[dktue]

Gain (planet, --store, 32 threads): wall 49.9 → 36.4 min (−27%), CPU 1,320 → 885 min (−33%). Germany 52.5 → 43.6 s (−17%), CPU −10%.

Needs the scaleRing fix from #933 (or #941): without it, any change that alters where a clipped ring starts can drop polygons through the scaleRing bug. (#949, which this was built on, is merged now.)

What changes

Validity precheck (multiPolygonDefinitelyValid). It gave up, so is_valid
ran anyway, on:

• multipolygons whose parts have overlapping bounding boxes (common after
  clipping a concave polygon). It now checks that no ring of one part touches a
  ring of another and that no part lies inside another, unless inside one of its
  holes.
• polygons with more than 256 holes. The cap is replaced by a work limit, and
  hole-against-hole tests use a sweep.

The precheck still answers "certainly valid" only when is_valid agrees. A
Germany run that compared every positive answer with is_valid found no
disagreement, and neither did the randomised test.

Shapefile polygons were stored without a validity flag, so every tile they
touched took the slow path: is_valid, then Boost intersection on the original
polygon. On the planet that was 21% of all CPU. They are now validated once when
stored, so tiles take the rect_clip path.

Results

Planet, 32 threads, --store, compared with #949 + #941. This run also included
the #949 fix above and an unrelated hash-map change, both speed-neutral on Germany.

	before	after
wall	49.9 min	36.4 min	-27%
CPU (user+sys)	1320 min	885 min	-33%
peak heap (anon)	25.3 GB	27.3 GB	+2 GB

Germany, 32 threads: 52.5 s -> 43.6 s wall (-17%), 1167 s -> 1045 s CPU (-10%).

The higher peak heap on the planet is not explained yet (Germany's peak goes
down slightly). I'm looking into it.

Output

Planet: the same 345.9M tiles. Of a 200k-tile sample, 89.4% are byte-identical,
10.0% have the same features in a different order, and 0.64% differ in content.
Feature counts match except poi (which already differs between identical runs)
and one landcover feature.

Germany, compared per tile, layer and tags: with simplification off, nothing is
lost. With simplification on, a few small polygons at z6–z12 differ by a few
percent, both ways, because rings now start at a different vertex.

Tests

New unit tests for the precheck: overlapping bounding boxes, a polygon in
another's hole, nested and crossing polygons, touching at a vertex, and 400
holes.

🤖 Generated with Claude Code

[systemed]

Very neat - Europe is down from 48m30 to 41m32 on my machine, output looks the same and RAM usage is unaffected. Thank you.

(Mostly looking at performance for now and I'll move onto correctness later!)