The nine-hour job runs for nine hours

The backfill takes nine hours. On platforms with execution ceilings, that fact becomes architecture: split the work into resumable chunks, checkpoint to external storage, write the resume logic, test the resume logic.

On your own box, the fact stays a fact. The job runs as a daemon program until it finishes. There's no ceiling to design around, because nothing is metering the run and nothing reclaims the machine.

If you want insurance anyway, take it at the phase boundaries: a snapshot is copy-on-write, close to instant, and records the disk as it stands: so a crash at hour seven restarts from hour six's on-disk state, not from zero.

The chunking layer, the checkpoint store, the resume tests: none of it exists.

You stopped architecting around a timeout you don't have.