Recover governed runtime provenance for legacy YOLO models

scripts/README.md

@@ -771,6 +771,22 @@ selected sample slugs and excluded sample slugs. Split validation still applies
 after filtering, so a manifest-backed holdout cannot be selected as training by
 omitting it from `--val-samples`.
 
+### Runtime provenance migration for surviving checkpoints
+
+`migrate_runtime_model_provenance.py` recovers the narrow runtime provenance
+contract for a legacy local YOLO checkpoint only when the active model, retained
+`best.pt`, base model, training summary, arguments, results, dataset summary and
+dataset YAML all match their recorded SHA-256 values. The default run is a
+read-only dry-run; `--apply` writes the neighbouring
+`.geointel-model.json` atomically and registers the exact model bytes as an
+immutable `model` source snapshot in Postgres.
+
+The generated receipt explicitly does not claim a missing historical commit,
+container digest, human review, protected-test independence, national validity
+or promotion. It downloads nothing and runs no inference. After applying, use
+`yolo_preflight.py --check-model-load --json` and a separate real inference
+smoke to validate the actual CUDA runtime.
+
 After rebuilding the all-in-one image, the operator scripts are available inside
 the container at `/app/scripts/...`. Before rebuilding, use the host checkout or
 temporarily copy scripts into the running container for one-off data prep.