R8a: Escape scoping and focus-trap hardening (findings #16, #17)

[dovvnloading]

Problem

Two related overlay-system bugs from the R8a UI/UX audit:

• Finding Refactor navigation pins and interaction surfaces #16 (Escape scoping): the document-level Escape-to-close listener ran in the capture phase and unconditionally called stopPropagation(), so it fired and closed the surface before any inner field's own Escape handler could run. Rename inputs, inline-edit textareas, and the Pins editor all had Escape silently discard their own local edit and close (or in PinOverlay's case, had no Escape handling at all — Escape closed the whole popover).
• Finding Refactor navigation pins and interaction surfaces #17 (focus trap): the Tab-trap's FOCUSABLE selector counted disabled buttons as valid stops. When the true last focusable element was disabled (e.g. Approve/Deny both disabled while busy), the wrap-boundary check never matched the real last element, and Tab walked focus straight out of the panel.

Change

• overlays.tsx: moved the Escape listener to the bubble phase and gated it on event.defaultPrevented, so an inner handler can claim Escape via preventDefault() — the standard DOM convention — without the outer surface closing underneath it.
• Applied the same preventDefault() fix to NoteNodeView, GroupNodeView (inline label editors), and PinOverlay (pin title/note editor, which had no Escape handling before this).
• overlays.tsx and CodeExecutionApprovalPanel.tsx: excluded [disabled] elements from the FOCUSABLE selector.
• Added a Tab-gated focus-escape backstop as defense in depth (useFocusEscapeBackstop, shared between Dialog and CodeExecutionApprovalPanel): if focus lands outside the panel immediately after a Tab keydown, it's redirected back in. Gated on the preceding key being Tab so it can't fight a dialog's own legitimate close()-triggered focus restoration, and so two simultaneously-open panels don't redirect focus into each other indefinitely.

Two related-but-distinct issues were flagged during review and are intentionally out of scope here: CommandPalette claims a focus trap it doesn't actually implement, and NodeMenu's global capture-phase Escape handler has no target check and can swallow unrelated inline edits elsewhere on the canvas. Both should be tracked as separate follow-ups.

Test plan

Layer	Result
npx tsc --noEmit	clean
npx vitest run (full suite)	868/868 passed
python -m pytest -q (backend, unaffected)	1033/1033 passed
npm run lint / npm run build	clean

New/updated coverage: overlays.test.tsx (Escape-claim scoping, disabled-button wrap, Tab-gated backstop, close()-restore race, mouse-click-after-Tab reset), NoteNodeView.test.tsx / GroupNodeView.test.tsx (Escape claims the event), ChatLibraryDialog.test.tsx (corrected a stale test that had codified the old buggy behavior as intended), PinOverlay.test.tsx (Escape cancels edit without closing the popover), CodeExecutionApprovalPanel.test.tsx (disabled-button Tab wrap, Tab-gated backstop).

Live-browser verification of PinOverlay was attempted but blocked by the preview pane's non-composited state, which prevents React Flow's viewport from initializing (a known environment limitation, unrelated to this change); the fix is covered by real-DOM, real-React-reconciliation unit tests instead.