vellum-memory-v3-migration

Run the migration skill, which snapshots memory/concepts/, freezes consolidation triggers, authors v3 articles into a staging tree, audits for lost content, and cuts over only after a retrieval eval gate passes. The live corpus is never edited until the final step.

Memory v3 retrieves at section grain: each article's lead acts as a retrieval card and individual ## sections are independently matchable. Flat v2 pages with bullets and summary frontmatter collapse into one bloated card under v3, which is why reshaping is required.

Skip it when the corpus is already in v3 article shape and memory.v3.live is true, or when memory/concepts/ is empty or near-empty. In the empty case, run the vellum-memory-v2-migration skill first since there is nothing to reorganize.

It works on a read-only snapshot, writes only to a staging tree, and records provenance mapping every staged article to its source pages. A mechanical quote-screen plus a semantic reader panel audit then verify no load-bearing fact was dropped before cutover.

Cutover is blocked. The failing turns identify clusters that under-retrieve, typically due to thin leads, over-merged articles, or missing links; you repair the staged articles, re-run the eval on the same pinned turn set, and re-judge until the gate passes.

Yes. The cutover step creates a timestamped backup of memory/concepts/ before deploying. Restore it with rsync, set memory.v3.live back to false, and the restored consolidation triggers resume v2-shape operation on the restored corpus.