[dktue]

Makes combine_polygons_below much cheaper (#868).

Cause
union_many grouped multipolygons by bbox overlap and ran a binary reduction of union_() within each group. Adjacent landuse parcels (farmland, forest, ...) share edges, so at low zooms a tile holds one group of ~20,000 polygons. In the last rounds of the reduction both sides are huge, and Boost's union_() tests every ring without intersections against all vertices of the other side (select_rings -> range_in_geometry, no bbox prefilter in Boost 1.83). That is quadratic: on Luxembourg with the OpenTopoMap config, 30% of all CPU time was in point_in_range, and two single union_many calls took ~10 s each.

Change (union_many in src/geom.cpp only)

• Split the inputs into single polygons and sort them along a Z-order curve, so each reduction step unions two neighbouring regions.
• At each step, pass only polygons whose bbox intersects a bbox on the other side to union_(); everything else can't overlap and is copied through.

Results
OpenTopoMap config with combine_polygons_below: 13 on its landuse layer (land_low), as in #868:

	before	after
Replay of all 2544 union_many calls from Luxembourg	45.3 s	11.4 s
Slowest single call (19,520 polygons, 498k vertices)	9.9 s	1.2 s
Luxembourg, whole run	14.5 s wall, 99 s CPU, 650 MB	5.8 s wall, 60 s CPU, 443 MB

Germany with --store --shard-stores (the setup from #868, on top of #943): 571 s without combine_polygons_below, 694 s with it. #868 reported ~10 min without and >2 h with.

Output
Boost's union is not exactly order-independent, so results differ slightly from before, within the variation master already has:

• Union results (replayed inputs): master vs this PR differ in 42 of 2544 calls, by at most 1.0% of the area (symmetric difference). Master vs master with the inputs in reverse order: 34 calls, up to 1.5%.
• Validity: master produced 1 invalid multipolygon (intersecting interiors), this PR none.
• Final Luxembourg tiles: only land polygon areas differ. Master vs master with reversed union order: 165 (tile, attributes) groups off by >0.1%, worst 6.3%; master vs this PR: 191, worst 6.0%. The larger differences come from simplification of rings that start at a different vertex: with simplification off on that layer, master vs this PR differ in a single group (0.76%). Points, lines and all other layers are identical, and so is output without combine_polygons_below.

🤖 Generated with Claude Code

[systemed]

Thank you! With this and #949 together, Europe (RAM-resident) on my machine falls from 58m06 to 48m30. I've eyeballed the usual failure points and can't see any regressions.