Build a resumable batch extraction job

The job is a loop over three idempotent steps (submit, wait, collect) with every decision recorded in SQLite, so a crash at any point leaves state that the next run can finish. The decision that matters most is the order in run(): collect every batch already submitted before submitting anything new. Deterministic custom_ids and INSERT OR IGNORE make replays harmless, but only collecting open batches first stops a restarted job from paying for the same documents twice. The model calls go to a fake service; a real run goes through OpenRouter's Batch API on your own key.