Raising retry_after stops one dispatch from being re-reserved, but it
does nothing against a *duplicate* dispatch. The ai queue runs two
concurrent workers, so a double "Enable AI" click (or re-enabling
consent mid-run) launches two backfills that read the same pending
snapshot and both bill the model — the very harm tries=1 was meant to
avoid but structurally cannot.
Implement ShouldBeUnique keyed on the user id, mirroring the sibling
RetryTransientAiCategorizationJob, so a user has at most one backfill in
flight. Add a test anchoring the dedup key.
6f7c60a8fe