Checkpoint recovery API
Inspect saved updates and explicitly restore a selected checkpoint. Update N is evaluated by pass N+1; the final update normally remains unevaluated. Evaluation health describes configured skip limits, nonzero processed counts and finite likelihood statistics, not recognition quality. Legacy telemetry without matching hashes is unverified.
Restoration never selects a checkpoint automatically. Stop concurrent training
and readers first. The previous model is backed up under recovery-history;
completion markers, provenance and cached sendump are invalidated before the
selected parameters and counts are published. Training diagnostics and original
checkpoints remain available. dry_run=True performs inspection only.
Public API for explicit checkpoint inspection and restoration.