…pe test

Bundle of low-risk improvements flagged by the independent review:

Popover keyboard + a11y
- Add `role="listbox"` to the focus-picker list and `role="option"` +
  `aria-selected` + `tabindex="-1"` to each item, matching the listbox
  pattern the popover already implies via aria-haspopup.
- Add Escape handler that closes the popover and returns focus to the
  trigger button — same pattern as the commit-picker.

Stable test attributes
- `data-label` on each picker item (lifts the label out of rendered DOM
  children, replacing brittle `evaluate(() => innerText)` reads in
  focus-picker spec).
- `data-focus-kind` on each item, replacing the substring-match selector
  `[data-focus-payload*="working_tree"]` which would false-positive on any
  payload that happened to contain that string.

Brittle test fixes
- comment-stale spec: replace `stackOrdered[1]` with a feat-b-relative
  lookup so fixture growth doesn't silently shift the test.

Net new coverage
- scope-toggle-files: new test asserts a layer-scoped comment is hidden
  in full-stack and reappears when the user flips back. This is the
  lossless-visibility invariant — the only regression test against
  accidentally serialising scope-filtered comments out to disk.

Skipped from the polish list (with reasons):
- Branches-section base_sha precompute (#2) — server-side change, owned
  by go-expert; would be a separate ticket.
- Pre-feature comment WT-only test (#6) and /api/focus error-rollback
  toast test (#7) — both require fixture or daemon-state changes that
  exceed the 30-min budget for "polish".

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>