feat(pick): preview the highlighted message while the picker is open

[Triyambak-CA]

The itch

Running several agents against one repo, herdr last regularly gives me a dozen-plus candidates. The picker lists them one line each, but the document behind it is always the newest, so choosing means: Enter, read, p, Enter on the next one, read, and so on. The list tells me a timestamp and a truncated first line, which often is not enough to tell two replies apart.

Moving the cursor now swaps the document underneath and stays in the picker, so the list becomes browsable.

What changed

• show_candidate(index) swaps the open document without leaving Mode::Pick; j/k/arrows call it.
• open_candidate is now show_candidate plus the mode change and status line, so Enter behaves exactly as before.
• Esc returns to whatever was open when the picker appeared. In the common case that is still the newest, so escaping_the_picker_keeps_the_newest_message passes unchanged; when p reopens the picker mid-review it now puts you back where you were instead of leaving a preview showing.
• Footer reads ↑↓ preview · enter open · esc cancel · q quit.

The one non-obvious bit

Previewing swaps Open, and Open owns the Store. A reply review's store is transient, so a naive swap would silently discard annotations the moment someone touched an arrow key. The document being left is therefore kept in a small map keyed by candidate index and restored on the way back.

That hole exists today on Enter too: annotate the newest, p, Enter on another, and the annotations are gone. This closes it as a side effect rather than widening it.

Memory is bounded by the candidate count, and only visited candidates are ever built.

Tests

Two, one per invariant, in app/tests.rs alongside the existing picker tests:

• moving_the_picker_cursor_previews_that_message - the document follows the cursor and the picker stays open
• previewing_away_and_back_keeps_annotations - annotate, p, j, k, and the annotation is still there

cargo fmt --all --check, cargo clippy --workspace --all-targets and cargo test --workspace are all clean on 1.97.1. 34 tests pass.

Note on performance

AGENTS.md says never reparse a whole document on an edit. This reparses on a selection change, not an edit, and only the first time each candidate is visited; after that it comes from the map. Agent replies are small. Happy to add a cap on cached documents if you would rather bound it explicitly.

Context for why I was in here: #53. This does not fix that issue, it is a separate annoyance I hit while testing it.

[backnotprop]

Merged, with the annotation-cache design appreciated in particular — closing an existing data-loss hole as a side effect of the feature is the good kind of side effect. I rebased over #55 (two trivial keep-both hunks in the App struct) using the maintainer-edit permission; the squash is otherwise your commit. Both picker features ship in the next release. Thanks for the #53 investigation too — the follow-up it earned lives in #57.