12,686 repositories fetched and emitted, zero queued for completion, every RocksDB row still not_started. The archive was healthy the whole time; the completion commit never fires. This corrects experiment 3's read. That failure was treated as a metric defect and the restart-stable gauge fixed the reporting -- the gauge then reported zero truthfully for 55 minutes. The defect underneath is the commit itself. It is also the exact property the differential oracle is documented as unable to confirm end to end. That row was called an evidence limit rather than a known defect. It was a real defect.