[rmkr-dev]

Adds a Skillbook entry and a Promptbook entry for keeping a RAG index in step with its sources: updates, deletes, permission changes, and embedding model upgrades.

Skill: rag-ingestion-index-lifecycle-guardrails

• Stable document IDs, deterministic chunk IDs, chunk metadata (content, ACL, chunker, and embedding versions), and a per-document manifest
• Change-feed sync with cursors saved only after writes succeed; chunk-set replacement instead of upsert-only; hash-skip and idempotent, order-safe workers
• Tombstones first, then deletes that reach vectors, staging, embedding cache, answer caches, and fixtures; scheduled reconciliation
• ACL principals on every chunk, one query-filter builder applied inside the search, fast path for permission changes, fail closed on missing ACLs
• One embedding model version per index, alias-based blue/green re-index with dual writes, checks, eval gate, and rollback window
• Freshness SLOs for content, ACL, and delete visibility; drift metrics; lifecycle tests

Prompt: harden-rag-ingestion-pipeline

• Inventory, identity and sync diff, delete/ACL diff, re-index script, metrics, and a test list, with placeholders for sources, store, embedding model, identity, data rules, and caches

Copilot, Claude Code, and Codex placement notes are included. Catalogs now 58/58; node scripts/site-check.mjs passes locally.