[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

tests/test_wake_vad_stt.py

@@ -116,9 +116,10 @@ class WakeVadSttTests(unittest.TestCase):
         provider.load()
         recorder = PrimarySpeakerVadRecorder(
             provider,
-            min_duration_ms=40,
+            min_duration_ms=250,
             end_silence_ms=1000,
             speaker_profile_ms=40,
+            speaker_profile_min_ms=40,
             speaker_absent_ms=40,
         )
         frames = [
@@ -142,6 +143,35 @@ class WakeVadSttTests(unittest.TestCase):
         self.assertEqual(consumed, 4)
         self.assertEqual(segment.metadata["transcript"], "你是谁")
 
+    def test_primary_speaker_endpoint_does_not_wait_for_vad_min_duration(self) -> None:
+        provider = EnergyVadProvider()
+        provider.load()
+        recorder = PrimarySpeakerVadRecorder(
+            provider,
+            min_duration_ms=1000,
+            end_silence_ms=1000,
+            speaker_profile_ms=120,
+            speaker_profile_min_ms=40,
+            speaker_absent_ms=40,
+        )
+        frames = [
+            make_frame(1, 0, speech=True, metadata={"speaker_id": "owner", "transcript": "你在做什么"}),
+            make_frame(2, 20, speech=True, metadata={"speaker_id": "owner"}),
+            make_frame(3, 40, speech=True, metadata={"speaker_id": "background"}),
+            make_frame(4, 60, speech=True, metadata={"speaker_id": "background"}),
+            make_frame(5, 80, speech=True, metadata={"speaker_id": "owner", "transcript": "第二次重复"}),
+        ]
+        segment = None
+        for item in frames:
+            result = recorder.feed(item)
+            if isinstance(result, AudioSegment):
+                segment = result
+                break
+        self.assertIsNotNone(segment)
+        assert segment is not None
+        self.assertEqual(segment.metadata["end_reason"], "primary_speaker_absent")
+        self.assertEqual(segment.metadata["transcript"], "你在做什么")
+
     def test_primary_speaker_endpoint_allows_short_pause(self) -> None:
         provider = EnergyVadProvider()
         provider.load()
@@ -150,6 +180,7 @@ class WakeVadSttTests(unittest.TestCase):
             min_duration_ms=40,
             end_silence_ms=1000,
             speaker_profile_ms=40,
+            speaker_profile_min_ms=40,
             speaker_absent_ms=60,
         )
         frames = [