[eric] dictation: the mic warms on hover so the first press of a launch is not cold (ENG-300)

frontend/src/app/pages/Dashboard/desktop/DesktopSpawnPill.tsx

@@ -40,7 +40,7 @@ function DesktopSpawnPill({
   const [menuOpen, setMenuOpen] = useState(false);
   const newAgentShortcut = useAppSelector((s) => s.settings.data.new_agent_shortcut);
   const rootRef = useRef<HTMLDivElement | null>(null);
-  const { state: voiceState, pct: voicePct, pressStart: voicePressStart, pressEnd: voicePressEnd } = useVoice();
+  const { state: voiceState, pct: voicePct, pressStart: voicePressStart, pressEnd: voicePressEnd, prewarm: voicePrewarm } = useVoice();
   const recording = voiceState === 'recording';
   const transcribing = voiceState === 'transcribing';
   const preparing = voiceState === 'preparing';
@@ -166,7 +166,10 @@ function DesktopSpawnPill({
           <Box
             role="button"
             aria-label="Voice dictation"
-            onPointerDown={(e) => { e.stopPropagation(); if (!voiceBusy) voicePressStart(); }}
+            // Open the mic on hover so the FIRST press of a launch is warm. Cold it costs ~2.5s and
+              // silently eats what you say, which reads as "the button did nothing" (ENG-300).
+              onPointerEnter={() => { if (!voiceBusy) voicePrewarm(); }}
+              onPointerDown={(e) => { e.stopPropagation(); if (!voiceBusy) voicePressStart(); }}
             onPointerUp={(e) => { e.stopPropagation(); voicePressEnd(); }}
             onPointerLeave={() => voicePressEnd()}
             onClick={(e) => e.stopPropagation()}

frontend/src/shared/voice/VoiceDictationContext.tsx

@@ -12,7 +12,7 @@ import VoiceOverlay from './VoiceOverlay';
 // out-of-sync state. Mounted once near the app root.
 
 export function VoiceDictationProvider({ children }: { children: React.ReactNode }): React.ReactElement {
-  const { state, lastText, error, pct, feedback, partial, target, toggle, start, stop, cancel, notify, volumeRef } = useVoiceDictation();
+  const { state, lastText, error, pct, feedback, partial, target, toggle, start, stop, cancel, notify, prewarm, volumeRef } = useVoiceDictation();
 
   // fn is the dictation key, but macOS may still have its own Globe action bound (emoji picker on a quick tap); say so once.
   const globeWarnedRef = useRef(false);
@@ -136,7 +136,7 @@ export function VoiceDictationProvider({ children }: { children: React.ReactNode
   const confirmRecording = useCallback((): void => { void stop(); }, [stop]);
 
   return (
-    <VoiceContext.Provider value={{ state, lastText, error, pct, feedback, partial, target, toggle, pressStart, pressEnd, confirmRecording, cancelRecording: cancel, holdMode, volumeRef }}>
+    <VoiceContext.Provider value={{ state, lastText, error, pct, feedback, partial, target, toggle, prewarm: () => { void prewarm(); }, pressStart, pressEnd, confirmRecording, cancelRecording: cancel, holdMode, volumeRef }}>
       {children}
       <VoiceOverlay />
     </VoiceContext.Provider>

frontend/src/shared/voice/useVoiceDictation.ts

@@ -169,6 +169,25 @@ export function useVoiceDictation() {
 
   useEffect(() => () => releaseWarmMic(), [releaseWarmMic]);
 
+  // ENG-300: the park above only helps AFTER a first session, so the FIRST press of every launch
+  // paid the full ~2.5s cold open and dropped whatever the user said into it. Measured by Eric: the
+  // button "never turns on until I unclick it", every time, right after startup. Arming on hover
+  // makes that first press warm like the rest, and costs nothing until the user reaches for the
+  // control, which is why this is not done at boot: an always-hot mic lights the OS indicator for
+  // people who never dictate.
+  const prewarm = useCallback(async (): Promise<void> => {
+    if (warmMicRef.current || stateRef.current !== 'idle') return;
+    try {
+      const stream = await navigator.mediaDevices.getUserMedia({
+        audio: { channelCount: 1, echoCancellation: true, noiseSuppression: false, autoGainControl: false },
+      });
+      parkWarmMic(stream, new AudioContext({ sampleRate: VOICE_SAMPLE_RATE }));
+    } catch (_) {
+      // No permission yet, or no device. Stay silent: the real press is what should surface that,
+      // and a hover must never raise an error the user did not ask for.
+    }
+  }, [parkWarmMic]);
+
   const teardown = useCallback((): Float32Array | null => {
     const rec = recRef.current;
     recRef.current = null;
@@ -460,5 +479,5 @@ export function useVoiceDictation() {
     setFeedback({ tone: 'warn', icon: 'info', text, at: Date.now() });
   }, []);
 
-  return { state, lastText, error, pct, feedback, partial, target, toggle, start, stop, cancel, notify, volumeRef };
+  return { state, lastText, error, pct, feedback, partial, target, toggle, start, stop, cancel, notify, prewarm, volumeRef };
 }

frontend/src/shared/voice/voiceContext.ts

@@ -13,6 +13,9 @@ export interface VoiceContextValue {
   // Where the transcript will land right now, in user words plus the surface's icon.
   target: InjectTargetInfo;
   toggle: () => void;
+  // Open and park the mic before the first press, so it is not cold when the user actually
+  // clicks (ENG-300). Safe to call repeatedly; a no-op once armed or while recording.
+  prewarm: () => void;
   // Mic-button press semantics that respect the hold/toggle setting: press starts (or toggles),
   // release stops only in hold mode. Buttons wire onPointerDown/Up to these and stay mode-agnostic.
   pressStart: () => void;
@@ -25,7 +28,7 @@ export interface VoiceContextValue {
 }
 
 const NOOP_REF = { current: 0 };
-const NOOP: VoiceContextValue = { state: 'idle', lastText: '', error: null, pct: 0, feedback: null, partial: null, target: { label: '', icon: null, composerId: null }, toggle: () => {}, pressStart: () => {}, pressEnd: () => {}, confirmRecording: () => {}, cancelRecording: () => {}, holdMode: true, volumeRef: NOOP_REF };
+const NOOP: VoiceContextValue = { state: 'idle', lastText: '', error: null, pct: 0, feedback: null, partial: null, target: { label: '', icon: null, composerId: null }, toggle: () => {}, prewarm: () => {}, pressStart: () => {}, pressEnd: () => {}, confirmRecording: () => {}, cancelRecording: () => {}, holdMode: true, volumeRef: NOOP_REF };
 
 export const VoiceContext = createContext<VoiceContextValue>(NOOP);