feat(tiles): serve facade_scan and the incident layers from Martin

[yorickdewid]

Moves the last five tilesets off tippecanoe and onto the Martin tileserver.

Why

Tileset	Features	Nightly tile build	Spaces objects
incident_district	1,016	26.5 min	155,335
incident_neighborhood	1,530	11.5 min	67,346
incident	2,728	2.0 min	7,373
facade_scan	3,202	1.0 min	1,090
incident_municipality	298	0.5 min	497

~40 of the ~50 minutes process_mapset spends every night, and 231,641 Spaces objects, to render 8,774 features. incident_district alone spends 26.5 minutes uploading 155k tile files for 1,016 polygons — the upload dominates, not the tiling. Nightly flow should go from ~101 min to ~60 min.

What

Two SQL files following the create_building_cluster_tiles.sql pattern: flat table in 3857 + GiST index, refresh_* procedure (TRUNCATE + INSERT), (z,x,y) function source Martin auto-publishes from maplayer, TileJSON COMMENT (with the mandatory fields map), and grants.

MVT layer names match the retired tileset names, so no layer config changes anywhere — every WebFront config already uses source == source-layer == tileset name, and the Report incident chapter is the same.

Zoom ranges come from maplayer.bundle: incident 12–16, district/neighborhood 10–16, municipality 7–11.

The views are not dropped

maplayer.facade_scan and the four incident views still feed the nightly GPKG export to s3://fundermaps-data/mapset/, which is the permanent static-model history. The refresh procedures repeat the views' joins rather than selecting from them, so the archived GPKG schema cannot drift as the tile attribute set evolves.

Geofence — please read

The four incident sources now emit neighborhood_id / district_id / municipality_id, which the static tiles never carried.

WebFront's geofence builds any(match(<id>, fence), !has(<id>)) — a feature missing the id is shown. Schiedam publiek is a public mapset fenced to GM0606 client-side (store/mapsets.ts) that includes the incident layer, so today it renders all 2,728 Dutch incidents instead of Schiedam's 216. Emitting the ids makes the existing fence actually bite.

Same failure mode as the z12–13 slimming fixed in #52. This is the one behaviour change in the PR — if you'd rather keep exact parity for now, drop the three id columns from the ST_AsMVT select in maplayer.incident() and nothing else changes.

Notes

• Aggregation polygons are detailed (860k vertices over 1,016 districts) and are drawn at zooms where that detail is sub-pixel, so they carry a geom_simple variant for z<12 — with separate query branches, because a CASE over two geometry columns in WHERE is not indexable.
• facade_scan and incident get no simplified variant on purpose: the whole of each is smaller than a single dense building_tiles z12 tile.
• Enum columns stored as text (ST_AsMVT emits the label either way, and it keeps these tables independent of enum DDL).
• Windmill: new f/fundermaps/data/refresh_layer_tiles script created, to be wired as step v after u in refresh_data_model.

Cutover order

1. Apply both SQL files to prod, restart Martin, verify the catalog + tile parity.
2. Wire the Windmill step.
3. Merge + deploy the WebFront and Report PRs.
4. Only then set generate_tileset=false on the five bundle rows — enabled stays true and upload_dataset stays true, so the nightly GPKG exports keep feeding the archive.
5. Purge the Spaces prefixes last, once the frontends are confirmed live.

🤖 Generated with Claude Code