feat: VAD ingestion buffer — Silero VAD, per-room accumulation

TDD: 3 tests covering silence passthrough, speech→silence segment emission,
and short-speech discard. Uses _speech_ms tracking (not total buffer length)
for accurate min_speech_ms enforcement. Silero VAD import is try/except'd
so tests run without torch via mocker.patch on vad_prob.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

tests/test_ingestion.py

@@ -0,0 +1,55 @@
+import numpy as np
+import pytest
+from datetime import datetime, timezone
+from unittest.mock import patch, MagicMock
+
+
+def _make_chunk(samples: int = 4000, amplitude: float = 0.0) -> bytes:
+    """Create a fake PCM int16 chunk."""
+    audio = (np.ones(samples) * amplitude * 32767).astype(np.int16)
+    return audio.tobytes()
+
+
+@pytest.fixture
+def buf(mocker):
+    """RoomVADBuffer with mocked VAD model."""
+    mocker.patch(
+        "app.pipeline.ingestion.vad_prob",
+        side_effect=lambda chunk: 0.9 if np.frombuffer(chunk, np.int16).max() > 100 else 0.1,
+    )
+    from app.pipeline.ingestion import RoomVADBuffer
+    return RoomVADBuffer(silence_threshold_ms=600, min_speech_ms=500, chunk_ms=250)
+
+
+def test_silence_produces_no_segment(buf):
+    ts = datetime(2026, 1, 1, tzinfo=timezone.utc)
+    for _ in range(10):
+        result = buf.process_chunk(_make_chunk(amplitude=0.0), ts)
+        assert result is None
+
+
+def test_speech_then_silence_produces_segment(buf):
+    ts = datetime(2026, 1, 1, tzinfo=timezone.utc)
+    # 4 speech chunks = 1 second of speech
+    for _ in range(4):
+        result = buf.process_chunk(_make_chunk(amplitude=0.5), ts)
+        assert result is None  # still accumulating
+    # 3 silence chunks = 750ms silence (> 600ms threshold)
+    segment = None
+    for _ in range(3):
+        segment = buf.process_chunk(_make_chunk(amplitude=0.0), ts)
+    assert segment is not None
+    assert isinstance(segment, tuple)
+    audio_bytes, start_time = segment
+    assert len(audio_bytes) > 0
+
+
+def test_short_speech_discarded(buf):
+    ts = datetime(2026, 1, 1, tzinfo=timezone.utc)
+    # 1 speech chunk = 250ms (below 500ms min)
+    buf.process_chunk(_make_chunk(amplitude=0.5), ts)
+    # Now silence to trigger emission
+    result = None
+    for _ in range(3):
+        result = buf.process_chunk(_make_chunk(amplitude=0.0), ts)
+    assert result is None  # discarded as too short