Echo loop fix: asistan sesi yeni turn tetiklemesin

- VAD threshold 0.5 -> 0.7 (daha kati: hoparlor sesi user gibi algilanmasin)
- silence_duration 700ms -> 1200ms (daha uzun bekleme)
- Frontend gate: asistan konusurken mikrofon audio'su OpenAI'a gonderilmez

realtime_relay.py

@@ -118,9 +118,9 @@ def _session_update_payload() -> dict:
                     "transcription": {"model": "gpt-4o-mini-transcribe", "language": "tr"},
                     "turn_detection": {
                         "type": "server_vad",
-                        "threshold": 0.5,
+                        "threshold": 0.7,            # Daha katı (0.5 -> 0.7) — asistan sesi user gibi algılanmasın
                         "prefix_padding_ms": 300,
-                        "silence_duration_ms": 700,
+                        "silence_duration_ms": 1200, # Sessizlik daha uzun (700 -> 1200ms)
                         "interrupt_response": False,
                         "create_response": True,
                     },

static/index.html

@@ -246,6 +246,8 @@ async function connect() {
     workletNode = new AudioWorkletNode(audioCtx, 'pcm-processor');
     workletNode.port.onmessage = (e) => {
       if (ws?.readyState !== WebSocket.OPEN) return;
+      // Asistan konusurken mikrofon audio'sunu OpenAI'a GONDERME — echo-feedback'i onler
+      if (assistantSpeaking) return;
       const b64 = arrayBufferToBase64(e.data.pcm);
       ws.send(JSON.stringify({ type: 'input_audio_buffer.append', audio: b64 }));
     };