fix(findings): replay task declares the eval's scorers so inspect-dataset's scorer-aware scanners judge applicability correctly

[MattFisher]

Summary

Closes #27. A task-mode dataset scan replays dumped samples through a stand-in task that had no scorer, so inspect-dataset 0.5.0 marked answer_length and inconsistent_format not applicable with "this task has no scorer" for every eval.

• The dump script records the task's scorer registry names in its meta, read the same way inspect-dataset reads them off a task.
• The adapter passes them to the replay in INSPECT_AUDIT_SCORERS and records them under inputs.dataset.scorers.
• The replay task declares one stub scorer per name, registered under the eval's exact scorer name. The stubs never score; the name is all inspect-dataset needs.
• The Inputs section says which scorers the scan assumed.

Verified on StereoSet

The scan now reports the task's real scorers, inspect_evals/multiple_choice_scorer and inspect_evals/stereoset_scorer, and skips the two verbatim-text scanners with "this task scores with ..." rather than "no scorer". That skip is correct: StereoSet is not scored by verbatim text comparison, and the earlier 2169 findings were mostly answer_length noise on it.

The other drop noted on #27, duplicate_questions 18 to 0, is explained and by design: inspect-dataset 0.5.0 identifies a sample by its question and its choices, and StereoSet's eight repeated contexts carry different candidate sentences. No change needed.

Checks

• uv run pre-commit run --all-files clean
• uv run basedpyright src 0 errors
• uv run pytest -q -p no:cacheprovider 575 passed, 5 skipped

🤖 Generated with Claude Code