fix(audio): address CoreML VAD review feedback

crates/chanora_audio/src/vad/apple_coreml.rs

@@ -94,6 +94,7 @@ pub struct AppleCoreMlVad {
     handle: *mut c_void,
     symbols: AppleSileroSymbols,
     accum: Vec<f32>,
+    frame_scratch: Box<[f32; SILERO_COREML_FRAME_16K]>,
     last_probability: f32,
 }
 
@@ -130,6 +131,7 @@ impl AppleCoreMlVad {
             handle,
             symbols,
             accum: Vec::with_capacity(SILERO_COREML_FRAME_16K),
+            frame_scratch: Box::new([0.0; SILERO_COREML_FRAME_16K]),
             last_probability: 0.0,
         })
     }
@@ -148,14 +150,13 @@ impl AppleCoreMlVad {
         }
     }
 
-    fn calc_level(&mut self, audio_frame: &[f32]) -> f32 {
-        debug_assert_eq!(audio_frame.len(), SILERO_COREML_FRAME_16K);
+    fn calc_level(&mut self) -> f32 {
         let mut probability = self.last_probability;
         let rc = unsafe {
             (self.symbols.process)(
                 self.handle,
-                audio_frame.as_ptr(),
-                audio_frame.len(),
+                self.frame_scratch.as_ptr(),
+                self.frame_scratch.len(),
                 &mut probability,
             )
         };
@@ -183,11 +184,10 @@ impl VoiceActivityDetector for AppleCoreMlVad {
 
         self.accum.extend_from_slice(samples);
         if self.accum.len() >= SILERO_COREML_FRAME_16K {
-            let audio_frame: Vec<f32> = self.accum[..SILERO_COREML_FRAME_16K].to_vec();
-            self.calc_level(&audio_frame);
-            let overflow: Vec<f32> = self.accum.drain(SILERO_COREML_FRAME_16K..).collect();
-            self.accum.clear();
-            self.accum.extend_from_slice(&overflow);
+            self.frame_scratch
+                .copy_from_slice(&self.accum[..SILERO_COREML_FRAME_16K]);
+            self.calc_level();
+            drop(self.accum.drain(..SILERO_COREML_FRAME_16K));
         }
 
         VadOutput {

crates/chanora_audio/src/vad/mod.rs

@@ -1,9 +1,9 @@
 //! Voice activity detection backends and helpers.
 //!
-//! iOS capture feeds VoiceProcessingIO-processed microphone frames into
-//! this module and uses the realtime-safe WebRTC fallback. Other
-//! platforms may use a model-backed detector when available so
-//! VoiceActivity mode never collapses back to Continuous transmit.
+//! Apple capture feeds VoiceProcessingIO/CoreAudio-processed microphone
+//! frames into this module and prefers Apple CoreML Silero VAD when the
+//! Swift bridge is linked. WebRTC VAD remains the realtime-safe fallback;
+//! non-Apple platforms may use ONNX-backed Silero when available.
 
 #[cfg(any(target_os = "ios", target_os = "macos"))]
 pub mod apple_coreml;