v0.2.0 - the toolkit can now test itself

The toolkit can now test itself, at every level: its code, its detection
scripts, and the behavior its prose skills are supposed to produce. Plus the
post-release audit fixes from a full-repo sweep.

Added

• Skill evaluation harness (eval.ts): plays a skill through headless
  claude -p inside a throwaway fixture, then verdicts it with mechanical
  gates derived from the skill's own exit conditions, plus an adjunct LLM
  judge for what stays subjective (a judge verdict without a verbatim
  quote as evidence is downgraded to a failure). Verdicts: PASS, FAIL(gate),
  JUDGE-FAIL(criterion), ERROR(infra, never a regression).
• An eval.yaml manifest for every skill (34/34), co-located with the
  skill. Closed gate vocabulary: file_exists, file_absent, grep_zero,
  grep_count, repo_clean, transcript_contains, transcript_absent,
  exit_ok. Manifests stay canonical and are never distributed to targets.
• npm run eval:dry validates every manifest with no agent call, no network
  and no auth - wired into CI. Real runs (eval.ts run [--only <skill>]) are
  on demand and before a release, never per push.
• Mutation harness (npm run test:mutation): 14 declared mutations must
  each turn their covering tests red; a surviving mutation fails CI.
• install.ts --pick <token...> installs exactly the named artifacts.
• Playground fixtures with planted defects and an EXPECTED.md ground truth:
  flawed-app, buggy-app, shipped-feature, design-system,
  audit-target.
• CI matrix on Node 18 (the documented minimum) and 22.
• docs/evaluating-skills.md.

Changed

• Test harness split by domain (tests/) plus one co-located test procedure
  per executable skill script (skills/<skill>/scripts/*.test.ts); 80 tests.
  *.test.ts files are never distributed to target repos.
• bin now points at install.ts. Anyone who linked the package got the
  deprecated rules.ts CLI before this release.
• The interactive selector's keyboard logic is a pure, unit-tested reducer;
  the TTY shell only does I/O.
• Em dashes replaced with plain punctuation across the toolkit.

Fixed

• CI now triggers on main, the public default branch. It previously only
  listened to master and therefore never ran on the public repository.
• Installer: pre-existing artifacts are backed up to *.pre-install.bak on
  first install instead of being silently overwritten (data loss); a
  settings.json whose hooks section has an unexpected shape no longer
  crashes mid-install; a truncated scan is announced instead of silently
  missing deep signals.
• Audit suite: every domain scores with the flat mean defined in
  scoring-model.md (dead weight formulas removed); orchestrator
  contradictions resolved; performance-frontend restored in $ARGUMENTS.
• worktree-env-setup hook: broken .env symlinks are repaired.
• subscription-leak-scan.ts and templates/.adopted.example translated to
  English; docs no longer recommend the deprecated rules.ts sync;
  package.json description translated and engines declared.