[低延迟端点]：完成首句保留和快速结束修正，包含ACK缓冲策略、批量读帧和主说话人端点回归测试

src/owner_voice_pet/vad.py

@@ -287,6 +287,7 @@ class VadRecorder:
 @dataclass(slots=True)
 class PrimarySpeakerVadRecorder(VadRecorder):
     speaker_profile_ms: int = 600
+    speaker_profile_min_ms: int = 120
     speaker_absent_ms: int = 300
     similarity_threshold: float = 0.70
     min_rms: float = 0.012
@@ -336,7 +337,7 @@ class PrimarySpeakerVadRecorder(VadRecorder):
                 self.primary_absent_ms = 0
             else:
                 self.primary_absent_ms += frame_ms
-            if self.primary_absent_ms >= self.speaker_absent_ms and duration >= self.min_duration_ms:
+            if self.primary_absent_ms >= self.speaker_absent_ms:
                 return self._build_segment("primary_speaker_absent")
         if result.silence_ms >= self.end_silence_ms and duration >= self.min_duration_ms:
             return self._build_segment("silence")
@@ -365,7 +366,7 @@ class PrimarySpeakerVadRecorder(VadRecorder):
             self.profile.vector = tuple(averaged)
 
     def _profile_ready(self) -> bool:
-        minimum_ms = min(self.speaker_profile_ms, max(120, self.min_duration_ms))
+        minimum_ms = min(self.speaker_profile_ms, self.speaker_profile_min_ms)
         return self.profile.speech_ms >= minimum_ms and (
             self.profile.speaker_id is not None or self.profile.vector is not None
         )