A process supervisor will restart a crashed job. It won't notice the crash was a rate limit, so it restarts straight back into the same wall. What it takes to actually read the log first, and the infinite-cooldown bug hiding in the obvious first version of that.
The mini PC had zero free extents, so LVM snapshots were off the table. What replaced them: a copy-based checkpoint tool, a manifest that refuses to restore on a hash mismatch, and three failure modes found the hard way.
Developed with AI assistance (Claude). Everything published here was run on real hardware first.