Fix issue #105: linear_extrude scale= heap corruption on non-finite values

[particlesector]

Summary

Fixes #105 — linear_extrude-parameter-tests.scad crashed with heap corruption (free(): invalid next size (fast)).

Root cause: CsgEvaluator::evalExtrusion() special-cases scale= to accept either a scalar or a 2-vector. Every other extrusion param (height, twist, slices, convexity, ...) routes through Interpreter::evalNumber(), which folds non-finite doubles (1/0, -1/0, 0/0) down to 0.0. angle= (for rotate_extrude) gets its own explicit std::isfinite() fallback. But the scale= branch assigned sv.asNumber() straight into scale_x/scale_y with no such guard — and 1/0/-1/0/0/0 are real IEEE-754 Number values per the interpreter's arithmetic propagation, so they sailed right past the isNumber() type check unclamped.

Those non-finite values then flowed into MeshEvaluator::evalExtrusion() → manifold::Manifold::Extrude(..., {scaleX, scaleY}), which lerps every extruded vertex toward the given scale across the profile's height — with an inf/nan scale, this corrupts the heap.

Fix: reject non-finite scale values (scalar or per-component in the 2-vector form), leaving scale_x/scale_y unset so MeshEvaluator falls back to its existing default of 1.0 (no scaling) — the same "malformed scale is ignored" behavior already used for e.g. a stray 3-element scale vector.

Verification

• Reproduced the exact reported crash locally: built chiselcad_cli against the pre-fix code and ran a linear_extrude(scale=1/0) square(10); file — free(): invalid next size (fast) / Aborted (core dumped), matching the issue.
• Rebuilt with the fix — same file now renders and exports cleanly (exit 0), including a full repro of all eight of the corpus file's odd param values (undef, 1/0, -1/0, 0/0, "", true, [1:3], 3) fed through scale=.
• Added CsgEval:linear_extrude non-finite scale is rejected, not passed through to tests/test_csg_evaluator.cpp, verifying the resolved IR params (the half of the pipeline that's unit-testable without linking Manifold — see the file's existing header comment on this).
• Full suite: ./chiselcad_tests → all 3699 assertions in 665 test cases pass.

Test plan

• ./chiselcad_tests (full suite) passes
• New regression test covers non-finite scalar scale (1/0, -1/0, 0/0) and non-finite 2-vector component
• Manually reproduced and confirmed-fixed the original crash via chiselcad_cli on both pre-fix and post-fix builds

---

Generated by Claude Code

[particlesector]

Reviewed. Fix is correct and narrowly scoped: scale= now goes through the same std::isfinite() guard already used for angle= (<cmath> was already included), and leaving scale_x/scale_y unset correctly falls back to MeshEvaluator's default of 1.0 (verified at src/csg/MeshEvaluator.cpp:426-427). Test case covers 1/0, -1/0, 0/0, the 2-vector form, and a finite-value regression check.

One nit: scale=[1/0, 2] is tested, but not the symmetric scale=[2, 1/0] (non-finite in the second component only) — probably not worth a separate case given the shared guard, just flagging.

No blocking issues.

---

Generated by Claude Code

---

Generated by Claude Code

[particlesector]

Good catch — added the symmetric scale=[2, 1/0] case in 81e8c7c. Full suite still green (3703 assertions, 665 test cases).

---

Generated by Claude Code

[particlesector]

Symmetric scale=[2, 1/0] case added, addressing the earlier nit. Fix logic unchanged and still correct. LGTM.

---

Generated by Claude Code

---

Generated by Claude Code