feat(voice): Implement continuous VAD capture loop for hands-free mode

[mmogr]

Summary

The voice pipeline supports two interaction modes: Push-to-Talk (PTT) and Voice Activity Detection (VAD). PTT works end-to-end, but VAD mode has no continuous capture loop — there is nothing driving the vad_process_frame() method with live microphone audio.

The pipeline correctly:

• Creates a VoiceActivityDetector when mode is VAD
• Provides vad_process_frame() which processes audio frames and emits transcripts
• Loads the Silero VAD model

But there is no Tauri command or background task that:

1. Continuously reads audio from the microphone
2. Splits it into frames
3. Feeds those frames to vad_process_frame()

As a result, VAD mode silently does nothing when activated.

What needs to change

Option A: Background capture loop in the pipeline

When start() is called in VAD mode, spawn a background task that:

1. Starts audio capture (capture.start_recording())
2. Reads accumulated samples periodically (e.g. every 30ms)
3. Feeds frames to vad_process_frame()
4. Stops when the pipeline is stopped

This keeps all logic inside the Rust pipeline.

Option B: Tauri-driven polling

Add a Tauri command (or timer-based event) that periodically:

1. Reads audio from the capture buffer
2. Calls vad_process_frame() on the pipeline
3. Emits transcript events when speech is detected

Option A is cleaner but requires the stream thread work from #178.

Either way

• Add a voice_vad_toggle or similar command if not already present
• Update the frontend to use VAD mode (currently PTT-only in the overlay)
• Consider documenting VAD as experimental / not-yet-functional until this is implemented

Files

• crates/gglib-voice/src/pipeline.rs — start(), vad_process_frame()
• src-tauri/src/commands/voice.rs — no VAD-driving command exists
• src/components/VoiceOverlay/VoiceOverlay.tsx — PTT button shown but no VAD UI

Priority

Medium — PTT mode works, but the VAD mode option in settings is misleading without this.