[rihokirss]

Fixes the walls still broken in #2269.

Revit exports in #2269 place openings 0.02–0.14 mm off the wall faces and tilt them by about 1.8e-5 rad. Because of the tilt, the distance between an opening face and the wall face varies along the face, for example 0.998e-4 … 1.38e-4 m, which straddles the 1e-4 coplanarity tolerance. Some vertices of the face count as lying on the wall plane and others don't. Depending on the wall, the difference comes out open, keeps the opening's flipped face as a cap across the hole, or skips the cut. One wall fails with offsets of 0.02–0.07 mm, fully inside the tolerance, because the tilted face is still classified inconsistently.

Before each difference, this PR moves cutter faces that lie within 2× TOLERANCE_PLANE_DEVIATION of a parallel operand face exactly onto that face's plane (SnapToOperandPlanes in IfcGeometryProcessor.cpp):

• Cutter positions are welded first, so every copy of a vertex moves together. A vertex on faces snapped to several planes goes to the closest point on all of them.
• The stored vertex normals of a snapped face follow it. buildPlanes prefers them over the triangle normal and would otherwise tilt the face back.
• Only planes that carry operand faces are used. Results of earlier differences can still list planes of former cutters, including non-finite ones.
• A face within reach of parallel planes at different offsets is left alone, because the operand has a sliver thinner than the snap distance and the target is ambiguous.
• Open cutters are left alone. The snapped cutter is used only if it is still closed after its planes are fitted.

Validation (element by element, with and without the change, against IfcOpenShell 0.8.5):

• The reduced model from [Bug]: Incorrect boolean subtraction creates triangular surfaces across wall openings #2269, on main dd44aa7 with COORDINATE_TO_ORIGIN on and off:

  wall	main	this branch	expected
  0jk7VvYE9D3uZXbw6Ypv5R	open, 104.73 m², cap across the back of the hole	13.166 m³, 96.34 m²	13.165 m³, 96.33 m² (IfcOpenShell)
  2so2LXaZX1ZQ6s1VN2AI1S	open, 56.99 m²	6.839 m³, 54.55 m²	6.839 m³, 54.55 m² (IfcOpenShell)
  3kzyxbyJn9FgaTZBR8AJx8	openings not cut, open	12.049 m³, 86.76 m²	12.050 m³ (wall minus openings; IfcOpenShell keeps a 0.1 mm skin)

  The other 48 elements are unchanged.

• tests/ifcfiles/public: 8 files change, and 11 elements get better:

  • 7 now match IfcOpenShell: 1 in ISSUE_159_kleine_Wohnung_R22.ifc (open mesh → closed, 2.4275 m³) and 6 in S_Office_Integrated Design Archi.ifc;
  • 4 get closer to it: 2 walls in AC20-FZK-Haus.ifc (also in ISSUE_005_haus.ifc) and 2 more in S_Office.

  The rest change triangulation only. One S_Office element (#396617) turns from open to closed with the same volume; IfcOpenShell differs from both by 2.6 m³ and 20 m².

• Six larger models, mostly private, with about 135,000 elements: 25 elements change. Seven are better: five now match IfcOpenShell and two are closer. The rest change triangulation only or by less than 0.01 %. None got worse.

• tests/functional/BooleanRobustness.spec.js gets three synthetic metre-unit walls with misaligned extruded and faceted brep openings. All three fail on main and pass here. The expected values are the nominal sizes, because IfcOpenShell keeps a skin for two of them.

• Jest: all 324 tests pass on this branch's freshly built Node WASM. npm run regression: all 30 models pass.

• Cost: about +3 % total geometry time on a 640 MB model.

git clone --single-branch --branch codex/pr-snap-flush-cutter-faces https://github.com/rihokirss/engine_web-ifc.git web-ifc-snap-flush-cutter-faces
cd web-ifc-snap-flush-cutter-faces
npm ci
npm run build-release
npm test -- --runTestsByPath tests/functional/BooleanRobustness.spec.js

Build environment: Node 24.20.0, Emscripten 4.0.23.

Notes for review:

• Behaviour change: a cutter face (or half-space plane) closer than 2× TOLERANCE_PLANE_DEVIATION to a parallel operand face (0.2 mm in metre models) now ends exactly on that face, so skins thinner than that are no longer kept.
• The step is self-contained before the kernel call, so it is easy to put behind a setting if you prefer it opt-in.