feat(voice): support April Pocket model bundles

[johnmatthewtennant]

Context

Pocket's April 2026 ONNX release changes the tokenizer and native inference loop. Buzz Voice needs a reusable, caller-provided model-bundle layer before Desktop and Mobile can adopt the release independently.

This PR is stacked directly on #2467 and intentionally excludes model acquisition, cache migration, platform packaging, UI, and playback policy.

Summary

Adds a platform-neutral April Pocket runtime with pinned FP32 and INT8 bundle metadata, deterministic validation, tokenizer-aware 50-token chunking, and callback-compatible synthesis.

Changes

• Add the April tokenizer, BOS handling, recurrent inference loop, and PCM synthesis path.
• Add explicit PocketPrecision, PocketModelSelection, and PocketLoadOptions APIs while preserving the existing loader and synthesis APIs.
• Expose immutable model information: revision, precision, artifact names, SHA-256 checksums, byte sizes, maximum token count, sample rate, and quantized graph names.
• Validate complete caller-provided bundles and return deterministic errors for missing or mismatched variants.
• Add resident-engine token-aware chunking and transparently handle synthesis input over the model's exact 50-token limit.
• Add a raw, untrimmed onset-stress example for reviewer reproduction.

Related issue

N/A

Testing

• just ci
  • Rust formatting, workspace checks, and lint
  • Desktop checks, 1,801 core tests, 3 audio mixer diagnostics, web tests/build
  • Mobile formatting, analysis, and full Flutter test suite
• cargo test --locked -p buzz-voice
  • 18 passed, 6 model-dependent tests ignored by default
• With January and April model directories configured, all 6 model-dependent tests passed:
  • unknown-word tokenizer behavior
  • exact token splitting
  • callback cancellation
  • FP32 and INT8 non-silent output
  • oversized public synthesis
  • January source compatibility
• Release onset synthesis completed for FP32 and INT8.
• macOS linkage inspection found one system C++ runtime and no duplicate ONNX Runtime image.
• Full review fanout completed cleanly: BuilderBot, Kalvin, diff minimization, NCMB, dead code, timeless comments, and Stacksmith placement against the exact stack.

Screenshots

N/A — runtime-only Rust change with no UI.

Reviewer-reproducible examples

Canonical base:

91eaee8804ec0c139a53782212df44a70aa994c1

Upstream bundle contract:

Model: KevinAHM/pocket-tts-onnx
Revision: 58a6d00cf13d239b6748cb0769f35c580a8f606c
Bundle: english_2026-04
License: CC-BY-4.0
FP32 bytes: 439555904
Hybrid INT8 bytes: 165232420
INT8 graphs: flow_lm_main, flow_lm_flow, mimi_decoder

Exact artifact names, byte sizes, and SHA-256 checksums are defined in crates/buzz-voice/src/pocket_models.rs.

Run the unit suite:

cargo test --locked -p buzz-voice

Run the model-backed gates:

BUZZ_POCKET_JANUARY_MODEL_DIR="$HOME/.buzz/models/pocket-tts" \
BUZZ_POCKET_APRIL_MODEL_DIR="/path/to/english_2026-04-fp32" \
BUZZ_POCKET_APRIL_INT8_MODEL_DIR="/path/to/english_2026-04-int8" \
cargo test --locked -p buzz-voice -- --ignored

Generate the raw, untrimmed onset-stress clip:

BUZZ_POCKET_APRIL_MODEL_DIR="/path/to/english_2026-04-fp32" \
cargo run --release -p buzz-voice --example pocket_april_onset -- \
  "/path/to/reference.wav" "/tmp/pocket-april-onset-raw.wav"

Locally generated FP32 evidence:

SHA-256: e2b8e7219f75a4924eca39e4b679500da4ecd824e14f6ea096f9965e8665511e
PCM: 24 kHz mono, 353280 samples, 14.720 seconds
Post-processing: none

Listening checklist:

• Is the first phoneme of every synthesis unit present and undistorted?
• Are phrase starts free of bursts, dropouts, or other onset artifacts?
• Are inserted unit boundaries free of clicks?
• Is speech intelligible and stable through the final unit?