test(evals): move repository metadata out of criteria.json

[martinfrancois]

Summary

• Problem: tessl eval lint (CLI 0.102 and later) rejects the metadata key that every criteria.json here carries and warns on the per-item category field. Publishing still works today, but the format is outside Tessl's schema. Tracked in the linked issue.
• Why it matters: a future CLI could turn the lint into a publish-time rejection, and the repository's own validators depend on that metadata.
• What changed: each scenario gets a criteria-meta.json with the metadata object and a categories map from checklist name to category; criteria.json keeps only context, type, and checklist items with name, description, max_score. validate_eval_criteria.py merges the two (and fails a criteria.json that still carries metadata or categories), classify_eval_result.py reads the sidecar, eval_impact.py fingerprints it, the test fixture writes it, and the agent docs describe the layout.
• What did not change: no task text, criterion name, description, or max_score changed anywhere (pure metadata move, so no hosted rerun is needed per docs/agents/evals.md); runtime skill untouched.

Change Type

• Skill behavior
• Evals or scoring
• Documentation
• CI, release, or dependency automation
• Repository metadata or contribution process
• Other maintenance

Linked Issue

• Closes test(evals): criteria.json metadata fails the current tessl eval lint #89

User-Visible Behavior

None. The published scenarios carry one extra small JSON file.

Validation

Checks most contributors can run:

• python3 scripts/validate_eval_criteria.py evals evals-reference evals-regression -> passed
• python3 scripts/test_validate_eval_criteria.py -> OK
• python3 -m py_compile scripts/*.py -> passed
• tessl eval lint ., tessl eval lint evals-reference, tessl eval lint evals-regression -> all scenarios valid, no errors, no warnings
• scripts/pre_submit_gate.sh --plan-only -> runs

Tessl-authenticated checks:

• Not needed: pure metadata move; the repository rule in docs/agents/evals.md exempts it from hosted reruns.

Human Verification

Diffed every criteria.json before and after the move: only the metadata object and category fields left; names, descriptions, and max_score are byte-identical.

Review Checklist

• The change is scoped to the sections, skill files, evals, or workflows described above.
• Validation that applies to this change is checked above, or any unavailable check is explained.
• If evals or benchmark claims changed, the eval scenarios remain fair and do not leak answer keys, run IDs, or fixed score claims into runtime references.
• New or moved eval scenarios follow the classifier recommendation, or the PR explains the maintainer-approved override. (None moved.)
• Every retained eval scenario has a 100% with-context result, or any below-100 result is documented as blocking follow-up rather than classified/reportable coverage. (Unchanged scoring.)
• PR title or squash title uses Conventional Commits.
• Redaction checked: no tokens, private links, private eval artifacts, local host paths, or proprietary Java source.

AI Assistance (if used)

• AI-assisted PR
• I confirm I understand and reviewed the change