Audit Belgian corpus splits and record CUDA evidence

backend/tests/test_building_label_normalization.py

@@ -15,6 +15,13 @@ assert SPEC and SPEC.loader
 module = importlib.util.module_from_spec(SPEC)
 SPEC.loader.exec_module(module)
 
+ASSEMBLER_SPEC = importlib.util.spec_from_file_location(
+    "assemble_building_corpus", ROOT / "scripts" / "assemble_belgium_building_corpus.py"
+)
+assert ASSEMBLER_SPEC and ASSEMBLER_SPEC.loader
+assembler = importlib.util.module_from_spec(ASSEMBLER_SPEC)
+ASSEMBLER_SPEC.loader.exec_module(assembler)
+
 
 def test_normalizer_retains_native_identity_and_records_rejections(tmp_path: Path) -> None:
     raster_path = tmp_path / "image.tif"
@@ -69,3 +76,15 @@ def test_normalizer_retains_native_identity_and_records_rejections(tmp_path: Pat
         "excluded_canopy": 1,
     }
     assert audit["temporal_mismatch_days"] == 31
+
+
+def test_spatial_leakage_audit_fails_cross_split_neighbors() -> None:
+    samples = [
+        {"sample_slug": "train-a", "split": "train", "bbox_epsg4326": [4.0, 50.0, 4.01, 50.01]},
+        {"sample_slug": "val-a", "split": "val", "bbox_epsg4326": [4.005, 50.005, 4.02, 50.02]},
+        {"sample_slug": "test-far", "split": "test", "bbox_epsg4326": [5.0, 51.0, 5.01, 51.01]},
+    ]
+    audit = assembler.audit_spatial_leakage(samples)
+    assert audit["status"] == "failed"
+    assert audit["findings"][0]["left"] == "train-a"
+    assert audit["findings"][0]["right"] == "val-a"