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, 59 insertions, 0 deletions

diff --git a/src/mic_clipper/vad.py b/src/mic_clipper/vad.py
new file mode 100644
index 0000000..705dfd8
--- /dev/null
+++ b/src/mic_clipper/vad.py
@@ -0,0 +1,59 @@
+"""Stateful, local Silero VAD ONNX inference."""
+
+from __future__ import annotations
+
+from importlib import resources
+from pathlib import Path
+
+import numpy as np
+import onnxruntime
+
+
+FRAME_SAMPLES = 512
+CONTEXT_SAMPLES = 64
+
+
+class SileroVad:
+    """Run the bundled Silero v6 model on exact 32 ms PCM frames."""
+
+    def __init__(self, model_path: Path | None = None) -> None:
+        self.model_path = Path(model_path) if model_path else self._default_model_path()
+        options = onnxruntime.SessionOptions()
+        options.inter_op_num_threads = 1
+        options.intra_op_num_threads = 1
+        options.enable_cpu_mem_arena = False
+        options.log_severity_level = 4
+        self.session = onnxruntime.InferenceSession(
+            self.model_path,
+            providers=["CPUExecutionProvider"],
+            sess_options=options,
+        )
+        self._context = np.zeros(CONTEXT_SAMPLES, dtype=np.float32)
+        self._h = np.zeros((1, 1, 128), dtype=np.float32)
+        self._c = np.zeros((1, 1, 128), dtype=np.float32)
+
+    def probability(self, samples: np.ndarray) -> float:
+        samples = np.asarray(samples, dtype=np.float32)
+        if samples.ndim != 1 or len(samples) != FRAME_SAMPLES:
+            raise ValueError("Silero VAD requires exactly 512 mono samples at 16 kHz")
+        model_input = np.concatenate((self._context, samples)).reshape(1, -1)
+        output, self._h, self._c = self.session.run(
+            None,
+            {"input": model_input, "h": self._h, "c": self._c},
+        )
+        self._context = samples[-CONTEXT_SAMPLES:].copy()
+        return float(output[0])
+
+    @staticmethod
+    def _default_model_path() -> Path:
+        bundled = resources.files("mic_clipper.assets").joinpath("silero_vad_v6.onnx")
+        if bundled.is_file():
+            return Path(bundled)
+        try:
+            import faster_whisper
+        except ImportError as error:
+            raise RuntimeError(
+                "Silero VAD model is missing; bundle silero_vad_v6.onnx or install the "
+                "pre-provisioned faster-whisper runtime asset"
+            ) from error
+        return Path(faster_whisper.__file__).parent / "assets" / "silero_vad_v6.onnx"