feat(transcription): ParakeetEngine via sherpa-rs transducer (Track-B step 2)

[farce1]

What — Track-B step 2 (stacked on #9)

Implements NVIDIA Parakeet-TDT-v3 behind the AsrEngine trait from #9 — and, after verifying the sherpa-rs API, does it with zero new dependencies.

Why sherpa-rs, not transcribe-rs

You asked for transcribe-rs; researching it surfaced a real cost: transcribe-rs pulls ort (ONNX Runtime 1.22), which collides with our sherpa-rs-sys's bundled ONNX Runtime (~1.17) — and we must keep sherpa (VAD + diarization). That meant a second ONNX Runtime + ort load-dynamic + dylib bundling. You opted to verify the sherpa path first — and it works: sherpa-rs 0.6.8 ships transducer::TransducerRecognizer (wraps SherpaOnnxOfflineRecognizer), which loads NeMo Parakeet-TDT directly. So Parakeet rides our existing sherpa-rs / single ONNX Runtime. No transcribe-rs, no dual-ORT, no packaging changes.

Changes

• engine.rs: ParakeetEngine wrapping TransducerRecognizer (model_type="nemo_transducer", decoding_method="greedy_search", 16 kHz / feature_dim 80, CPU), implementing AsrEngine → AsrOutput { text, language: "" }.
• engine.rs: load_engine(engine: &str, model_dir) -> Box<dyn AsrEngine> — string dispatch mirroring the existing meetings.asr_engine value; default "whisper".
• worker.rs/mod.rs: worker builds the engine via the factory from WorkerConfig.asr_engine (set to "whisper" for now → behavior unchanged).

Verification

• cargo check + clippy -D warnings pass — confirms Parakeet compiles against sherpa-rs (the key de-risking) with no dead-code. Full cargo test: only the pre-existing model_archive_consts_tests (REPLACE_WITH_ placeholders) fail.
• ⚠️ Runtime not yet validated — needs the sherpa-onnx Parakeet-v3 int8 model (encoder/decoder/joiner.int8.onnx + tokens.txt from csukuangfj/sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8) + a real recording. Same untestable-in-CI nature as the existing Whisper FFI path.

Model (next step)

Use the sherpa-onnx export sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8 (NOT Handy's transcribe-rs tarball — different layout). Next steps: per-engine model-dir resolution in model.rs, the download (parallel to the Whisper download), the engine-selection setting (default Parakeet-v3, Whisper-turbo fallback) + picker UI.

Merge notes

Stacked on #9 (base = feat/asr-engine-trait); retarget to main once #9 merges. Overlaps worker.rs/mod.rs with #4/#6/#8.

CI red on the usual maintainer-only blockers (vendor binary + REPLACE_WITH_), not this PR. Attribution: Parakeet-TDT-0.6B-v3 is CC-BY-4.0 (credit NVIDIA in the app's licenses screen).