feat(provenance): govern source snapshots and data inputs

scripts/build_regional_yolo_expert_dataset.py

@@ -6,8 +6,23 @@ from __future__ import annotations
 import argparse
 import hashlib
 import json
+import sys
 from pathlib import Path
 
+SCRIPT_DIR = Path(__file__).resolve().parent
+if str(SCRIPT_DIR) not in sys.path:
+    sys.path.insert(0, str(SCRIPT_DIR))
+
+from training_dataset_eligibility import (  # noqa: E402
+    TrainingEligibilityError,
+    assert_frozen_manifest_training_eligible,
+)
+from training_release_manifest import (  # noqa: E402
+    TrainingReleaseError,
+    assert_yolo_summary_bound_to_embedded_training_release,
+    create_training_release_manifest,
+)
+
 
 def sha256(path: Path) -> str:
     digest = hashlib.sha256()
@@ -31,12 +46,35 @@ def main() -> int:
         default=0,
         help="Include each train tile outside the expert region this many times.",
     )
+    parser.add_argument(
+        "--fixture-mode",
+        action="store_true",
+        help="Accept only an explicitly fixture-only corpus manifest; never use for operational training data.",
+    )
+    parser.add_argument(
+        "--review-audit",
+        type=Path,
+        help="Accepted corpus-audit evidence; mandatory for an operational training release.",
+    )
     args = parser.parse_args()
     if args.positive_repeat < 1 or args.negative_repeat < 1 or args.other_region_repeat < 0:
         raise SystemExit("regional repeats must be positive and other-region repeat non-negative")
 
     summary = json.loads(args.summary.read_text(encoding="utf-8"))
     manifest = json.loads(args.corpus_manifest.read_text(encoding="utf-8"))
+    try:
+        assert_frozen_manifest_training_eligible(
+            args.corpus_manifest,
+            fixture_mode=args.fixture_mode,
+            verify_live=True,
+        )
+        assert_yolo_summary_bound_to_embedded_training_release(
+            summary_path=args.summary,
+            corpus_manifest=args.corpus_manifest,
+            fixture_mode=args.fixture_mode,
+        )
+    except (TrainingEligibilityError, TrainingReleaseError) as exc:
+        raise SystemExit(str(exc)) from exc
     samples = {item["sample_slug"]: item for item in manifest["samples"]}
     paths: list[str] = []
     selected_samples: set[str] = set()
@@ -69,6 +107,15 @@ def main() -> int:
         f"val: {args.summary.parent / 'images' / 'val'}\nnames:\n  0: building\n",
         encoding="utf-8",
     )
+    try:
+        release_paths = create_training_release_manifest(
+            train_yaml=dataset_yaml,
+            corpus_manifest=args.corpus_manifest,
+            review_audit_path=args.review_audit,
+            fixture_mode=args.fixture_mode,
+        )
+    except TrainingReleaseError as exc:
+        raise SystemExit(str(exc)) from exc
     evidence = {
         "schema_version": 1,
         "status": "ok",
@@ -88,6 +135,11 @@ def main() -> int:
         "protected_samples_in_training": [],
         "train_list": str(train_list),
         "dataset_yaml": str(dataset_yaml),
+        "training_release_manifest": str(release_paths["release_manifest"]),
+        "training_release_manifest_sha256": sha256(release_paths["release_manifest"]),
+        "training_release_freeze": str(release_paths["release_freeze"]),
+        "training_asset_manifest": str(release_paths["asset_manifest"]),
+        "fixture_mode": bool(args.fixture_mode),
     }
     evidence_path = args.output_dir / "regional-expert-dataset.json"
     evidence_path.write_text(json.dumps(evidence, indent=2), encoding="utf-8")