WebKit: Dashboard-tree row Tab-order lands on INPUT/BODY, not pencil/menu

[BorisTyshkevich]

The gap

tests/e2e/dashboard-tree.spec.js:316 — "Tab walks the row, its chevron and its
trailing actions, each ringed differently" — fails on WebKit on current main
(434643e, the tip after PR #495/#493), independent of any change on this branch.

Expected: ["workspace:sales", "chevron", "pencil", "menu"]
Received: ["workspace:sales", "chevron", "INPUT", "BODY"]

Tab-focus lands on INPUT/BODY instead of the row's pencil/menu trailing
actions on WebKit only; chromium passes the same spec. Reproduced twice,
deterministically (not a timing flake):

npx playwright install chromium webkit   # browsers were not preinstalled
npm run test:e2e -- --project=webkit -g "Tab walks the row"

Why deferred

Surfaced while shipping #466 (tab-close unsaved-change guard), which never
touches dashboard-tree.ts, dialog-shell.ts, or focus/tab-order code.
Confirmed present on a clean main checkout (stashed #466's changes, reran —
same failure), so it predates and is unrelated to that work. Likely connected
to the #429 phase 3 rename-pencil trailing action recently added to the row
(the expected order names pencil/menu — the two newest trailing actions).

Also worth noting for whoever picks this up: CI does not run test:e2e on a
PR (chromium/webkit only run locally), so this has had no green/red signal
since #429 phase 3 merged.

[BorisTyshkevich]

Reproduced again while validating #464 on current main base 3259b54.

The full Playwright run completed with 518 passes and five WebKit-only failures,
all in tests/e2e/dashboard-tree.spec.js:

• hierarchy-row Tab walk lands on INPUT/BODY;
• Library assignment fails to focus “Add to dashboard…”;
• Dashboard pencil Enter/Space misses the edit action/dialog;
• keyboard-driven panel delete misses the remove action.

The hierarchy-row failure reproduces in a detached worktree at untouched commit
3259b54, so it is independent of #464. Chromium and Firefox were clean.
#464's focused tab/variable suites subsequently passed 73 tests across all three
engines.

This remains the native Dashboard-tree traversal gap tracked here, distinct
from the anchored-dialog trap fixed by #439/#442.