lgtm

pipeline/transform/crime_hotspot_tiles.py

@@ -53,12 +53,13 @@ def _write_geojsonseq(csvs: list[Path], output_path: Path) -> tuple[int, int]:
     to a shared "map point" anchor, so many incidents land on the exact same
     coordinate. Collapsing them into one feature carrying ``count`` (the number
     of incidents) keeps the per-crime-type and per-month filters intact while
-    turning each hotspot into a single high-weight point. That matters because
-    tippecanoe's ``--drop-densest-as-needed`` thins *feature density*, not
-    weight: with one feature per row the busiest streets were silently deleted;
-    with one weighted feature per anchor those hotspots survive and the dropped
-    detail is only redundant duplicate points. The heatmap reads ``count`` as
-    its weight.
+    turning each hotspot into a single high-weight point. That matters for the
+    heatmap weight: each anchor becomes one high-weight point, and tippecanoe's
+    ``--cluster-densest-as-needed`` (with ``--accumulate-attribute=count:sum``)
+    merges any still-too-dense low-zoom features by *summing* their counts
+    rather than dropping them, so the total heat weight is conserved across zoom
+    levels and the surface no longer jumps at tile-zoom boundaries. The heatmap
+    reads ``count`` as its weight.
     """
     grouped = (
         pl.scan_csv(
@@ -144,7 +145,18 @@ def build_crime_hotspot_tiles(
                 str(min_zoom),
                 "--maximum-zoom",
                 str(max_zoom),
-                "--drop-densest-as-needed",
+                # Merge (don't delete) the densest features at low zoom and sum
+                # their incident counts into the surviving point, so total heat
+                # weight is conserved across zoom levels. With --drop-densest the
+                # z14 tile lost hotspots that z15 kept, so the heatmap visibly
+                # collapsed into smaller spots when crossing the z14<->z15 tile
+                # boundary. Clustering is spatial only (it can merge different
+                # crime_types into one representative point), so per-type
+                # filtering is slightly approximate in the densest z14 tiles;
+                # the all-types surface and every zoom >= 15 stay accurate.
+                "--cluster-densest-as-needed",
+                "--accumulate-attribute=count:sum",
+                "--accumulate-attribute=weight:sum",
                 "--extend-zooms-if-still-dropping",
                 "--temporary-directory",
                 tmp,