fix: TTS 발음/타이밍 대규모 개선 (숫자, 단위, 방위, 줄바꿈 pause)

- 숫자를 완전한 한글 표기로 변환 (사이노-한국어 기본, 시각은 순우리말)
- °C/°F/%/m/s/mm/km 등 단위를 한글로 스펠아웃
- 괄호 앞뒤 pause 강화, 방위 약어(SSW 등) 한글 변환
- 마크다운 표/리스트를 줄바꿈 전에 분할 후 정제하도록 파이프라인 재설계
- 구조적 개행(리스트/표 셀) 유래 조각은 병합 금지, 대화체 문장만 병합
- 소수점(27.7) 오탐 문장경계 버그 수정
- num_step 16, speed 1.15로 튜닝 (품질 손실 없이 생성속도 개선)
- 대기열 위치 표시 (tts_queue 메시지)

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

src/tools/voice_engine.py

@@ -69,6 +69,25 @@ def crash_and_restart(where: str, e: Exception):
     os._exit(1)
 
 
+# Every TTS request (call streaming + the single-shot text button) serializes
+# on the same GPU (measured: one request already saturates it, see
+# project_tts_edge memory), so this counter doubles as an honest queue
+# position — no separate scheduler needed.
+_gpu_pending_lock = threading.Lock()
+_gpu_pending = 0
+
+def gpu_queue_enter() -> int:
+    global _gpu_pending
+    with _gpu_pending_lock:
+        _gpu_pending += 1
+        return _gpu_pending
+
+def gpu_queue_exit():
+    global _gpu_pending
+    with _gpu_pending_lock:
+        _gpu_pending -= 1
+
+
 def load_models(args):
     log.info('Loading faster-whisper model=%s device=%s ...', args.stt_model, args.device)
     from faster_whisper import WhisperModel
@@ -137,6 +156,8 @@ class Engine:
     def synthesize_full(self, text: str, speed=None, num_step=None):
         if num_step is None:
             num_step = 16
+        if speed is None:
+            speed = 1.15
         audio = self.tts_model.generate(
             text=text[:4000],
             language='Korean',
@@ -221,13 +242,29 @@ async def handle_tts_start(ws, engine: Engine, state: ConnectionState, msg: dict
             loop.call_soon_threadsafe(queue.put_nowait, ('error', str(e)))
             if is_unrecoverable_cuda_error(e):
                 crash_and_restart('handle_tts_start', e)
+        finally:
+            gpu_queue_exit()
 
+    position = gpu_queue_enter()
+    if position > 1:
+        # Someone else's synthesis is already running on the GPU (it's
+        # effectively serialized — see project_tts_edge memory); let the
+        # client show a queue indicator instead of silently hanging.
+        await ws.send(json.dumps({'type': 'tts_queue', 'aheadCount': position - 1, 'id': req_id}))
     threading.Thread(target=produce, daemon=True).start()
-    await ws.send(json.dumps({'type': 'tts_stream_start', 'sampleRate': engine.tts_sample_rate, 'id': req_id}))
     try:
+        stream_started = False
         while True:
             kind, payload = await queue.get()
             if kind == 'chunk':
+                if not stream_started:
+                    # synthesize_stream yields the whole clip as one chunk
+                    # (see its docstring) once GPU generation actually
+                    # finishes, so this is the right moment to tell the
+                    # client "now speaking" — sending it eagerly up front
+                    # would stomp the tts_queue status above.
+                    await ws.send(json.dumps({'type': 'tts_stream_start', 'sampleRate': engine.tts_sample_rate, 'id': req_id}))
+                    stream_started = True
                 await ws.send(payload)
             elif kind == 'error':
                 await ws.send(json.dumps({'type': 'error', 'message': payload, 'id': req_id}))
@@ -268,7 +305,11 @@ async def handle_connection(ws, engine: Engine):
                 speed = msg.get('speed')
                 num_step = msg.get('num_step')
                 loop = asyncio.get_event_loop()
-                pcm, sample_rate = await loop.run_in_executor(None, engine.synthesize_full, text, speed, num_step)
+                gpu_queue_enter()
+                try:
+                    pcm, sample_rate = await loop.run_in_executor(None, engine.synthesize_full, text, speed, num_step)
+                finally:
+                    gpu_queue_exit()
                 await ws.send(json.dumps({
                     'type': 'tts_result',
                     'audioBase64': base64.b64encode(pcm).decode('ascii'),