Add direct polygon label QA

docs/CODEX_EXECUTION_LOG.md

@@ -12047,3 +12047,28 @@ Open:
   checkpoint (`iteration-005`, SHA-256
   `be8c5a4d27dfcd03e29e59ce78ea66d3c772c3e612c08b71fd457d3006001b71`)
   and trains against the expanded V44 corpus on the NVIDIA server.
+## 2026-07-30 - Direct polygon QA and rejected automatic roof-alignment routes
+
+- Revalidated the three independent V66 Flemish AOIs against their immutable
+  EPSG:31370 rasters and CRS84 GRB responses. Raster bounds, OGC `Content-Crs`,
+  feature coordinates and tile transforms are internally consistent; no CRS or
+  WMS axis-order defect was found.
+- Proved that the existing YOLO contact sheet can make rotated and concave GRB
+  polygons look substantially worse by showing only their axis-aligned boxes.
+  Added `scripts/render_operator_polygon_label_qa.py` so governed polygon/raster
+  alignment is reviewed directly before lossy bounding-box export.
+- Rejected streamed automatic SAM2 masks: the broad trial matched 365/413 source
+  objects but selected trees, roads, parking and open ground; a strict trial kept
+  only 42/413 and still contained false roofs.
+- Rejected local polygon-edge registration: 27/73 Zutendaal, 5/66 Zoersel and
+  32/110 Landen objects passed numerical uniqueness gates, but direct overlays
+  still contained vegetation, road and shadow-edge matches.
+- Rejected a swapped Lambert WMS-axis hypothesis and a live `most_recent` image
+  refresh after isolated visual trials. Neither output entered a training corpus.
+
+### Gate consequence
+
+- V66 remains candidate-only and is not promotion or protected-test evidence.
+- No new training is authorized from these rejected outputs. The next corpus
+  revision must use independently image-visible roof annotations or an official
+  roof-surface product, with direct polygon QA before YOLO box generation.