fix: replace PCM16 capture with MediaRecorder (Opus/webm)

PCM16 capture via AudioContext was streaming raw audio continuously,
causing massive accumulated buffers that took ~20s to transcribe.
Replaced with MediaRecorder which records compressed Opus/webm and
sends a single blob on release — much smaller, faster to transcribe.

Also removed all unused PCM16/WAV helper functions from both frontend
and backend.

backend/main.py

@@ -96,17 +96,12 @@ async def run_conversation(text: str, user_id: str) -> str:
 
 
 async def transcribe_audio(audio_bytes: bytes) -> str | None:
-    """Transcribe audio bytes using cheapest STT model.
-    
-    Accepts raw PCM16 mono 24kHz data — wraps in WAV container automatically.
-    """
+    """Transcribe Opus/webm audio using cheapest STT model."""
     client = get_openai()
     try:
-        # Wrap raw PCM16 in WAV container for the API
-        wav_bytes = _pcm16_to_wav(audio_bytes)
         transcript = await client.audio.transcriptions.create(
             model="gpt-4o-mini-transcribe",
-            file=("audio.wav", wav_bytes, "audio/wav"),
+            file=("audio.webm", audio_bytes, "audio/webm"),
             response_format="text",
         )
         return transcript.strip() if transcript and transcript.strip() else None
@@ -211,8 +206,8 @@ async def conversation_ws(websocket: WebSocket):
                 continue
 
             # ── Conversation ──
-            if msg_type == "audio":
-                # Accumulate PCM16 audio chunks
+            if msg_type == "audio_chunk":
+                # Single Opus/webm blob from MediaRecorder
                 chunk = base64.b64decode(msg["data"])
                 audio_buffer.extend(chunk)