feat(recorder): add offline voice clip pipeline

[变更性质]
- 本提交新增本机离线的语音片段录音能力，不涉及网络服务或语音识别。

[新增功能]
- 通过 Pulse 默认输入持续采集音频，以本地 Silero VAD 触发片段。
- 以私有目录和权限写入 24 kbps Ogg/Opus 录音，并提供 systemd 用户服务安装器。

[实现方案]
- 使用有限前置缓冲、静音封段和最长段滚动状态机，避免静音落盘及内存无限增长。
- 增加真实 Opus 编解码、服务 dry-run 和纯逻辑状态转换测试；记录模型归属和部署前置条件。

[影响范围]
- 新增 Python CLI、运行时模块、中文运维文档和自动测试。
- 未安装、启用或启动任何用户 systemd 服务；不删除或修改录音数据。

1 files changed, 96 insertions, 0 deletions

diff --git a/tests/test_segmenter.py b/tests/test_segmenter.py
new file mode 100644
index 0000000..d1f7ee5
--- /dev/null
+++ b/tests/test_segmenter.py
@@ -0,0 +1,96 @@
+from datetime import datetime, timedelta, timezone
+
+import numpy as np
+
+from mic_clipper.segmenter import Audio, End, Segmenter, Start
+
+
+SAMPLE_RATE = 16_000
+FRAME_SAMPLES = 512
+FRAME_DURATION = timedelta(seconds=FRAME_SAMPLES / SAMPLE_RATE)
+START = datetime(2026, 9, 29, 23, 59, 59, tzinfo=timezone.utc)
+
+
+def frame(value: float = 0.0) -> np.ndarray:
+    return np.full(FRAME_SAMPLES, value, dtype=np.float32)
+
+
+def push(segmenter: Segmenter, index: int, speech: bool, value: float = 0.0):
+    return segmenter.push(frame(value), speech, START + index * FRAME_DURATION)
+
+
+def test_speech_starts_segment_with_half_second_preroll():
+    segmenter = Segmenter(pre_roll_seconds=0.5, silence_seconds=1.5)
+
+    for index in range(16):
+        assert push(segmenter, index, False) == ()
+
+    events = push(segmenter, 16, True, 1.0)
+
+    assert len(events) == 1
+    assert isinstance(events[0], Start)
+    assert events[0].started_at == START + 16 * FRAME_DURATION - timedelta(seconds=0.5)
+    assert events[0].samples.shape == (8_512,)
+    assert np.all(events[0].samples[:8_000] == 0.0)
+    assert np.all(events[0].samples[8_000:] == 1.0)
+
+
+def test_short_silence_is_merged_into_active_segment():
+    segmenter = Segmenter(pre_roll_seconds=0.5, silence_seconds=1.5)
+
+    assert isinstance(push(segmenter, 0, True, 1.0)[0], Start)
+    assert isinstance(push(segmenter, 1, False)[0], Audio)
+    assert isinstance(push(segmenter, 2, True, 1.0)[0], Audio)
+    assert segmenter.active
+
+
+def test_one_point_five_seconds_of_silence_closes_segment():
+    segmenter = Segmenter(pre_roll_seconds=0.5, silence_seconds=1.5)
+
+    push(segmenter, 0, True, 1.0)
+    events = ()
+    for index in range(1, 48):
+        events = push(segmenter, index, False)
+
+    assert isinstance(events[0], Audio)
+    assert isinstance(events[-1], End)
+    assert not segmenter.active
+
+
+def test_maximum_segment_rolls_without_losing_audio():
+    segmenter = Segmenter(
+        pre_roll_seconds=0.5,
+        silence_seconds=1.5,
+        max_segment_seconds=FRAME_SAMPLES * 2 / SAMPLE_RATE,
+    )
+
+    first = push(segmenter, 0, True, 1.0)
+    second = push(segmenter, 1, True, 2.0)
+    third = push(segmenter, 2, True, 3.0)
+
+    assert isinstance(first[0], Start)
+    assert isinstance(second[0], Audio)
+    assert isinstance(third[0], End)
+    assert isinstance(third[1], Start)
+    assert np.all(third[1].samples == 3.0)
+
+
+def test_segment_start_date_is_retained_across_midnight():
+    segmenter = Segmenter(pre_roll_seconds=0.5, silence_seconds=1.5)
+
+    events = push(segmenter, 16, True, 1.0)
+
+    assert isinstance(events[0], Start)
+    assert events[0].started_at.date().isoformat() == "2026-09-29"
+
+
+def test_speech_after_a_closed_segment_keeps_terminal_silence_as_preroll():
+    segmenter = Segmenter(pre_roll_seconds=0.5, silence_seconds=1.5)
+
+    push(segmenter, 0, True, 1.0)
+    for index in range(1, 48):
+        push(segmenter, index, False)
+    events = push(segmenter, 48, True, 1.0)
+
+    assert isinstance(events[0], Start)
+    assert events[0].samples.shape == (8_512,)