A retry is a snapshot restore, not a second sandbox

Your migration agent fails about one run in five. That's fine. Agents fail; the loop retries.

What isn't fine is what a retry costs on a metered sandbox platform. Every attempt provisions a fresh environment, reinstalls the toolchain, clones the repo, and starts the meter again. Five attempts is five boots. The retry budget becomes a line on the invoice.

On Hoody the agent takes a snapshot before it touches anything. Snapshots are copy-on-write, so the snapshot is metadata, not a copy. When the run goes sideways, the container restores to it and the agent tries again on the same disk, in the same container. No second environment exists. No second meter starts.

The fourth attempt costs the same as the first: nothing beyond the box.