feat(skills): reorganize skills/ into domain suites

[chr6192]

Summary

Regroups the 9 flat non-eval_pipeline skills into three folder-based domain suites, applying the paradigm eval_pipeline (#187) already established: three-layer structure, directory-as-grouping, every <NN-name>/SKILL.md independently installable (Anthropic Agent Skill protocol), no code-level registry.

• skills/academic-eval/ ← paper-review, bib-verify, + a copy of ref-hallucination-arena, plus a new 00-academic-router entry point
• skills/arena-eval/ ← auto-arena, + a second independent copy of ref-hallucination-arena, plus a new 00-arena-router entry point
• skills/openjudge-core/ ← openjudge, rl-reward (no router — sub-skills are mutually exclusive by description, see the suite's README.md)
• skills/README.md — new top-level suite index + the breaking-change path migration table
• claude-authenticity / mmx-cli / find-skills-combo stay standalone (no other skill references them)

ref-hallucination-arena is intentionally duplicated (not shared or symlinked) rather than "fixed" — the independent-installability constraint means a skill can't rely on content living outside its own folder, so the two copies are free to diverge over time.

This is a breaking change for 6 previously published skill paths (see the migration table in skills/README.md); migrations used git mv to preserve file history.

Full design rationale: docs/superpowers/specs/2026-07-09-skills-domain-suite-proposal.md (Option D, §7).

Testing

Added actor+judge functional-test coverage for the two new suites, reusing eval_pipeline's runner (parametrized via --skill-root/--cases/--report-prefix) instead of forking it per suite — the runner isn't subject to the installability constraint (that applies to SKILL.md content, not internal test tooling), and the change is backward-compatible (verified eval_pipeline's own default invocation produces byte-identical output filenames/report title).

• academic-eval: 12/12 test cases pass (skills/academic-eval/tests/)
• arena-eval: 9/9 test cases pass (skills/arena-eval/tests/)
• 3x-repeat variance check on the 5 trickiest routing-ambiguity cases across both suites: all pass, no variance observed

See each suite's tests/*_skill_audit.md for full results and a caveat on why a clean first-run pass rate shouldn't be over-read (same author wrote both the router content and the test cases).

Test plan

• git mv-based renames preserve file history (git log --follow)
• No repo file still links to a pre-migration skill path (paper-review, bib-verify, ref-hallucination-arena, auto-arena, openjudge, rl-reward)
• Every migrated/new SKILL.md frontmatter has only name + description
• Relative links depth-adjusted for the new nesting (../../ → ../../../ where a skill moved one level deeper)
• eval_pipeline's existing functional-test invocation still produces identical default output (regression-checked)
• Full functional-test suite run for academic-eval (12 cases) and arena-eval (9 cases), plus --repeat 3 on ambiguous routing cases

Made with Cursor

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.