SPlit up

pipeline/transform/test_merge.py

@@ -1609,6 +1609,37 @@ def test_best_listing_match_numbered_query_cannot_subset_inflate_across_numbers(
     assert result is None
 
 
+def test_best_listing_match_letter_suffix_flats_do_not_cross_match() -> None:
+    # Regression: the gate uses fuzzy_join's suffix-aware tokens, so "8A" and
+    # "8B" are different numbers. Under the old digit-only tokens both looked
+    # like {8} and token_sort scored ~93, attaching the wrong flat's record
+    # whenever the true candidate was absent from the bucket.
+    candidates = [{"pp_address": "8B HIGH STREET"}]
+    result = _best_listing_match(
+        listing_uprn=None,
+        query="8A HIGH STREET",
+        uprn_index={},
+        bucket_candidates=candidates,
+        addressed_fields=["pp_address"],
+    )
+    assert result is None
+
+
+def test_best_listing_match_building_listing_cannot_absorb_single_flat() -> None:
+    # Regression: set equality (not subset) over number tokens, so a whole-
+    # building listing "188 GREAT NORTH WAY" no longer matches "FLAT 1 188
+    # GREAT NORTH WAY" (token_set would have scored the pair 100).
+    candidates = [{"pp_address": "FLAT 1 188 GREAT NORTH WAY"}]
+    result = _best_listing_match(
+        listing_uprn=None,
+        query="188 GREAT NORTH WAY",
+        uprn_index={},
+        bucket_candidates=candidates,
+        addressed_fields=["pp_address"],
+    )
+    assert result is None
+
+
 def test_finalize_listings_promotes_overlay_columns_and_filters_to_listing_rows() -> (
     None
 ):