Launch V45 expanded corpus training loop

docs/CODEX_EXECUTION_LOG.md

@@ -11963,3 +11963,25 @@ Open:
   protected test and pure-background sets.
 - Obtain final representative human contact-sheet approval, promote the exact
   checksummed model, and redeploy from the canonical Tower directory.
+## 2026-07-29 - V44 Flemish corpus expansion and V45 training loop
+
+- Acquired twelve independent Flemish train-only AOIs covering dense coastal
+  development, ribbon development, industry, dunes, docks and pure-empty
+  agricultural background. The immutable expansion manifest SHA-256 is
+  `62fdb4adee46ff67a40636f30a22fc97b8c9bd14516cbe0a24ed0de2db56f4fb`.
+- Refined 2,391 labels with SAM2 on `cuda:0`; 124 unsafe refinements fell back
+  to their source geometry. Lossless overlap retiling retained all 2,515
+  unique source labels in 192 tiles.
+- Composed the 168-AOI V44 corpus and retained the exact V42 calibration,
+  protected-test, background-test and internal-validation assignments. The
+  rotated manifest SHA-256 is
+  `448bb3ec426a6af129f01d29732160f84b9936c9871a033de13a7d98b9d25112`.
+- Corpus evidence reports 30,113 accepted features, zero temporal-unknown
+  samples, zero spatial leakage and zero protected samples in training. The
+  YOLO dataset-quality audit passed. Automated corpus checks have no failures;
+  final human contact-sheet review remains deliberately deferred until every
+  model release gate passes.
+- Added the reproducible V45 command. It starts from the strongest V43 Flemish
+  checkpoint (`iteration-005`, SHA-256
+  `be8c5a4d27dfcd03e29e59ce78ea66d3c772c3e612c08b71fd457d3006001b71`)
+  and trains against the expanded V44 corpus on the NVIDIA server.

scripts/tower-v45-v44-loop-command.json

@@ -0,0 +1,73 @@
+[
+  "docker",
+  "exec",
+  "-d",
+  "geointel",
+  "/opt/geointel/venv/bin/python",
+  "/app/scripts/run_belgium_building_training_loop.py",
+  "--initial-model",
+  "/app/storage/training/building-be-v43-v42-closed-loop-r1/iteration-005/candidate.pt",
+  "--train-yaml",
+  "/app/storage/operator-data/building-be-v44-rotated-holdouts-r1/train/dataset.yaml",
+  "--train-summary",
+  "/app/storage/operator-data/building-be-v44-rotated-holdouts-r1/train/yolo_tile_dataset_summary.json",
+  "--dataset-audit",
+  "/app/storage/training/building-be-v44-corpus-audit-r1/belgium-building-corpus-audit.json",
+  "--train-quality-audit",
+  "/app/storage/training/building-be-v44-yolo-quality-audit-r1/operator_yolo_dataset_quality_audit.json",
+  "--calibration-summary",
+  "/app/storage/operator-data/building-be-v44-rotated-holdouts-r1/calibration/yolo_tile_dataset_summary.json",
+  "--test-summary",
+  "/app/storage/operator-data/building-be-v44-rotated-holdouts-r1/test/yolo_tile_dataset_summary.json",
+  "--background-summary",
+  "/app/storage/operator-data/building-be-v44-rotated-holdouts-r1/background-test/yolo_tile_dataset_summary.json",
+  "--corpus-manifest",
+  "/app/storage/operator-data/building-be-v44-rotated-holdouts-r1/operator_samples_manifest.json",
+  "--output-dir",
+  "/app/storage/training/building-be-v45-v44-closed-loop-r1",
+  "--iterations",
+  "20",
+  "--epochs",
+  "80",
+  "--patience",
+  "8",
+  "--batch",
+  "8",
+  "--workers",
+  "0",
+  "--max-det",
+  "1000",
+  "--imgsz",
+  "640",
+  "--optimizer",
+  "AdamW",
+  "--lr0",
+  "0.0001",
+  "--mosaic",
+  "0",
+  "--scale",
+  "0.15",
+  "--translate",
+  "0.05",
+  "--degrees",
+  "180",
+  "--flipud",
+  "0.5",
+  "--fliplr",
+  "0.5",
+  "--warmup-epochs",
+  "1",
+  "--warmup-bias-lr",
+  "0.01",
+  "--hsv-h",
+  "0.01",
+  "--hsv-s",
+  "0.2",
+  "--hsv-v",
+  "0.15",
+  "--seed",
+  "20260944",
+  "--yolo",
+  "/opt/geointel/venv/bin/yolo",
+  "--evaluate-initial-model"
+]