A reference check that could not tell (a failed server read) dropped
the job's mark. A staged job has no other way out, so it then waited
forever. The check now tells deleted, still referenced, and unknown
apart: unknown keeps a staged job's mark and lets a job that can
upload go on, as before.

Two retries could each run the check on the same job; one that could
not tell let it start uploading while the other deleted it. Checks now
run one pass at a time, and every retry waits for the running pass
before starting an upload.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>