[aroff]

Summary

Integrate eval authoring into the existing fastskill/ skill, rather than distributing a second skill from the CLI repository. The entrypoint routes to the authoring workflow, supporting references, and a copyable invoice example. Version the expanded package as 2.1.0.

Retain the useful repository-only evals/ harness, generator, fixtures, and historical measurements. Move authoring scenarios into evals/authoring/fixtures/, outside the released ZIP. Document the assessment and exact changed-file coverage inventory in specs/002-eval-authoring-consolidation.md.

Repair staging that could erase an existing destination, propagate runtime failures from subset/full sweeps, and correct stale documentation and CI comments. Add deterministic package, preservation, invalid-input, and runner tests to CI.

Companion CLI cleanup: gofastskill/fastskill#333. These are coordinated draft changes; the linked main-branch documentation becomes available after the companion changes land. Do not merge or enable auto-merge.

Verification

• server02, pinned FastSkill 0.9.230: 12 Python tests passed; namespace check, generator parity, vacuity guard, and diff check passed.
• Real ZIP installation as 2.1.0; copied example and all three preserved v2 suites validated.
• Installed-package Claude Sonnet 4.5 suite extension completed; independent validation and byte/row preservation checks passed. This is file authoring, not a new target-runtime sweep.
• Existing positive/negative v1 fixtures exercised; failure detected for its intended reason.
• Bash line coverage: run.sh 93.1%, stage.sh 100%; exact embedded Python 100%. Bash branch instrumentation is unavailable and not claimed.
• Companion CLI: six engine contract and 13 documentation/related tests passed after relocation; prior full Linux, two-tool authoring, and calibrated native-judge evidence retained.

Limitations

No new full live model baseline is claimed. Existing baseline numbers are historical. The existing negative-control script remains useful but its assumptions and missing budget-invariance assertion are documented. Native-judge placeholders require owner configuration. GitHub CI status is reported separately from local evidence.