Harden the Voice Foundation pipeline

[logannye]

Summary

• replaces the prototype 10 Hz speech extractor and five prototype.speech.* values with a continuous AudioWorklet to voice-worker path and task-specific prototype.voice.* measurements
• adds a microphone-only Voice Foundation protocol with quiet/speech preflight and five completion-gated task phases
• centralizes audio quality, per-metric abstention, uncertainty, provenance, accepted-interval clipping, and exact trajectory compatibility
• adds a disabled-by-default, loopback-only Python 3.11 FastAPI WavLM Large representation service with pinned model provenance and deterministic CI adapter
• updates encounter/evidence contracts, protocol fixtures, privacy assertions, architecture, safety, validation, and operator documentation
• keeps Facial Foundation on the existing MediaPipe pipeline and preserves its six facial measurements; face-voice fusion remains deferred

Voice architecture

• mono audio requested at ideal 48 kHz with echo cancellation, noise suppression, and AGC disabled
• reusable 20 ms PCM blocks transferred through a MessageChannel
• bounded 30-second worker ring; 40 ms windows on a 10 ms hop; discontinuities are never synthesized as silence
• 50-700 Hz F0 estimation with estimator agreement and octave checks
• cepstral peak prominence, HNR, perturbation, LPC formants, timing, and estimated syllabic/DDK features with independent quality gates
• two sustained /a/ trials, standardized reading, rapid /pa-ta-ka/, and spontaneous response; elapsed time never advances a task

Privacy and safety

• no audio recording, transcript, voiceprint, native audio arrays, or WavLM embeddings enter observations, events, evidence, diagnostics payloads, reports, or trajectory artifacts
• PCM and representation arrays are in-memory and cleared or released at lifecycle boundaries
• WavLM accepts only finite 16 kHz mono PCM over loopback, writes no audio, logs no request body, and returns only transient layer summaries plus provenance
• all measurements remain descriptive, deterministic engineering features with clinicalValidation: "none"

Validation

• pnpm test - 198 unit tests, structure/privacy validation, typecheck, production build
• pnpm test:browser - 25 Playwright scenarios
• pnpm test:python - 4 sidecar schema/service tests

Real WavLM checkpoint extraction remains a documented manual smoke because model weights are intentionally absent from required CI.