feat(#429): split the Dashboard tree row into chevron, name and trailing targets (phase 2)

[BorisTyshkevich]

What & why

Part of #429 — phase 2, which absorbs #472: the Dashboard-tree row becomes three
independent targets.

Target	Action
Disclosure chevron (now a real <button>)	expand / collapse this Dashboard's hierarchy, and nothing else
Dashboard name / primary row content (icon included)	open in View, immediately
Shift-click on the name	open Edit
Trailing ⋯	acts without navigating or expanding

Before this, a Dashboard row's single click only expanded it — deferred ~300ms through
the click arbiter — and opening needed a double-click. Expansion and navigation competed
for one gesture; now they don't, so the arbitration is gone from Dashboard rows entirely.
A panel row is untouched: click opens its query, double-click and Shift-click focus
the tile, and it still arbitrates.

The decision this phase owed the issue

"Once the name target opens View and Shift-click opens Edit, the Dashboard row's
Open in View / Open in Edit menu items are redundant. Either drop them or keep them
deliberately … and say which."

Split, not all-or-nothing:

• Dropped Open in View — it is the row's primary press now (pointer and Enter),
  so nothing about it is hidden any more, which was the menu item's only justification.
• Kept Open in Edit — its only forms are Shift-click and Shift+Enter, both hidden
  modifier gestures. The menu is where a keyboard or first-time user finds Edit mode at
  all; dropping it would trade a fixed redundancy for a real discoverability regression.

That also leaves the Dashboard row with one trailing control going into phase 3, which
is what the issue's "a third icon on a narrow sidebar row is the reason to think twice"
warns about — phase 3 can still choose a pencil icon or a second menu row.

Implementation notes

• dashboard-tree-model.ts: the Dashboard row emits single: open-dashboard(view),
  double: null, and keeps shift. double === null is the view's new arbitration
  switch — no new field. single is now unconditional, so a Dashboard the search is
  holding open can be opened (it still cannot be collapsed — that is toggleable),
  where before that click did nothing at all.
• The chevron carries aria-expanded, an Expand <name> / Collapse <name> accessible
  name, native Enter/Space that is stopped from reaching the tree's own Enter (which now
  navigates), a focus ring in a different channel from the row's, and a roving
  tabindex so the tree stays one composite tab stop instead of one stop per chevron.
  The treeitem keeps its own aria-expanded; both are painted from row.expanded.
• Keyboard activation returns focus to the chevron across the repaint its own toggle
  causes — otherwise one Space moved focus to the row and the next Enter navigated.
• Three guards the split made unreachable were removed, not left as dead code: the
  chevron's cancelFor, the Add Library drag assignment to Dashboard panels and variables #428 drop handler's cancelFor (only a row with a double
  is ever arbitrated → a panel row → which accepts no drop and has no chevron), and
  toggleRow's Dashboard arm (the model now emits {kind:'toggle'} for group rows only).
• No duplicate history entry on a double-click needed no new tree code: openDashboard
  already collapses a repeated same-id/same-mode open. That guarantee is now pinned by a
  test in tests/unit/app.test.ts rather than assumed — the tree deliberately does not
  duplicate the guard, since it cannot see the route.

The bug only a real browser could show (second commit)

Hit-testing and getBoundingClientRect use the transformed box. The new button is
10×24 (stretched to the row height so a 10px glyph is worth aiming at), so a collapsed
row's rotate(-90deg) turned its clickable band into 24 wide × 10 tall — spilling 7px
each side and swallowing clicks meant for the row icon, which then expanded instead of
opening. The schema tree's chevron survives rotating its own box only because it is a
10×10 square. The rotation now lives on the glyph inside the control, and an e2e case
asserts the control's box stays in its own slot and never reaches the icon (verified to
fail with the rotation put back on the button). happy-dom has no layout, so nothing below
e2e could hold this.

The accessibility regression the review pass caught (third commit)

A treeitem names itself from its contents, and those contents now include a
labelled button. Measured with Chromium's own accname engine
(Accessibility.getPartialAXTree), not assumed:

collapsed   "Expand Sales revenue Sales revenue 2"
expanded    "Collapse Sales revenue Sales revenue 2 Actions for Sales revenue"

The chevron's name was being folded into the row's — the opposite of the criterion this
PR claims. The ⋯ leak was pre-existing and only while hovered/focused; the permanent
Expand … prefix on every Dashboard and group row was this phase's regression.

The row now states its own name, composed from the model's strings in the order it paints
them (name, count, status word, meta, marker label), so it cannot drift from the screen
and nothing previously announced is lost. Held by exact-name unit assertions and by
getByRole in e2e — which resolves real accessible names in every engine, where
happy-dom cannot compute one at all. Sabotage-checked: dropping the attribute fails both.

Two smaller review items came with it: the fourth corner of #472's matrix (expanded
and current) is now covered, a real dblclick is asserted to dispatch one repeated
command, and a CSS comment that called a three-class selector "equal-specificity" is
corrected.

Acceptance criteria claimed

• The Dashboard row's chevron, name and trailing action are three independent targets,
  each keyboard-operable and separately announced (Separate Dashboard-tree expansion from opening the Dashboard #472).
• Action buttons do not interfere with tree navigation gestures — for every control that
  exists after this phase (chevron, ⋯, the Replace curated Dashboard filters with inferred Variables and batched option queries #447 orphan trash). Phases 4/5 re-assert it as
  they add the pencil and the panel trash.

Deferred by design: the Dashboard pencil and the one create path (phase 3), panel
metadata edit (4), atomic panel+query delete and ownership (5), the repair planner (6).
The ## Regressions bullet "row click/double-click/Shift-click contract remains
unchanged" is scoped to the rows this issue does not revise — #472's own section
overrides it for Dashboard rows, and panel/variable/group rows are genuinely untouched.
The non-goal "Moving or redesigning the trailing Dashboard action menu as a control" is
respected: the control is unchanged, only one item of its contents went.

#438 / layout engines: phase 2 adds no tile-focus gesture — a Dashboard row navigates
to a Dashboard, never to a member — so there was nothing to verify on the flow engine. The
obligation stands for phases 4/5.

Verification

• npm test — 5827 passed / 177 files, per-file gate held; tsc --noEmit clean;
  npm run build clean.
• npm run test:e2e — 442 passed, 0 failed on chromium + webkit + firefox.
  (Firefox launches here again after npx playwright install firefox; the dependabot
  Playwright bump had left all three browser binaries stale, which reads as ~435 failures
  at 1ms each.)
• An independent read-only audit against the issue's phase-2 contract came back clean on
  under-delivery, scope creep, non-goal breaches and rule-softening; its one substantive
  finding is the accessible-name regression fixed above.
• Sabotage-checked, six ways: letting Enter/Space bubble past the chevron, dropping the
  chevron focus restore, not moving the keyboard owner, making every chevron tabbable,
  reverting the model split, and restoring Open in View — each is caught by 1–6 tests and
  nothing else. The repeated-open history guard and the chevron box were sabotage-checked
  too, as was the accessible name.
• Driven in the real served app (not the e2e fixture) against a live ClickHouse: the
  chevron is a BUTTON labelled Expand Sales revenue with aria-expanded=false and
  tabindex=0; Tab walks row → chevron → ⋯; the three focus treatments are a box-shadow
  ring, a 2px accent outline and the trailing button's own; a chevron click expands 2 → 4
  rows without touching the URL; a name click routes to
  ?surface=dashboard&mode=view. No page errors.

Checklist

• npm test passes (the per-file coverage gate is non-negotiable)
• Tests added/updated in the same change as the code
• npm run build succeeds (single-file dist/sql.html)
• Layers kept honest: pure logic in src/core/, network in src/net/ (injected fetch), DOM in src/ui/
• No new runtime dependency (or it's a deliberate, justified addition — see CONTRIBUTING)
• README / CHANGELOG.md ([Unreleased]) updated if behavior or the deployed surface changed
• Reconciled affected tracked work (roadmap Roadmap to 1.0.0 #68, the issue body, ADR/CHANGELOG) if this change reshaped it

🤖 Generated with Claude Code

https://claude.ai/code/session_018mujm1kW7jDGpTEfscndcU