Decision card: presence faces and tooltips

[MaggieAppleton]

Stacked on #231 (base design/decision-card-restyle). Retarget to main after #231 merges.

Summary

The @handle pills in the decision card header become a face stack with design-system tooltips.

• QuestionView gains renderPeople(people); without it the old pills still render. The editor host supplies PresenceFaces: 24px Face (ring="page"), overlapping by 0.375rem, deduped by handle (case-insensitive), max 3 then +N (--text-xs 600 tertiary), centred on the 1.5rem mark row.
• IconTooltip now works on any element with data-tooltip (not only buttons); data-tooltip-verbatim skips capitalisation so handles keep their casing. Faces tooltip the handle, +N lists the hidden handles, comma separated.
• Face: radius scales with size (18px and under --radius-sm, larger --radius-md), and a new titled={false} prop suppresses the native title where a tooltip is used (no DOM stripping).
• Room header presence faces get the same tooltip (trivial, same prop).

Decisions made

• The group is labelled "Editing this question: a, b, c" and lists every person including those behind +N; the +N itself is not a separate a11y node beyond its tooltip.
• Dedupe is case-insensitive and keeps the first spelling.
• +N counts unique hidden handles, not connections.
• Radius threshold: size <= 18 is small. Implemented as a static class map (FACE_RADIUS_CLASS) to avoid a new dynamic design-contract exception; sourceHash renewed for face.tsx, question-view.tsx, questionnaire.tsx (no new dynamic data flow).
• Tooltip pure logic (tooltipText) is unit tested; hover/DOM behaviour is in Playwright since bun test has no DOM.
• Existing gap, not fixed here: no production client currently sends question:presence with a question (only clears it), so faces only appear when such frames arrive. The e2e injects peer presence frames through routeWebSocket. Wiring focus-to-presence is a separate change.

Before / after (2x)

	Before	After
1 person
5 people
+N tooltip

Checks

• bun run fix, bun test (1741 pass, 2 skip, 0 fail), bun run types (exit 0), bun run ci (exit 0).
• E2E (--project=fixtures): sidecar.e2e.ts (38 passed, including the new face tooltip test), plus responsive-decisions, collab, toolbar (40 passed in the combined run). One responsive-decisions 390x844 run showed a transient strict-mode duplicate card that did not reproduce on rerun.

🤖 Generated with Claude Code

[MaggieAppleton]

Review and fixes (pushed to this branch).

Gap closed: faces now appear in real sessions. QuestionView gains onQuestionFocus(question | undefined): fired on focus within the current question (choosing an option or typing both focus a control), cleared on blur out of the question, question change, and unmount. QuestionnaireController.focusQuestion feeds a pure FocusReporter that sends question:presence once per change (no per-keystroke traffic), waits until the draft is open, resends after a reconnect, and clears on submit, cancel and stop. The existing protocol frame and server relay were sufficient; no protocol change.

Tests. Unit: FocusReporter dedupe/online/reconnect, controller presence frames. E2E face test now uses four real browser contexts (Bo, cy, Di, ed) focusing the question, asserts three faces + +1, verbatim tooltips, no native title, clearing on blur and on a peer closing. Flake fix: the tooltip hides on scroll and scroll events arrive a frame late, so the test settles the page (scrollIntoView + two frames) and arrives with stepped mouse movement; passed 4/4 repeats plus full runs.

Review notes, no change needed. Tooltip is supplementary: the group's aria-label names everyone (including those behind +N) and faces keep img names; non-button tooltip targets are not focusable, which is acceptable as the information is not tooltip-only. Tokens/static class maps OK. Design-contract hashes renewed (no new dynamic data flow).

Checks: fix, test (1745 pass), types, ci all green; e2e fixtures sidecar + responsive-decisions + collab: 41 passed after merging #231. Unrelated: responsive-decisions 320x568 once hit a transient strict-mode duplicate card (two step panels mid-swap), not reproduced on rerun.