Treat GPU workspaces as caches, not training prerequisites

Reusable training-recovery lesson: a full ephemeral GPU-volume copy should not block a run when four durable inputs exist independently—an exact source commit, a sealed dataset with immutable splits and hashes, a hash-pinned starting checkpoint, and reproducibility receipts/configuration. Rebuild a clean, collision-free run directory from those inputs and verify hashes before execution. This does not apply when any required artifact exists only on the old volume; audit that boundary first rather than assuming either full-copy dependence or full recoverability.