the queue is keyed by a random index id so the worker spreads its work, which means it can't be paged by did the way /repos is. pending_dids holds `{DID} 00 {ID}` for every entry, written in the same batch as `pending` by stage_pending_insert and stage_pending_remove, and v12 builds it for existing databases. the NUL keeps it in the same order as `resync`, so the two can be merged by did. nothing point-reads it, so it builds no filters, pins nothing outside the block cache and gets a 4 MB memtable. tests that move the queue now check both keyspaces still match. v11's migration test rewinds to 9 now, since rewinding to 10 trips the deliberate rejection of an unmarked v10 database.