perf(worker): preload VAD at startup, skip Qwen3 when using Mimo ASR

asr_agent/worker.py

@@ -464,9 +464,16 @@ class Worker:
         self.hist: list[dict] = []
 
     async def run(self):
-        logger.info("Loading ASR model...")
-        self.asr = QwenASREngine(model_id=ASR_MODEL, language=None)
-        logger.info("ASR ready")
+        # Pre-load Silero VAD at startup (not lazy)
+        _get_vad_session()
+        # Only load Qwen3 if using local ASR
+        if ASR_PROVIDER != "mimo":
+            logger.info("Loading ASR model...")
+            self.asr = QwenASREngine(model_id=ASR_MODEL, language=None)
+            logger.info("ASR ready")
+        else:
+            self.asr = None
+            logger.info("ASR: using Mimo API (no local model)")
 
         self.room.on("track_subscribed", self._on_track)
         self.room.on("participant_disconnected", self._on_off)