Deploy: Stack A (NIM brain + Maverick judge + Sarvam voice). D-019.

frontend/src/app/page.tsx

@@ -442,7 +442,11 @@ export default function Page() {
       // Small wait so useLiveConversation tears down its stream before we
       // open a fresh one; without this, two AudioContexts can briefly grab
       // the same input device.
-      await new Promise((r) => setTimeout(r, 120));
+      // KI-134 (2026-05-15) — bumped 120ms → 400ms because AudioContext.close()
+      // returns a Promise that's discarded; on HF Space hardware the previous
+      // context can still hold the device when getUserMedia fires, producing
+      // a silent stream with no error.
+      await new Promise((r) => setTimeout(r, 400));
     }
     try {
       const stream = await navigator.mediaDevices.getUserMedia({ audio: true });
@@ -461,7 +465,14 @@ export default function Page() {
         // does its one turn and leaves Live off. They'll keep using PTT
         // until they click the dot back to green themselves.
         const maybeResumeLive = () => { if (userPrefersLive) live.setLive(true); };
-        if (blob.size < 1000) { maybeResumeLive(); return; }
+        // KI-134 (2026-05-15) — surface the silent-recording case to the user
+        // instead of returning quietly. Previously, holding PTT briefly and
+        // releasing produced no feedback at all; now they at least see why.
+        if (blob.size < 1000) {
+          pushAssistant("Didn't catch any audio — try holding the mic button while speaking.");
+          maybeResumeLive();
+          return;
+        }
         setBusy(true);
         setVoicePhase("transcribing"); // KI-038 — STT in flight on PTT
         try {

frontend/src/lib/api.ts

@@ -131,13 +131,18 @@ export async function postTranscribe(
   const fd = new FormData();
   // Use blob's mime to derive extension; default to wav
   const mime = blob.type || "audio/wav";
+  // KI-134 (2026-05-15) — iOS Safari MediaRecorder produces audio/mp4 (no
+  // webm support). Without explicit mapping the file was sent as audio.wav
+  // with mp4 bytes inside, breaking the backend's mime/ext whitelist.
   const ext = mime.includes("webm")
     ? "webm"
-    : mime.includes("mp3")
-      ? "mp3"
-      : mime.includes("ogg")
-        ? "ogg"
-        : "wav";
+    : mime.includes("mp4") || mime.includes("m4a")
+      ? "m4a"
+      : mime.includes("mp3")
+        ? "mp3"
+        : mime.includes("ogg")
+          ? "ogg"
+          : "wav";
   fd.append("file", blob, `audio.${ext}`);
   if (language_code) fd.append("language_code", language_code);
 

frontend/src/lib/useLiveConversation.ts

@@ -119,12 +119,12 @@ const DEFAULTS = {
   // segment. Avoids bot's TTS attack transient bleeding through even
   // with echoCancellation on.
   postUtteranceCooldownMs: 700,
-  // KI-113 — raised 0.35 → 0.50 minimum fraction of total FFT energy
-  // that must sit in the voice band (bins 2-22 ≈ 190-2150 Hz at 48 kHz).
-  // Broadband HVAC / fan / traffic noise typically scores 0.20-0.30;
-  // voiced speech typically scores 0.40-0.70. Lifting to 0.50 puts the
-  // gate squarely above noise and just below the bottom of speech.
-  voiceBandMinProp: 0.50,
+  // KI-113 raised 0.35 → 0.50 to gate out broadband HVAC / fan / traffic.
+  // KI-134 (2026-05-15) — backed off to 0.35 because some laptop built-in
+  // mics with aggressive noiseSuppression flatten the voice-band proportion
+  // to 0.35-0.45, so the VAD never opens and the green pill renders but
+  // no audio is ever posted — matches the live-test symptom exactly.
+  voiceBandMinProp: 0.35,
 };
 
 // AudioWorklet processor source — inlined as a Blob URL so we don't need
@@ -489,6 +489,24 @@ export function useLiveConversation(opts: LiveConversationOptions): LiveConversa
           window.AudioContext;
         const ctx = new AudioCtx();
         audioCtxRef.current = ctx;
+        // KI-134 (2026-05-15) — Chrome/Safari autoplay policy starts new
+        // AudioContexts in state='suspended' when the click that toggled
+        // Voice on has already been consumed by React state propagation.
+        // Without resume(), the worklet's process() never runs, no PCM
+        // frames arrive, and the green pill renders forever with zero
+        // audio posted. This is THE canonical "voice on but nothing
+        // happens" trap. Resume + bail visibly if the context refuses.
+        if (ctx.state === "suspended") {
+          try {
+            await ctx.resume();
+          } catch (e) {
+            // eslint-disable-next-line no-console
+            console.error("[live-mode] AudioContext resume failed", e);
+            setMicPermissionDenied(true);
+            setLive(false);
+            return;
+          }
+        }
         sampleRateRef.current = ctx.sampleRate;
 
         const source = ctx.createMediaStreamSource(stream);