fairtask-eval skill: run the evaluation tooling from inside an agent session

[mnkprs]

Why

The screening skill (/fairtask) covers the product use, but everything else a reader might want to see — the evaluation set, scores, the comparison report, the evidence audit, one instance laid out — still required a terminal. Task authors and judges work inside agents; the evaluation should be reachable there too, with numbers that come from the scripts rather than from the model's memory.

What changed

• skills/fairtask-eval/SKILL.md: maps a request to one of the repository's evaluation commands (data:eval-set, score, report, audit, code-check, data:annotations --check, show, trajectory), runs it from the repository root, prints the output verbatim, and adds one or two sentences on how to read it. Offline except the pinned data fetches; makes no model calls; refuses to start paid reproduction runs and points at REPRODUCE.md instead.
• Plugin now carries two skills; CI's inventory check expects Skills (2); the manifest validator checks both skills' frontmatter.
• README §2b mentions the skill; plugin description updated.

Verified by invoking the skill for "show the evaluation set": it ran npm run data:eval-set, printed the 30-row table, and git status stayed clean.

[mnkprs]

Superseded by a clean branch: this one accidentally carried a 6-byte change to data/raw/ensembled_annotations_public.csv, which the provenance check in CI caught.