Docs ↔ code drift audit

Verify every factual claim against the source. Anything unverifiable or wrong gets fixed, or listed if --report-only was passed ($ARGUMENTS).

Checks

1. Commands: every pnpm <script> mentioned in README.md, CONTRIBUTING.md, CLAUDE.md, .claude/**, and .github/** exists in package.json scripts.
2. Files and paths: every path mentioned in docs exists (ls/Glob). Pay special attention to the architecture trees in README.md and CLAUDE.md.
3. CI/workflows: descriptions of .github/workflows/*.yml (triggers, jobs, what runs) match the YAML.
4. Env vars: .env.example lists exactly the variables read by nuxt.config.ts, server/, playwright.config.ts, and workflows, each with an accurate description.
5. Product facts: palette, font, shade, and radius lists in app/data/*.md, content/learn/**, and server/utils/aiSystemPrompt.ts match shared/constants/theme.ts. AI model lists in docs match app/types/ai.ts.
6. Dependencies: libraries named in docs (e.g. SECURITY.md's "DOMPurify", "Dependabot") are actually installed or configured and used.
7. Claude setup: skills and agents named in CLAUDE.md and .claude/README.md exist; .claude/rules/* paths: globs match real files; hooks referenced in .claude/settings.json exist.
8. Config hygiene: .prettierignore, .gitignore, the knip key in package.json, eslint.config.mjs, .github/labeler.yml, and .github/CODEOWNERS don't reference paths that no longer exist.

Output

A table of file:line | claim | reality | action. Fix what you can (docs only; if the code is wrong, report it rather than change behavior), run pnpm format, and summarize the fixes and open questions.