feat(PROJ-000): add skill hardening loop

[HenryCordes]

Summary

Adopts a "small always-loaded context, self-hardening skills" strategy: skills move to the
current folder format, failures now become permanent skill fixes via a new improve-skill
meta-skill, and the always-on context (CLAUDE.md, docs) shrinks to only what a model can't
infer. Net −13 lines while adding a skill.

Changes

• Migrate all skills to <name>/SKILL.md folders (git mv, history preserved); scaffold-module gains colocated templates/ for the types/implementation/test trio
• Add improve-skill: the hardening loop — name the failure, find the owning skill/doc, add the smallest preventing rule, log it in a ## Lessons changelog
• Seed scaffold-module's Lessons with the three real failures from docs/CASE_STUDY.md (fake-timer rejection race, non-enumerable Error.cause in toMatchObject, ESLint scoping for harness scripts)
• Shrink CLAUDE.md ~150 → ~80 words; drop listings that skill discovery and AGENTS.md already provide
• Prune model-known content from REACT_GUIDELINES.md (−45%) and DEVELOPMENT.md; both now state they hold project conventions only
• README: new "Small context, self-hardening skills" section documenting the strategy; refreshed skill references and repo tour

Testing

• pre-commit (lint-staged) and commit-msg hooks passed on the commit itself
• Verified no broken relative links across README/AGENTS/CLAUDE/docs/skills; only historical specs still mention old skill paths (left as point-in-time records)
• Markdown + .template files only — pnpm test unaffected; CI runs the full four checks

Related

PROJ-000

Checklist

• pnpm lint:fix && pnpm typecheck && pnpm test pass locally
• Tests added/updated for the change — N/A: docs/tooling only, no runtime code touched
• Screenshots or a recording attached for UI changes — N/A: no UI
• Branch name follows {username}/{TICKET}-{description}