' } });
+    assert.notEqual(r.status, 0, '死端口请求必须失败(转码在前, 网络在后)');
+    const newEntries = fs.readdirSync(path.join(proj, 'audio')).filter(f => !before.has(f));
+    assert.deepEqual(newEntries, [], `输入目录不得出现转码产物: ${newEntries.join(', ')}`);
+    assert.equal(fs.readFileSync(canaryFile, 'utf8'), 'ORIGINAL', 'canary 文件不得被写穿');
+    assert.equal(fs.lstatSync(planted).isSymbolicLink(), true, '预置链接不得被替换成普通文件');
+    assert.deepEqual(fs.readdirSync(os.tmpdir()).filter(f => f.startsWith('asr-16k-')), [],
+      'os.tmpdir() 的转码临时目录必须已清理(asr 的 finally 整目录删除)');
+  });
+});
+
+describe('二审 · 网络边界: 拒绝的目标必须一个请求都收不到', () => {
+  test('被策略拒绝的目标: 本地服务器收到的请求数必须是 0', async () => {
+    let hits = 0;
+    const srv = http.createServer((req, res) => { hits++; res.end('x'); });
+    await new Promise(r => srv.listen(0, '127.0.0.1', r));
+    const port = srv.address().port;
+    try {
+      // --url 直下模式(不需要 playwright): 目标是 loopback → 策略层就该拦住
+      const r = await runSkillAsync('fetch-official-images.mjs', ['--url', `http://127.0.0.1:${port}/x.png`, '--out-dir', 'assets'], { cwd: tmpdir() });
+      assert.equal(hits, 0, '被拒目标不得收到任何请求(闸门必须在连接之前)');
+      assert.notEqual(r.status, 0);
+      // 公网域名形式但解析回环(rebinding): 字符串层放行, DNS 层必须拦
+      const r2 = await runSkillAsync('fetch-official-images.mjs', ['--url', `http://localtest.me:${port}/x.png`, '--out-dir', 'assets'], { cwd: tmpdir() });
+      if (/localtest\.me/.test(r2.stdout + r2.stderr)) {           // 解析成功才可能有结论
+        assert.equal(hits, 0, 'DNS 解析回环也必须拦在连接之前');
+        assert.notEqual(r2.status, 0);
+      }
+    } finally { srv.close(); }
+  });
+
+  test('assertResolvedHost: 注入假解析器即可测(不需要真 DNS)', async () => {
+    const lookup = async () => [{ address: '127.0.0.1', family: 4 }];
+    await assert.rejects(() => assertResolvedHost('https://cdn.example.com/a.png', { lookup }), (e) => e instanceof PolicyError);
+    const publicLookup = async () => [{ address: '93.184.216.34', family: 4 }];
+    await assert.doesNotReject(() => assertResolvedHost('https://cdn.example.com/a.png', { lookup: publicLookup }));
+    // 云元数据地址同样要拦
+    await assert.rejects(() => assertResolvedHost('https://evil.example.com/', { lookup: async () => [{ address: '169.254.169.254', family: 4 }] }));
+    // 裸主机名(无点)本就被字符串层拦, 这里直接放行不查
+    await assert.doesNotReject(() => assertResolvedHost('http://localhost/', { lookup }));
+  });
+});
+
+describe('二审 · 放映页属性解析(单引号/无引号不得产生重复属性)', () => {
+  test('单引号 style/class: 就地改写, 不得出现重复属性', () => {
+    const html = `<html lang="zh" class='theme-a' style='--t2:800ms'>`;
+    const out = injectHtmlVars(html, '--t2:4866ms');
+    assert.equal((out.match(/style=/g) || []).length, 1, `不得出现第二个 style 属性: ${out}`);
+    assert.ok(out.includes('--t2:800ms;--t2:4866ms'), `应就地在原值后追加: ${out}`);
+    const nofx = addNoFx(html);
+    assert.equal((nofx.match(/class=/g) || []).length, 1, `不得出现第二个 class 属性: ${nofx}`);
+    assert.ok(/class='theme-a no-fx'|class="theme-a no-fx"/.test(nofx), `应在原类后追加: ${nofx}`);
+  });
+
+  test('无引号属性与已有 no-fx: 也要认出来', () => {
+    const out = injectHtmlVars('<html class=deck style=color:red>', '--t1:0ms');
+    assert.equal((out.match(/style=/g) || []).length, 1, out);
+    assert.ok(out.includes('color:red;--t1:0ms'), out);
+    assert.equal(addNoFx('<html class=deck>'), '<html class="deck no-fx">');
+    const once = addNoFx('<html class="deck no-fx">');
+    assert.equal(addNoFx(once), once, '已带 no-fx 时不重复添加');
+    assert.equal(addNoFx(`<html class='deck no-fx'>`).match(/no-fx/g).length, 1);
+  });
+});
+
+describe('二审 · ASR 请求错误必须算失败并非零退出', () => {
+  // asr.mjs 启动第一件事就是 requireTool('ffmpeg')(切音频要用), 无 ffmpeg 时走不到请求那一步。
+  // 主仓 CI(npm run check → 根目录 node --test)没有 ffmpeg —— 这条例外必须像其它渲染用例一样
+  // 带原因跳过, 否则整个插件把主仓 CI 染红(第三轮复查实测踩到)。
+  test('转写请求全失败(503) → 退出码非零且不计入通过', { skip: !findTool('ffmpeg') && '无 ffmpeg(asr 启动即需要; 真跑由 scoped smoke workflow 覆盖)' }, async () => {
+    // 本地假服务器: 一律 503(不需要真 Key, 也不需要真 ASR)
+    const srv = http.createServer((req, res) => { res.statusCode = 503; res.end('{"error":"busy"}'); });
+    await new Promise(r => srv.listen(0, '127.0.0.1', r));
+    const port = srv.address().port;
+    const proj = tmpdir();
+    mkproj(proj, { slides: [{ id: '01', audio: '01.mp3', clauses: [{ stage: 1, text: '一。' }] }] });
+    // asr.mjs 读的是 build-video --asr 切好的 asr/part-*.mp3, 这里直接造一个占位文件
+    fs.mkdirSync(path.join(proj, 'asr'), { recursive: true });
+    fs.writeFileSync(path.join(proj, 'asr', 'part-01-1.mp3'), 'not-really-mp3');
+    try {
+      const r = await runSkillAsync('asr.mjs', [proj, '--provider', 'api', '--allow-any-endpoint', '--base-url', `http://127.0.0.1:${port}`], {
+        cwd: proj, env: { ...process.env, MINIMAX_API_KEY: 'fake-key-for-test', MINIMAX_REGION: 'cn' },
+      });
+      assert.notEqual(r.status, 0, `请求失败不能被当成通过: ${r.stdout}${r.stderr}`);
+      assert.match(r.stdout + r.stderr, /请求失败/, '要点名"请求失败"而不是混进通过数');
+    } finally { srv.close(); }
+  });
+
+  test('asr --out: 越出项目目录 / 已存在不覆盖 → 在任何网络请求之前拒绝(1.7.5 复查③)', () => {
+    // --out 校验前置到 transcribe 之前, 所以这两条拒绝路径不需要网络也不需要 ffmpeg
+    const proj = tmpdir();
+    const outside = path.dirname(proj);
+    mkproj(proj, { slides: [] });
+    fs.writeFileSync(path.join(proj, 'asr-src.mp3'), 'stub-audio');
+    fs.writeFileSync(path.join(proj, 'exists.srt'), 'OLD');
+    const env = { ...process.env, MINIMAX_API_KEY: 'fake-key-for-test' };
+    // ① 相对路径穿越
+    const r1 = runSkill('asr.mjs', [proj, '--file', path.join(proj, 'asr-src.mp3'), '--out', '../escaped.txt'], { env });
+    assert.notEqual(r1.status, 0, '穿越项目目录的 --out 必须被拒');
+    assert.match(r1.stdout + r1.stderr, /越出项目目录|符号链接/, r1.stdout + r1.stderr);
+    assert.equal(fs.existsSync(path.join(outside, 'escaped.txt')), false, '拒绝必须发生在写入之前');
+    // ② 绝对路径
+    const r2 = runSkill('asr.mjs', [proj, '--file', path.join(proj, 'asr-src.mp3'), '--out', path.join(outside, 'abs.txt')], { env });
+    assert.notEqual(r2.status, 0, '绝对路径 --out 必须被拒');
+    assert.match(r2.stdout + r2.stderr, /相对路径/, r2.stdout + r2.stderr);
+    // ③ 项目内已存在 → 不覆盖
+    const r3 = runSkill('asr.mjs', [proj, '--file', path.join(proj, 'asr-src.mp3'), '--out', 'exists.srt'], { env });
+    assert.notEqual(r3.status, 0, '已存在的 --out 必须先拒绝');
+    assert.match(r3.stdout + r3.stderr, /不覆盖/, r3.stdout + r3.stderr);
+    assert.equal(fs.readFileSync(path.join(proj, 'exists.srt'), 'utf8'), 'OLD', '旧内容不得被改写');
+  });
+});
+
+describe('二审 · ffmpeg 发现(项目内 node_modules)', () => {
+  test('findTool: 从项目目录调用时能找到项目内的 ffmpeg-static', async () => {
+    const { findTool } = await import('../scripts/tools.mjs');
+    const proj = tmpdir();
+    const exe = process.platform === 'win32' ? 'ffmpeg.exe' : 'ffmpeg';
+    const dir = path.join(proj, 'node_modules', 'ffmpeg-static');
+    fs.mkdirSync(dir, { recursive: true });
+    fs.writeFileSync(path.join(dir, exe), 'stub');
+    // 模拟"在项目目录里执行": KIT_PROJECT_DIR 未设, 靠 cwd 找到
+    const prev = process.env.KIT_PROJECT_DIR;
+    delete process.env.KIT_PROJECT_DIR;
+    const cwd = process.cwd();
+    try {
+      process.chdir(proj);
+      assert.ok(findTool('ffmpeg', process.cwd()), '应能在项目目录里发现 ffmpeg-static');
+    } finally {
+      process.chdir(cwd);
+      if (prev !== undefined) process.env.KIT_PROJECT_DIR = prev;
+    }
+  });
+
+  test('prep-image / asr 会把项目目录交给工具发现(源码级断言, 防回退)', () => {
+    const pre = fs.readFileSync(skill('prep-image.mjs'), 'utf8');
+    assert.match(pre, /requireTool\('ffmpeg',\s*process\.cwd\(\)\)/, 'prep-image 必须把项目目录(此处为 cwd)传给 requireTool');
+    const asr = fs.readFileSync(skill('asr.mjs'), 'utf8');
+    assert.match(asr, /requireTool\('ffmpeg',\s*dir\)/, 'asr 必须把项目目录传给 requireTool');
+  });
+});
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/review-round3.test.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/review-round3.test.mjs
new file mode 100644
index 00000000..a54ff152
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/review-round3.test.mjs
@@ -0,0 +1,521 @@
+// 第三轮 review(外部评审 + /code-review)修复的回归: D1–D6 功能缺陷 + 陈旧闸门
+// 每条都对应一个"能静默产生错误产物"的具体路径, 不是风格问题:
+//   D1 check-timing --calibrate 混着写回半份实测 + 谎报匹配方法
+//   D2 `--topic X <项目目录>` 把 X 当项目目录, 在 cwd 静默建骨架
+//   D3 fetch-official-images 的 --json 吃掉位置参数 / --min 只比宽度 / 全失败还退 0
+//   D4 capture --mode/--dsf 与 grab-frames --at 的非法值一路带到 playwright/ffmpeg
+//   D5 check-slides 字幕带闸门误报 padding-bottom(且不要求绝对定位)
+//   D6 transition 裸字符串在闸门与渲染器里结论不一致
+import { test, describe } from 'node:test';
+import assert from 'node:assert/strict';
+import fs from 'node:fs';
+import path from 'node:path';
+import { spawnSync } from 'node:child_process';
+import { runSkill, mkproj, tmpdir, SCRIPTS, LEGACY_TOKENS } from './helpers.mjs';
+import { pathToFileURL } from 'node:url';
+import { findTool, loadPackage } from '../scripts/tools.mjs';
+
+const FFMPEG = findTool('ffmpeg');
+const { positionals, readTransition, positionalDir, VALUE_FLAGS, matchBoundaries } = await import(
+  pathToFileURL(path.join(SCRIPTS, 'tools.mjs')).href);
+
+const SLIDE = `<!doctype html><html data-theme="a"><head><meta charset="utf-8"><link rel="stylesheet" href="tokens.css"></head>
+<body><div class="stage"><h1 class="fx-rise" data-stage="1">标题</h1></div></body></html>`;
+
+describe('第 25 轮 · matchBoundaries 对齐用发音形态(say ?? text)', () => {
+  test('say/text 标点不一致时精确模式按 say 对齐; 按 text 会降级成 nearest(判据可区分)', () => {
+    const gaps = [1.0, 2.0, 3.0, 4.0].map(end => ({ end }));
+    const clauses = [
+      { start: 0, text: '开场白,要点。', say: '开场白,要点,补充。' },   // 标点: text 2 / say 3
+      { start: 4, text: '结尾。', say: '结尾,致谢。' },                 // 标点: text 1 / say 2
+    ];
+    const r = matchBoundaries(gaps, clauses);
+    // say 总标点 5, gaps=4 = 5-1 → 精确模式成立; 第 0 边界 = gaps[say 标点累积(3)-1] = gaps[2] = 3.0
+    // 按 text(总 3)则精确模式不成立 → nearest, meas[0]=4.0 —— 两个结果形状不同, 断言可区分
+    assert.equal(r.method, 'exact', '应按 say 的标点进精确模式');
+    assert.equal(r.meas[0], 3.0, '边界应对齐 say 标点累积对应的静音段');
+  });
+});
+
+describe('D1 · check-timing --calibrate 不写回半份实测', () => {
+  // 造一段"两句话 + 句间静音"的音频: 0–1.0s 噪声(第 1 句) → 1.0–1.6s 静音 → 1.6–2.6s 噪声(第 2 句)
+  // (silencedetect 阈值为 -38dB/0.18s, 所以句间用真静音、句子用可闻噪声)
+  const makeAudio = (proj) => {
+    const out = path.join(proj, 'audio', '01.mp3');
+    const g = spawnSync(FFMPEG, ['-v', 'error',
+      '-f', 'lavfi', '-i', 'anoisesrc=d=1.0:c=pink:a=0.4:r=32000',
+      '-f', 'lavfi', '-i', 'anullsrc=r=32000:cl=mono:d=0.6',
+      '-f', 'lavfi', '-i', 'anoisesrc=d=1.0:c=pink:a=0.4:r=32000',
+      '-filter_complex', '[0:a][1:a][2:a]concat=n=3:v=0:a=1[out]', '-map', '[out]',
+      '-c:a', 'libmp3lame', '-b:a', '64k', '-y', out], { windowsHide: true });
+    assert.equal(g.status, 0, g.stderr ?? '');
+    return out;
+  };
+  const mk = (clauses) => {
+    const proj = mkproj(tmpdir(), {
+      slides: [{ id: '01', html: '01.html', audio: '01.mp3', clauses }],
+    });
+    fs.writeFileSync(path.join(proj, 'slides', '01.html'), SLIDE);
+    fs.writeFileSync(path.join(proj, 'script.json'), JSON.stringify({
+      topic: 't', fps: 30, width: 1920, height: 1080,
+      slides: [{ id: '01', html: '01.html', audio: '01.mp3', title: '开场', clauses }],
+    }, null, 2));
+    return proj;
+  };
+
+  test('静音段不足(有缺口) → 报 sparse, 一个字节都不写回', { skip: !FFMPEG && '无 ffmpeg' }, () => {
+    const clauses = [
+      { stage: 1, start: 0, text: '第一句。' },
+      { stage: 1, start: 1.2, text: '第二句。' },
+      { stage: 1, start: 2.2, text: '第三句。' },   // 需要 2 个句边界, 音频里只有 1 个静音段
+    ];
+    const proj = mk(clauses);
+    makeAudio(proj);
+    const tp = path.join(proj, 'build', 'timings.json');
+    fs.mkdirSync(path.dirname(tp), { recursive: true });
+    fs.writeFileSync(tp, JSON.stringify({ fps: 30, total: 2.6, tts: 2.6, slides: [{ id: '01', tts: 2.6, duration: 2.6, clauses }] }, null, 2));
+    const before = fs.readFileSync(tp, 'utf8');
+    const r = runSkill('check-timing.mjs', [proj, '--calibrate']);
+    assert.equal(r.status, 0, r.stderr);
+    assert.match(r.stdout, /静音段不足\(未写回\)/, '要如实报 sparse: ' + r.stdout);
+    assert.equal(fs.readFileSync(tp, 'utf8'), before, '半份实测不得写回 timings.json');
+  });
+
+  test('全命中 → 写回, 且每个值都来自实测(不是估算)', { skip: !FFMPEG && '无 ffmpeg' }, () => {
+    const clauses = [
+      { stage: 1, start: 0, text: '第一句。' },
+      { stage: 2, start: 1.0, text: '第二句。' },
+    ];
+    const proj = mk(clauses);
+    makeAudio(proj);
+    const tp = path.join(proj, 'build', 'timings.json');
+    fs.mkdirSync(path.dirname(tp), { recursive: true });
+    fs.writeFileSync(tp, JSON.stringify({ fps: 30, total: 2.6, tts: 2.6, slides: [{ id: '01', tts: 2.6, duration: 2.6, clauses }] }, null, 2));
+    const r = runSkill('check-timing.mjs', [proj, '--calibrate']);
+    assert.equal(r.status, 0, r.stderr);
+    assert.match(r.stdout, /(精确\(标点对齐\)|最近邻\(全命中\))|已写回/, r.stdout);
+    const after = JSON.parse(fs.readFileSync(tp, 'utf8'));
+    const c2 = after.slides[0].clauses[1];
+    assert.ok(Math.abs(c2.start - 1.6) < 0.35, `第 2 句开口应≈1.6s(实测静音段末端), 实际 ${c2.start}(${r.stdout.slice(-200)})`);
+    assert.notEqual(c2.start, 1.0, '不能原样留着估算值');
+    assert.ok(after.slides[0].stages && Object.keys(after.slides[0].stages).length > 0, 'stage 时刻也要写回');
+  });
+});
+
+describe('D2 · 位置参数解析: 取值型 flag 的值不是项目目录', () => {
+  test('positionalDir 跳过取值型 flag 的值', () => {
+    assert.ok(VALUE_FLAGS.has('--topic'));
+    assert.equal(positionalDir(['--topic', '我的主题', './proj']), path.resolve('./proj'));
+    assert.equal(positionalDir(['./proj', '--topic', 'x']), path.resolve('./proj'));
+    assert.equal(positionalDir(['--ids', '01,03', './proj']), path.resolve('./proj'));
+    assert.equal(positionalDir(['--check-css']), path.resolve('.'));
+  });
+
+  test('init-project --topic X <项目目录> → 只在目标目录建骨架, cwd 不得被污染', () => {
+    const cwd = tmpdir();
+    const proj = path.join(cwd, 'proj');
+    const r = runSkill('init-project.mjs', ['--topic', '我的主题', proj], { cwd });
+    assert.equal(r.status, 0, r.stderr);
+    assert.ok(fs.existsSync(path.join(proj, 'slides', 'tokens.css')), '目标目录要有骨架');
+    assert.deepEqual(fs.readdirSync(cwd), ['proj'], `cwd 只能多出 proj/, 实际 ${fs.readdirSync(cwd).join(',')}`);
+    const tp = fs.readFileSync(path.join(proj, 'script.json'), 'utf8');
+    assert.ok(tp.includes('我的主题'), '主题名要写进 script.json: ' + tp.slice(0, 200));
+  });
+
+  test('--topic 写在项目目录之后也照样生效', () => {
+    const cwd = tmpdir();
+    const proj = path.join(cwd, 'p2');
+    const r = runSkill('init-project.mjs', [proj, '--topic', 'V2'], { cwd });
+    assert.equal(r.status, 0, r.stderr);
+    assert.ok(fs.readFileSync(path.join(proj, 'script.json'), 'utf8').includes('V2'));
+    assert.deepEqual(fs.readdirSync(cwd), ['p2']);
+  });
+});
+
+describe('D3 · fetch-official-images 的参数与退出码', () => {
+  const proj = () => { const d = tmpdir(); fs.mkdirSync(path.join(d, 'assets'), { recursive: true }); return d; };
+
+  test('--json 是布尔 flag: 放在网址前后都不吃掉位置参数', () => {
+    const d = proj();
+    for (const args of [['--json', 'http://127.0.0.1/x'], ['http://127.0.0.1/x', '--json']]) {
+      const r = runSkill('fetch-official-images.mjs', args, { cwd: d });
+      // 两种情况都必须走到"网址校验"(内网被拒), 而不是报用法错误
+      assert.ok(!r.stderr.includes('用法:'), `--json 位置 ${args[0]} 不该导致缺位置参数: ${r.stderr}`);
+      assert.equal(r.status, 1);
+      assert.ok(r.stderr.includes('内网') || r.stderr.includes('拒绝') || r.stderr.includes('127.0.0.1'), r.stderr);
+    }
+  });
+
+  // 终轮复查抓到: --json 在末尾(文档写的用法)时, flagValue 取"下一个参数"得到 undefined,
+  // 分支为假 → 静默退化成人类可读列表。上面的用例只断言"走到了网址校验",
+  // 观察不到这个退化 —— 必须端到端断言真的输出了 JSON。
+  test('--json 真的输出 JSON(两种位置都对 file:// 夹具页生效)', async t => {
+    const playwright = await loadPackage('playwright');
+    if (!playwright) return t.skip('无 playwright(列表路径要开浏览器; 真跑由 scoped smoke workflow 覆盖)');
+    try { const b = await playwright.chromium.launch({ headless: true }); await b.close(); } catch {
+      return t.skip('chromium 未安装(npx playwright install chromium)');
+    }
+    const d = proj();
+    fs.writeFileSync(path.join(d, 'pic.png'),
+      Buffer.from('iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mP8z8BQDwAEhQGAhKmMIQAAAABJRU5ErkJggg==', 'base64'));
+    fs.writeFileSync(path.join(d, 'page.html'),
+      '<!doctype html><html><body><img src="pic.png" alt="fixture"></body></html>');
+    const pageUrl = pathToFileURL(path.join(d, 'page.html')).href;   // 临时路径含空格/非 ASCII 也安全
+    for (const args of [[pageUrl, '--allow-file', '--json'], ['--json', pageUrl, '--allow-file']]) {
+      const r = runSkill('fetch-official-images.mjs', args, { cwd: d });
+      assert.equal(r.status, 0, `--json 位置 ${args[0]}: ${r.stdout}${r.stderr}`);
+      let arr = null;
+      try { arr = JSON.parse(r.stdout); } catch { /* 下面断言报出实际输出 */ }
+      assert.ok(Array.isArray(arr) && arr.length >= 1 && arr[0].src,
+        `--json 必须输出候选 JSON 数组, 实际: ${(r.stdout || r.stderr).slice(0, 200)}`);
+      assert.ok(!r.stdout.includes('候选图'), 'JSON 模式不得再混人类可读列表');
+    }
+  });
+
+  test('--min 非法写法 → 退出 1 并说清支持的形式', () => {
+    const d = proj();
+    for (const bad of ['abc', '800x', 'x', '']) {
+      const r = runSkill('fetch-official-images.mjs', ['https://example.com/', '--min', bad], { cwd: d });
+      assert.equal(r.status, 1, `--min ${JSON.stringify(bad)} 必须被拒`);
+      assert.ok(r.stderr.includes('--min'), r.stderr);
+    }
+  });
+
+  test('--url 全部失败 → 退出 1(不再"全失败还退 0")', () => {
+    const d = proj();
+    const r = runSkill('fetch-official-images.mjs',
+      ['--url', 'http://127.0.0.1/a.png,http://169.254.169.254/b.png', '--out-dir', path.join(d, 'assets')], { cwd: d });
+    assert.equal(r.status, 1);
+    assert.deepEqual(fs.readdirSync(path.join(d, 'assets')), [], '拒绝时不得落盘');
+  });
+
+  test('--url 重定向循环守卫: 只存在于 followRedirects 单一实现(--url 与 --get 共用; 十三轮 review 去重)', () => {
+    // 行为级环测试无法本地构造: --url 路径的 SSRF 策略拒绝 loopback(设计使然, 没有测试例外),
+    // 公网 URL 在 CI 不可靠。退而求其次钉住结构: 逐跳循环只有一份, 环守卫/0 字节守卫在共享核心里,
+    // 两条下载路径(--url 直下与 --get 浏览器)都必须走它 —— 第二份拷贝再出现就红。
+    const src = fs.readFileSync(path.join(SCRIPTS, 'fetch-official-images.mjs'), 'utf8');
+    assert.equal((src.match(/for \(let hop = 0/g) || []).length, 1, '逐跳循环必须只有一份实现');
+    assert.match(src, /async function followRedirects\(/, '共享核心必须存在');
+    assert.match(src, /seen\.has\(next\)/, '环守卫(ping-pong)必须在共享核心里');
+    assert.match(src, /await followRedirects\(c\.src/, '--get(浏览器)路径必须走共享核心');
+    assert.match(src, /await followRedirects\(src, u => policyGet\(/, '--url 直下路径必须走共享核心(1.9.11 起出网走 policyGet, 建连期否决)');
+    assert.equal((src.match(/policyGet\(u, \{/g) || []).length, 2, '--url 与 --get 两条下载路径的出网都必须走 policyGet(连接期绑定, 不许退回 fetch/context.request)');
+    assert.equal((src.match(/响应为空\(0 字节\)/g) || []).length, 2, '--url 与 --get 两条下载路径必须各有一条 0 字节守卫(此前 --get 路径缺这条)');
+  });
+});
+
+describe('D4 · --mode / --dsf / --at 的非法值必须在入口被拒', () => {
+  const proj = () => {
+    const p = mkproj(tmpdir(), { slides: [{ id: '01', html: '01.html', audio: '01.mp3' }] });
+    fs.writeFileSync(path.join(p, 'slides', '01.html'), SLIDE);
+    return p;
+  };
+
+  test('capture 的 script.width/height 必须是 16–16384 整数(与 preview-page/build-video 同一道门)', () => {
+    // 第八轮复查 F3: capture 曾对同一字段裸用(非整数一路带进 viewport)。闸门在 playwright
+    // 加载之前, 所以本用例不需要任何外部工具 —— 无工具 CI 上照样真跑。
+    const p = proj();
+    fs.mkdirSync(path.join(p, 'build'), { recursive: true });
+    fs.writeFileSync(path.join(p, 'build', 'timings.json'), JSON.stringify(
+      { fps: 30, total: 1, slides: [{ id: '01', duration: 1, stages: { 1: 0 }, clauses: [] }] }));
+    const setWH = (w, h) => {
+      const s = JSON.parse(fs.readFileSync(path.join(p, 'script.json'), 'utf8'));
+      s.width = w; s.height = h;
+      fs.writeFileSync(path.join(p, 'script.json'), JSON.stringify(s, null, 2));
+    };
+    for (const [w, h] of [[99999, 1080], [8, 1080], [1920, 1079.5], [0, 0]]) {
+      setWH(w, h);
+      const r = runSkill('capture.mjs', [p]);
+      assert.equal(r.status, 1, `width=${JSON.stringify(w)}, height=${JSON.stringify(h)} 必须被拒`);
+      assert.ok(r.stderr.includes('width/height'), r.stderr);
+    }
+  });
+
+  test('capture --mode 只收 still/motion', () => {
+    const p = proj();
+    for (const bad of ['motions', 'Still', 'video']) {
+      const r = runSkill('capture.mjs', [p, '--mode', bad]);
+      assert.equal(r.status, 1, `--mode ${bad} 必须被拒`);
+      assert.ok(r.stderr.includes('--mode'), r.stderr);
+    }
+  });
+
+  test('capture --dsf 只收 1–4 的整数', () => {
+    const p = proj();
+    for (const bad of ['0', '9', '2.5', 'abc', '-1']) {
+      const r = runSkill('capture.mjs', [p, '--dsf', bad]);
+      assert.equal(r.status, 1, `--dsf ${bad} 必须被拒`);
+      assert.ok(r.stderr.includes('--dsf'), r.stderr);
+    }
+  });
+
+  test('grab-frames --at 只收 (0,1] 的小数', { skip: !FFMPEG && '无 ffmpeg' }, () => {
+    const p = mkproj(tmpdir());
+    fs.mkdirSync(path.join(p, 'build'), { recursive: true });
+    fs.mkdirSync(path.join(p, 'out'), { recursive: true });
+    fs.writeFileSync(path.join(p, 'build', 'timings.json'), JSON.stringify({
+      fps: 30, total: 2.2, slides: [{ id: '01', duration: 1.1 }, { id: '02', duration: 1.1 }],
+    }));
+    const g = spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'testsrc=size=320x240:rate=10', '-t', '2.6',
+      '-pix_fmt', 'yuv420p', '-y', path.join(p, 'out', 'final.mp4')], { windowsHide: true });
+    assert.equal(g.status, 0, g.stderr ?? '');
+    for (const bad of ['0', '-1', '2', 'abc', '']) {
+      const r = runSkill('grab-frames.mjs', [p, '--at', bad]);
+      assert.equal(r.status, 1, `--at ${JSON.stringify(bad)} 必须被拒`);
+      assert.ok(r.stderr.includes('--at'), r.stderr);
+    }
+    // 合法值照常出帧
+    const ok = runSkill('grab-frames.mjs', [p, '--at', '0.5']);
+    assert.equal(ok.status, 0, ok.stdout + ok.stderr);
+    assert.ok(fs.existsSync(path.join(p, 'build', 'introspect', 'frame-01-end.png')));
+    // --ids 一个都不匹配 → 零帧 = 退出 1(别静默成功)
+    const none = runSkill('grab-frames.mjs', [p, '--ids', '99']);
+    assert.equal(none.status, 1, '零帧必须退出 1: ' + none.stdout);
+    assert.ok(none.stderr.includes('一帧都没抽到'), none.stderr);
+    // 部分匹配不再静默少做: 没命中的 id 点名告警(2026-09-28 Mavis 交接 §4)
+    const mixed = runSkill('grab-frames.mjs', [p, '--ids', '01,99']);
+    assert.equal(mixed.status, 0, mixed.stdout + mixed.stderr);
+    assert.ok((mixed.stdout + mixed.stderr).includes('已忽略: 99'), '没命中的 id 必须点名: ' + (mixed.stdout + mixed.stderr).slice(-300));
+    assert.ok(fs.existsSync(path.join(p, 'build', 'introspect', 'frame-01-end.png')), '命中的 01 照常抽帧');
+  });
+});
+
+describe('D5 · 字幕带闸门只对"绝对定位的 bottom"报警', () => {
+  const mk = html => {
+    const p = mkproj(tmpdir(), { slides: [{ id: '01', html: '01.html', audio: '01.mp3' }] });
+    fs.writeFileSync(path.join(p, 'slides', '01.html'), html);
+    return p;
+  };
+  test('padding-bottom / margin-bottom 不再误报(实测误报点)', () => {
+    const r = runSkill('check-slides.mjs', [mk(`<!doctype html><html data-theme="a"><head><meta charset="utf-8">
+<style>.cap { padding-bottom: 40px; margin-bottom: 24px; }</style></head>
+<body><div class="stage"><p class="fx-fade" data-stage="1">x</p><div class="cap">图:来源</div></div></body></html>`)]);
+    assert.equal(r.status, 0, r.stdout + r.stderr);
+    assert.ok(!/字幕带/.test(r.stdout), 'padding/margin-bottom 不该被当 bottom: ' + r.stdout.slice(-400));
+  });
+  test('流内元素写 bottom 不报(不产生位移), 绝对定位才报', () => {
+    const stat = runSkill('check-slides.mjs', [mk(`<!doctype html><html data-theme="a"><head><meta charset="utf-8">
+<style>.cap { bottom: 96px; }</style></head>
+<body><div class="stage"><p class="fx-fade" data-stage="1">x</p><div class="cap">图:来源</div></div></body></html>`)]);
+    assert.ok(!/字幕带/.test(stat.stdout), '没有 position:absolute 就不该报: ' + stat.stdout.slice(-300));
+    const abs = runSkill('check-slides.mjs', [mk(`<!doctype html><html data-theme="a"><head><meta charset="utf-8">
+<style>.cap { position: absolute; bottom: 96px; }</style></head>
+<body><div class="stage"><p class="fx-fade" data-stage="1">x</p><div class="cap">图:来源</div></div></body></html>`)]);
+    assert.match(abs.stdout, /字幕带/, '绝对定位贴底要报');
+    assert.equal(abs.status, 0, '这是提示级, 不阻塞流水线');
+    // 行内 style 也要能认出来
+    const inline = runSkill('check-slides.mjs', [mk(`<!doctype html><html data-theme="a"><head><meta charset="utf-8"></head>
+<body><div class="stage"><p class="fx-fade" data-stage="1">x</p><div class="cap" style="position:absolute;bottom:80px">图:来源</div></div></body></html>`)]);
+    assert.match(inline.stdout, /字幕带/, '行内 style 的 absolute+bottom 同样要报');
+  });
+  test('2026-09-28 Mavis 交接 · 字幕带按 SUB_GEOMETRY 实算(单行 170 / 双语双行档)', () => {
+    // 旧常量 168: bottom:169 漏报(实测单行 170); 且折行后 230px —— 62px 碰撞区永远不报
+    const edge = runSkill('check-slides.mjs', [mk(`<!doctype html><html data-theme="a"><head><meta charset="utf-8">
+<style>.cap { position: absolute; bottom: 169px; }</style></head>
+<body><div class="stage"><p class="fx-fade" data-stage="1">x</p><div class="cap">图:来源</div></div></body></html>`)]);
+    assert.match(edge.stdout, /字幕带/, 'bottom:169 落在实测单行带 170 内, 必须报(旧常量 168 会漏)');
+    const safe = runSkill('check-slides.mjs', [mk(`<!doctype html><html data-theme="a"><head><meta charset="utf-8">
+<style>.cap { position: absolute; bottom: 185px; }</style></head>
+<body><div class="stage"><p class="fx-fade" data-stage="1">x</p><div class="cap">图:来源</div></div></body></html>`)]);
+    assert.ok(!/字幕带/.test(safe.stdout), 'bottom:185 在单行带之外, 不该报: ' + safe.stdout.slice(-300));
+    // 双语(text2): 第二行 .kit-sub-2 叠高, 带更厚 —— bottom:185 此时也要报
+    const p2 = mkproj(tmpdir(), { slides: [{ id: '01', html: '01.html', audio: '01.mp3', clauses: [{ stage: 1, text: '一', text2: 'one' }] }] });
+    fs.writeFileSync(path.join(p2, 'slides', '01.html'), `<!doctype html><html data-theme="a"><head><meta charset="utf-8">
+<style>.cap { position: absolute; bottom: 185px; }</style></head>
+<body><div class="stage"><p class="fx-fade" data-stage="1">x</p><div class="cap">图:来源</div></div></body></html>`);
+    const bi = runSkill('check-slides.mjs', [p2]);
+    assert.match(bi.stdout, /字幕带/, '双语张的带按双行档算, bottom:185 仍落在带内');
+  });
+  test('2026-09-28 Mavis 交接 · 流内贴底(margin-top:auto)给复核提示(闸门只扫绝对定位)', () => {
+    const flow = runSkill('check-slides.mjs', [mk(`<!doctype html><html data-theme="a"><head><meta charset="utf-8">
+<style>.stage { display: flex; flex-direction: column; } .cap { margin-top: auto; }</style></head>
+<body><div class="stage"><p class="fx-fade" data-stage="1">x</p><div class="cap">底部免责</div></div></body></html>`)]);
+    assert.match(flow.stdout, /流内贴底/, '流内贴底写法要给 grab-frames 复核提示(08/11 压住内容而闸门全绿的形状)');
+    assert.equal(flow.status, 0, '提示级, 不阻塞');
+  });
+});
+
+describe('柱状图结构闸门: 百分比柱体不得裸放 flex column(2026-09-22 域外实测踩坑)', () => {
+  const mk = html => {
+    const p = mkproj(tmpdir(), { slides: [{ id: '01', html: '01.html', audio: '01.mp3' }] });
+    fs.writeFileSync(path.join(p, 'slides', '01.html'), html);
+    return p;
+  };
+  test('坑形态: flex column + 百分比柱体 + 无 .chart-plot-cell → 提示(不阻塞)', () => {
+    const r = runSkill('check-slides.mjs', [mk(`<!doctype html><html data-theme="a"><head><meta charset="utf-8">
+<style>.pillar { background: var(--accent); }</style></head>
+<body><div class="stage"><div style="display:flex;flex-direction:column;height:300px">
+<div class="pillar" style="height:77.5%"></div><div class="pillar" style="height:60%"></div>
+</div></div></body></html>`)]);
+    assert.match(r.stdout + r.stderr, /没包 \.chart-plot-cell/, r.stdout + r.stderr);
+    assert.equal(r.status, 0, '提示级不阻塞流水线: ' + r.stdout + r.stderr);
+  });
+  test('规范形态: .chart-plot-cell.grid 包柱体 → 不报', () => {
+    const r = runSkill('check-slides.mjs', [mk(`<!doctype html><html data-theme="a"><head><meta charset="utf-8">
+<style>.chart-bar-wrap { background: var(--accent); }</style></head>
+<body><div class="stage"><div class="chart-col" style="height:300px"><div class="chart-plot-cell grid">
+<div class="chart-bar-wrap" style="height:77.5%"></div></div></div></div></body></html>`)]);
+    assert.ok(!/没包 \.chart-plot-cell/.test(r.stdout), r.stdout.slice(-300));
+  });
+  test('误报防线: 柱状语义类 + 百分比高度但张内没有 flex column → 不报', () => {
+    const r = runSkill('check-slides.mjs', [mk(`<!doctype html><html data-theme="a"><head><meta charset="utf-8">
+<style>.pillar { background: var(--accent); }</style></head>
+<body><div class="stage"><div style="display:grid;height:300px">
+<div class="pillar" style="height:77.5%"></div><div class="pillar" style="height:60%"></div>
+</div></div></body></html>`)]);
+    assert.ok(!/没包 \.chart-plot-cell/.test(r.stdout), '非 flex column 是合法用法: ' + r.stdout.slice(-300));
+  });
+});
+
+describe('D6 · transition 单一解析(闸门与渲染器结论一致)', () => {
+  test('readTransition 接受裸字符串与对象, 结论相同', () => {
+    assert.deepEqual(readTransition('xfade'), { type: 'xfade', dur: 0.4, label: '交叉溶解 xfade 0.4s' });
+    assert.deepEqual(readTransition({ type: 'xfade' }), readTransition('xfade'));
+    assert.deepEqual(readTransition('cut'), { type: 'cut', dur: 0, label: '硬切 cut' });
+    assert.deepEqual(readTransition(null), readTransition(undefined));
+    assert.deepEqual(readTransition({ type: 'xfade', duration: 1.2 }).dur, 1.2);
+    for (const bad of ['fade', 42, { type: 'fade' }, { type: 'xfade', duration: 9 }, { type: 'xfade', duration: 0 }]) {
+      assert.throws(() => readTransition(bad), /非法/, `${JSON.stringify(bad)} 必须报错`);
+    }
+  });
+
+  test('check-slides 与 build-video 对 "xfade" 字符串给同一结论', { skip: !FFMPEG && '无 ffmpeg(用 check-slides 的标签确认)', }, () => {
+    const p = mkproj(tmpdir(), { slides: [{ id: '01', html: '01.html', audio: '01.mp3' }] });
+    fs.writeFileSync(path.join(p, 'slides', '01.html'), SLIDE);
+    // build-video 读 timings.json 在 --dry-run 之前, 所以最小骨架要补一份
+    fs.mkdirSync(path.join(p, 'build'), { recursive: true });
+    fs.writeFileSync(path.join(p, 'build', 'timings.json'), JSON.stringify({
+      fps: 30, total: 3.0, tts: 3.0, slides: [{ id: '01', duration: 3.0, tts: 3.0, clauses: [] }],
+    }, null, 2));
+    // build-video 的输入检查(帧序列 / preview/<id>.png)在 --dry-run 之前也会做, 所以给一张占位静帧
+    const png = path.join(p, 'preview', '01.png');
+    fs.mkdirSync(path.dirname(png), { recursive: true });
+    const g = spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'color=c=#123456:s=1920x1080', '-frames:v', '1', '-y', png], { windowsHide: true });
+    assert.equal(g.status, 0, '应能造出占位静帧');
+    // 音轨也要在(缺音频会在自检之前就退出, 看不到我们要断言的转场行)
+    const au = spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'anullsrc=r=32000:cl=mono', '-t', '3',
+      '-c:a', 'libmp3lame', '-b:a', '64k', '-y', path.join(p, 'audio', '01.mp3')], { windowsHide: true });
+    assert.equal(au.status, 0, '应能造出静音音轨');
+    const setT = v => {
+      const s = JSON.parse(fs.readFileSync(path.join(p, 'script.json'), 'utf8'));
+      s.transition = v;
+      fs.writeFileSync(path.join(p, 'script.json'), JSON.stringify(s, null, 2));
+    };
+    setT('xfade');
+    const str = runSkill('check-slides.mjs', [p]);
+    assert.match(str.stdout, /交叉溶解 xfade/, str.stdout.slice(-300));
+    setT({ type: 'xfade' });
+    const obj = runSkill('check-slides.mjs', [p]);
+    assert.equal(obj.stdout.match(/转场 = ([^;]*)/)?.[1], str.stdout.match(/转场 = ([^;]*)/)?.[1],
+      '裸字符串与对象在闸门里必须一样(旧实现只认对象里的 type, 字符串会显示成非法值)');
+    setT('fade');
+    const bad = runSkill('check-slides.mjs', [p]);
+    assert.match(bad.stdout, /✗/, '非法值要显式报出来');
+  });
+});
+
+describe('复查补充 · 取值型 flag 清单 + 落盘收监 + transition 两端一致', () => {
+  test('VALUE_FLAGS 不得包含布尔开关(否则它后面那个位置参数会被当值吃掉)', () => {
+    // 成因: 清单里混进 --json / --allow-stale-css 这类开关后, `capture --allow-stale-css <项目>`
+    // 会把 <项目> 当开关的值跳过 → 位置参数落到 cwd, 对着调用目录干活(2026-09-18 复查实测)。
+    const DUAL = new Set(['--transition']);   // 既能当开关又能带值(实测里有 argv.includes 也有取值)
+    const used = new Set();
+    for (const f of fs.readdirSync(SCRIPTS)) {
+      if (!f.endsWith('.mjs')) continue;
+      const src = fs.readFileSync(path.join(SCRIPTS, f), 'utf8');
+      for (const m of src.matchAll(/argv\.includes\(['"](--[a-z-]+)['"]\)/g)) used.add(m[1]);
+    }
+    const bad = [...used].filter(f => !DUAL.has(f) && VALUE_FLAGS.has(f));
+    assert.deepEqual(bad, [], `这些开关被当成取值型 flag, 会把后面的位置参数吃掉: ${bad.join(', ')}`);
+    // 真实用法逐条: 开关后面的目录必须仍被认成项目目录
+    for (const sw of ['--allow-stale-css', '--no-subs', '--json', '--both', '--force']) {
+      assert.equal(positionalDir([sw, './proj']), path.resolve('./proj'), `${sw} 后面跟目录时不得被当值吃掉`);
+    }
+    // 取值型照旧要跳过一个值
+    for (const vf of ['--topic', '--ids', '--mode', '--dsf', '--at', '--transition']) {
+      assert.equal(positionalDir([vf, 'X', './proj']), path.resolve('./proj'), `${vf} 的值必须被跳过`);
+    }
+    assert.deepEqual(positionals(['--check', 'a.png', 'b.png']), ['a.png', 'b.png'], 'prep-image 的文件名是位置参数');
+  });
+
+  test('init-project --upgrade-css 也要落盘收监: slides/ 是悬空链接时不得写到项目外', (t) => {
+    // 复查 B4: 计划里点名 init-project 与 fetch-official-images, 但两者此前都没有 safeOut ——
+    // CHANGELOG 却宣称"每个写入点都收监", 属于文档与代码不符(第三轮 review 抓到的)。
+    const proj = mkproj(tmpdir(), { tokens: LEGACY_TOKENS });
+    const outside = tmpdir();
+    const slidesReal = path.join(proj, 'slides');
+    fs.rmSync(slidesReal, { recursive: true, force: true });
+    const made = (() => {
+      try { fs.symlinkSync(outside, slidesReal, 'dir'); return true; } catch { /* 试 junction */ }
+      const r = spawnSync('cmd', ['/c', 'mklink', '/J', slidesReal, outside], { encoding: 'utf8', windowsHide: true });
+      return r.status === 0 && fs.existsSync(slidesReal);
+    })();
+    if (!made) return t.skip('当前环境建不了符号链接/junction(收监拒绝无从观察)');
+    // 让 slides 指向项目外(外面先放一份 tokens.css, 好走到"升级"分支)
+    fs.writeFileSync(path.join(outside, 'tokens.css'), LEGACY_TOKENS);
+    const r = runSkill('init-project.mjs', [proj, '--upgrade-css']);
+    assert.equal(r.status, 1, '项目内的目录段指向项目外时必须拒绝: ' + r.stdout + r.stderr);
+    assert.ok(/符号链接|越出|收监/.test(r.stderr + r.stdout), r.stderr + r.stdout);
+    assert.equal(fs.readFileSync(path.join(outside, 'tokens.css'), 'utf8'), LEGACY_TOKENS, '项目外的文件不得被改写');
+    assert.equal(fs.existsSync(path.join(outside, 'tokens.css.bak')), false, '备份也不得写到项目外');
+  });
+
+  test('transition: 渲染器(build-video)与闸门用同一解析 —— 裸字符串在两端同义', { skip: !FFMPEG && '无 ffmpeg' }, () => {
+    const p = mkproj(tmpdir(), { slides: [{ id: '01', html: '01.html', audio: '01.mp3' }] });
+    fs.writeFileSync(path.join(p, 'slides', '01.html'), SLIDE);
+    // build-video 在 --dry-run 之前就读 timings.json, 所以最小骨架要补一份
+    fs.mkdirSync(path.join(p, 'build'), { recursive: true });
+    fs.writeFileSync(path.join(p, 'build', 'timings.json'), JSON.stringify({
+      fps: 30, total: 3.0, tts: 3.0, slides: [{ id: '01', duration: 3.0, tts: 3.0, clauses: [] }],
+    }, null, 2));
+    const setT = v => {
+      const s = JSON.parse(fs.readFileSync(path.join(p, 'script.json'), 'utf8'));
+      if (v === null) delete s.transition; else s.transition = v;
+      fs.writeFileSync(path.join(p, 'script.json'), JSON.stringify(s, null, 2));
+    };
+    // --dry-run 会打印解析结果, 不需要真编码; 但 build-video 的输入检查(帧序列 / preview/<id>.png /
+    // 音轨)在 dry-run 之前也会做, 所以占位静帧与静音音轨都得在
+    const png = path.join(p, 'preview', '01.png');
+    fs.mkdirSync(path.dirname(png), { recursive: true });
+    assert.equal(spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'color=c=#123456:s=1920x1080',
+      '-frames:v', '1', '-y', png], { windowsHide: true }).status, 0, '应能造出占位静帧');
+    assert.equal(spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'anullsrc=r=32000:cl=mono', '-t', '3',
+      '-c:a', 'libmp3lame', '-b:a', '64k', '-y', path.join(p, 'audio', '01.mp3')], { windowsHide: true }).status, 0, '应能造出静音音轨');
+    setT('xfade');
+    const dry = runSkill('build-video.mjs', [p, '--dry-run']);
+    assert.equal(dry.status, 0, dry.stdout + dry.stderr);
+    assert.match(dry.stdout, /交叉溶解 0\.4s/, '裸字符串 xfade 必须被渲染器认成溶解: ' + dry.stdout);
+    setT({ type: 'xfade', duration: 1.2 });
+    assert.match(runSkill('build-video.mjs', [p, '--dry-run']).stdout, /交叉溶解 1\.2s/);
+    setT(null);
+    assert.match(runSkill('build-video.mjs', [p, '--dry-run']).stdout, /转场: 硬切/, '缺省 = 硬切');
+    setT('fade');
+    const bad = runSkill('build-video.mjs', [p, '--dry-run']);
+    assert.equal(bad.status, 1, '非法值渲染器也要拒');
+    assert.match(bad.stderr, /transition/, bad.stderr);
+    // 来源要点名: --transition 写错时不该说成 script.transition 写错(2026-09-18 复查抓到)
+    const badFlag = runSkill('build-video.mjs', [p, '--transition', 'bogus', '--dry-run']);
+    assert.equal(badFlag.status, 1);
+    assert.match(badFlag.stderr, /--transition/, `要点名 --transition: ${badFlag.stderr}`);
+    assert.ok(!/script\.transition/.test(badFlag.stderr), `不该甩锅给 script.json: ${badFlag.stderr}`);
+  });
+
+  test('init-project 骨架写点也收监: assets/ 指向项目外时 --force 重建不得写到项目外', (t) => {
+    // 与上一条同源(复查 B4): 收监是逐个写点做的, --upgrade-css 只是其中之一。骨架那次创建
+    // 8 个目录 + 5 个文件, 这些路径此前都是裸 path.join —— 目标目录里预置一个指向项目外的
+    // assets 链接(用户共享素材的常见做法), 生成物就会落到项目外。
+    const proj = tmpdir();
+    const outside = tmpdir();
+    const assets = path.join(proj, 'assets');
+    const made = (() => {
+      try { fs.symlinkSync(outside, assets, 'dir'); return true; } catch { /* 试 junction */ }
+      const r = spawnSync('cmd', ['/c', 'mklink', '/J', assets, outside], { encoding: 'utf8', windowsHide: true });
+      return r.status === 0 && fs.existsSync(assets);
+    })();
+    if (!made) return t.skip('当前环境建不了符号链接/junction(骨架写点收监无从观察)');
+    const r = runSkill('init-project.mjs', [proj, '--force']);
+    assert.equal(r.status, 1, `项目内的目录段指向项目外时必须拒绝: ${r.stdout}${r.stderr}`);
+    assert.match(r.stderr + r.stdout, /越出|符号链接|收监/, r.stderr + r.stdout);
+    assert.deepEqual(fs.readdirSync(outside), [], '项目外目录里不得出现任何生成物');
+  });
+});
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/safe-paths.test.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/safe-paths.test.mjs
new file mode 100644
index 00000000..82d91be9
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/safe-paths.test.mjs
@@ -0,0 +1,330 @@
+// 安全边界 · 路径收监(script.json 派生路径不得越出项目目录)
+// 对应评审意见 1: Input-derived IDs and paths are not contained.
+import { test, describe } from 'node:test';
+import assert from 'node:assert/strict';
+import fs from 'node:fs';
+import path from 'node:path';
+import { spawnSync } from 'node:child_process';
+import { inside, findTool } from '../scripts/tools.mjs';
+import { runSkill, probeHelper, mkproj, tmpdir } from './helpers.mjs';
+
+describe('inside() 判定', () => {
+  test('根内相对路径 → true; 越出 → false', () => {
+    const root = path.resolve('/proj');
+    assert.equal(inside(root, path.resolve('/proj/a/b.txt')), true);
+    assert.equal(inside(root, path.resolve('/proj-a/b.txt')), false, '前缀相似的兄弟目录不算在内');
+    assert.equal(inside(root, path.resolve('/victim')), false);
+    assert.equal(inside(root, path.resolve('/proj')), false, '根本身不算 inside(要求严格子路径)');
+  });
+});
+
+describe('safeId / safeRel(退出型, subprocess 探测)', () => {
+  test('safeId: 合法 id 原样返回', () => {
+    for (const id of ['01', 'slide-2', 'A_b', 'x'.repeat(64)]) {
+      const r = probeHelper('safeId', `[${JSON.stringify(id)}]`);
+      assert.equal(r.status, 0, r.stderr);
+      assert.ok(r.stdout.startsWith('RETURN:'), `safeId(${id}) 应返回原值`);
+    }
+  });
+  test('safeId: 穿越/绝对/空/超长/换行 全部退出 1', () => {
+    for (const id of ['../../victim', '..\\..\\victim', '/abs', 'C:\\x', '', 'a/b', 'a\nb', 'x'.repeat(65), null, 42]) {
+      const r = probeHelper('safeId', `[${JSON.stringify(id)}]`);
+      assert.notEqual(r.status, 0, `safeId(${JSON.stringify(id)}) 必须拒绝`);
+    }
+  });
+  test('safeRel: 根内相对路径返回绝对路径', () => {
+    const r = probeHelper('safeRel', `["/proj", "a/b.txt", {}]`);
+    assert.equal(r.status, 0, r.stderr);
+    assert.ok(r.stdout.startsWith('RETURN:'), '根内路径应放行');
+  });
+  test('safeRel: 绝对路径 / ../ 穿越 / 空串 拒绝', () => {
+    for (const rel of ['/etc/passwd', '../../victim', '', 'a/../../..']) {
+      const r = probeHelper('safeRel', `["/proj", ${JSON.stringify(rel)}, {}]`);
+      assert.notEqual(r.status, 0, `safeRel(${JSON.stringify(rel)}) 必须拒绝`);
+    }
+  });
+  test('safeRel: 符号链接逃逸拒绝(建不了链接则 skip)', (t) => {
+    const root = tmpdir();
+    const outside = tmpdir();
+    fs.writeFileSync(path.join(outside, 'secret.txt'), 'x');
+    const link = path.join(root, 'assets-link');
+    // symlink 需要开发者模式, junction(mklink /J)不需要 —— 两条路都试, 尽量真跑而不是跳过
+    const made = (() => {
+      try { fs.symlinkSync(outside, link, 'dir'); return true; } catch { /* 退回 junction */ }
+      const r = spawnSync('cmd', ['/c', 'mklink', '/J', link, outside], { encoding: 'utf8', windowsHide: true });
+      return r.status === 0 && fs.existsSync(link);
+    })();
+    if (!made) return t.skip('当前环境建不了符号链接/junction');
+    const r = probeHelper('safeRel', `["${path.resolve(root).replace(/\\/g, '/')}", "assets-link/secret.txt", {}]`);
+    assert.notEqual(r.status, 0, '指向项目外的符号链接必须拒绝');
+  });
+
+  test('safeRel: 悬空链接(目标已删除)必须拒绝 —— existsSync 看不出它, 旧回溯会当"不存在"跳过', (t) => {
+    // 复查 B3: 旧实现用 existsSync 回溯祖先 —— 悬空链接的 existsSync 是 **false**, 于是它被跳过、
+    // 链接本身从未被复核, 写进去就落到项目外。现在改用 lstat 停住 + realpath 解不开即 fail closed。
+    const root = tmpdir();
+    const target = tmpdir();
+    const link = path.join(root, 'slides-link');
+    const made = (() => {
+      try { fs.symlinkSync(target, link, 'dir'); return true; } catch { /* 退回 junction */ }
+      const r = spawnSync('cmd', ['/c', 'mklink', '/J', link, target], { encoding: 'utf8', windowsHide: true });
+      return r.status === 0 && fs.existsSync(link);
+    })();
+    if (!made) return t.skip('当前环境建不了符号链接/junction(悬空链接场景无从构造)');
+    fs.rmSync(target, { recursive: true, force: true });   // 目标消失 → 链接悬空
+    assert.equal(fs.existsSync(link), false, '前置: 悬空链接的 existsSync 必须是 false(旧实现的盲区正在这里)');
+    assert.ok(fs.lstatSync(link).isSymbolicLink(), '前置: lstat 仍能看到这个 reparse point');
+    const r = probeHelper('safeRel', `["${path.resolve(root).replace(/\\/g, '/')}", "slides-link/x.css", {}]`);
+    assert.notEqual(r.status, 0, '悬空链接必须拒绝(解不开就无法证明它落在项目内)');
+  });
+
+  test('safeOut 单元级: 项目内的目录段是指向项目外的链接 → 直接调用即拒绝, 且外面没有留下任何文件', (t) => {
+    // 复查 B4 收尾: 收监此前只有"跑某个脚本"的集成用例。这里直接调 safeOut(不经任何流水线),
+    // 确认拒绝发生在这一层 —— 集成用例可能因为别的校验先退出, 从而给出假绿灯。
+    const root = tmpdir();
+    const outside = tmpdir();
+    const linked = path.join(root, 'out');
+    const made = (() => {
+      try { fs.symlinkSync(outside, linked, 'dir'); return true; } catch { /* 退回 junction */ }
+      const r = spawnSync('cmd', ['/c', 'mklink', '/J', linked, outside], { encoding: 'utf8', windowsHide: true });
+      return r.status === 0 && fs.existsSync(linked);
+    })();
+    if (!made) return t.skip('当前环境建不了符号链接/junction(safeOut 拒绝无从观察)');
+    const r = probeHelper('safeOut', `["${path.resolve(root).replace(/\\/g, '/')}", "out", "x.txt"]`);
+    assert.notEqual(r.status, 0, 'safeOut 必须自己拒绝, 而不是把判断留给调用方');
+    assert.match(r.stderr, /越出项目目录|符号链接/, `要点名是链接越界, 实际: ${r.stderr.slice(0, 200)}`);
+    assert.equal(fs.existsSync(path.join(outside, 'x.txt')), false, '拒绝必须发生在写入之前');
+  });
+
+  test('safeRel: 项目内链接指向项目"父目录"时, 链接下的新路径必须拒绝(1.7.5 复查②, 实弹复现过的洞)', (t) => {
+    // 形状: link → 项目的父目录。目标**已存在**时游走能走到完整路径, 正确拒绝;
+    // 目标**尚不存在**时旧回退 inside(real, realRoot) 被满足(根在父目录之内)→ 放行 → 写穿到项目外。
+    const root = tmpdir();
+    const parent = path.dirname(root);          // tmpdir() 每个都是独立目录, root 的父就是公共层
+    fs.writeFileSync(path.join(parent, 'canary.txt'), 'KEEP');
+    const link = path.join(root, 'link');
+    const made = (() => {
+      try { fs.symlinkSync(parent, link, 'dir'); return true; } catch { /* 退回 junction */ }
+      const r = spawnSync('cmd', ['/c', 'mklink', '/J', link, parent], { encoding: 'utf8', windowsHide: true });
+      return r.status === 0 && fs.existsSync(link);
+    })();
+    if (!made) return t.skip('当前环境建不了符号链接/junction(祖先链接形状无从构造)');
+    t.after(() => fs.rmSync(path.join(parent, 'canary.txt'), { force: true }));   // canary 落在共享 %TEMP% 层, 测完即清
+    const mk = rel => probeHelper('safeRel', `["${path.resolve(root).replace(/\\/g, '/')}", ${JSON.stringify(rel)}, {}]`);
+    const rNew = mk('link/new.txt');
+    assert.notEqual(rNew.status, 0, `链接指向父目录 + 新目标必须拒绝, 实际: ${(rNew.stderr || rNew.stdout).slice(-200)}`);
+    const rOld = mk('link/canary.txt');
+    assert.notEqual(rOld.status, 0, '已存在目标穿过链接同样必须拒绝');
+    assert.equal(fs.existsSync(path.join(parent, 'new.txt')), false, '拒绝必须发生在写入之前');
+    assert.equal(fs.readFileSync(path.join(parent, 'canary.txt'), 'utf8'), 'KEEP', 'canary 必须完好');
+  });
+});
+
+describe('消费者脚本: 恶意 script.json 必须在干坏事之前退出', () => {
+  test('capture: id=../../canary → 退出 1 且 canary 完好(递归删除被拦)', () => {
+    const proj = tmpdir();
+    const canary = path.join(path.dirname(proj), 'canary-' + path.basename(proj));
+    fs.mkdirSync(path.join(canary, 'deep', 'deeper'), { recursive: true });
+    fs.writeFileSync(path.join(canary, 'deep', 'deeper', 'keep.txt'), 'do not delete');
+    mkproj(proj, { slides: [{ id: '../../' + path.basename(canary), html: 'x.html', audio: 'a.mp3' }] });
+    const r = runSkill('capture.mjs', [proj]);
+    assert.notEqual(r.status, 0, '恶意 id 必须被拒绝');
+    assert.ok(fs.existsSync(path.join(canary, 'deep', 'deeper', 'keep.txt')), 'canary 必须完好');
+  });
+
+  test('check-slides: html=../outside.html → 退出 1', () => {
+    const proj = tmpdir();
+    mkproj(proj, { slides: [{ id: '01', html: '../outside.html', audio: '01.mp3', clauses: [] }] });
+    const r = runSkill('check-slides.mjs', [proj]);
+    assert.notEqual(r.status, 0);
+    assert.ok((r.stdout + r.stderr).includes('越出') || (r.stdout + r.stderr).includes('非法'), `应指明越界原因, 实际: ${r.stderr.slice(-300)}`);
+  });
+
+  test('check-slides: <img src> 越出项目目录 → 记为 error 并退出 1', () => {
+    const proj = tmpdir();
+    mkproj(proj, { slides: [{ id: '01', html: '01.html', audio: '01.mp3' }] });
+    fs.writeFileSync(path.join(proj, 'slides', '01.html'),
+      `<img src="../../neighbor.png"><div class="fx-fade" data-stage="1">x</div>`);
+    const r = runSkill('check-slides.mjs', [proj]);
+    assert.notEqual(r.status, 0);
+    assert.ok(r.stdout.includes('越出项目目录'), `应报图片越界, 实际: ${r.stdout.slice(-300)}`);
+  });
+
+  test('plan-timings: audio=../../secret.mp3 → 退出 1(需 ffprobe)', { skip: !findTool('ffprobe') && '无 ffprobe, 跳过(CI 主工作流无 ffmpeg)' }, () => {
+    const proj = tmpdir();
+    mkproj(proj, { slides: [{ id: '01', audio: '../../secret.mp3', clauses: [{ stage: 1, text: 'x' }] }] });
+    const r = runSkill('plan-timings.mjs', [proj]);
+    assert.notEqual(r.status, 0);
+  });
+
+  test('build-video: bgm.file 为绝对路径 → 退出 1(需 ffmpeg+ffprobe)', { skip: (!findTool('ffmpeg') || !findTool('ffprobe')) && '无 ffmpeg, 跳过' }, () => {
+    const proj = tmpdir();
+    mkproj(proj, { slides: [{ id: '01', audio: '01.mp3' }] });
+    const scriptPath = path.join(proj, 'script.json');
+    const script = JSON.parse(fs.readFileSync(scriptPath, 'utf8'));
+    script.bgm = { file: path.join(tmpdir(), 'evil-bgm.mp3') }; // 绝对路径直接逃逸
+    fs.writeFileSync(scriptPath, JSON.stringify(script));
+    const r = runSkill('build-video.mjs', [proj]);
+    assert.notEqual(r.status, 0);
+    assert.ok(r.stdout.includes('bgm') || r.stderr.includes('bgm'), '应点名 bgm.file');
+  });
+
+  // ── 2026-09-18 三维审计发现的一致性缺口: 同一信任级的入参, 别处管了这里没管 ──
+  // 这些用例只喂 script.json + timings.json: 坏值必须在编码开始前就被拒(不依赖 ffmpeg),
+  // 否则就要等整条流水线跑完才报错 —— 原 width 用例没写 timings.json, 读文件就退出了, 属假绿
+  const withTimings = (scriptPatch, timingsPatch = {}) => {
+    const proj = tmpdir();
+    mkproj(proj, { slides: [{ id: '01', audio: '01.mp3' }] });
+    fs.writeFileSync(path.join(proj, 'build', 'timings.json'),
+      JSON.stringify({ fps: 30, total: 1, slides: [{ id: '01', duration: 1 }], ...timingsPatch }));
+    const sp = path.join(proj, 'script.json');
+    const script = JSON.parse(fs.readFileSync(sp, 'utf8'));
+    Object.assign(script, scriptPatch);
+    fs.writeFileSync(sp, JSON.stringify(script));
+    return proj;
+  };
+
+  test('build-video: bgm.volume/fadeIn/fadeOut 非数值 → 退出 1(与 width 同一注入面, 此前漏管)', () => {
+    for (const bad of [
+      { volume: "0.5,amovie='C:/x',volume=0.5" },   // 追加 filter 节点(ffmpeg filter 可读本地文件进输出音频)
+      { fadeIn: '1.5,volume=9' },
+      { fadeOut: 'x' },
+      { volume: 999 },
+    ]) {
+      const proj = withTimings({ bgm: { file: 'assets/bgm.mp3', ...bad } });
+      fs.mkdirSync(path.join(proj, 'assets'), { recursive: true });
+      fs.writeFileSync(path.join(proj, 'assets', 'bgm.mp3'), 'x');
+      const r = runSkill('build-video.mjs', [proj]);
+      assert.notEqual(r.status, 0, `bgm ${JSON.stringify(bad)} 必须被拒绝`);
+      assert.ok(/bgm\./.test(r.stdout + r.stderr), `要点名是 bgm 的哪个字段非法, 实际: ${(r.stdout + r.stderr).slice(-200)}`);
+    }
+  });
+
+  test('build-video: fps 非数值 → 退出 1(拼进 -r/-framerate; timings.fps 优先于 script.fps)', () => {
+    const r = runSkill('build-video.mjs', [withTimings({}, { fps: '30 -vf scale=1:1' })]);
+    assert.notEqual(r.status, 0);
+    assert.ok(/fps/.test(r.stdout + r.stderr), `要点名 fps, 实际: ${(r.stdout + r.stderr).slice(-200)}`);
+  });
+
+  test('build-video: width 非法字符串 → 退出 1(ffmpeg filter 注入面)', () => {
+    const r = runSkill('build-video.mjs', [withTimings({ width: '1920,drawtext=text=pwned' })]);
+    assert.notEqual(r.status, 0);
+    assert.ok(/width/.test(r.stdout + r.stderr), `要点名 width, 实际: ${(r.stdout + r.stderr).slice(-200)}`);
+  });
+
+  test('preview-page: script.width 非法字符串 → 退出 1(该值插进放映页 CSS/JS, 此前直接杀死整页脚本)', () => {
+    const proj = tmpdir();
+    mkproj(proj, { slides: [{ id: '01', html: '01.html', audio: '01.mp3' }] });
+    fs.writeFileSync(path.join(proj, 'slides', '01.html'), '<html><head></head><body><p class="fx-fade" data-stage="1">x</p></body></html>');
+    const scriptPath = path.join(proj, 'script.json');
+    for (const bad of ['1920px', '1920; } body{background:url(x)} /*']) {
+      const script = JSON.parse(fs.readFileSync(scriptPath, 'utf8'));
+      script.width = bad;
+      fs.writeFileSync(scriptPath, JSON.stringify(script));
+      const r = runSkill('preview-page.mjs', [proj]);
+      assert.notEqual(r.status, 0, `width=${JSON.stringify(bad)} 必须被拒绝(与 build-video 的 clampDim 同一道门)`);
+    }
+  });
+
+  test('preview-page: 超大 tokens.css → 退出 1(与 check-slides/init-project 同一道门, 此前实测跑 129s)', () => {
+    const proj = tmpdir();
+    mkproj(proj, { slides: [{ id: '01', html: '01.html', audio: '01.mp3' }] });
+    fs.writeFileSync(path.join(proj, 'slides', '01.html'), '<html><head></head><body><p class="fx-fade" data-stage="1">x</p></body></html>');
+    // 反复的开标记但无收标记 —— 受管块扫描的二次方放大输入
+    fs.writeFileSync(path.join(proj, 'slides', 'tokens.css'),
+      ':root{--accent:#111}\n' + '/* >>> html2video:nofx rev=deadbeef >>> */\n'.repeat(60000) + 'x'.repeat(3_000_000));
+    const t0 = Date.now();
+    const r = runSkill('preview-page.mjs', [proj], { timeout: 90000 });
+    const ms = Date.now() - t0;
+    assert.notEqual(r.status, 0, '超大 tokens.css 必须被上限拒绝');
+    assert.ok(/tokens\.css/.test(r.stdout + r.stderr), '要点名 tokens.css');
+    assert.ok(ms < 30000, `应在秒级拒绝, 实际 ${ms}ms`);
+  });
+
+  // ── 2026-09-18 实测踩坑(PITFALLS.md)固化成闸门 ──
+  test('check-slides 5e: 绝对定位落进字幕带(bottom < 168px@1080) → 点名警告', () => {
+    const proj = tmpdir();
+    mkproj(proj, { slides: [{ id: '01', html: '01.html', audio: '01.mp3' }] });
+    fs.writeFileSync(path.join(proj, 'slides', '01.html'),
+      `<html><head><style>.cap{position:absolute;bottom:96px}</style></head><body>
+       <p class="fx-fade" data-stage="1">x</p><div class="cap">图:某来源</div></body></html>`);
+    const r = runSkill('check-slides.mjs', [proj]);
+    assert.equal(r.status, 0, '提示级不得阻塞流水线');
+    assert.match(r.stdout, /字幕带/, '要点名落在字幕带');
+    assert.match(r.stdout, /bottom:96px/, '报出具体值');
+    // 安全区之上的图注不该报
+    mkproj(proj, { slides: [{ id: '01', html: '01.html', audio: '01.mp3' }] });
+    fs.writeFileSync(path.join(proj, 'slides', '01.html'),
+      `<html><head><style>.cap{position:absolute;bottom:196px}</style></head><body>
+       <p class="fx-fade" data-stage="1">x</p><div class="cap">图:某来源</div></body></html>`);
+    const r2 = runSkill('check-slides.mjs', [proj]);
+    assert.ok(!/字幕带/.test(r2.stdout), `bottom:196px 在安全区之上, 不该报: ${r2.stdout.slice(-200)}`);
+  });
+
+  test('check-slides 5d: .fx-stagger 与 data-stage 同张 → 提示确认(PITFALLS #2)', () => {
+    const proj = tmpdir();
+    mkproj(proj, { slides: [{ id: '01', html: '01.html', audio: '01.mp3' }] });
+    fs.writeFileSync(path.join(proj, 'slides', '01.html'),
+      `<html><head></head><body>
+       <ul class="fx-stagger" style="--stagger-base:var(--t1)"><li>a</li><li>b</li></ul>
+       <p class="fx-fade" data-stage="3">自己管入场</p></body></html>`);
+    const r = runSkill('check-slides.mjs', [proj]);
+    assert.equal(r.status, 0, '提示级不得阻塞流水线');
+    assert.match(r.stdout, /fx-stagger/, '要点名 stagger 与 data-stage 共存');
+    assert.match(r.stdout, /动画窗/, '要给出可自查的信号(动画窗变短)');
+  });
+
+  test('check-slides 5: 带 data-stage 却没有 fx 类 → 警告(stagger 不再兜底)', () => {
+    const proj = tmpdir();
+    mkproj(proj, { slides: [{ id: '01', html: '01.html', audio: '01.mp3' }] });
+    fs.writeFileSync(path.join(proj, 'slides', '01.html'),
+      `<html><head></head><body><ul class="fx-stagger"><li data-stage="2">以为 stagger 会管</li></ul></body></html>`);
+    const r = runSkill('check-slides.mjs', [proj]);
+    assert.match(r.stdout, /没有 fx-\* 类/, '要警告元素会停在 opacity:0');
+    assert.ok(!/可豁免/.test(r.stdout), '旧措辞"放进 .fx-stagger 容器可豁免"已不成立 —— stagger 规则带 :not([data-stage])');
+  });
+});
+
+// 七审 blocker 2: root 侧用未规范化的 path.resolve、probe 侧用 realpath —— macOS 上
+// /var/… 与 /private/var/… 是同一目录的两种写法, root 尚不存在时合法路径被 exit 1
+// (官方 CI 的 macOS 实测: preview-page 拒绝了合法的缩略图/音频路径)。
+// macOS 跑不到, 这里用 junction 造同构别名: root = <alias>/proj/preview 未生成,
+// probe 停在 <alias>/proj(存在), 两侧指向同一目录却是两个字符串。
+describe('七审 blocker 2 · 别名形态的 root(root 尚未生成)不得误判越界', () => {
+  const linkDir = (target, linkPath) => {
+    try { fs.symlinkSync(target, linkPath, 'dir'); return true; } catch { /* 试 junction */ }
+    const r = spawnSync('cmd', ['/c', 'mklink', '/J', linkPath, target], { encoding: 'utf8', windowsHide: true });
+    return r.status === 0 && fs.existsSync(linkPath);
+  };
+  const slash = p => path.resolve(p).split(path.sep).join('/');
+
+  test('safeRel / assertContained: 项目内路径放行(junction 别名 + 未生成的 root)', (t) => {
+    const real = tmpdir();
+    const alias = path.join(tmpdir(), 'alias');
+    fs.mkdirSync(path.join(real, 'proj'), { recursive: true });
+    if (!linkDir(real, alias)) return t.skip('当前环境建不了符号链接/junction(别名形态无从构造)');
+    assert.notEqual(path.resolve(alias), fs.realpathSync(alias), '前置: alias 必须真的指向别处(否则两条分支相同)');
+    const root = path.join(alias, 'proj', 'preview');          // 尚不存在(= macOS 上未生成的目录)
+    assert.equal(fs.existsSync(root), false, '前置: root 必须真的不存在, 否则走的是另一条分支');
+    const rel = probeHelper('safeRel', `[${JSON.stringify(slash(root))}, "01.png", {}]`);
+    assert.equal(rel.status, 0, `safeRel 必须放行(root 就在项目内), 实得 stderr: ${rel.stderr.slice(0, 200)}`);
+    const out = probeHelper('assertContained',
+      `[${JSON.stringify(slash(root))}, ${JSON.stringify(slash(path.join(root, '01.png')))}, {}]`);
+    assert.equal(out.status, 0, `assertContained 必须放行, 实得 stderr: ${out.stderr.slice(0, 200)}`);
+  });
+
+  test('同形态下的真逃逸仍必须拒绝(规范化修复不得变成放行洞)', (t) => {
+    // root 是真目录(正常存在), 它**里面**有一段链接指向项目外 —— 这才是收监承诺拦的形状。
+    // (root 自身是链接不在威胁模型内: root 由调用方给定, 其 canonical 即基准。)
+    const proj = path.join(tmpdir(), 'proj');
+    const alias = path.join(tmpdir(), 'alias');
+    const outside = tmpdir();
+    fs.mkdirSync(proj, { recursive: true });          // junction 的目标必须先存在
+    if (!linkDir(proj, alias)) return t.skip('当前环境建不了符号链接/junction');
+    fs.writeFileSync(path.join(outside, 'secret.txt'), 'x');
+    if (!linkDir(outside, path.join(proj, 'preview'))) return t.skip('建不出第二层链接');
+    const r = probeHelper('assertContained',
+      `[${JSON.stringify(slash(alias))}, ${JSON.stringify(slash(path.join(alias, 'preview', 'secret.txt')))}, {}]`);
+    assert.notEqual(r.status, 0, 'root 内的一段链接指向项目外时必须拒绝');
+  });
+});
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/sec-audit-2026-10-10.test.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/sec-audit-2026-10-10.test.mjs
new file mode 100644
index 00000000..adf5a5e8
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/sec-audit-2026-10-10.test.mjs
@@ -0,0 +1,262 @@
+// 2026-10-10 独立安全审计的守卫(解析与生成这一维, 前七轮维护者评审未覆盖)。
+// 判据原则(沿用本仓既有纪律):
+//   · H1 的检测必须**解析属性名**而不是搜 `on*=` 文本 —— 修复后文本仍在属性值里, 文本搜索会假红;
+//     所以每个"抓到注入"的用例都配一个检测器自检(未转义的形态必须被认出), 否则整组恒绿。
+//   · H2 的不变量是"行首时间码数 == cue 数"(SRT 的结构边界), 不是全文 "-->" 计数。
+//   · 闸门类(M2/M3/M4)每个绕过用例都配一个合法对照组, 防"收紧到全报错"这种假修。
+import { test, describe } from 'node:test';
+import assert from 'node:assert/strict';
+import fs from 'node:fs';
+import path from 'node:path';
+import net from 'node:net';
+import os from 'node:os';
+import { spawnSync } from 'node:child_process';
+import { addNoFx, injectHtmlVars, openCmdArgv } from '../scripts/preview-page.mjs';
+import { assertConcatPathSafe, buildSrt, sanitizeCaptionText, stripControlChars } from '../scripts/tools.mjs';
+import { isBlockedHost } from '../scripts/url-policy.mjs';
+import { mkproj, runSkill, tmpdir } from './helpers.mjs';
+
+// ── 属性解析器(判据的基础: 尊重引号) ────────────────────────────────
+function attrsOf(tag) {
+  const body = tag.replace(/^<\s*[a-zA-Z][\w:-]*/, '').replace(/\/?>$/, '');
+  const attrs = [];
+  let i = 0;
+  while (i < body.length) {
+    while (i < body.length && /\s/.test(body[i])) i++;
+    const s = i;
+    while (i < body.length && !/[\s=]/.test(body[i])) i++;
+    if (i === s) break;
+    const name = body.slice(s, i);
+    while (i < body.length && /\s/.test(body[i])) i++;
+    if (body[i] !== '=') { attrs.push({ name, value: null }); continue; }
+    i++;
+    while (i < body.length && /\s/.test(body[i])) i++;
+    let value;
+    if (body[i] === '"' || body[i] === "'") {
+      const q = body[i++], end = body.indexOf(q, i);
+      if (end < 0) return { malformed: true, attrs };
+      value = body.slice(i, end); i = end + 1;
+    } else {
+      const vs = i;
+      while (i < body.length && !/\s/.test(body[i])) i++;
+      value = body.slice(vs, i);
+    }
+    attrs.push({ name, value });
+  }
+  return { malformed: false, attrs };
+}
+const htmlTag = html => /<html\b(?:[^>"']|"[^"]*"|'[^']*')*>/i.exec(html)?.[0] ?? '';
+const handlersIn = html => attrsOf(htmlTag(html)).attrs.filter(a => /^on/i.test(a.name));
+
+describe('H1 · 放映页属性注入: 重新序列化前必须转义', () => {
+  test('addNoFx: 单引号值里的 " 不能闭合属性并追加 on* 处理器', () => {
+    const out = addNoFx(`<html class='x" onmouseover=window.FIRED=1 data-x='><head><title>t</title></head><body></body></html>`);
+    assert.deepEqual(handlersIn(out), [], '注入的事件处理器: ' + JSON.stringify(attrsOf(htmlTag(out)).attrs));
+    const cls = attrsOf(htmlTag(out)).attrs.find(a => a.name === 'class');
+    assert.match(cls.value, /no-fx/, 'no-fx 仍要加上');
+    assert.match(cls.value, /&quot;/, '引号应被转义成实体, 而不是原样闭合属性');
+  });
+  test('injectHtmlVars: 同样的注入形态(style 属性)', () => {
+    const out = injectHtmlVars(`<html style='--t2:800ms" onload=window.FIRED=1 data-x='><head></head></html>`, '--t1:400ms');
+    assert.deepEqual(handlersIn(out), [], '注入的事件处理器: ' + htmlTag(out));
+    assert.match(attrsOf(htmlTag(out)).attrs.find(a => a.name === 'style').value, /--t1:400ms/);
+  });
+  test('检测器自检: 未转义的形态必须被认出来(否则上面两条是恒绿)', () => {
+    const raw = attrsOf('<html class="x" onmouseover=window.FIRED=1 data-x=" no-fx">');
+    assert.ok(raw.attrs.some(a => /^on/i.test(a.name)), '未转义形态应被解析出 on* 属性');
+  });
+  test('data-style / data-class 不得遮蔽真属性(L3)', () => {
+    const out = injectHtmlVars(`<html data-style="--t2:9999ms" style="--t2:800ms"><head></head></html>`, '--t1:400ms');
+    const style = attrsOf(htmlTag(out)).attrs.find(a => a.name === 'style');
+    assert.ok(style, '真 style 属性必须还在');
+    assert.match(style.value, /--t2:800ms/, '真 style 必须保留');
+    assert.match(style.value, /--t1:400ms/, '注入值必须挂到真 style 上');
+  });
+  test('注释里的 <html> 是诱饵: 改真标签, 不改注释(L3)', () => {
+    const out = addNoFx(`<!doctype html><html><head><!-- <html class='decoy'> --><title>t</title></head><body><h1>1</h1></body></html>`);
+    const real = /<html\b[^>]*>/i.exec(out.slice(out.indexOf('<html>', 5) > -1 ? out.indexOf('>', out.indexOf('<html>', 5)) + 1 : 0))?.[0] ?? '';
+    assert.match(real, /class="[^"]*no-fx/, `真 <html> 应被加上 no-fx, 实际: ${out}`);
+    assert.match(out, /<!-- <html class='decoy'> -->/, '注释内容不该被改写');
+  });
+  test('属性值里的 > 不该把标签截断(L3)', () => {
+    const out = injectHtmlVars(`<html style="--x: calc(1 > 0)" data-k="v"><head></head></html>`, '--t1:400ms');
+    assert.match(htmlTag(out), /data-k="v"/, `标签应完整读到 data-k, 实际: ${htmlTag(out)}`);
+  });
+});
+
+describe('H2 · SRT: clause 文本不得伪造 cue 结构', () => {
+  const TC = /^\d{2}:\d{2}:\d{2},\d{3} --> \d{2}:\d{2}:\d{2},\d{3}$/;
+  test('注入的时间码行只是单行文本, 行首时间码数仍等于 cue 数', () => {
+    const { text, count } = buildSrt({ slides: [{ duration: 10, clauses: [
+      { start: 0, text: 'benign\n\n99\n00:59:00,000 --> 01:00:00,000\nFORGED' },
+      { start: 5, text: 'second' },
+    ] }] });
+    const lines = text.split('\n');
+    assert.equal(lines.filter(l => TC.test(l)).length, count, '行首时间码数必须等于 cue 数');
+    assert.ok(!lines.some(l => /-->/.test(l) && !TC.test(l)), '不应有"含 --> 但不是合法 cue 头"的行');
+  });
+  test('sanitizeCaptionText: 换行折平 / 控制字符与 bidi 剔除 / --&gt; 中和 / 空白折叠', () => {
+    assert.equal(sanitizeCaptionText('a\r\nb'), 'a b');
+    assert.equal(sanitizeCaptionText('a\n\n\nb'), 'a b');
+    assert.equal(sanitizeCaptionText('a\u0007\u001b[31m b'), 'a[31m b');
+    assert.equal(sanitizeCaptionText('IMDb\u202Etest'), 'IMDbtest');
+    assert.equal(sanitizeCaptionText('x --> y'), 'x → y');
+    assert.equal(sanitizeCaptionText('\n\n'), '');
+    assert.equal(sanitizeCaptionText('x'.repeat(500)).length, 400);
+  });
+  test('双语行仍保留(净化不许吃掉正常内容)', () => {
+    const { text, count } = buildSrt({ slides: [{ duration: 4, clauses: [{ start: 0, text: '主行', text2: 'second line' }] }] });
+    assert.equal(count, 1);
+    assert.match(text, /主行\nsecond line/);
+  });
+});
+
+describe('M1 · --open 打开浏览器的命令行不得起第二条命令(Windows)', () => {
+  const isWin = process.platform === 'win32';
+  test('显式引用后 & 不再是命令分隔符; 未引用则确实会被执行', { skip: !isWin && 'Windows 专属(cmd 语义)' }, async () => {
+    const tmp = fs.mkdtempSync(path.join(os.tmpdir(), 'm1-'));
+    const marker = path.join(tmp, 'MARKER.txt');
+    const target = `C:/a>${marker.split(path.sep).join('/')}&rem`;
+    const run = (quoted, verbatim) => new Promise(resolve => {
+      // 注意: 参数必须是**表达式**而不是 `const args = [...]` 声明 —— 声明塞进调用里
+      // 会让子进程语法错误、命令根本没执行, 于是"未引用也不建 marker"的假绿就是这么来的。
+      const argsExpr = quoted
+        ? `['/c','echo','', '"' + target + '"']`
+        : `['/c','echo','', target]`;
+      const opts = verbatim ? `, windowsVerbatimArguments: true` : '';
+      const code = `import { spawn } from 'node:child_process';
+        const target = ${JSON.stringify(target)};
+        await new Promise(r => { const p = spawn('cmd', ${argsExpr}, { windowsHide: true${opts} });
+          p.stdout.on('data',()=>{}); p.on('close',()=>r()); p.on('error',()=>r()); });`;
+      spawnSync(process.execPath, ['--input-type=module', '-e', code], { encoding: 'utf8', windowsHide: true, timeout: 20000 });
+      resolve(fs.existsSync(marker));
+    });
+    const createdQuoted = await run(true, true);
+    assert.equal(createdQuoted, false, '显式引用后不得创建第二条命令的产物');
+    const createdUnquoted = await run(false, false);
+    assert.equal(createdUnquoted, true, '对照组: 未引用时确实会执行(证明本用例会分辨)');
+    fs.rmSync(tmp, { recursive: true, force: true });
+  });
+});
+
+describe('M2/M3/M4 · 三条闸门绕过 + 合法对照组', () => {
+  const slide = body => `<!doctype html><html><head><title>t</title></head><body>${body}</body></html>`;
+  const check = (html, extra = () => {}) => {
+    const proj = tmpdir();
+    mkproj(proj, { slides: [{ id: '01', html: '01.html' }] });
+    fs.writeFileSync(path.join(proj, 'slides', '01.html'), html);
+    extra(proj);
+    const r = runSkill('check-slides.mjs', [proj], { cwd: proj });
+    return String(r.stdout + r.stderr);
+  };
+  const slideId = [{ id: '01', html: '01.html' }];
+
+  test('M2: 只借 fx- 前缀的假类被点名', () => {
+    assert.match(check(slide(`<div class="fx-notreal" data-stage="1">x</div>`)), /fx 类没有对应动画声明/);
+  });
+  test('M2: 本张 <style> 里自定义 fx 类的关键帧缺 opacity 被点名(tokens.css 之外的本地类也查)', () => {
+    assert.match(check(slide(`<style>.fx-mine{animation:k-mine .6s}@keyframes k-mine{from{transform:scale(.9)}to{transform:scale(1)}}</style><div class="fx-mine" data-stage="1">x</div>`)), /没声明 opacity/);
+  });
+  test('M3: 注释里的 --var 不能充当定义', () => {
+    assert.match(check(slide(`<p style="color:var(--c-fake)">x</p><!-- --c-fake: 1 -->`)), /未定义且无 fallback/);
+  });
+  test('M4: 大写 SRC= / 协议相对 / @import / CSS url() 都被点名', () => {
+    assert.match(check(slide(`<IMG SRC=https://evil.example/x.png>`)), /外链资源/);
+    assert.match(check(slide(`<img src="//host.example/x.png">`)), /外链资源/);
+    assert.match(check(slide(`<style>@import url(https://cdn.example/x.css);</style>`)), /CSS 外链/);
+    assert.match(check(slide(`<div style="background:url(https://cdn.example/y.png)">x</div>`)), /CSS 外链/);
+  });
+  test('对照组: 合法幻灯片零误报(含正文里的可见链接与真实 fx 类)', () => {
+    const proj = tmpdir();
+    mkproj(proj, { slides: slideId });
+    fs.writeFileSync(path.join(proj, 'slides', '01.html'), slide(`<div class="fx-up" data-stage="1">标题</div><p>见 https://example.com 说明</p>`));
+    fs.writeFileSync(path.join(proj, 'slides', 'pic.png'), Buffer.from('iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mP8z8BQDwAEhQGAhKmMIQAAAABJRU5ErkJggg==', 'base64'));
+    const r = runSkill('check-slides.mjs', [proj], { cwd: proj });
+    const out = String(r.stdout + r.stderr);
+    assert.equal(r.status, 0, out);
+    assert.ok(!/外链资源|CSS 外链|fx 类没有对应动画声明|未定义且无 fallback/.test(out), '不应误报: ' + out);
+  });
+});
+
+describe('LOW · L1/L5/L6 与控制字符', () => {
+  test('L1: slides/ 下的链接指向项目外 → 图片闸门按真实路径判越界(建不了链接则 skip)', (t) => {
+    const proj = tmpdir();
+    mkproj(proj, { slides: [{ id: '01', html: '01.html' }] });
+    const outside = tmpdir();
+    fs.writeFileSync(path.join(outside, 'secret.png'), 'x');
+    const link = path.join(proj, 'slides', 'assets-link');
+    let made = false;
+    try { fs.symlinkSync(outside, link, 'dir'); made = true; } catch {
+      const r = spawnSync('cmd', ['/c', 'mklink', '/J', link, outside], { encoding: 'utf8', windowsHide: true });
+      made = r.status === 0 && fs.existsSync(link);
+    }
+    if (!made) return t.skip('当前环境建不了符号链接/junction');
+    fs.writeFileSync(path.join(proj, 'slides', '01.html'), `<!doctype html><html><head></head><body><img src="assets-link/secret.png" alt="x"></body></html>`);
+    const r = runSkill('check-slides.mjs', [proj], { cwd: proj });
+    const out = String(r.stdout + r.stderr);
+    assert.notEqual(r.status, 0, '经链接指向项目外必须报错');
+    assert.match(out, /越出项目目录|指向项目外/);
+  });
+  test('L5: 展开写法的 IPv4-mapped IPv6(0:0:0:0:0:ffff:7f00:1)也判为被拦', () => {
+    assert.equal(net.isIP('0:0:0:0:0:ffff:7f00:1'), 6, '前置: 这是合法 IPv6 文本');
+    assert.equal(isBlockedHost('0:0:0:0:0:ffff:7f00:1'), true);
+    assert.equal(isBlockedHost('0:0:0:0:0:ffff:a9fe:a9fe'), true, '元数据地址的展开写法同样要拦');
+  });
+  test('L6: concat 列表的路径守卫(单引号/换行 → 明确失败)', () => {
+    assert.equal(assertConcatPathSafe('C:/ok/path.mp4'), 'C:/ok/path.mp4');
+    assert.throws(() => assertConcatPathSafe("C:/it's/x.mp4"), /不能有单引号或换行/);
+    assert.throws(() => assertConcatPathSafe('C:/x\ny.mp4'), /不能有单引号或换行/);
+  });
+  test('L4: stripControlChars 剔除控制字符与 bidi, 保留普通内容', () => {
+    assert.equal(stripControlChars('a\u0007\u001b[0m\u202Eb'), 'a[0mb');
+    assert.equal(stripControlChars('正常 文本\r\n第二行'), '正常 文本\n第二行');
+    assert.equal(stripControlChars('纯 ASCII'), '纯 ASCII');
+  });
+});
+
+// M1 返工: 上一版的用例拿 `echo` 当 `start` 的替身 —— echo 没有"标题"语义, 于是
+// start 首个带引号 token 会被吃掉这一条根本测不到。下面的用例直接测 start 本身。
+describe('M1 返工 · start 的空标题(Windows)', () => {
+  const isWin = process.platform === 'win32';
+
+  test('argv: 空标题必须是字面量 "" 且在路径之前; verbatim 下空串会消失', () => {
+    const { cmd, args, verbatim } = openCmdArgv('C:\\work&calc\\proj\\preview\\play\\index.html', 'win32');
+    assert.equal(cmd, 'cmd');
+    assert.equal(verbatim, true, 'Windows 必须 verbatim —— 否则 Node 只给含空格的参数加引号, & 会被执行');
+    assert.deepEqual(args.slice(0, 2), ['/c', 'start']);
+    assert.equal(args[2], '""', '空标题必须写成字面量; windowsVerbatimArguments 下 "" 不会消失');
+    assert.equal(args[3], '"C:\\work&calc\\proj\\preview\\play\\index.html"', '路径自带双引号');
+
+    // 红证明: 修复前那一份(空标题写成空串)必须过不了同一条判据, 否则本用例分辨不了
+    const preFix = ['/c', 'start', '', '"C:\\x\\index.html"'];
+    assert.notEqual(preFix[2], '""', '修复前的形状确实没有空标题');
+    assert.throws(() => assert.equal(preFix[2], '""', '空标题'),
+      '修复前的形状应当过不了这条判据 —— 少了它, 本用例就是永真');
+  });
+
+  test('非 Windows 分支不带 verbatim, 也不需要空标题', () => {
+    const m = openCmdArgv('/tmp/index.html', 'darwin');
+    assert.deepEqual(m, { cmd: 'open', args: ['/tmp/index.html'], verbatim: false });
+    const l = openCmdArgv('/tmp/index.html', 'linux');
+    assert.equal(l.cmd, 'xdg-open');
+    assert.equal(l.verbatim, false);
+  });
+
+  test('实测: start 的首个带引号 token 是标题 —— 空标题因此是承重的', { skip: !isWin && 'Windows 专属(cmd 语义)' }, () => {
+    const tmp = fs.mkdtempSync(path.join(os.tmpdir(), 'starttitle-'));
+    const mk = n => path.join(tmp, n);
+    // 必须带 /b: start 的本职是拉起新进程, 不加 /b 每次跑都会闪一个黑窗口出来骚扰使用者。
+    // /b 只影响"要不要开新窗口", 标题的解析规则完全不变 —— 这正是本用例要测的东西。
+    const run = line => {
+      spawnSync('cmd', ['/c', line], { encoding: 'utf8', windowsHide: true, timeout: 20000, windowsVerbatimArguments: true });
+      return fs.existsSync(mk(line.includes('"TITLE"') ? 'mQuoted' : 'mBare'));
+    };
+    // 带引号 → 标题, 后续命令真的执行 → 目录出现
+    const quotedRan = run(`start /b "TITLE" cmd /c mkdir "${mk('mQuoted')}"`);
+    // 不带引号 → 命令(找不到该程序) → 后续命令根本没跑
+    const bareRan = run(`start /b TITLE cmd /c mkdir "${mk('mBare')}"`);
+    assert.equal(quotedRan, true, '带引号的首 token 必须被 start 当作标题(后续命令被执行)');
+    assert.equal(bareRan, false, '对照组: 不带引号才当命令 —— 本用例确实在分辨 start 的语义, 不是永真');
+    fs.rmSync(tmp, { recursive: true, force: true });
+  });
+});
\ No newline at end of file
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/subtitles-invalidate.test.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/subtitles-invalidate.test.mjs
new file mode 100644
index 00000000..02ddee51
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/subtitles-invalidate.test.mjs
@@ -0,0 +1,228 @@
+// 字幕静帧必须随字幕一起作废(复查 B2): 成片是"帧序列 + 逐句字幕静帧"拼出来的, 静帧留在磁盘上;
+// 删掉一句 clauses 或改用 --no-subs 重建时, 如果旧静帧没被清掉、旧清单没被核对, 成片会把**已经不存在
+// 的那句字幕**拼回来 —— 画面里出现一句没人说过的字幕, 而全链路零报错。
+// 这里用像素级判据: 末段实帧与"当前该显示的静帧"几乎一致, 与"旧的第 2 句静帧"明显不同。
+import { test } from 'node:test';
+import assert from 'node:assert/strict';
+import fs from 'node:fs';
+import path from 'node:path';
+import { spawnSync } from 'node:child_process';
+import { findTool, loadPackage, subtitleKeyframes, subtitleWindows } from '../scripts/tools.mjs';
+import { runSkill, tmpdir, FRESH_TOKENS } from './helpers.mjs';
+
+const FFMPEG = findTool('ffmpeg');
+const FFPROBE = findTool('ffprobe');
+
+// 两图逐像素差(YAVG): 本文件唯一的成对差异实现(两个用例共用)
+const imgDiff = (a, b) => {
+  const o = spawnSync(FFMPEG, ['-hide_banner', '-loglevel', 'error', '-i', a, '-i', b,
+    '-lavfi', '[0:v][1:v]blend=all_mode=difference,signalstats,metadata=print:file=-', '-frames:v', '1', '-f', 'null', '-'],
+    { encoding: 'utf8', windowsHide: true });
+  const m = /YAVG=([0-9.]+)/.exec((o.stdout || '') + (o.stderr || ''));
+  return m ? parseFloat(m[1]) : null;
+};
+
+test('字幕静帧随字幕作废: 删掉一句 / --no-subs 重建后, 成片不得再出现旧字幕(像素级)', async t => {
+  if (!FFMPEG || !FFPROBE) return t.skip('无 ffmpeg/ffprobe(主 CI 环境; 由 scoped smoke workflow 覆盖)');
+  const playwright = await loadPackage('playwright');
+  if (!playwright) return t.skip('无 playwright');
+  try { const b = await playwright.chromium.launch({ headless: true }); await b.close(); } catch {
+    return t.skip('chromium 未安装(npx playwright install chromium)');
+  }
+
+  const proj = tmpdir();
+  assert.equal(runSkill('init-project.mjs', [proj, '--topic', 'SubInvalidate']).status, 0);
+  const g = spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'anullsrc=r=32000:cl=mono', '-t', '3',
+    '-c:a', 'libmp3lame', '-b:a', '64k', '-y', path.join(proj, 'audio', '01.mp3')], { windowsHide: true });
+  assert.equal(g.status, 0, g.stderr ?? '');
+  fs.writeFileSync(path.join(proj, 'slides', '01-title.html'),
+    `<!doctype html><html data-theme="a"><head><meta charset="utf-8"><link rel="stylesheet" href="tokens.css"></head>
+<body><div class="stage"><h1 class="fx-rise" data-stage="1">字幕作废检查</h1></div></body></html>`);
+
+  const setClauses = list => {
+    const p = path.join(proj, 'script.json');
+    const s = JSON.parse(fs.readFileSync(p, 'utf8'));
+    s.slides = s.slides.filter(x => x.id === '01');
+    s.slides[0].clauses = list;
+    fs.writeFileSync(p, JSON.stringify(s, null, 2));
+  };
+  const man = () => JSON.parse(fs.readFileSync(path.join(proj, 'build', 'substills', '01.json'), 'utf8'));
+  const subFile = k => path.join(proj, 'build', 'substills', '01', `s${k}.png`);
+  // 字幕条: 模板里 .kit-sub 是 bottom: 7.78% × 1080 ≈ 84px、40px 字、居中 → 取正下方居中一块,
+  // 比整条字幕带敏感得多(整带 1920×260 会把字幕只占的一小块平均掉)
+  const box = 'crop=600:140:660:890';
+  const YAVG = imgDiff;
+  const crop = (src, out) => { spawnSync(FFMPEG, ['-y', '-hide_banner', '-loglevel', 'error', '-i', src, '-vf', box, out], { windowsHide: true }); return out; };
+  // 字幕带亮度统计: YMAX=文字峰值(满亮 235 档, 截在淡变上掉到 ~130), YMIN=底板地板(叠框会下陷)
+  const STATS = img => {
+    const o = spawnSync(FFMPEG, ['-hide_banner', '-loglevel', 'error', '-i', img,
+      '-lavfi', `[0:v]${box},signalstats,metadata=print:file=-`, '-frames:v', '1', '-f', 'null', '-'],
+      { encoding: 'utf8', windowsHide: true });
+    const txt = (o.stdout || '') + (o.stderr || '');
+    const g = k => { const m = new RegExp(k + '=([0-9.]+)').exec(txt); return m ? parseFloat(m[1]) : null; };
+    return { YMIN: g('YMIN'), YAVG: g('YAVG'), YMAX: g('YMAX') };
+  };
+  const YMAX = img => STATS(img).YMAX;
+  const YFLOOR = img => STATS(img).YMIN;
+  const grab = (t0, out) => { spawnSync(FFMPEG, ['-y', '-hide_banner', '-loglevel', 'error', '-ss', String(t0), '-i', path.join(proj, 'out', 'final.mp4'), '-frames:v', '1', '-vf', box, out], { windowsHide: true }); return out; };
+  const dur = () => parseFloat((spawnSync(FFPROBE, ['-v', 'error', '-show_entries', 'format=duration', '-of', 'csv=p=0', path.join(proj, 'out', 'final.mp4')], { encoding: 'utf8' }).stdout || '').trim());
+  const work = path.join(proj, 'build', 'subcheck');
+  fs.mkdirSync(work, { recursive: true });
+
+  // ① 两句: 出片并留下证据(第 2 句的静帧 = 旧字幕长什么样)
+  setClauses([{ stage: 1, text: '第一句。' }, { stage: 2, text: '第二句展开内容。' }]);
+  assert.equal(runSkill('plan-timings.mjs', [proj]).status, 0);
+  assert.equal(runSkill('capture.mjs', [proj, '--mode', 'still']).status, 0);
+  assert.ok(fs.existsSync(subFile(1)), '前置: 第 2 句的静帧应已生成');
+  assert.equal(man().stills.length, 2, '前置: 清单里应有 2 句');
+  // 拼接静帧必须满亮(2026-09-22 灰字幕回归): 静帧曾截在关键帧淡变的半山腰上,
+  // 与帧序列段拼出"字幕出现不到一秒就跳灰"。标定: 满亮峰值实测 235(细笔画 AA 余量), 灰 bug 实测 138。
+  for (const k of [0, 1]) {
+    const y = YMAX(subFile(k));
+    console.log(`    实测(第 ${k + 1} 句静帧字幕带 YMAX, 满亮=235 档): ${y}`);
+    assert.ok(y != null && y >= 200, `第 ${k + 1} 句的拼接静帧必须满不透明(字幕带峰值 ${y})`);
+  }
+  // 跨句残留防漏(2026-09-22): 同一页逐张截图, 钉过的字幕若漏进后续静帧, 会叠出深色底板补丁 ——
+  // 两张静帧的暗部地板必须一致(同夹具成对比较; 暗主题下叠框差会缩小, 夹具固定 theme a 才够判别力)
+  const floor = [YFLOOR(subFile(0)), YFLOOR(subFile(1))];
+  console.log(`    实测(两句静帧字幕带暗部地板, 应一致): ${floor.join(' / ')}`);
+  assert.ok(Math.abs(floor[0] - floor[1]) <= 12, `上一句字幕不得漏进下一张静帧(暗部地板 ${floor.join(' / ')})`);
+  // 封面与静帧基底不得烙字幕(2026-09-22 修复的两条回归): 两者都是"无字幕终态图" —— 封面曾被
+  // 钉死残留烙进最后一句, 基底曾被末句关键帧终值烙进最后一句。与带字幕静帧在字幕带上必须有
+  // 明显差异, ≈0 即说明字幕被烙进了无字幕图。
+  for (const [label, img] of [['封面', path.join(proj, 'preview', 'cover.png')], ['静帧基底', path.join(proj, 'preview', '01.png')]]) {
+    const d = YAVG(crop(img, path.join(work, `${label}-band.png`)), crop(subFile(1), path.join(work, 'leak-s1-band.png')));
+    console.log(`    实测(${label} 与带字幕静帧的差异, ≈0=字幕被烙进): ${d}`);
+    assert.ok(d != null && d > 3, `${label} 不得烙进字幕(与带字幕静帧差异仅 ${d})`);
+  }
+  const oldS1 = path.join(work, 'old-s1.png');
+  fs.copyFileSync(subFile(1), oldS1);
+
+  // ② 删掉第 2 句: 重建后旧静帧必须消失, 且成片里不能再出现那句字幕
+  setClauses([{ stage: 1, text: '第一句。' }]);
+  assert.equal(runSkill('plan-timings.mjs', [proj]).status, 0);
+  assert.equal(runSkill('capture.mjs', [proj, '--mode', 'still']).status, 0);
+  assert.equal(fs.existsSync(subFile(1)), false, '旧的第 2 句静帧必须被作废(否则会被拼回成片)');
+  assert.equal(man().stills.length, 1, '清单必须只剩当前这一句');
+  assert.equal(runSkill('build-video.mjs', [proj]).status, 0);
+  const lateFrame = grab(Math.max(0.3, dur() * 0.85), path.join(work, 'late.png'));
+  const dOldS1 = YAVG(lateFrame, crop(oldS1, path.join(work, 'old-s1-band.png')));
+  const dCurS0 = YAVG(lateFrame, crop(subFile(0), path.join(work, 'cur-s0-band.png')));
+  assert.ok(dOldS1 != null && dCurS0 != null, '应能算出字幕带差异');
+  console.log(`    实测(字幕条差异, 越小越像): 当前字幕 ${dCurS0} · 被删掉的旧字幕 ${dOldS1}`);
+  assert.ok(dCurS0 < 2.5, `末段画面应贴合当前唯一那句字幕, 实际差异 ${dCurS0}`);
+  assert.ok(dOldS1 > 4.0, `末段画面不得还是被删掉的那句(与旧静帧差异只有 ${dOldS1})`);
+
+  // ③ --no-subs 重建: 静帧与清单都要清掉, 成片不再有字幕
+  const withSubs = path.join(work, 'with-subs-s0.png');   // 先留一份"有字幕"的参照(下面会把源文件清掉)
+  fs.copyFileSync(subFile(0), withSubs);
+  assert.equal(runSkill('capture.mjs', [proj, '--mode', 'still', '--no-subs']).status, 0);
+  assert.equal(fs.existsSync(path.join(proj, 'build', 'substills', '01.json')), false, '--no-subs 时清单必须清掉');
+  assert.equal(fs.existsSync(path.join(proj, 'build', 'substills', '01')), false, '--no-subs 时静帧目录必须清掉');
+  assert.equal(runSkill('build-video.mjs', [proj]).status, 0);
+  const lateNoSubs = grab(Math.max(0.3, dur() * 0.85), path.join(work, 'late-nosubs.png'));
+  const dWithSubs = YAVG(lateNoSubs, crop(withSubs, path.join(work, 'gone-s0-band.png')));
+  assert.ok(dWithSubs != null, '应能算出差异');
+  console.log(`    实测(--no-subs 后与"有字幕"参照的差异): ${dWithSubs}`);
+  assert.ok(dWithSubs > 4.0, `--no-subs 之后画面不该还有字幕(与有字幕静帧的差异只有 ${dWithSubs})`);
+});
+
+// 第三轮复查(1.7.1): 清单的 framesCover 必须与"当前帧序列的实际时长"对齐 —— B2 点名的第三项。
+// 时长与句序都对、但动画窗重截过(帧数变了)的清单, 会把字幕静帧从错误的时刻拼进成片。
+// 判据用整帧差异: 静帧是纯红、帧序列是纯灰, 拼没拼进画面一眼可辨(不需要字幕带裁剪)。
+test('清单帧覆盖与当前帧序列不一致 → 忽略清单(像素级: 陈旧静帧不得再进画面)', async t => {
+  if (!FFMPEG || !FFPROBE) return t.skip('无 ffmpeg/ffprobe(主 CI 环境; 由 scoped smoke workflow 覆盖)');
+
+  const proj = tmpdir();
+  fs.mkdirSync(path.join(proj, 'slides'), { recursive: true });
+  fs.writeFileSync(path.join(proj, 'slides', 'tokens.css'), FRESH_TOKENS);
+  fs.writeFileSync(path.join(proj, 'slides', '01.html'), '<!doctype html><html><body></body></html>');
+  fs.writeFileSync(path.join(proj, 'script.json'), JSON.stringify({
+    topic: 't', fps: 30, width: 1920, height: 1080,
+    slides: [{ id: '01', html: '01.html', audio: '01.mp3', clauses: [{ stage: 1, text: '唯一一句。' }] }],
+  }, null, 2));
+  fs.mkdirSync(path.join(proj, 'audio'), { recursive: true });
+  assert.equal(spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'anullsrc=r=32000:cl=mono', '-t', '3',
+    '-c:a', 'libmp3lame', '-b:a', '64k', '-y', path.join(proj, 'audio', '01.mp3')], { windowsHide: true }).status, 0);
+  fs.mkdirSync(path.join(proj, 'build'), { recursive: true });
+  fs.writeFileSync(path.join(proj, 'build', 'timings.json'), JSON.stringify({
+    fps: 30, total: 3.0, tts: 3.0, slides: [{ id: '01', duration: 3.0, tts: 3.0, clauses: [{ stage: 1, text: '唯一一句。' }] }],
+  }, null, 2));
+  // 手工帧序列: 30 帧(= 1.0s)纯灰 —— 清单却声称帧覆盖 2.0s(旧动画窗), 差 1s 必须被识破
+  // (image2 输出序列默认从 f00001 起编号, 而 build-video 认 f00000 → 必须 -start_number 0)
+  const fdir = path.join(proj, 'build', 'frames', '01');
+  fs.mkdirSync(fdir, { recursive: true });
+  assert.equal(spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'color=c=0x808080:s=1920x1080:r=30', '-t', '1',
+    '-start_number', '0', '-y', path.join(fdir, 'f%05d.png')], { windowsHide: true }).status, 0, '应能造出 30 帧灰底序列');
+  assert.ok(fs.existsSync(path.join(fdir, 'f00000.png')), '帧序列应从 f00000 开始(build-video 的输入模式)');
+  const stills = path.join(proj, 'build', 'substills');
+  fs.mkdirSync(path.join(stills, '01'), { recursive: true });
+  assert.equal(spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'color=c=red:s=1920x1080', '-frames:v', '1',
+    '-y', path.join(stills, '01', 's0.png')], { windowsHide: true }).status, 0, '应能造出红色字幕静帧');
+  fs.mkdirSync(path.join(proj, 'preview'), { recursive: true });
+  assert.equal(spawnSync(FFMPEG, ['-v', 'error', '-f', 'lavfi', '-i', 'color=c=0x808080:s=1920x1080', '-frames:v', '1',
+    '-y', path.join(proj, 'preview', 'cover.png')], { windowsHide: true }).status, 0, '应能造出封面占位');
+
+  const diff = imgDiff;
+  const grab = (t0, out) => { spawnSync(FFMPEG, ['-y', '-hide_banner', '-loglevel', 'error', '-ss', String(t0),
+    '-i', path.join(proj, 'out', 'final.mp4'), '-frames:v', '1', out], { windowsHide: true }); return out; };
+  const manPath = path.join(stills, '01.json');
+
+  // ① 陈旧清单: 时长对、句数对, 但 framesCover=2 ≠ 当前 1.0 → 必须点名并忽略
+  fs.writeFileSync(manPath, JSON.stringify({ fps: 30, duration: 3, framesCover: 2, stills: [{ start: 2, end: 3, k: 0 }] }));
+  let r = runSkill('build-video.mjs', [proj]);
+  assert.equal(r.status, 0, r.stdout + r.stderr);
+  assert.match(r.stdout + r.stderr, /帧覆盖 2s ≠ 当前 1s.*忽略该清单/, '要点名帧覆盖不符并忽略: ' + r.stdout + r.stderr);
+  const dStale = diff(grab(2.55, path.join(proj, 'build', 'cov-stale.png')), path.join(stills, '01', 's0.png'));
+  console.log(`    实测(陈旧清单被忽略后, 末段与红色静帧的差异): ${dStale}`);
+  assert.ok(dStale != null && dStale > 4.0, `陈旧静帧不得再进画面(与红色静帧差异仅 ${dStale})`);
+
+  // ② 对齐清单: framesCover=1 与当前帧序列一致 → 同一张静帧正常拼入(末段 = 红色)
+  fs.writeFileSync(manPath, JSON.stringify({ fps: 30, duration: 3, framesCover: 1, stills: [{ start: 1, end: 3, k: 0 }] }));
+  r = runSkill('build-video.mjs', [proj]);
+  assert.equal(r.status, 0, r.stdout + r.stderr);
+  const dOk = diff(grab(2.55, path.join(proj, 'build', 'cov-ok.png')), path.join(stills, '01', 's0.png'));
+  console.log(`    实测(清单对齐后, 末段与红色静帧的差异): ${dOk}`);
+  assert.ok(dOk != null && dOk < 2.5, `对齐清单的静帧应当拼入画面(与红色静帧差异 ${dOk})`);
+});
+
+// 2026-09-22 灰字幕回归(工具层, 无工具门槛)。形状契约四条, 前两条各抓过一次真回归:
+// ①两端隐藏: fill both 下 finish() 后回到 opacity:0 —— 封面/静帧基底图靠它保持无字幕
+//   (1.9.4 的"末句无尾帧"形状让终值停在 1, 最后一句被烙进封面与基底);
+// ②无同偏移撞车: 同偏移多次声明时后者生效, 平台会被尾帧吃掉(灰字幕根因);
+// ③末句零可见淡出(平台保持到片尾) / 非末句保留切换淡出; ④偏移不越 [0,100] 且单调。
+test('字幕关键帧形状: 两端隐藏契约、末句零可见淡出、无同偏移撞车(工具层)', () => {
+  const parse = css => (css.match(/@keyframes kit-sub-\d+\{(?:[^{}]|\{[^{}]*\})*\}/g) ?? []).map(blk => {
+    const groups = [...blk.matchAll(/([\d.,%]+)\{opacity:([\d.]+)\}/g)].map(m => ({
+      offsets: [...new Set(m[1].split(',').map(parseFloat))],   // 同一选择器列表里的重复偏移是别名(如 100.000%,100%), 不算撞车
+      opacity: parseFloat(m[2]),
+    }));
+    return { blk, groups, offsets: groups.flatMap(g => g.offsets) };
+  });
+
+  const suites = [
+    { name: '单句(即末句)', css: subtitleKeyframes([{ start: 0.1 }], 3), lasts: [true] },
+    { name: '两句', css: subtitleKeyframes([{ start: 0 }, { start: 1.2 }], 3), lasts: [false, true] },
+    { name: '末句起点 99.8%', css: subtitleKeyframes([{ start: 0 }, { start: 2.994 }], 3), lasts: [false, true] },
+  ];
+  for (const { name, css, lasts } of suites) {
+    const blocks = parse(css);
+    assert.equal(blocks.length, lasts.length, `${name}: 块数应等于句数`);
+    for (let i = 0; i < blocks.length; i++) {
+      const { blk, groups, offsets } = blocks[i];
+      assert.ok(offsets.length > 0 && groups.length >= 3, `应解析出关键帧组: ${blk}`);
+      assert.ok(offsets.every(o => o >= 0 && o <= 100), `偏移越界(CSS 关键帧选择器只认 0–100%): ${blk}`);
+      for (let j = 1; j < offsets.length; j++) assert.ok(offsets[j] >= offsets[j - 1], `偏移乱序: ${blk}`);
+      assert.equal(new Set(offsets).size, offsets.length, `同偏移声明两次(后者会吃掉前者, 灰字幕根因): ${blk}`);
+      assert.equal(groups[0].opacity, 0, `起点必须隐藏: ${blk}`);
+      assert.equal(groups.at(-1).opacity, 0, `100% 终值必须归 0(finish() 后自动隐藏的契约): ${blk}`);
+      assert.ok(groups.at(-1).offsets.includes(100), `尾帧应落在 100%: ${blk}`);
+      const plateau = groups.find(g => g.opacity === 1);
+      const fade = groups.at(-1).offsets[0] - plateau.offsets.at(-1);
+      if (lasts[i]) assert.ok(fade < 0.1, `末句不得有可见淡出(淡出窗 ${fade}%): ${blk}`);
+      else assert.ok(fade >= 0.1, `非末句应保留切换淡出(淡出窗 ${fade}%): ${blk}`);
+    }
+  }
+  // b 夹紧: 末句起点晚于 99.5% 时窗口终点不得越过 100
+  assert.equal(subtitleWindows([{ start: 0 }, { start: 2.994 }], 3)[1].b, 100);
+});
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/table-kit.test.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/table-kit.test.mjs
new file mode 100644
index 00000000..d84a4947
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/table-kit.test.mjs
@@ -0,0 +1,68 @@
+// 表格工具箱: 原语必须真在生成的 tokens.css 里, 且可读性硬指标不能被改回去。
+// 覆盖过的真问题: 表格文字用 caption(24px) 在手机上读不清; 行高压到 68px; 涨跌色自造色。
+import { test, describe } from 'node:test';
+import assert from 'node:assert/strict';
+import fs from 'node:fs';
+import path from 'node:path';
+import { runSkill, tmpdir } from './helpers.mjs';
+import { TABLE_CSS, hasTableKit } from '../scripts/table-css.mjs';
+
+const proj = tmpdir();
+const r = runSkill('init-project.mjs', [proj, '--topic', 'T']);
+assert.equal(r.status, 0, r.stderr);
+const css = fs.readFileSync(path.join(proj, 'slides', 'tokens.css'), 'utf8');
+
+describe('表格工具箱 · 模板内容', () => {
+  test('四种形态的原语都在(.tbl 数据表 / .kv 规格表 / .matrix 对比矩阵 / .rank 排名表)', () => {
+    for (const sel of ['.tbl {', '.kv {', '.matrix {', '.rank {']) {
+      assert.ok(css.includes(sel), `缺 ${sel}`);
+    }
+    assert.match(css, /\.tbl tbody tr\s*\{[^}]*height:\s*72px/, '数据表行高必须 ≥72px');
+    assert.match(css, /\.rank td\s*\{[^}]*height:\s*72px/, '排名表行高必须 ≥72px');
+    // 高亮列必须连表头一起覆盖: 只写 td.hi 时 <th class="hi"> 静默不生效(2026-09-18 排查抓到)
+    assert.match(css, /\.matrix th\.hi/, '矩阵高亮列要覆盖表头');
+  });
+
+  test('主数据用正文号(视频在手机上要读得清), 表头才降到 caption', () => {
+    for (const sel of ['.tbl {', '.kv {', '.matrix {', '.rank {']) {
+      const i = css.indexOf(sel);
+      assert.match(css.slice(i, css.indexOf('}', i)), /font-size:\s*var\(--fs-body\)/, `${sel} 主数据必须是 --fs-body`);
+    }
+    assert.match(css, /\.tbl th\s*\{[^}]*font-size:\s*var\(--fs-caption\)/, '表头用 caption');
+  });
+
+  test('涨跌走令牌, 高亮/合计各有语义类', () => {
+    assert.match(css, /\.tbl \.up\s*\{[^}]*var\(--up\)/);
+    assert.match(css, /\.tbl \.down\s*\{[^}]*var\(--down\)/);
+    assert.match(css, /\.tbl tr\.key td\s*\{/, '缺高亮行 .key');
+    assert.match(css, /\.tbl tr\.sum td\s*\{/, '缺合计行 .sum');
+    assert.match(css, /\.tbl \.num\s*\{[^}]*text-align:\s*right[^}]*tabular-nums/, '数值列必须右对齐 + 等宽数字');
+  });
+
+  test('只画横线(不画竖线/斑马纹): 表格原语里不得出现 border-left/right 或 nth-child 斑马纹', () => {
+    const block = css.slice(css.indexOf('── 表格原语'), css.indexOf('依次入场容器'));
+    assert.ok(!/border-(left|right)\s*:/.test(block), '表格原语不该画竖线');
+    assert.ok(!/nth-child\((odd|even)\)/.test(block), '表格原语不该有斑马纹');
+  });
+});
+
+describe('表格工具箱 · 老项目升级', () => {
+  test('--upgrade-css 会补表格原语, 且幂等', () => {
+    const old = tmpdir();
+    fs.mkdirSync(path.join(old, 'slides'), { recursive: true });
+    fs.writeFileSync(path.join(old, 'slides', 'tokens.css'), ':root { --accent: #111; }\n');
+    assert.equal(hasTableKit(':root { --accent: #111; }'), false);
+    const r1 = runSkill('init-project.mjs', [old, '--upgrade-css']);
+    assert.equal(r1.status, 0, r1.stderr);
+    const after = fs.readFileSync(path.join(old, 'slides', 'tokens.css'), 'utf8');
+    assert.match(after, /\.tbl tbody tr\s*\{/);
+    assert.match(after, /\.matrix th\s*\{/);
+    const r2 = runSkill('init-project.mjs', [old, '--upgrade-css']);
+    assert.equal(fs.readFileSync(path.join(old, 'slides', 'tokens.css'), 'utf8'), after, '第二次跑不得再追加');
+    assert.ok(r2.stdout.includes('无需升级'));
+  });
+
+  test('TABLE_CSS 自身通过 hasTableKit 判定(防止判定与内容脱节)', () => {
+    assert.equal(hasTableKit(TABLE_CSS), true);
+  });
+});
diff --git a/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/tokens-fx.test.mjs b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/tokens-fx.test.mjs
new file mode 100644
index 00000000..7a538814
--- /dev/null
+++ b/plugins/Wzdhehe/html2video-for-mcode/skills/html2video-for-mcode/tests/tokens-fx.test.mjs
@@ -0,0 +1,123 @@
+// 模板自检: 生成出来的 tokens.css 里, 每个"入场类" fx 的关键帧都必须声明 opacity。
+// 背景: [data-stage] 的基础态是 opacity:0, 入场靠 animation 的 both 填充把 opacity 抬回 1;
+// 关键帧若不碰 opacity, 该元素入场后永远不可见(2026-09-18 实测踩过 fx-grow-x/y 与 fx-spotlight)。
+// 这条规则以前只写在 check-slides 的逐张检查里 —— 只有用户刚好用到那个类才报; 这里对模板本身断言。
+import { test, describe } from 'node:test';
+import assert from 'node:assert/strict';
+import fs from 'node:fs';
+import path from 'node:path';
+import { runSkill, tmpdir, SCRIPTS } from './helpers.mjs';
+
+function parseFx(css) {
+  const classes = new Map(); // .fx-name → { anim, infinite }
+  for (const m of css.matchAll(/\.(fx-[a-z-]+)\s*\{([^}]*)\}/g)) {
+    const decl = m[2];
+    const a = /animation\s*:\s*([^;}]+)/.exec(decl);
+    if (!a) continue; // 容器类(如 .fx-stagger)不算
+    classes.set(m[1], { anim: a[1].trim().split(/\s+/)[0], infinite: /\binfinite\b/.test(a[1]) });
+  }
+  // 关键帧体要按括号配对取(单行写法 @keyframes x { from {…} to {…} } 不能靠"行尾 }"切)
+  const kf = new Map(); // 关键帧名 → 是否声明了 opacity
+  const re = /@keyframes\s+([\w-]+)\s*\{/g;
+  let m;
+  while ((m = re.exec(css))) {
+    let depth = 0, end = -1;
+    for (let j = m.index + m[0].length - 1; j < css.length; j++) {
+      if (css[j] === '{') depth++;
+      else if (css[j] === '}' && --depth === 0) { end = j; break; }
+    }
+    if (end < 0) continue;
+    kf.set(m[1], /(^|[;{\s])opacity\s*:/.test(css.slice(m.index + m[0].length, end)));
+    re.lastIndex = end;
+  }
+  return { classes, kf };
+}
+
+describe('模板 tokens.css: 入场类必须能从 opacity:0 抬回来', () => {
+  const proj = tmpdir();
+  const r = runSkill('init-project.mjs', [proj, '--topic', 'T']);
+  assert.equal(r.status, 0, r.stderr);
+  const css = fs.readFileSync(path.join(proj, 'slides', 'tokens.css'), 'utf8');
+  const { classes, kf } = parseFx(css);
+
+  test('解析到了 fx 类与关键帧(防止正则失效后测试假绿)', () => {
+    assert.ok(classes.size >= 10, `只解析到 ${classes.size} 个 fx 类`);
+    assert.ok(kf.size >= 8, `只解析到 ${kf.size} 组关键帧`);
+  });
+
+  test('非无限(入场)动画的关键帧都声明了 opacity', () => {
+    const bad = [];
+    for (const [cls, { anim, infinite }] of classes) {
+      if (infinite) continue;                  // 氛围类不承载入场, 不受基础态影响
+      if (!kf.has(anim)) { bad.push(`${cls} → 找不到关键帧 ${anim}`); continue; }
+      if (!kf.get(anim)) bad.push(`${cls} → @keyframes ${anim} 没有 opacity(该元素会永远隐形)`);
+    }
+    assert.deepEqual(bad, []);
+  });
+
+  test('no-fx 规则含 opacity 重置(只关动画不管基础态 = 整片空白)', () => {
+    assert.match(css, /\.no-fx\s+\[data-stage\][^{]*\{[^}]*opacity\s*:\s*1\s*!important/);
+  });
+
+  test('.fx-stagger 入场规则全部排除 [data-stage](2026-09-18 实测: nth-child 延迟特异度更高, 会覆盖子元素自己的时刻)', () => {
+    // 根因: .fx-stagger > *:nth-child(n) 的特异度(0,2,0) > .fx-up 等(0,1,0), 且写在文件末尾 →
+    // 容器内带 data-stage 的子元素会被 stagger 的 --stagger-base 时刻接管(实测提前 3 秒冒头)。
+    // 注意排除 .no-fx 前缀的重置规则 —— 它们本就该作用于所有子元素(含带 data-stage 的, 那是"关动画")
+    // 看整行选择器: no-fx 重置写成多选择器(.no-fx [data-stage], .no-fx .fx-stagger > * {),
+    // 从 .fx-stagger 处起匹配拿不到前缀, 所以按行判定
+    const staggerRules = [...css.matchAll(/\.fx-stagger\s*>\s*([^{]+)\{/g)]
+      .filter(m => {
+        const lineStart = css.lastIndexOf('\n', m.index) + 1;
+        return !css.slice(lineStart, m.index + m[0].length).includes('.no-fx');
+      })
+      .map(m => m[1].trim());
+    assert.ok(staggerRules.length >= 9, `应解析到 9 条入场 stagger 规则, 实际 ${staggerRules.length}: ${staggerRules.join(' | ')}`);
+    const bad = staggerRules.filter(sel => !sel.includes(':not([data-stage])'));
+    assert.deepEqual(bad, [], '每条入场 stagger 选择器都必须带 :not([data-stage])');
+  });
+});
+
+describe('init-project --upgrade-css: 老项目补 no-fx 规则', () => {
+  const stale = () => {
+    const proj = tmpdir();
+    fs.mkdirSync(path.join(proj, 'slides'), { recursive: true });
+    fs.writeFileSync(path.join(proj, 'slides', 'tokens.css'), ':root { --accent: #111; }\n');
+    fs.writeFileSync(path.join(proj, 'script.json'), '{"slides":[]}');
+    return proj;
+  };
+
+  test('缺规则 → 追加, 且不碰其他文件', () => {
+    const proj = stale();
+    const r = runSkill('init-project.mjs', [proj, '--upgrade-css']);
+    assert.equal(r.status, 0, r.stderr);
+    const css = fs.readFileSync(path.join(proj, 'slides', 'tokens.css'), 'utf8');
+    assert.match(css, /\.no-fx\s+\[data-stage\][^{]*\{[^}]*opacity\s*:\s*1\s*!important/);
+    assert.ok(css.startsWith(':root { --accent: #111; }'), '原有内容必须在前');
+    assert.equal(fs.readFileSync(path.join(proj, 'script.json'), 'utf8'), '{"slides":[]}', 'script.json 不得被动');
+    assert.ok(!fs.existsSync(path.join(proj, 'slides', '_template.html')), '不得生成新文件');
+  });
+
+  test('幂等: 再跑一次不重复追加', () => {
+    const proj = stale();
+    runSkill('init-project.mjs', [proj, '--upgrade-css']);
+    const once = fs.readFileSync(path.join(proj, 'slides', 'tokens.css'), 'utf8');
+    const r = runSkill('init-project.mjs', [proj, '--upgrade-css']);
+    assert.equal(r.status, 0);
+    assert.equal(fs.readFileSync(path.join(proj, 'slides', 'tokens.css'), 'utf8'), once);
+    assert.ok(r.stdout.includes('无需升级'));
+  });
+
+  test('没有 tokens.css → 退出 1', () => {
+    const r = runSkill('init-project.mjs', [tmpdir(), '--upgrade-css']);
+    assert.equal(r.status, 1);
+  });
+});
+
+describe('2026-09-28 Mavis 交接 · .disclaimer 不再内置 max-width:1240px', () => {
+  test('生成物里 .disclaimer 无 1240 上限(那会卡死多行整宽免责块, 项目覆写 width:100% 也压不过)', async () => {
+    const { generateTokensCss } = await import('file://' + path.join(SCRIPTS, 'tokens-template.mjs').split(path.sep).join('/'));
+    const css = generateTokensCss();
+    assert.ok(css.includes('.disclaimer'), '原语要在');
+    assert.ok(!css.includes('max-width: 1240px'), '不得再内置 1240px 上限 —— 需要收窄的项目自己覆写');
+  });
+});