OneCAD/docs/cam-crates.md

CAM crate assessment

Research date: 2026-10-15. This assessment supplements the CAM implementation status. Versions below are the releases inspected, not a promise that every upstream latest page is current. The initial assessment changed no dependencies. Subsequent implementation pinned Parry 0.31.1 for collision queries and robust 1.2.0 for independent projected-coverage predicates. Other candidates remain proposals; see the current quality-gate results for target compilation evidence.

Recommended order

Candidate	Inspected version / license	Proposed use and decision
parry3d-f64	0.31.1; Apache-2.0	Implemented behind a private adapter: distance and swept collision queries refine fixture, holder and shank checks. Conservative fallback remains.
proptest	1.12.0; MIT OR Apache-2.0	Add as a development dependency for generated cases and shrinking failures to small reproducible examples.
stl_io	0.11.0; MIT	Evaluate for bounded ASCII/binary import and binary stock export. This simplifies file handling; our independent conforming stock boundary extractor supplies the geometry.
robust	1.2.0; MIT OR Apache-2.0	Implemented as a pinned direct dependency for independent stock-mesh coverage and sidedness checks. Exact predicate signs prevent near-collinear coverage errors; guarded plane intervals still account for dimensional uncertainty.
static_aabb2d_index	2.1.0; MIT OR Apache-2.0	Already present through Cavalier. Benchmark against the verifier's current XY index before replacing it. Keep unsafe optimizations disabled.
gcode	0.7.0; MIT OR Apache-2.0	Useful as a second parser in differential tests. It handles syntax and source spans, not modal machine semantics. Keep SolidCraft's independent interpreter and rejection of unsupported behavior.

Keep cavalier_contours 0.9.0, already pinned in solidcraft-cam. Its Rust core supports line/arc offsets and regions with islands under MIT OR Apache-2.0. It also documents limitations: rounded offset joins, arc segments of at most a half-circle, and nonintersecting closed loops for multi-loop regions. Our bounded adapter and invalid-region tests remain necessary. The adapter now uses explicit 1e-8 mm position/join and 1e-7 mm offset-distance tolerances, with a round-island regression that exposed a slice-merging failure under upstream defaults. The six-case corpus also exposed degenerate tangencies in multi-loop offsets. Pocket offsets reserve an additional 0.000001 mm of clearance, within the existing post error budget, while the independent verifier checks the unchanged requested geometry and allowance. Debug builds exercise the concave/two-island recipes as part of cam-validate. This is a bounded numerical mitigation, not a general proof of the upstream stitcher: the 0.9.0 release notes say the shape offset algorithm has not received the polyline stitching update, and issue 66 records threshold/tangency limitations. Keep the pin, native error guard and impossible-region checks; do not adopt a new geometry dependency or disable debug assertions to hide these failures.

A 2026-10-16 original 10,000-vertex circle also exposed a short-segment cross-product threshold failure. The adapter now normalizes coordinates by an exact power of two and scales physical distance tolerances with them, then restores output to millimetres. It retains every input vertex. Native debug regressions cover dense convex/concave offsets, shifted/scaled coordinates and independent distance expectations. The release-Wasm dense concave probe still measured a 21.2 ms atomic offset batch, an explicit scheduling limit recorded in the acceptance audit.

Portability and integration

• Parry: use double precision and a private adapter from SolidCraft geometry types. The inspected API uses its own math types; avoid exposing these in persisted models. Start with sequential queries and no optional Rayon feature. Native and wasm32-unknown-unknown compilation, dependency/license review, and bounded execution measurements are adoption gates. The upstream manifest exposes optional parallelism and enhanced determinism; exact release features must be pinned and tested.
• Proptest: keep it out of runtime dependencies. Native CI can run the broad randomized suite. Browser tests need a separate feature configuration: its defaults include process-based fork/timeout functionality. Save seeds and minimized regressions.
• STL: use memory readers/writers in the portable layers. Enforce our byte/triangle caps and finite-value checks around the library. Measure float serialization error against the artifact tolerance budget. A portable I/O abstraction alone is not evidence of a passing Wasm build.
• Existing geometry crates: robust 1.2.0, spade 2.15.1, and static_aabb2d_index 2.1.0 are already in this worktree's lockfile and were included in the previous workspace Wasm gate. Any new direct adapter still needs its own tests.
• G-code: the crate explicitly targets no-std/WebAssembly. Treat any parser diagnostic as a rejection in validation; error recovery must never silently remove machine behavior.

Other candidates

spade 2.15.1 provides constrained 2D triangulation and already exists transitively. The 2026-10-16 meshing profile justified using its bulk_load_cdt API for empty boundary triangulations, replacing expensive sequential point location. The vendor manifest now requires at least the inspected 2.15.1 release. No constraints are passed during bulk loading; rounded positions (including duplicates and signed zero) map back to vertex handles before the existing constraint insertion loop. Nonempty or failed loads retain the sequential path. The installed release's API documentation and implementation were inspected; workspace native/Wasm gates and analytic kernel tests exercise the adapter. Combined with an original conservative boundary candidate index, the original large-cylinder browser measurement fell from about 1087 ms to 91 ms without changing its measured chord error or triangle count. This is a concrete reuse of an existing dependency, not a new mesh Boolean or solid-boundary library. It does not supply stock side-wall stitching; the remaining atomic browser latency is recorded in the acceptance audit.

rstar is an alternative spatial index. Defer an additional index dependency until profiling shows a need beyond the existing XY index or Parry's structures.

csgrs is MIT licensed and worth a separate mesh-Boolean experiment, but is not a recommended dependency for this milestone yet. The inspected published 0.20.1 documentation describes BSP operations, while the current repository describes a substantially different Hypermesh/Hypercurve stack and 0.23.0 API. Resolve the exact published release and audit its dependency graph before evaluating it. Neither upstream feature claims nor a successful Boolean operation establish our CAM error budget or performance at hundreds of thousands of cells.

Concrete adoption experiments

1. Collision refinement: cast the cutter, shank and holder against fixture triangles. Cover a clear near miss, grazing contact, an initial overlap, a thin obstacle crossed between endpoints, shifted setup frames, and degenerate triangles. Linear shape casts fit fixed-axis motions; arc subdivision needs conservative inflation for chord error. Unsupported shape pairs and numerical failures remain inconclusive. Preserve operation, motion and NC-line IDs.
2. Generated tests: exercise winding reversal, translation, inch/mm equivalence, post round trips, cancellation boundaries, non-finite values, malformed NC, and budget exhaustion. Cross-check indexed candidate sets against brute force on small fixtures.
3. Stock artifacts: round-trip ASCII and binary STL against analytic solids, reject malformed and oversized files, and compare sizes. Independently implement or qualify conforming surface extraction; verify edge incidence, orientation, volume and local surface error. A valid STL container is insufficient evidence of a watertight stock mesh.
4. Interpreter independence: compare parser tokens/spans on the supported dialect, then use hand-calculated machine-state expectations and deliberate faults. Never reuse toolpath generation algorithms as the semantic oracle.

These crates reduce implementation effort. Broader dimensional and uncertainty coverage, and complete Fusion reference captures still require project-specific work.