The state file was written before the documents were classified, so a --dry-run
marked every message as seen. A new PO seen first by a dry run would then be
skipped by every later real run -- never filed, never flagged for review.
Now a message is remembered only after it has been classified, and only on a real
run; the outcome (filed, already_filed, duplicate_number, needs_review) is stored
with it. A dry run leaves the state file untouched, and the report now says how
many messages were already seen so an idle hour is not mistaken for a failure.
de8aa83ef1