feat: 운영 안정성과 세션 음성 경험 개선

apps/api/app/routes/voice.py

@@ -158,14 +158,26 @@ _PROVIDER_EVENT_TYPE_FIELDS = ("event_type", "type", "kind", "label")
 async def voice_health() -> JSONResponse:
     """Return voice service readiness."""
     available = voice_service.is_available()
+    stt_available = voice_service.stt_available()
+    tts_available = voice_service.tts_available()
     tts_provider = voice_service.tts_provider()
     body = {
         "status": "ok" if available else "degraded",
         "available": available,
+        "stt_available": stt_available,
+        "tts_available": tts_available,
         "stt_model": voice_svc.STT_MODEL,
-        "tts_model": voice_svc.TTS_MODEL,
+        "tts_model": (
+            voice_svc.HIGGS_TTS_MODEL
+            if tts_provider == "higgs"
+            else voice_svc.TTS_MODEL
+        ),
         "tts_provider": tts_provider,
-        "reason": None if available else "OPENAI_API_KEY is not configured",
+        "reason": (
+            None
+            if available
+            else "STT 또는 TTS provider가 준비되지 않았습니다"
+        ),
     }
     return JSONResponse(body, status_code=200 if available else 503)
 
@@ -192,12 +204,6 @@ async def voice_speech(
     access_error = await _practice_access_error(learner)
     if access_error is not None:
         raise HTTPException(status_code=status.HTTP_403_FORBIDDEN, detail=access_error)
-    if not voice_service.is_available():
-        raise HTTPException(
-            status_code=status.HTTP_503_SERVICE_UNAVAILABLE,
-            detail="OPENAI_API_KEY is not configured",
-        )
-
     sess, err = await turn_runtime.load_owned_session(
         body.session_id,
         learner,
@@ -224,6 +230,11 @@ async def voice_speech(
         persona_code=sess.persona.code,
         explicit_preset=None,
     )
+    if not voice_service.tts_available(voice_preset):
+        raise HTTPException(
+            status_code=status.HTTP_503_SERVICE_UNAVAILABLE,
+            detail="TTS provider is not configured",
+        )
     try:
         chunks = [
             chunk.audio
@@ -248,11 +259,11 @@ async def voice_speech(
         )
     return Response(
         content=audio,
-        media_type="audio/mpeg",
+        media_type=voice_service.tts_media_type_for_voice(voice_preset),
         headers={
             "Cache-Control": "no-store",
-            "X-Vignette-TTS-Model": voice_svc.TTS_MODEL,
-            "X-Vignette-TTS-Provider": voice_service.tts_provider(),
+            "X-Vignette-TTS-Model": voice_service.tts_model_for_voice(voice_preset),
+            "X-Vignette-TTS-Provider": voice_service.tts_provider_for_voice(voice_preset),
         },
     )
 
@@ -314,7 +325,7 @@ async def voice_ws(websocket: WebSocket) -> None:
             "session_id": session_id,
             "voice": voice_preset.openai_voice,
             "preset": voice_preset.preset,
-            "tts_provider": voice_service.tts_provider(),
+            "tts_provider": voice_service.tts_provider_for_voice(voice_preset),
             "state": "idle",
             **bind_meta,
         },
@@ -751,7 +762,7 @@ async def _run_turn_and_speak(
             "type": "state",
             "state": "speaking",
             "voice": context.voice_preset.openai_voice,
-            "tts_provider": voice_service.tts_provider(),
+            "tts_provider": voice_service.tts_provider_for_voice(context.voice_preset),
         },
     )
     try:
@@ -787,15 +798,35 @@ async def _load_voice_session(
 
 
 def _client_turn_text_for_speech(sess: InProcSession, turn_seq: int) -> str | None:
-    """Return the persisted client-visible reply for one completed turn."""
-    for turn in reversed(sess.turns):
-        if (
-            turn.turn_seq == turn_seq
-            and turn.speaker == "client"
-            and turn.is_visible_to("client")
-        ):
-            text = (turn.text_masked or turn.text).strip()
-            return text or None
+    """완료된 상담 턴 번호에 대응하는 client-visible 응답을 반환한다.
+
+    degraded 인메모리 미러는 상담자/내담자 한 쌍이 같은 논리 turn_seq를 쓰지만,
+    DB의 app.turns.seq는 발화마다 1씩 증가한다. DB 스냅샷에서 논리 1턴을
+    그대로 seq=1로 찾으면 상담자 발화만 잡혀 TTS가 404가 되므로 두 저장 형태를
+    명시적으로 구분한다.
+    """
+    visible_clients = [
+        turn
+        for turn in sess.turns
+        if turn.speaker == "client" and turn.is_visible_to("client")
+    ]
+    counselor_sequences = {
+        turn.turn_seq for turn in sess.turns if turn.speaker == "counselor"
+    }
+    paired_sequences = counselor_sequences.intersection(
+        turn.turn_seq for turn in visible_clients
+    )
+    if paired_sequences or not counselor_sequences:
+        match = next(
+            (turn for turn in reversed(visible_clients) if turn.turn_seq == turn_seq),
+            None,
+        )
+    else:
+        index = turn_seq - 1
+        match = visible_clients[index] if 0 <= index < len(visible_clients) else None
+    if match is not None:
+        text = (match.text_masked or match.text).strip()
+        return text or None
     return None