Audio: bounded voice and signal candidate screening

[codingshreyash]

Summary

Adds bounded candidate screening for matched harmful and benign spoken requests:

• PairedDirect varies a fixed request across deterministic speed, echo, band-pass, and seeded-noise recipes.
• AcousticVoiceBestOfN varies configured TTS voice presets while keeping the request fixed.

Each candidate records its group, index, recipe, synthesis identity, and audio hash. An optional validation manifest can fail closed on incomplete or unscoreable candidate pairs. Without that manifest, no independent-ASR intelligibility claim is made.

This is a bounded diagnostic inspired by Best-of-N Jailbreaking, not a reproduction of its random composed search or stop-on-success procedure. Voice results are meaningful only when hashes and on-target qualification show that the backend honoured the requested preset.

Relationship to other work

The upstream duplicate search found no open bounded audio candidate-screening PR. This is a focused child of the spoken tool-risk PR and does not include generated audio, reports, local endpoint configuration, or experiment artifacts.

Verification

Run locally with Python 3.12:

XDG_CACHE_HOME=/private/tmp/garak-pr6-cache \
XDG_DATA_HOME=/private/tmp/garak-pr6-data \
python -m pytest -q \
  tests/probes/test_probes_audio_acoustic.py \
  tests/probes/test_probes_audio_bon.py \
  tests/test_docs.py

Result: 713 passed.

Plugin/translation integration selection: 2 passed.

python -m black --check <changed Python files>
git diff --check

Result: clean.

No new dependency is added.

AI assistance

ChatGPT was used to review this change.

[codingshreyash]

Closing — BoN deferred to a fast-follow after the first-wave stack (#9–#11). Its acoustic-robustness surface overlaps the existing ToolRiskAcousticBestOfN, and its candidate-validation-manifest logic is being folded into the shared PETTS base rather than shipped as a standalone module.