feat(desktop): upgrade Pocket TTS to April model

[johnmatthewtennant]

Context

Buzz Desktop currently installs the January 2026 Pocket TTS model. Pocket TTS v2.1.0 ships an April English model with a different SentencePiece tokenizer, learned BOS embedding, prompt policy, and recurrent-state manifest, so changing download URLs alone is not compatible.

Summary

This PR replaces the January Pocket TTS model with the pinned full-precision english_2026-04 export. Buzz keeps sherpa-onnx 1.13.4 as its shared ONNX runtime and adds the April-specific native frontend and state loop needed by the new bundle.

The PR remains draft pending manual listening of the raw, untrimmed onset-stress clip.

Changes

• Pins the April ONNX bundle, Mary reference voice, CC BY 4.0 license, and SHA-256 checksums to immutable upstream commits.
• Loads April bundle schema 2, SentencePiece tokens, learned BOS conditioning, and declared recurrent state tensors.
• Applies April prompt behavior without the January model's leading-space prefix.
• Splits oversized prompts at the bundle's exact 50-token boundary in the desktop worker, preserving per-chunk fades and barge-in checks.
• Bumps the TTS model manifest from 3 to 4 and atomically replaces a verified January cache with April, while retaining the previous usable cache if installation is interrupted.
• Adds a standalone onset-stress generator that preserves raw phrase beginnings without trimming, fade, gain, or normalization.
• Removes the obsolete desktop probes that were hard-coded to January model filenames and precision comparisons.

Related issue

None found.

Testing

• cargo test --locked -p buzz-voice --lib: 12 passed, 2 model-dependent tests ignored.
• Model-dependent SentencePiece parity and exact 50-token splitting tests: passed, including quizzaciously xyzzy.
• cargo test --locked --manifest-path desktop/src-tauri/Cargo.toml huddle::models::tests: 2 passed.
• cargo test --locked --manifest-path desktop/src-tauri/Cargo.toml huddle::tts::tests: 43 passed.
• cargo check --locked --manifest-path desktop/src-tauri/Cargo.toml: passed.
• cargo check --locked --manifest-path desktop/src-tauri/Cargo.toml --all-targets: passed.
• cargo clippy --locked -p buzz-voice --all-targets -- -D warnings: passed.
• Release-mode synthesis loaded the pinned FP32 April bundle and produced 362,880 samples (15.120 seconds) at 24 kHz.
• Raw clip SHA-256: 4002a93db353e8770ac6ceba4969b3dd5bc590f43bc533d3c9747e8e3f527e1b.

Manual listening is intentionally still pending. Automated checks do not satisfy that gate.

Screenshots

N/A. This changes model installation and speech synthesis, not a visual surface.

Reviewer-reproducible examples

Before and after model identity

From the repository root:

git show origin/jt/buzz-voice-refactor:desktop/src-tauri/src/huddle/models.rs \
  | sed -n '39,105p' \
  | grep -E 'sherpa-onnx-pocket-tts|TTS_MODEL_VERSION'

sed -n '39,105p' desktop/src-tauri/src/huddle/models.rs \
  | grep -E 'english_2026-04|TTS_MODEL_VERSION'

Target-branch output identifies the January bundle and manifest version 3:

"https://huggingface.co/csukuangfj2/sherpa-onnx-pocket-tts-2026-01-26/..."
const TTS_MODEL_VERSION: &str = "3";

PR-branch output identifies the pinned April bundle and manifest version 4:

"https://huggingface.co/KevinAHM/pocket-tts-onnx/resolve/58a6d00.../onnx/english_2026-04";
const TTS_MODEL_VERSION: &str = "4";

End-to-end raw synthesis

Install the Pocket model from Buzz Desktop Settings on this branch, then run:

. ./bin/activate-hermit
cargo run --locked --release -p buzz-voice --example pocket_april_onset -- \
  /path/to/installed/pocket-tts \
  /tmp/pocket-tts-april-onset-stress-raw.wav
shasum -a 256 /tmp/pocket-tts-april-onset-stress-raw.wav

The generator makes each sentence a separate model call and inserts only 250 ms of digital silence between outputs. It does not trim or post-process onset artifacts. A successful run prints all nine explicit boundaries and writes a mono 24 kHz PCM WAV.

Listen with headphones and record any affected boundary numbers:

• Confirm no phrase start is missing.
• Confirm no first phoneme is distorted or smeared.
• Confirm no click, pop, or burst occurs at a generation boundary.
• Pay particular attention to Hello, What, I'm, I've, Try, Please, Start, and Sounds.

The PR must remain draft until John records this listening result and separately authorizes ready-for-review.