The review queue mixed every kind of card together. ADR-0022 orders it by note
type: the bucket derivation, then the drain, then the re-insertion of an again
card, then the persistence of a pending card across launches.
It landed in stages, one small piece per commit, each with its own test. The staging matters because the queue is the part of a spaced-repetition app that a learner actually feels. A wrong interleave is a subtle daily annoyance, not a crash.
The again re-insertion was the one that took thought: a card you fail should
come back in the same sitting, but not immediately, and not at the expense of
the cards you haven’t seen yet.