[dktue]

Gain: the two slowest planet relations 43.9 → 3.9 s (−91%) and 53.8 → 3.2 s (−94%); extract user CPU −20%. On the planet, measured together with the ring-assignment PR #975: relations phase 105 → 61 s (−42%), total wall 1659 → 1620 s (−2.4%). Output identical.

make_valid repairs self-intersecting rings with geometry::correct() in include/geometry/correct.hpp. dissolve_generate_rings traces each repaired ring point by point, and after every new point is_closed scanned the whole ring traced so far for an earlier point equal to the new one. That is quadratic in the ring's size, and on a few huge planet relations it dominates.

This PR keeps a hash map from each traced point to its position in the ring. A ring's points are distinct until it closes, so a point found in the map closes it, and the points before the earlier occurrence are dropped as before. The test was exact equality (comparable_distance == 0) and stays exact; + 0.0 makes -0.0 and 0.0 hash alike, as they compared equal before.

Equivalence

The same test program, compiled once with the old and once with the new header, repaired 2,000 random self-intersecting polygons (continuous and grid coordinates, with and without holes; 1,999 non-empty results). The WKT output of both is byte-identical. Germany output (compare_area, per-tile area/length per layer and tag group) is identical too; only poi differs, as between identical runs.

Speed, extract with the five slowest planet relations

relation	before	after
r20516020 wood (193 outer, 210 inner ways)	43.9 s	3.9 s (−91%)
r13985343 Lake Nasser (878 outer, 6,758 inner)	53.8 s	3.2 s (−94%)

User CPU of the extract run −20%. For the planet, see the ring-assignment PR #975, measured together: the relations phase drops from 105 s to 61 s.

🤖 Generated with Claude Code