feat(desktop): import local Pocket voices

[johnmatthewtennant]

Context

Pocket TTS currently offers only bundled reference voices. People need a local, private way to add another voice without sending audio through the renderer or a cloud service.

Summary

This adds a minimal Pocket voice import flow to the existing Voice settings card. Buzz opens the native file picker, validates and canonicalizes the selected WAV in the reusable buzz-voice crate, stores it in app data under a content-derived identity, selects it, and lets the user safely delete it later.

Changes

• Adds a native WAV picker and validates mono PCM or float WAV files between 2 and 30 seconds.
• Adds a reusable PocketVoiceLibrary to crates/buzz-voice; Desktop remains a thin Tauri picker and app-data adapter.
• Canonicalizes accepted audio to mono 32 kHz PCM16 before hashing and storage.
• Uses stable pocket:imported:<sha256> voice identities and content-addressed app-data files.
• Keeps absolute file paths inside the native process and exposes only voice metadata to React.
• Adds local imported voices to Pocket preview and live huddle playback.
• Adds Add voice and delete controls to Voice settings, with Mary as the confirmed deletion fallback.
• Preserves bundled voices if optional import metadata is unreadable and keeps failed deletion retryable.
• Adds deterministic synthesis evidence, BlackHole loopback coverage, and a desktop E2E import, select, and delete flow.

Related issue

N/A. No related issue was supplied.

Testing

Passed:

• cargo test -p buzz-voice imported --lib: 3 passed.
• BUZZ_POCKET_MODEL_DIR="$HOME/.buzz/models/pocket-tts" cargo test -p buzz-voice --test pocket_import_audio objective_import_synthesis_delete_and_mary_fallback -- --ignored --nocapture: 1 passed. Imported synthesis produced 2.48 seconds of non-silent PCM, interruption discarded partial PCM, deletion selected Mary, and Mary fallback produced 2.56 seconds of non-silent PCM.
• BUZZ_POCKET_MODEL_DIR="$HOME/.buzz/models/pocket-tts" BUZZ_VOICE_AUDIO_DEVICE="BlackHole 2ch" cargo test -p buzz-voice --test pocket_import_audio blackhole_playback_captures_generated_non_silent_pcm -- --ignored --nocapture: 1 passed. BlackHole captured 15,806 non-silent samples.
• cd desktop && pnpm run build:e2e.
• cd desktop && pnpm exec playwright test tests/e2e/voice-settings.spec.ts --project=smoke: 5 passed.
• Final exact-head dead-code, timeless-comments, and stacked-diff review passes reported no findings.

Manual listening validation remaining:

1. In Buzz, import a local WAV through the native picker and listen to Preview and a live huddle reply.
2. Delete that selected voice and listen to Mary afterward.

Screenshots

Focused Voice settings screenshot. The image was captured after a fresh E2E build, visually inspected for clipping and layout regressions, and has SHA-256 3cee8ad0f42c0a69ab5fa927af16ba283c6b5c6f06c86c89d68040f2a9e247df.

Reviewer-reproducible examples

With Pocket TTS already downloaded to Buzz's normal model directory, exercise the full library workflow: import, persist, reload, synthesize, interrupt, delete, and synthesize with Mary fallback.

BUZZ_POCKET_MODEL_DIR="$HOME/.buzz/models/pocket-tts" \
  cargo test -p buzz-voice --test pocket_import_audio \
  objective_import_synthesis_delete_and_mary_fallback \
  -- --ignored --nocapture

Observed result:

1 passed
importedPreview: 2.48 seconds, 33,308 non-silent samples
interruption: partial PCM discarded
afterDelete: pocket:mary
maryFallbackPreview: 2.56 seconds, 53,885 non-silent samples

On macOS with BlackHole 2ch installed, verify that generated imported-voice PCM reaches the audio device and can be captured:

BUZZ_POCKET_MODEL_DIR="$HOME/.buzz/models/pocket-tts" \
BUZZ_VOICE_AUDIO_DEVICE="BlackHole 2ch" \
  cargo test -p buzz-voice --test pocket_import_audio \
  blackhole_playback_captures_generated_non_silent_pcm \
  -- --ignored --nocapture

Observed result:

1 passed
captured: 274,432 samples, 15,806 non-silent samples

Build the desktop E2E surface and exercise the settings import, selection, preview dispatch, deletion, cancellation, and invalid-file states:

cd desktop
pnpm run build:e2e
pnpm exec playwright test tests/e2e/voice-settings.spec.ts --project=smoke

Observed result:

5 passed
import_pocket_voice
delete_pocket_voice { voiceKey: "pocket:imported:<sha256>" }

The Playwright flow mocks the native picker boundary. The two manual listening steps above remain required to validate the real file dialog and perceived audio quality.

[johnmatthewtennant]

🤖 ### Pocket voice import

The Voice settings card keeps the two bundled Pocket voices and adds the local Add voice action.