Add NonEngagement detector (#975)

[azrabano23]

Summary

Implements #975 — a non-engagement detector. Where mitigation.MitigationBypass and mitigation.Prefixes detect genuine safety refusals, this detector flags responses that deflect or change the subject without taking a safety stance — the issue's canonical "ma'am, this is a Wendy's" non-answer. This is a complementary eval signal: used as an extended detector alongside jailbreak/refusal probes, it separates "refused on safety grounds" from "simply did not engage", which the existing detectors conflate or miss.

Changes

• garak/detectors/mitigation.py — new NonEngagement(StringDetector). Mirrors the existing Prefixes structure (StringDetector, lang_spec="en", matchtype="str", hit_desc/pass_desc, RST docstring). Unlike the refusal detectors it does not invert detect() — a hit means a deflection phrase is present. The trigger list is deliberately conservative and topic-change oriented (explicit subject changes, declining-to-discuss without a safety rationale, evasive non-answers), grouped with inline comments so it does not overlap the safety-refusal phrasing in the other detectors. Tagged avid-effect:performance:P0204.
• tests/detectors/test_detectors_mitigation.py — added test_non_engagement_init, test_non_engagement_strings (deflections → 1.0; a genuine answer and a plain safety refusal → 0.0, confirming it doesn't overlap MitigationBypass), and test_non_engagement_none_output (None → None).

Not a duplicate

I checked garak/detectors/mitigation.py and the rest of garak/detectors/: the existing detectors (MitigationBypass, Prefixes, ModernBERTRefusal) all target refusals and invert to score compliance. None detect non-safety deflection. No open PR matches non-engagement/engagement (the only "engagement"-adjacent open PR, #1742, is an unrelated adaptive-attacks probe).

Testing

pytest tests/detectors/test_detectors_mitigation.py tests/test_docs.py -q   # 667 passed
pytest tests/detectors/ -q                                                  # 713 passed, 34 skipped

The generic tests/detectors/ and tests/test_docs.py suites auto-validate the new detector's tags, docstring, and plugin loading. All tests are offline (no model/network access required for this detector).

Note

AI assistance (Claude) was used in developing this change; I reviewed and validated every line, ran the tests above, and am able to defend the implementation. The detector is purely defensive — it classifies a target's output and ships no attack payloads.