feat: add video support for intercom lines (WHIP/WHEP, H264)

[bjowe]

Adds in-call participant video: publish the local camera via WHIP and consume remote video via WHEP (H264/VP8 over SMB), with self-preview, per-tile video grid, source pinning, return-feed and fullscreen controls, plus the video-enabled toggle in production setup.

Closes #692

Companion PR (backend): Eyevinn/intercom-manager#315

[birme]

Automated code-reviewer verdict (daily-backlog-pr Phase 3): NEEDS CHANGES

Reviewed the full +4752/−261 diff across 69 files with the branch checked out. Feature-scoped work that mirrors the existing audio patterns well, with good coverage of the new pure logic. Two concrete media-lifecycle bugs should be fixed before shipping a camera feature:

High

1. Camera capture is never stopped on call exit (privacy/hardware leak). production-line.tsx:149 destructures only const [inputVideoStream] = useVideoInput({...}), discarding the third tuple element — the reset() that stops the camera tracks. useVideoInput.reset (use-video-input.ts:67-72) is the only place videoInput.getTracks().forEach(t => t.stop()) runs, and it's never called. The effect cleanup (:62-64) only sets aborted = true; REMOVE_CALL (global-state-reducer.ts:123) doesn't stop mediaStreamVideoInput; and rtcPeerConnection.close() doesn't stop local getUserMedia tracks. Net: camera stays live + recording indicator on after leaving a video line, until the tab closes. Capture and call the video reset on exit/unmount.
2. Frame-monitor setIntervals leak on unmount. attachShowWhenReady (video-element-factory.ts:470-490) starts a 500ms interval per remote video element, cleared only by stopFrameMonitor, which is called only from removeThisElement on track ended/removetrack. useVideoElements cleanup (use-video-elements.ts:11-24) nulls srcObject but doesn't call stopFrameMonitor, nor does the useRtcConnection teardown. On abrupt teardown/navigation the interval + rVFC loop keep running against detached elements. Call stopFrameMonitor for every element in cleanUpVideo.

Medium

3. checkbox.tsx:76-87 — input made readOnly + tabIndex={-1} with toggle delegated to wrapper onClick, synthesizing a fake ChangeEvent via as unknown as. Removes keyboard focus/space-to-toggle from a shared primitive (a11y regression + scope creep). Prefer a native disabled prop.
4. PC held in useState (never recreated) but stream refs are in the effect deps — a stream-ref change would close() the PC and the re-run no-ops behind signalingState === "closed" guards, dead-ending the call. Latent today (no device-switch re-acquire); add a guard/comment.
5. No test asserting camera tracks stop / frame monitors clear on exit — exactly where the leaks live.

Nits

• Raw console.warn/error in prod paths vs the project logger util; video-element-factory.ts builds ~540 lines of tile chrome via imperative createElement+inline styles vs Emotion styled() used elsewhere.

Notes: no user-input URL construction (all via backend API_URL), no new bundle secrets, reducer ERROR change well-contained, CI green. Both High items are small localized fixes.

[birme]

This PR looks like part of a multi-repo/multi-PR dependency stack together with Eyevinn/intercom-manager#315 (backend video support). Holding off on automated review/merge here — this needs a human (or the relevant implementation agent) to assess the whole stack together, not a per-PR pass.