[lachlangrose]

Stack 2 of 3. Based on #24, and #22 is based on this. Merge order: #24 → this → #22.

What & why

Follow-on work to the mesh review.

Implicit-cut snapping. cut_with_implicit_function now snaps a crossing onto an
existing vertex when it lands within snap_tol (as a fraction of the edge) instead of
emitting a sliver an infinitesimal distance from it, and a per-edge dedup guard stops the
second visit to a shared edge pushing a duplicate crossing vertex — that was tearing the
mesh along every seam and compounding over sequential cuts.

The snap is recorded per slot, not by writing the isovalue into the shared per-vertex
array. The per-vertex form leaks a per-edge fact into every sub-triangle incident to that
vertex: a legitimate keep-side sub-triangle then reads as entirely on-seam, and an
inclusive keep/discard test drops it from both sides. Measured on a 2-triangle case, the
per-vertex form kept 0.25 of a correct 0.375 — a third of the retained surface silently
deleted. The filter now asks whether any original (non-crossing) vertex lies strictly on
the keep side, and applies only to the one-sided modes so PRESERVE_INTERSECTION still
keeps everything.

does_self_intersect() exposes PMP::does_self_intersect. A self-intersecting
surface is outside the domain of the boolean operations, so this is the check to run
before handing a mesh to them.

Threading contract documented. Most methods release the GIL, so concurrent access to
one TriMesh is a data race in C++ — corrupt memory, not merely a wrong answer. corefine
additionally mutates its other argument, which was undocumented.

perf on the kernel converters: flat vector remap keyed on the dense Vertex_index
instead of std::map, preallocation to the source's live counts, an rvalue set_mesh so
the converters' temporaries move rather than deep-copy, and interior_only on the
constraint snapshot to skip border edges the caller re-derives anyway.

How to verify

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release && cmake --build build -j
pytest

82 passing. New: the snap partition (both sides sum to the input area; snapping does not
move the cut) and three for does_self_intersect including that it does not mutate.

Risk

• Geometry — implicit-cut output changes where snapping now fires or where the
  duplicate-crossing tear used to happen. That is the point of the change; both directions
  are covered by tests.
• Kernel paths — the converter perf work is on the exact round trip; behaviour is
  unchanged, only allocation.
• API — does_self_intersect is new public surface. No signatures changed.

Checklist

• A test that fails without this change and passes with it
• pytest passes locally
• Both kernels covered where the change touches clipping or conversion

🤖 Generated with Claude Code