[EpicDima]

Fixes #219

When filterPoints removes the bridge copy of a vertex, the outer ring runs straight through that vertex. A later hole's ray can then hit the vertex and the healed edge at the same x, and findHoleBridge picks the vertex although the hole is outside its angle. With this PR, if a point was filtered out during hole elimination and the hit vertex doesn't face the hole, the bridge goes to the edge through that vertex.

polygon	main	this PR
example from #219	7 triangles, deviation 0.089	11 triangles, deviation 0
the same scaled by 0.12 in x and 3 in y	7 triangles, deviation 0.089	11 triangles, deviation 0
noisy raster region, 65426 vertices, 9113 holes	triangles cover 1.3% more area than the polygon	exact area
noisy raster region, 161064 vertices, 22662 holes	0.85% more area	exact area

The new test covers both small polygons. It fails on main and passes with this PR.

Details

• The output is identical to main's on the fixtures, all tile polygons from the benchmark and generated shapes. On a million random small polygons it differs on 50, all of them invalid.
• The tile benchmark is within 1% of main. Small polygons with holes, where the check costs the most relative to the work, stay within main's run-to-run spread over six runs.

[mourner]

Thanks! Also fixed with a much smaller change — I want to keep the code small and not add new complexity if possible. 77019d3

[EpicDima]

Nice fix, much simpler than mine 👍
Keeping bridge2 stops the problem where it starts, and it also handles a few float cases my PR still got wrong. There's still a mirror case with bridge itself, I opened #221 for it