Add local spoken passages with Google retrieval and isolated runtime setup

crates/searchi-media/src/lib.rs

@@ -770,6 +770,27 @@ pub fn decode_audio(
     path: &Path,
     window: AudioWindow,
     limits: &MediaLimits,
+) -> Result<Vec<f32>> {
+    decode_audio_inner(tools, path, window, limits, false)
+}
+
+/// Preserve the source clock across seek offsets and discontinuities for speech citations.
+/// Silence is derived from decoded frame timestamps, never semantic guesses.
+pub fn decode_audio_aligned(
+    tools: &FfmpegTools,
+    path: &Path,
+    window: AudioWindow,
+    limits: &MediaLimits,
+) -> Result<Vec<f32>> {
+    decode_audio_inner(tools, path, window, limits, true)
+}
+
+fn decode_audio_inner(
+    tools: &FfmpegTools,
+    path: &Path,
+    window: AudioWindow,
+    limits: &MediaLimits,
+    aligned: bool,
 ) -> Result<Vec<f32>> {
     ensure!(
         window.duration_ms > 0 && window.duration_ms <= 30_000,
@@ -794,7 +815,12 @@ pub fn decode_audio(
         .arg(path)
         .arg("-af")
         .arg(format!(
-            "aresample=16000,atrim=end_sample={},asetpts=PTS-STARTPTS",
+            "{},atrim=end_sample={},asetpts=PTS-STARTPTS",
+            if aligned {
+                "aresample=16000:async=1:first_pts=0"
+            } else {
+                "aresample=16000"
+            },
             u64::from(window.duration_ms) * 16
         ))
         .args([