Test transcription shielding while retaining clean audio

[ComputelessComputer]

Add an isolated local experiment with model and codec comparisons, listening reports, and measurement checks. ANLG-393.

---

Summary by cubic

Implements ANLG-393 with an isolated, offline transcription-shield experiment. It fits a perturbation to one complete utterance while retaining the clean audio, then compares transcription and codec effects through reproducible artifacts.

Details

• Adds WER, audio validation, SNR and length checks, deterministic noise control, safe WAV writing, and escaped HTML report generation.
• Makes noise-control comparisons reproducible with a fixed seed and by rejecting shuffled perturbations that would clip instead of changing their distribution.
• Compares tiny.en, base.en, or small.en models before and after a 24 kbps Opus round trip.
• Runs with uv, requires ffmpeg, downloads Whisper models locally, and refuses existing output directories.
• Adds tests for metric edge cases, clipping, WAV round trips, deterministic perturbations, and report escaping.
• Limits the experiment to one utterance and Whisper-based listening and measurement; it does not evaluate real-time calls or human intelligibility.

^{Written for commit 8190a5b. Summary will update on new commits.}

[cubic-dev-ai]

All reported issues were addressed across 2 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>