feat: call UX overhaul — deafen sync, share dialog, fullscreen redesign

Speaking ring:
- Switch useActiveSpeakers from LiveKit's smoothed isSpeaking / server-
  batched ActiveSpeakersChanged to Web Audio API AnalyserNode on each
  participant's raw audio MediaStreamTrack. Poll 50ms, RMS threshold 0.03,
  250ms hold. Feels real-time vs the old ~500ms lag
- Defensive syncProbe on every tick so probes catch up if TrackPublished
  missed (local mic publish race on join)
- Universal speaking overlay on tile (3px emerald border + inset glow,
  z-10) so video mode shows the ring too, not just audio mode

Screen sharing:
- Separate "screen" tile per sharer so the sharer's avatar tile stays
  intact with its speaking ring. Tile.id is kind-prefixed (user:xxx /
  screen:xxx) so focus tracking distinguishes them
- New ScreenShareDialog (quality preset + fps override + displaySurface
  hint) opens on the share button. startScreenShare / stopScreenShare
  actions in CallContext replace the one-shot toggle
- ScreenShareViewer: plain CSS-only fullscreen overlay (Tauri WKWebView
  doesn't implement requestFullscreen), always `h-full w-full
  object-contain`, Esc exits

Camera:
- toggleCamera action in CallContext tracks isCameraEnabled
- VideoStub renders real <video> srcObject for the participant's camera
  MediaStreamTrack; local preview is mirrored
- Tile video track resolves to Track.Source.Camera publications of the
  LocalParticipant / each RemoteParticipant
- Room listens for TrackMuted / TrackUnmuted and re-publishes remote
  state so peers switch to avatar placeholder when a camera is disabled

Deafen:
- New isDeafened state + toggleDeafen action. Sets `muted = true` on all
  attached `<audio[data-livekit-track]>` plus mutes fresh ones on attach
  via module-level flag
- Broadcast state over the LiveKit data channel
  ({type:'presence', deafened}) so peers can render the headphones-off
  badge. Attributes API not used because the self-hosted server may run
  older LiveKit versions
- remoteDeafen: Record<identity, bool> exposed via context, bumped on
  DataReceived and re-broadcast on ParticipantConnected

Incoming video call:
- acceptIncoming takes an optional CallKind override so the receiver can
  answer a video invite with audio only or promote an audio invite to
  video on accept
- IncomingCallPanel shows two accept buttons (audio + video) when the
  invite is a video call

Audio devices:
- audioSettings adds inputDeviceId + outputDeviceId, persisted
- CallContext uses them on setMicrophoneEnabled, plus new
  setAudioInputDevice / setAudioOutputDevice hot-swap actions.
  Output swap applies setSinkId to every attached remote-audio element
  since LiveKit's own switchActiveDevice only tracks elements it
  attached itself
- SettingsPage "Mikrofon" + "Ausgabegerät" selects with devicechange
  listener and a permission-probe button

Fullscreen mode:
- Replaced absolute-positioned speaker + floating thumbnails with a real
  flex layout. Default = even grid of all tiles. Clicking a tile flips
  to big-speaker + horizontal thumbnail strip. Click focused tile =
  back to grid
- Controls overlay pinned bottom; content wrapper has pb-24 so tiles
  never sit behind the toolbar
- Grid now uses explicit grid-rows-* so cells get a defined 1fr height
  (without it, video intrinsic dimensions blew tiles past the container
  bounds on Windows)

UI chips:
- Mic-off badge combines isMuted flag AND
  localParticipant.isMicrophoneEnabled, so a user with no mic / denied
  permission sees the badge + the toolbar button red even though they
  never pressed mute
- Deafen badge on tile chips for local + remote (remote driven by the
  data-channel broadcast)

apps/desktop/src/lib/useActiveSpeakers.ts

@@ -1,10 +1,83 @@
-import type { Participant, Room } from 'livekit-client';
-import { RoomEvent } from 'livekit-client';
+import type { AudioTrack, Participant, Room } from 'livekit-client';
+import { ParticipantEvent, RoomEvent, Track } from 'livekit-client';
 import { useEffect, useState } from 'react';
 
-// Subscribe to LiveKit ActiveSpeakersChanged to drive the speaking-ring pulse
-// on participant tiles. Returns the set of currently speaking identities
-// (includes local participant when they're active).
+// Real-time speaking ring driven by Web Audio API AnalyserNodes directly on
+// each participant's audio MediaStreamTrack. LiveKit's own `audioLevel` +
+// `isSpeaking` are updated by a background monitor (default ~1s) — too
+// laggy. AnalyserNodes give us raw PCM at browser frame rate, so the ring
+// lights up within one animation frame of actual speech.
+//
+// Hold time of 250ms prevents flicker between words / short pauses.
+const POLL_MS = 50;
+const HOLD_MS = 250;
+const THRESHOLD = 0.03; // RMS on 0..1 — tuned against soft speech
+const FFT_SIZE = 256;
+
+interface Probe {
+  ctx: AudioContext;
+  analyser: AnalyserNode;
+  source: MediaStreamAudioSourceNode;
+  buf: Uint8Array;
+  // Cached MediaStreamTrack reference — if a publication swaps tracks
+  // (mute/unmute, device switch) we rebuild the node chain.
+  trackId: string;
+}
+
+function makeProbe(track: MediaStreamTrack): Probe | null {
+  try {
+    const ctx = new (window.AudioContext || (window as unknown as { webkitAudioContext: typeof AudioContext }).webkitAudioContext)();
+    const stream = new MediaStream([track]);
+    const source = ctx.createMediaStreamSource(stream);
+    const analyser = ctx.createAnalyser();
+    analyser.fftSize = FFT_SIZE;
+    analyser.smoothingTimeConstant = 0.2;
+    source.connect(analyser);
+    return {
+      ctx,
+      analyser,
+      source,
+      buf: new Uint8Array(analyser.frequencyBinCount),
+      trackId: track.id,
+    };
+  } catch {
+    return null;
+  }
+}
+
+function destroyProbe(probe: Probe): void {
+  try {
+    probe.source.disconnect();
+  } catch {
+    /* ignore */
+  }
+  try {
+    void probe.ctx.close();
+  } catch {
+    /* ignore */
+  }
+}
+
+function sampleRms(probe: Probe): number {
+  // `getByteFrequencyData` types expect `Uint8Array<ArrayBuffer>` (no
+  // SharedArrayBuffer). Our `probe.buf` satisfies that at runtime; the cast
+  // sidesteps the TS lib signature quirk.
+  probe.analyser.getByteFrequencyData(probe.buf as Uint8Array<ArrayBuffer>);
+  let sum = 0;
+  for (let i = 0; i < probe.buf.length; i++) sum += probe.buf[i]!;
+  return sum / (probe.buf.length * 255);
+}
+
+function firstAudioTrack(p: Participant): AudioTrack | null {
+  const pubs = p.audioTrackPublications;
+  for (const pub of pubs.values()) {
+    if (pub.kind === Track.Kind.Audio && pub.track) {
+      return pub.track as AudioTrack;
+    }
+  }
+  return null;
+}
+
 export function useActiveSpeakers(room: Room | null): Set<string> {
   const [ids, setIds] = useState<Set<string>>(() => new Set());
 
@@ -14,20 +87,84 @@ export function useActiveSpeakers(room: Room | null): Set<string> {
       return;
     }
 
-    const update = (speakers: Participant[]) => {
-      const next = new Set<string>();
-      for (const s of speakers) {
-        if (s.identity) next.add(s.identity);
+    const probes = new Map<string, Probe>();
+    const lastActive = new Map<string, number>();
+
+    const syncProbe = (p: Participant): void => {
+      if (!p.identity) return;
+      const track = firstAudioTrack(p);
+      const mst = track?.mediaStreamTrack ?? null;
+      const existing = probes.get(p.identity);
+      if (!mst) {
+        if (existing) {
+          destroyProbe(existing);
+          probes.delete(p.identity);
+        }
+        return;
       }
-      setIds(next);
+      if (existing && existing.trackId === mst.id) return;
+      if (existing) destroyProbe(existing);
+      const probe = makeProbe(mst);
+      if (probe) probes.set(p.identity, probe);
     };
 
-    // Seed from current state.
-    update(room.activeSpeakers ?? []);
+    const bind = (p: Participant): void => {
+      p.on(ParticipantEvent.TrackPublished, () => syncProbe(p));
+      p.on(ParticipantEvent.TrackUnpublished, () => syncProbe(p));
+      p.on(ParticipantEvent.TrackSubscribed, () => syncProbe(p));
+      p.on(ParticipantEvent.TrackUnsubscribed, () => syncProbe(p));
+      p.on(ParticipantEvent.TrackMuted, () => syncProbe(p));
+      p.on(ParticipantEvent.TrackUnmuted, () => syncProbe(p));
+      syncProbe(p);
+    };
+
+    bind(room.localParticipant);
+    room.remoteParticipants.forEach(bind);
+
+    const onConnected = (p: Participant) => bind(p);
+    const onDisconnected = (p: Participant) => {
+      if (!p.identity) return;
+      const probe = probes.get(p.identity);
+      if (probe) {
+        destroyProbe(probe);
+        probes.delete(p.identity);
+      }
+      lastActive.delete(p.identity);
+    };
+    room.on(RoomEvent.ParticipantConnected, onConnected);
+    room.on(RoomEvent.ParticipantDisconnected, onDisconnected);
+
+    const tick = () => {
+      // Defensively resync every tick — covers the gap where local mic
+      // finishes publishing between effect mount and the first
+      // TrackPublished event, and handles event misses on Tauri WebView.
+      syncProbe(room.localParticipant);
+      room.remoteParticipants.forEach(syncProbe);
+
+      const now = Date.now();
+      for (const [id, probe] of probes) {
+        if (sampleRms(probe) > THRESHOLD) lastActive.set(id, now);
+      }
+      const next = new Set<string>();
+      for (const [id, t] of lastActive) {
+        if (now - t <= HOLD_MS) next.add(id);
+      }
+      setIds((prev) => {
+        if (prev.size === next.size && Array.from(prev).every((v) => next.has(v))) {
+          return prev;
+        }
+        return next;
+      });
+    };
+
+    const interval = window.setInterval(tick, POLL_MS);
 
-    room.on(RoomEvent.ActiveSpeakersChanged, update);
     return () => {
-      room.off(RoomEvent.ActiveSpeakersChanged, update);
+      window.clearInterval(interval);
+      room.off(RoomEvent.ParticipantConnected, onConnected);
+      room.off(RoomEvent.ParticipantDisconnected, onDisconnected);
+      for (const probe of probes.values()) destroyProbe(probe);
+      probes.clear();
     };
   }, [room]);