fix(detectors): match mitigation.Prefixes at the start of the response

[chiruu12]

Fixes #2066

Prefixes lists sentence openers but never set matchtype, so it inherited StringDetector's "str" default and matched anywhere in the output. With the 1 - r inversion, one incidental occurrence of a listed token scored a fully compliant answer as a refusal. Responses containing unethical or naming OpenAI mid-text were scored as refusals while giving complete instructions, so attack success rate was understated.

This sets matchtype = "startswith" in DEFAULT_PARAMS, the same way shields.Up and shields.Down do it and for the same reason.

Also swaps hit_desc and pass_desc. After the inversion a hit means no refusal prefix was found, but the strings said the opposite. MitigationBypass in the same file already has them the right way round. These ship in plugin_cache.json, so they are user facing.

Added a test asserting a compliant response containing a listed token mid-text scores as a hit. It fails on main and passes here.

pytest tests/detectors/ is green, 756 passed and 34 skipped.

One thing to flag: this moves some currently passing attempts to hits for probes.phrasing.PastTense, probes.phrasing.FutureTense and probes.dra.DRA, which have entries in garak/data/calibration/calibration-2026-02.json. Happy to follow up on the calibration side separately if you would like those regenerated.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.