Audio: spoken-channel reliability precursors

[codingshreyash]

Summary

Adds two inactive, benign calibration probes for the PETTS audio path:

• TTSReliability checks whether short words and symbol boundaries survive synthesis and the target's audio-input path.
• SemanticReliability checks whether simple spoken requests retain their meaning.

Each probe has a dedicated detector and explicit attempt metadata. Missing target text remains unscored; it is not counted as a safe response. This PR contains no tool-risk payloads, overlays, ASR service, or model-as-judge.

Relationship to other work

The upstream duplicate search found NVIDIA#1770 for the underlying PETTS technique and no open PR for these reliability precursors. This PR is a focused child of the provider/provenance refactor in #2.

Verification

Run locally with Python 3.12:

XDG_CACHE_HOME=/private/tmp/garak-pr-audit-cache \
XDG_DATA_HOME=/private/tmp/garak-pr-audit-data \
python -m pytest -q \
  tests/probes/test_probes_audio_tts_reliability.py \
  tests/probes/test_probes_audio_semantic_reliability.py \
  tests/detectors/test_detectors_audio_tts_reliability.py \
  tests/detectors/test_detectors_audio_semantic_reliability.py \
  tests/test_docs.py

Result: 695 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]

Superseded by the rebased, cleaned 3-PR stack based on current feature/technique_intent: #9 (foundation), #10 (probes+detectors), #11 (injection). Content preserved there; closing to consolidate.