refactor: replace ffmpeg shell-out with pure Rust audio decoding

[robdefeo]

Summary

• Removes the hard runtime dependency on the system ffmpeg binary — voxscribe now ships as a single self-contained binary on all platforms
• symphonia decodes mp3, aac, alac, ogg/vorbis, flac, wav, and isomp4 (m4a/mp4) natively; rubato resamples any source rate to 16 kHz
• Downmix uses arithmetic mean for stereo and ITU-R BS.775 coefficients for surround content, preserving the centre channel (dialogue) for Whisper accuracy
• AppError variants FfmpegNotFound/FfmpegFailed replaced by DecodeFailed/ResampleFailed
• hound and tempfile moved to [dev-dependencies]

Test plan

• just lint — clippy + fmt clean
• just test — 18 tests pass, including 5 new audio tests: identity decode, 44.1 kHz resample, stereo downmix, chunk-aligned tail (empty-tail rubato edge case), and sub-chunk-length input
• Manual smoke test against a real .m4a Voice Memo — transcription successful
• Single-binary check: works without ffmpeg in PATH

Closes #5

[greptile-apps]

Greptile Summary

This PR replaces the hard runtime dependency on the system ffmpeg binary with a pure-Rust pipeline using symphonia for demuxing/decoding and rubato (FftFixedIn) for resampling, making voxscribe a self-contained single binary. All concerns flagged in prior review threads (stale spec on ResetRequired, SampleBuffer capacity growth for variable-frame-size codecs, debug_assert for interleaved alignment) have been addressed in the current HEAD.

Confidence Score: 5/5

Safe to merge — no correctness or security issues remain; the one open comment is a minor efficiency P2.

All prior P1 concerns are resolved. The only remaining finding is a redundant channel_count == 1 guard that causes an extra to_vec() clone for multi-channel 16 kHz sources — a P2 style issue that does not affect correctness. Eight new unit tests cover identity decode, resample, stereo downmix, chunk-aligned tail, sub-chunk input, 5.1 surround downmix, and corrupt-file rejection.

No files require special attention.

Important Files Changed

Filename	Overview
src/audio.rs	Core audio pipeline replaced with symphonia + rubato; decode loop handles ResetRequired correctly (spec/sample_buf both cleared), SampleBuffer is grown on capacity increase, downmix covers stereo (arithmetic mean) and surround (ITU-R BS.775 with fallback), resample tail flush handles both chunk-aligned and sub-chunk inputs; 8 new unit tests covering all new code paths.
src/error.rs	FfmpegNotFound/FfmpegFailed replaced by DecodeFailed/ResampleFailed; straightforward and correct.
Cargo.toml	symphonia (0.5) and rubato (0.16) added as production deps; hound and tempfile correctly moved to dev-dependencies, eliminating them from the release binary.
Cargo.lock	Lock file updated to reflect new transitive deps (rubato, rustfft, realfft, symphonia-, arrayvec, bytemuck, num-, etc.); all checksums are from the crates.io registry.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[load input path] --> B[validate extension & existence]
    B --> C[decode_to_mono_16k]
    C --> D[Open file + MediaSourceStream]
    D --> E[symphonia probe format]
    E --> F[Find first decodable track]
    F --> G[Create decoder]
    G --> H{next_packet}
    H -->|UnexpectedEof| I[break]
    H -->|ResetRequired| J[Recreate decoder\nspec=None\nsample_buf=None]
    J --> H
    H -->|other track| H
    H -->|Ok packet| K{decode packet}
    K -->|DecodeError| H
    K -->|UnexpectedEof| I
    K -->|Ok decoded| L[copy_interleaved_ref\nextend interleaved]
    L --> H
    I --> M{interleaved empty?}
    M -->|yes| N[DecodeFailed error]
    M -->|no| O[downmix_to_mono]
    O --> P{Already 16kHz?}
    P -->|yes| Q[Return mono samples]
    P -->|no| R[resample_to_target FftFixedIn rubato]
    R --> S[Chunked resample loop]
    S --> T[process_partial_into_buffer tail flush]
    T --> U[Return resampled Vec f32]

Loading

Prompt To Fix All With AI

This is a comment left during a code review.
Path: src/audio.rs
Line: 135-137

Comment:
**Redundant `channel_count == 1` guard causes an extra clone for multi-channel 16 kHz input**

After `downmix_to_mono`, `mono` is always single-channel regardless of the source layout. The `channel_count == 1` check therefore only fast-paths when the source was *already* mono; for stereo (or surround) input that is already at 16 kHz the code falls through into `resample_to_target`, which immediately hits its own `src_rate == TARGET_RATE` guard and returns `Ok(mono.to_vec())` — an unnecessary allocation and copy.

The condition can simply be `spec.rate == TARGET_RATE`:

```suggestion
    if spec.rate == TARGET_RATE {
        return Ok(mono);
    }
```

How can I resolve this? If you propose a fix, please make it concise.

_{Reviews (2): Last reviewed commit: "fix(audio): address review feedback on d..." | Re-trigger Greptile}