#solidcore issues found while building tscad

Pinned mirror rev: e7627f7f18176460ad4fb64fd1376fd66e1fe3b3 (github.com/tangency-labs/solidcore). Each entry: what happens, a minimal tscad IR repro, and the tscad-side workaround (if any). These have NOT been filed upstream.

#1. STEP writer emits cylinders as zero-angle cones

• solidcore_step::write::product::solidcore_bodies writes fillet rounds (cylinders) as CONICAL_SURFACE('',#n,2.0,0.0). The semi-angle 0 is refused by solidcore's own STEP reader ("cone with a negative radius or a semi angle outside (0, 90°)") and is likely refused by other CAD tools.
• Repro: {"type":"fillet","radius":2,"edges":{"type":"max","axis":"z"},"children":[{"type":"box","size":[20,20,10]}]} → Kernel.step().
• Workaround: engine::cones_to_cylinders rewrites them to CYLINDRICAL_SURFACE after writing.

#2. STEP read: no pcurve of a NURBS edge on a spline surface

• Reading back a filleted box (blend corner patches) fails: pcurves: ... not supported yet: no exact pcurve of a NURBS on a spline surface (solidcore_ops::pcurve::pcurves_within). SolidCraft gets these pcurves from its separate smlib crate, which is not in the solidcore mirror.
• Repro: write the model from issue 1 to STEP, then evaluate {"type":"step","data":<text>}.
• Workaround: none; <import> of such files fails with that message.

#3. Union of overlapping spheres is slow

• Six spheres r=4 spaced 7 apart along X, unioned: ~4.5 s in wasm (O3), vs ~0.2 s for a filleted plate with a hole. Not wrong, just slow (sphere/sphere SSI).

#4. Inward shell across tangent walls fails

• Shell (thickness -1, top face open, closed, or bottom open) of an extruded rounded rectangle → "the result is not a closed solid". The same body shelled outward (+1) works, and an inward shell of a cylinder works.
• Repro: crates/tscad-wasm/tests/select_debug.rs (rounded_rect_top_fillet, "shell" cases).
• Workaround: none yet; use outward or build the cavity with a boolean.

#5. Sweep along a helix fails or is slow

• sweep (frame follow) of a circle along a helix NURBS: 10–23 s on the dev wasm, then "the result is not a closed solid"; frame: "translate" gives a near-zero volume.
• sweep along a polyline path with a sharp corner fails with status 1001.

#6. Booleans are ~1 s per cut on the dev wasm

• A plate with 24 slot holes: 18–21 s cut one by one, ~7 s as one difference(outer, union(holes)).

#7. Transforming a torus leaves a non-finite tolerance (confirmed on the raw ops API)

• solidcore_ops::make_torus([0;3], 20, 3) then Body::transform with a pure translation (row 3 = [5,0,0,1]): volume is still right, but Body::to_record() fails with "the body's tolerance is not finite and positive", so the body can't be meshed, picked or exported. Rotations fail the same way, and a rotated torus then fails tessellation ("pcurve is 3.6 mm off the edge").
• Repro: crates/tscad-wasm/tests/fuzz_triage.rs.
• Workaround: a transform directly over a torus (after composing nested transforms) is built in place as a circle revolved about the moved axis, which has no such problem and also fixes #8 for that case. A torus inside a moved union is still affected.

#8. Boolean of a rotated torus with itself is wrong

• difference(T, T) with T = torus(20, 3) rotated 30° about X returns all of T (3553 mm³) instead of nothing; union and intersection give the same volume. Unrotated, it is right.
• Found by the fuzzer (docs/FUZZ.md has 31 distinct findings, minimal repros in fuzz/findings/).

#9. More fuzzer findings (see docs/FUZZ.md)

• box − torus evaluates "ok" with a kernel volume of 0 (mesh says 1397 mm³);
• torus ∪ cylinder: mesh volume 52% below the kernel volume;
• side-by-side touching cylinders: union mesh not watertight; one extruded star polygon too;
• solids booleaned with a copy moved by 1e-5..1e-3 mm: "not a closed solid" / status 1001;
• ~10 boolean failures on revolves / rotated extrudes ("SmFace::PointClassify not supported yet", "proto edge is deleted", "mass::enclosure", empty unions).

#10. union_all refuses a union that pairwise unions (in the right order) accept

• Three extruded rounded rectangles: two disjoint plates (z 0..0.8 and 2.6..5) and a third touching the second at z = 5. union_all → "the result is not a closed solid"; folding left to right fails too (disjoint ∪ disjoint, then the touching one). Unioning the large touching one first, then the rest, works. Boxes in the same layout work.
• Workaround: when union_all refuses, tscad unions greedily, largest body first, retrying any union that fails after the others (eval::union).

#11. Extruding the DejaVu Sans glyph "s" at 11–13 mm fails

• "the result is not a closed solid", only at those sizes; the glyph's many short Bézier edges.
• Workaround (core): consecutive glyph Béziers are joined into one C0 B-spline (same curve, fewer edges).

#12. solidcore-ops serialises every operation process-wide

• solidcore_ops::session holds one static KERNEL: Mutex<Option<Kernel>>; every op locks it, so two threads evaluating different bodies in one process run one after the other (measured in tscad's dev server: two native part evaluations on two bun threads took 3.1 s and 4.0 s vs 1.9 s alone). The kernel crate itself is "one Kernel per thread"; the ops layer is not.
• Workaround: parallelism across processes (or wasm instances), not threads.

#13. An oversized fillet is accepted instead of refused

• fillet radius 8 on all edges of a 10 mm cube returns ok with volume 962 mm³; a radius above half the edge length cannot be built, so a refusal (TooLarge / NoRoom) is expected.
• Found by MCP dogfooding (docs/mcp-dogfood.md).

#14. Intersecting a nested lip with its rim: PointClassify unsupported

• Board case base ∩ its lid (the lid's locating lip nested inside the rim, faces touching) → "not supported yet: SmFace::PointClassify", the same unsupported path as the fuzzer's PointClassify findings (docs/upstream/pointclassify-cut-rounded-prism.md). Interference checks of board-vs-base and board-vs-lid work.

#15. Extrude depends on the profile's winding

• A closed profile of one cubic Bézier + two lines (an involute gear's root fillet) extruded with ops::sweep::extrude → "the result is not a closed solid"; the same loop traversed the other way extrudes fine (0.3044 mm³ either way once built). Found by the example param sweep: it sank spur-gear (7/33 combos) and the gear-pair recipe (19/33).
• Repro: crates/tscad-wasm/tests/fuzz_triage.rs (gear_fillet_extrude_orientation).
• Workaround: on failure tscad retries the extrude with the loop reversed (eval::reversed).

#16. Union of a 6×3 grid of overlapping cylinders fails

• r 4.85, h 21, pitch 11.72 × 8.3 (rows overlap): "the result is not a closed solid" after ~18 s; 5×3, 6×2 and 3×3 work. Found by the example sweep.

#17. Gear root gaps still fail to extrude after the #15 retry

• Two cubic root fillets, the short root arc between them and two lines through the origin (one root gap of an 8-tooth module-0.5 gear) → "the result is not a closed solid" in both windings. The profile is clean: endpoint gaps ~1e-14, no self-crossings.
• Whole gears: examples/spur-gear.tsx still fails at most tooth counts above ~35 (and some small fine-module ones), and examples/cookbook/gear-pair.tsx whenever a gear has that many. Joining the Béziers into one B-spline, more pieces, or cubics/lines instead of the arcs don't help; dropping the root fillets does.
• Repro: fuzz/sweep-regressions.json → sweep-extrude-gear-root.

#18. Extruding a profile with 0.008 mm arcs fails

• A circle minus 34 overlapping flute circles (exact 2D difference) leaves 0.008 mm arcs of the big circle between flutes; extruding it → "the result is not a closed solid".
• Hit by examples/agent/d-shaft-knob.tsx with many flutes on a small knob; the knob now caps the flute count so a land is left. Repro: sweep-extrude-tiny-arcs.
• Also seen by the sweep, not minimised: chamfer along a 41-tooth module-1.5 gear outline reports a "largest Distance that fits" of ~0.095 mm, well under the outline's tightest bend (0.22 mm) and shortest edge (0.5 mm); the vase shell with extreme radius swings gives "mass::enclosure"; examples/enclosure.tsx with large corner radii and thin walls hits the PointClassify gap of #14.

#19. Extruding the DejaVu Sans glyph "t" at 1–3 mm fails

• "the result is not a closed solid" at text sizes 1, 1.25 and 1.75–3 (1.5 and ≥ 3.25 work); the other letters of "tscad" extrude at every size. 16 edges: lines and two quadratic B-splines.
• Hit by examples/cookbook/name-tag.tsx on small tags. Repro: sweep-extrude-glyph-t.

#Note on #15, #17, #18, #19: two constructions, same refusal

• tscad now retries a refused extrude reversed, then through ops::prism::extrude_planar_regions (a separate construction). That rescued #18 (tiny arcs). The gear root (#17) and glyph "t" (#19) are refused by both constructions with the same "the result is not a closed solid", which comes from ops' closed-solid check (mass::enclosure, also ~61% of kernel time in docs/PERF.md). That check may be returning false negatives on these bodies; worth checking it first.

#20. Fillet "no room" on an obround arm

• fillet r 1.5 on edges.parallel("z").not() of an extruded obround (end bosses R 28 and 12.5, 54 apart, tangent lines) → "no room"; the same arm with R 20 works.
• Hit by examples/sketch-link-arm.tsx (1 combo of 33). Repro: tests/fixtures/kernel-noroom-obround-fillet.json.

#21. A union's result depends on how its children are grouped

• union(union(padLow, padHigh), hollowBody) → "the result is not a closed solid", where the two pads are disjoint and the hollow body sits on padHigh's top face (z = 5) with the same rounded-square outline. union(low, high, body), union(body, pads), union(low, union(high, body)) and a solid body all work. Related to #10.
• Hit by examples/converted/grid-bin.tsx at 1 × 1 units. Repro: tests/fixtures/kernel-union-nested-not-closed.json.

#22. Shell of a smooth loft fails

• A loft through three rounded-square sections (30 / 40 / 20 mm, corner r 0.15 · w), shelled 0.2 mm with the top open → "the result is not a closed solid"; it works at corner r ≥ 0.35 · w and fails erratically at larger sizes.
• examples/sketch-vase.tsx now hollows with a second, inset loft instead. Repro: tests/fixtures/kernel-shell-loft-not-closed.json.

#Notes (tscad bugs found and fixed, not kernel bugs)

• solidcore_ops::blend takes the kernel's native edge order, not record order (its SetBack docs say so; the function's own docs read as record indices). tscad now uses blend_ordered with the record's Order. Before that, fillet rounded the wrong edges.
• EdgeNotBlendable { edge } reports the native position, which callers of the ordered API can't map back; tscad pre-filters seam and tangent (smooth) edges from broad selectors instead.