feat(mobile): add on-device Pocket TTS

[johnmatthewtennant]

Context

Desktop Buzz can read conversations aloud with an on-device Pocket TTS model. Mobile did not have the same private, offline voice path.

Summary

Adds on-device Pocket TTS to Buzz Mobile using the shared April 2026 runtime. The model is downloaded after install, synthesized on a warm worker isolate, and played through native iOS and Android audio.

Changes

• Adds a minimal buzz-voice-mobile C ABI with persistent engine reuse, FP32 model selection, cancellation, and reset support.
• Shares desktop text preparation and PCM shaping, including the resident April runtime's exact 50-token chunking behavior.
• Pins official sherpa-onnx 1.13.4 mobile artifacts and checksums separately from Pocket model revision 58a6d00cf13d239b6748cb0769f35c580a8f606c.
• Downloads and verifies 439,555,904 bytes of core model artifacts, 440,213,643 bytes including voice and license files, into a versioned v4 cache with atomic promotion, retries, cancellation, and disk-space checks.
• Integrates native iOS audio session and Android audio focus, routing, and playback. The controller enters speaking state only after native playback actually starts.
• Preserves message ordering, steering, cancellation, and no-repeat behavior while excluding forum channels.
• Exposes structured model-load, synthesis, and playback failures for future native-system fallback PRs.

Related issue

N/A

Testing

• just ci
• Shared voice crate: 45 passed, 6 ignored; all 5 real-model ignored gates also passed with the April artifacts.
• Mobile ABI crate: 3 passed, 1 ignored; the real FP32 ABI test passed.
• Flutter: 845 passed, 1 skipped.
• iOS Runner tests: 13 passed, 1 model probe skipped. Native audio-session coverage passed.
• Built official sherpa-onnx 1.13.4 for iOS device arm64, simulator arm64/x86_64, and Android arm64/x86_64.
• Verified Android arm64/x86_64 libraries export create, cancel, and reset-cancel symbols.
• Ran the production Android integration on an isolated arm64 emulator with the real downloader, model, native playback, queued cancellation, and in-flight cancellation.
• Affected re-reviews from Builderbot, Kalvin, Dead Code, and Timeless Comments found no remaining issues.

Screenshots

The current Pocket Voice settings screenshot is attached in a PR comment.

Reviewer-reproducible examples

The production Android integration completed a full 440,213,643-byte download and reported:

BUZZ_POCKET_METRICS first_pcm_ms=951 ttfap_ms=3137 synthesis_ms=937 audio_s=2.820 rtf=0.333 playback_ms=2869 warm_pcm_ms=951 queued_cancel_ms=5 cancel_ms=241 download_ms=75838 rss_bytes=1612513280 peak_rss_bytes=1768235008

The integration exercises cold download, warm synthesis, native playback, immediate queued cancellation, and cancellation during synthesis.

Notes

• The shared runtime preserves the current stochastic seed behavior because sherpa-onnx does not expose an equivalent seed control.
• Model-size selection remains a separate follow-up.
• Lowest-memory physical-device validation remains a final hardware gate.
• iOS and Android native system TTS fallbacks remain separate follow-up PRs.

[johnmatthewtennant]

🤖 ### Pocket Voice settings

The mobile settings surface now describes the April 2026 on-device model and its approximately 440 MB post-install download.