Resolve #86: non-planar polyhedron() mismatch was already fixed, add regression test

[particlesector]

Closes #86

Summary

• Re-investigated issue [Medium] Non-planar polyhedron() faces triangulate differently than OpenSCAD, wrong volume #86 (polyhedron-nonplanar-tests.scad reportedly producing volume 1.29 vs. real OpenSCAD's 2.94) using a live oracle: apt-installed OpenSCAD 2021.01 + a from-source Manifold v3.5.2 build (per tests/tools/README.md).
• Bisecting the file's three polyhedron() calls individually showed CsgEvaluator::evalPolyhedron's fan-triangulation was never actually wrong — two of the three already matched the oracle to floating-point noise.
• The real culprit was the third case: scale(0.02) polyhedron(...). scale(0.02) is a bare scalar, and at the commit the issue was filed against, CsgEvaluator::makeMatrix still had the pre-v3.10 scale(<scalar>) bug that scaled only the Z axis, distorting the mesh into a self-intersecting shape whose volume collapsed to ~0. That bug was already fixed in 18f7a54 under unrelated issue [Medium] Parser fails on several 3D/features corpus files (for-loop / transform-argument grammar edge cases) #90 — nobody re-ran the volumetric corpus check afterward to notice it also closed [Medium] Non-planar polyhedron() faces triangulate differently than OpenSCAD, wrong volume #86.
• No change needed to CsgEvaluator::evalPolyhedron/PrimitiveGen; this file now matches the oracle at floating-point-noise level (rel_error ~1e-6) on main.
• Added a regression test locking this in, so it can't silently regress independently of the existing scale(scalar) unit test that covers the actual root cause.
• Documented the investigation in docs/roadmap.md (v3.11).

Test plan

• Added tests/fixtures/headless/polyhedron_nonplanar_scale.scad (the exact reproducing case from OpenSCAD's own test corpus)
• Added a chiselcad_tests case in tests/test_headless_build.cpp pinning its volume to the oracle's measured 1.6544281372
• Ran the full test suite against a real CMake + Manifold v3.5.2 build: 625 test cases / all passing (up from 624 baseline)
• Verified via scad_to_stl/stl_diff against a live OpenSCAD 2021.01 oracle: full file rel_error=3.5e-6, isolated scale(0.02) case rel_error=1.4e-6

🤖 Generated with Claude Code

---

Generated by Claude Code

[particlesector]

Reviewed. The root-cause finding checks out: CsgEvaluator's bare-scalar scale(n) broadcast (line ~547 in src/csg/CsgEvaluator.cpp, on main) does correctly expand to {n,n,n}, confirming the pre-v3.10 Z-only-scale bug is what actually caused issue #86's mismatch, not fan-triangulation. Roadmap edit is accurate and well-placed (new ## v3.11 section slots cleanly between v3.10 and v4), and the new regression test locks in the right invariant. One minor nit left inline on the test margin.

---

Generated by Claude Code

---

Generated by Claude Code