Voronoi: replace in-house tessellation with a real geometry library (CGAL / Voro++ / boost::polygon)

[petlenz]

Motivation

Both Voronoi generators are currently hand-rolled, with documented complexity / scale ceilings:

• voronoi_generator_2d (include/rvegen/generators/voronoi_generator_2d.h): Sutherland–Hodgman half-plane clipping. Each cell starts as the bounding rectangle and is clipped against the bisector of every other seed — O(N²) per build, fine for hundreds of seeds, no external dep.
• voronoi_generator_3d (include/rvegen/generators/voronoi_generator_3d.h): brute-force half-space-triplet vertex enumeration with Eigen PartialPivLU for the 3×3 solves. O(N⁵) — caps out around ~50 seeds. The header explicitly flags this: "beyond that, switch to a real Voronoi library (Voro++, CGAL)."

This sits in tension with our general preference for established geometry libraries over re-inventing primitives. The current code is correct (partition-of-unity invariant is tested in 2D and 3D), but it's a known scaling cliff and a maintenance burden when more sophisticated cases come up (weighted Voronoi / power diagrams, anisotropic seeds, periodic Voronoi on a torus, Laguerre cells for polydisperse grain packing, etc.).

Options

Library	2D	3D	Bounded cells	License	Notes
CGAL	✅ Voronoi_diagram_2	✅ via Periodic_3_regular_triangulation_3 + dual	✅	GPL / commercial dual	Industrial-strength; periodic-3D support is the standout feature for RVE workflows. Heavyweight build dep.
Voro++	—	✅ native	✅ native (this is its whole point)	BSD-3	Header + small .cc; designed for exactly this use case; supports walls, periodicity, polydisperse (radical) cells. No 2D.
boost::polygon	✅ voronoi_diagram	—	Requires our own clipping	Boost (already in tree)	Drop-in for 2D, integer-coord flavored — needs a small wrapper for double-precision unit-box coords. No 3D.
GTE (already vendored)	Delaunay only (Delaunay2/3)	Delaunay only	—	Boost-like	Would still leave us implementing the dual + clipping ourselves. Not a real fit.

Proposed scope

Not a single PR. Suggest splitting:

1. 3D first (highest payoff — replaces the O(N⁵) ceiling). Most likely Voro++ for the BSD license and surgical scope. CGAL's periodic 3D Voronoi is the alternative if we want periodic polycrystal RVEs.
2. 2D as a follow-up. Lower priority since current Sutherland–Hodgman is correct and scales fine for typical RVE sizes. boost::polygon is the obvious swap since Boost is already a tree dep.
3. Keep the current implementations behind a compile-time / runtime knob for the first round so we can A/B them and confirm the new backends reproduce the partition-of-unity tests bit-for-bit before deleting the in-house code.

Acceptance criteria

• Existing 2D and 3D partition-of-unity tests pass against the new backend.
• New regression test exercising N = 200+ seeds in 3D in under a second (validates we actually escaped the O(N⁵) cliff).
• voronoi_to_shapes(gen) / convex_polygon output shape stays the same — downstream svg_writer / vtk_legacy_writer paths don't change.
• The dep choice doesn't break the header-only contract for non-polycrystal users — gate the new generator behind a CMake option (e.g. RVEGEN_WITH_VORO_PP=ON) so consumers who only need circles/spheres don't pay the dep cost.

Out of scope here

• Weighted / Laguerre / periodic Voronoi — separate follow-up once the library is in.
• Migration of the test suite to a different convex-polytope representation — keep voronoi_cell / convex_polygon as the public shape.