Constrain visible roof grouping to compact clusters

scripts/normalize_belgium_building_labels.py

@@ -91,7 +91,7 @@ def _polygonal(geometry: Any) -> Any | None:
 
 
 def merge_touching_roof_instances(features: list[dict[str, Any]], source_name: str) -> list[dict[str, Any]]:
-    """Dissolve only touching/overlapping footprints into imagery-visible roof instances."""
+    """Dissolve only compact touching groups into imagery-visible roof instances."""
     if not features:
         return []
     source_geometries = [(feature, shape(feature["geometry"])) for feature in features]
@@ -105,6 +105,17 @@ def merge_touching_roof_instances(features: list[dict[str, Any]], source_name: s
             if geometry.intersects(component)
         ]
         source_ids = sorted(str(feature["properties"]["source_feature_id"]) for feature in contributors)
+        envelope_area = component.envelope.area
+        fill_ratio = component.area / envelope_area if envelope_area else 0.0
+        # Large connected blocks and irregular chains are administratively
+        # adjacent but not one reliably box-shaped roof target. Preserve their
+        # native instances instead of creating a giant ambiguous detector box.
+        if len(contributors) > 12 or fill_ratio < 0.55:
+            for feature in contributors:
+                retained = json.loads(json.dumps(feature))
+                retained["properties"]["label_semantics"] = "native_instance_complex_touch_group"
+                merged.append(retained)
+            continue
         properties = dict(contributors[0]["properties"])
         properties.update(
             {
@@ -112,6 +123,7 @@ def merge_touching_roof_instances(features: list[dict[str, Any]], source_name: s
                 "source_feature_ids": source_ids,
                 "source_feature_count": len(source_ids),
                 "source_feature_id": source_ids[0],
+                "roof_group_fill_ratio": fill_ratio,
             }
         )
         digest = hashlib.sha256(component.normalize().wkb).hexdigest()[:24]