[aroff]

Summary

Keep FastSkill evaluation lightweight while making the results easier to interpret:

• Add a six-case, one-trial smoke command without native-judge prerequisites. Preserve the full 42-case benchmark as an opt-in run.
• Rename trace-fragment matching so it is not presented as semantic answer accuracy. Remove the incorrect prohibition that rejected correct refusals mentioning a nonexistent publish command.
• Add labelled judge-calibration contrasts without claiming they have been calibrated against a live judge.
• Restrict hosted CI to its configured Claude authentication route, forward the target model explicitly, and require negative fixtures to fail for their intended check.
• Preserve original metadata and per-trial paths in the consultation negative control, asserting the specific consultation flip and unchanged tool-budget verdict.
• Add create/repair/extend authoring exercises with actual validation and file/row preservation checks.