feat: restore high-signal reflection workflow

[Disdjj]

Summary

• restore a bounded Reflector role for high-signal user corrections, verified failures, major rework, and instruction violations
• store privacy-safe candidates under .llmdoc-tmp/reflections/pending without persisting full transcripts or adding a tracked reflection kind
• let pending candidates independently trigger SessionStart and Stop update signals, while preserving lesson_candidates across compaction
• add the --reflection update quality gate, deduplication, stable-doc promotion, and resolved lifecycle
• keep Claude and Codex role and skill surfaces in sync, with automated surface checks
• update the dogfood llmdoc knowledge and promote the rebase-before-editing lesson into the existing development workflow owner

Validation

• npm test: 42/42
• npm run lint
• npm run typecheck
• npm run check:prompts
• node scripts/check-codex-surface.mjs
• npx -y @tokenroll/llmdoc validate
• real dogfood: a pending-only candidate triggered update with zero code delta, then Stop returned to silent after promotion and resolution

Merge requirement

This PR contains llmdoc fingerprint commits whose revisions must remain reachable. Merge with a merge commit; do not squash or rebase merge. If history is rewritten, rerun fingerprint --all on the rewritten history before merging.