Fix issue #88: seven volumetric corpus mismatches, root-caused individually

[particlesector]

Summary

Closes #88. Rebuilt the v3.9 volumetric-comparison toolchain (real OpenSCAD 2021.01 oracle, Manifold v3.5.2, headless chiselcad_core, scad_to_stl/stl_diff — see tests/tools/README.md) and bisected each of the issue's seven untriaged corpus files individually. All seven turned out to be real, distinct bugs, and all seven are now fixed — each file's sym_diff_volume against the live oracle is at floating-point-noise level:

• ifelse-tests.scad (175% → exact): cube()/square()'s positional center argument (cube(2, true)) was silently dropped by the parser — never an if/else bug at all.
• primitive-inf-tests.scad (83% → exact): non-finite cube/sphere/cylinder dimensions were folded to 0.0 instead of rejected, so e.g. cylinder(r2=1/0) silently became a valid zero-radius cone instead of empty geometry; a non-finite $fn also escaped its minimum-segment clamp via undefined int cast behavior (empirically INT_MAX on this platform).
• surface-simple.scad (~82% → exact): surface()'s base was clamped to 0 instead of real OpenSCAD's actual minHeight - 1, center= incorrectly re-centered Z (real OpenSCAD only offsets X/Y), and the row-to-Y mapping was flipped relative to real OpenSCAD's own convention.
• module-recursion.scad (31% → exact): rotate([x,y,z]) composed its three axis rotations in the reverse order from real OpenSCAD's X-then-Y-then-Z convention.
• resize-tests.scad (1.5% → exact): resize()'s negative-newsize and auto= handling was a generalization that didn't match real OpenSCAD's actual getResizeTransform() algorithm — rewritten as a verbatim port.
• intersection-tests.scad (6.4% → exact): a child that itself produced no geometry (e.g. bare render();) was silently dropped instead of nullifying the whole intersection(); carved out non-geometric statements (echo()/assert()) and %-backgrounded children so they're still just skipped, not treated as empty operands.
• cylinder-tests.scad (13% → exact): when both r and r1/r2 were given, r unconditionally overwrote both instead of only filling in the unspecified slot.

Full root-cause narrative and verification details for each are in docs/roadmap.md's new v3.14 section.

Regression safety net

Re-verified against the live OpenSCAD oracle across the entire tests/data/scad/3D/features corpus subdirectory (72 of 73 files — see below), not just the seven files this issue named. That wider sweep caught a real regression the intersection() fix introduced in background-modifier2.scad (a %-backgrounded child inside an intersection() was being treated as an empty operand instead of simply excluded from the main tree) before it landed — fixed in the same commit.

Filed #105 for a pre-existing, unrelated crash found during that sweep (linear_extrude-parameter-tests.scad — heap corruption, confirmed present before this PR's changes too via a clean worktree build of the base commit).

Test plan

• Full chiselcad_tests suite passes: 3679 assertions, 664 test cases (9 new regression tests added)
• All seven issue [Medium] Untriaged geometry corpus mismatches: intersection(), cylinder() edge case, primitive-inf, ifelse, module-recursion, resize(), surface() #88 files verified at floating-point-noise level against the live OpenSCAD oracle
• Full tests/data/scad/3D/features corpus (72/73 files) re-verified against the oracle — no regressions, two additional pre-existing gaps (rotate-parameters, transform-tests) improved by the rotation-order fix

---

Generated by Claude Code

[particlesector]

Reviewed the diff (excluding docs/roadmap.md and fixtures). Each of the seven fixes is narrowly scoped, well-justified in comments, and backed by a targeted regression test:

• Parser.cpp's positional true/false handling is correctly confined to parseParamList (cube/sphere/cylinder/square/circle + discarded-params call sites), not the generic user-module-call path, so it can't misinterpret a custom module's positional bool arg as center.
• evalBoolean's new intersection-emptiness check correctly re-inspects astModifiers(*child) directly to distinguish "genuinely empty child" from "disabled/backgrounded child," since evalNode already consumes those modifiers before returning nullptr — no double-counting.
• MeshEvaluator::checkStatus now discarding the manifold on error (not just the 7 named files) is the broadest behavioral change here, but it's covered by the described 72/73-file oracle resweep with no regressions, so it looks safe.

One minor note (non-blocking): the checkStatus change affects every call site (booleans, resize, primitives), not just the primitives touched by the issue's seven files — worth keeping an eye on if any downstream corpus file relied on a "degenerate but non-empty" manifold surviving a non-NoError status.

Nothing here blocks merging as far as I can tell.

---

Generated by Claude Code