fix: persist precomputed t/s so reloaded sessions match live rate

[Blankeos]

Live header showed ~145t/s but reopening the same session showed ~3292t/s.

Root cause: tokens_per_sec was computed live from generation samples but never persisted, and the reload fallback divided token_count (billed total incl. prompt + cache) by decode time instead of output tokens.

Changes:

• Add messages.tokens_per_sec (migration v4), round-tripped through history DAO + session conversions (plus 2 regression tests).
• Backfill output_tokens from billed usage.output buckets for pre-existing rows; never derive from tokens_used total.
• Render (chat.rs:6633) + notification (app.rs:1732) paths: t/s fallback uses output tokens only; total-only legacy rows show duration without t/s.

Note: test_edit_tool_renders_codex_style_diff_summary fails in worktrees (hardcodes main checkout path) — pre-existing, unrelated.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[cloudflare-workers-and-pages]

Deploying crabcode with    Cloudflare Pages

Latest commit:	eb0a66f
Status:	✅  Deploy successful!
Preview URL:	https://f94becfb.crabcode.pages.dev
Branch Preview URL:	https://fix-tps-reload-parity.crabcode.pages.dev

View logs

[Blankeos]

PR Review #47 — fix/tps-reload-parity -> main

Branch: fix/tps-reload-parity -> main
Merge confidence: 4/5
PR: #47 — "fix: persist precomputed t/s so reloaded sessions match live rate" (MERGEABLE, CLEAN)

PR Comment Body (paste to GitHub)

Fixes live (~145t/s) vs reload (~3292t/s) mismatch by persisting precomputed tokens_per_sec (migration v4) and restricting all t/s fallbacks to output tokens only. Finalized rate now matches opencode upstream #46108 (billed output / decode, 250ms floor). Legacy total-only rows show duration without t/s. Targeted tests + fmt --check + cargo check pass locally.

What Changed (5 files, +216/-39)

• src/persistence/migrations.rs:189 — v4: ALTER TABLE messages ADD COLUMN tokens_per_sec REAL (idempotent via let _ =), fresh v1 schema also includes it.
• src/persistence/history.rs:65,442,493,561,584 — Message.tokens_per_sec, round-tripped on both INSERT paths; SELECT uses row.get(14).unwrap_or(None) so v3 DBs still read.
• src/persistence/conversions.rs:113,189,242 — persist tokens_per_sec (finite+>0 filter); backfill output_tokens from usage.output buckets, never from tokens_used total. +2 tests.
• src/ui/components/chat.rs:458,485,2660,6648,6682 — new upstream_tokens_per_sec(); finalize_streaming_metrics prefers billed recorded_usage().output over sample estimate; render uses output-only; total-only legacy shows duration, no t/s. Test expectations 49→50, 99→100 + new test_finalize_prefers_billed_output_over_sample_estimate.
• src/app.rs:1711,1731,1747 — notification path: same upstream formula, 250ms floor, output-only, total-only legacy guard.

Regressions?

No blocking regressions found. Expected behavior shifts (intentional, not bugs):

• All t/s numbers shift slightly up: n/duration vs old (n-1)/duration (e.g. 49→50, 99→100).
• Fast responses (<250ms decode) now show no t/s where old code showed one if tokens≥2. Matches upstream floor.
• Reloaded sessions with only token_count+duration (no output tokens) lose their t/s display (duration-only). This is the fix for the 3292t/s inflation.

No dead-code issue: MIN_TPS_SAMPLE_TOKENS (chat.rs:409) still used by GenerationSample (chat.rs:446); new path correctly uses the 250ms floor instead.

Migrations?

Yes — v4. Safe:

• ADD COLUMN is nullable, idempotent, no backfill needed (fallback recomputes from billed buckets).
• Old binaries reading new DB: unknown column ignored (explicit column lists). New binary reading old DB: ALTER runs on open; SELECT tolerates missing col.
• No data loss, no downgrade path needed.

Pre-Merge Checklist

• Run full cargo test (only 3 targeted tests run here; PR notes test_edit_tool_renders_codex_style_diff_summary fails in worktrees — confirm pre-existing on main)
• Manual: stream a response, note live t/s → /new → reopen session → confirm same t/s
• Manual: open a pre-v4 session with large prompt + small output → confirm no inflated t/s, duration-only or correct rate
• Confirm notification (app.rs) rate matches chat header on completion

Confidence Bumps

4/5 → 5/5 if: full cargo test green (minus known worktree failure) + one manual live-vs-reload screenshot/numbers posted on the PR. No code changes needed — this is verification-only.

Checks Run (read-only)

• gh pr view/checks: OPEN, MERGEABLE, CLEAN; CI plan pass (others skipped by design)
• cargo fmt --check: clean (exit 0)
• cargo check: zero warnings/errors
• cargo test -- precomputed_tps billed_output finalize_prefers_billed: 3 passed