feat(kon): add audio crate — cpal capture, VAD, rubato resample, symphonia decode

- File decode via symphonia (mp3, aac, flac, wav, ogg) with mono mixdown
- Sinc interpolation resampling via rubato 0.15 (48kHz/44.1kHz → 16kHz)
- WAV I/O via hound (read/write with f32→i16 conversion)
- Microphone capture via cpal 0.17 (WASAPI on Windows) with mpsc channel output
- Voice activity detection via Silero VAD V5 (voice_activity_detector 0.2)
- Async decode_and_resample() for file transcription pipeline
- 3 tests passing (resample passthrough, duration preservation, WAV roundtrip)
- clippy clean, no ort version conflicts

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

crates/audio/src/vad.rs

@@ -0,0 +1,50 @@
+use voice_activity_detector::VoiceActivityDetector;
+
+use kon_core::constants::{VAD_SPEECH_THRESHOLD, WHISPER_SAMPLE_RATE};
+use kon_core::error::{KonError, Result};
+
+/// VAD chunk size for 16kHz audio (required by Silero VAD V5).
+const VAD_CHUNK_SIZE: usize = 512;
+
+/// Wraps Silero VAD V5 for speech detection.
+pub struct SpeechDetector {
+    vad: VoiceActivityDetector,
+    threshold: f64,
+}
+
+impl SpeechDetector {
+    /// Create a new speech detector with the default threshold.
+    pub fn new() -> Result<Self> {
+        let vad = VoiceActivityDetector::builder()
+            .sample_rate(WHISPER_SAMPLE_RATE as i64)
+            .chunk_size(VAD_CHUNK_SIZE)
+            .build()
+            .map_err(|e| KonError::AudioCaptureFailed(format!("VAD init failed: {e}")))?;
+
+        Ok(Self {
+            vad,
+            threshold: VAD_SPEECH_THRESHOLD,
+        })
+    }
+
+    /// Predict whether a chunk of 16kHz f32 audio contains speech.
+    /// Returns the speech probability (0.0 to 1.0).
+    pub fn predict(&mut self, samples: &[f32]) -> f32 {
+        self.vad.predict(samples.iter().copied())
+    }
+
+    /// Returns true if the chunk contains speech above the threshold.
+    pub fn is_speech(&mut self, samples: &[f32]) -> bool {
+        self.predict(samples) > self.threshold as f32
+    }
+
+    /// Reset the internal VAD state (call between recording sessions).
+    pub fn reset(&mut self) {
+        self.vad.reset();
+    }
+
+    /// The required chunk size for this VAD instance.
+    pub fn chunk_size(&self) -> usize {
+        VAD_CHUNK_SIZE
+    }
+}