improve fixtures and relability

[justinmclean]

Summary

Hardens the skill-eval suite so cases stop failing on legitimate model
variation, while still catching genuine skill and model regressions. Also
fixes the real skill gaps the evals surfaced along the way.

The guiding rule throughout: only descriptive, formatting, and set-valued
fields were loosened. Verdict fields (classification, decision, severity,
boolean flags, enum values) stay exact, so a wrong answer still fails.

Grading robustness (harness + fixtures)

• Scalar normalization in the runner: non-prose string comparison now
  ignores case and collapsed/surrounding whitespace, so a correct verdict
  written invalid or request_changes matches INVALID / REQUEST_CHANGES.
  Different values (READY vs NEAR-MISS, advisory vs blocking) still fail.
• Judged set/list fields via per-step grading-schema.json: contribution
  tracks, blocking_factors, failing_criteria, violations, review
  findings, render sections, inventory blocks, confirm items, and other
  descriptive lists are graded for "same conclusion" rather than exact set
  equality, tolerating order and granularity differences.
• Redundant derived fields dropped from assertions where they only
  restated a verdict (e.g. axes_without_findings, the subjective
  tracks_thin_or_absent).
• Permissive review-findings fixtures: the pairing review cases now assert
  the essential finding(s) with axis and severity and explicitly allow
  additional findings, since a thorough reviewer legitimately returns a
  superset. A model that misses or misclassifies the key finding still fails.

Skill fixes (real gaps the evals found)

• contributor-nomination: validate the GitHub login before any API call and
  reject path-traversal / malformed handles; scope the mentoring track to the
  off-GitHub row rather than the community-interaction narrative.
• good-first-issue-author (readiness): R4 rejects a prose summary posing as
  acceptance criteria; R6 degrades gracefully when the link cannot be fetched or
  matched to config.
• good-first-issue-sweep: G3 excludes a bare command/CLI name as a code
  pointer, with an independence rule (one failing criterion is a NEAR-MISS) and
  a worked example.
• issue-reproducer: clarify E-vague vs E-precise (a named error is not
  enough if reproducing needs invented setup).
• pairing-multi-agent-review: a silent invariant violation is a blocking
  correctness finding, not advisory.
• pairing-self-review: add a prompt-injection guard (injection content is a
  blocking security finding, treated as data).

Harness context the tests were withholding

• Feed the Step 4 health-rating thresholds to the step-3-aggregate case.
• Define pool_name for explicit-key selectors, and runtime_version as the
  full runtime stack.
• Treat ordering as vacuously satisfied on an empty dependency-audit report.

Test-case fix

• Rewrote the no-code-pointer classify case to name nothing locatable, so it
  cleanly tests the intent instead of arguing over whether a command name
  counts as a pointer.

Type of change

• Skill change (.claude/skills/<name>/) — eval fixtures updated below
• Tool / bridge contract (tools/<system>/*.md)
• Python package (tools/*/ with pyproject.toml)
• Groovy reference impl
• Cross-cutting (RFC, AGENTS.md, sandbox, privacy-LLM)
• Documentation (docs/, README.md, CONTRIBUTING.md)
• Project template (projects/_template/)
• CI / dev loop (prek, workflows, validators)
• Other:

Test plan

• prek run --all-files passes
• For Python packages touched: uv run pytest / ruff check / mypy passes
• For Groovy bridges touched: command-line invocation tested end-to-end
• For skill changes: eval suite passes for the affected skill
  (PYTHONPATH=tools/skill-evals/src python3 -m skill_evals.runner tools/skill-evals/evals/<skill>/)
• For skill behaviour changes: a new or updated eval fixture is included in this PR
  (a regression test for the bug fixed / the behaviour added — see CONTRIBUTING.md)
• Other: