record mutations hold their DID's write-stripe lock from read through commit, and every multi-DID path acquires its complete lock set in one canonical order before lifecycle counts: the crawler groups each page by lock index instead of capturing nearly all 256 stripes behind one long redaction, and resync/track pre-acquire sorted stripe sets. writers re-check filter exclusion under the lock so a drop_repo racing in-flight work cannot be undone by a stale write; backfill tasks drop orphan pending keys for excluded DIDs before doing resolver or network work, and the gone/error retry scans take a per-repo transaction instead of one global pass. issue: hydrant-vvq, hydrant-ara