Tessellation quality: NativeCAD vs free STEP-to-mesh converters

• Triangles. How many facets the tessellator emitted for a solid. More triangles means a closer approximation to the exact surface, at the cost of memory and downstream processing time. Measured by counting the emitted mesh's facets directly — no sampling involved.
• Time (ms). Wall-clock milliseconds the tessellator spent on that one solid, measured on the machine that produced this report (—). It moves with CPU, load on the machine at measurement time, and the mesh density chosen — a slower number is not necessarily a regression without a same-machine, same-commit comparison.
• Max sag / chordal deviation. The worst-case distance between a mesh facet and the exact analytic or spline surface it approximates — how far the mesh can sit from the true shape. This is the number a viewer or a downstream tolerance check actually cares about, not triangle count by itself. Reported here as maxSag (absolute, model units) and maxSagRatio (relative to the tessellator's target deviation); meanSag/p95Sag show the typical and near-worst-case facet, since a single outlier facet can dominate the max.
• Watertight. Whether every edge of the mesh is shared by exactly two triangles in opposite winding, with no non-manifold or duplicate-directed edges — the property a solid needs before volume, 3D printing, or boolean operations on the mesh are meaningful. Measured by an edge census over the emitted triangles (§3's cascadio row explains why that census has to weld coincident vertices first).
• Volume error vs anchor. How far a solid's shipped-default mesh's enclosed volume moves from the same kernel's own volume at a forced —mm chord tolerance (the "anchor" — see §1's note). This corpus has no independently published analytic volume, so the anchor is a self-referential ceiling, not ground truth; §2's primitives are the one comparison on this page with an independent, closed-form answer.
• Aspect-ratio distribution. Not measured by this harness. The report JSON carries no per-triangle aspect-ratio field, so no aspect-ratio chart or number appears on this page — inventing one would not trace back to benchmarks/out/cadsf-778/report.json.

Provenance for every figure on this page: measured — on — / —. The report JSON does not record a commit hash or build id, so this page does not claim one — reproduce with node benchmarks/tessellation_quality_bench.js against the commit you have checked out if you need that link.

• Watertightness on this corpus. §1's summary line gives the count of solids that mesh watertight at the shipped default on these GD&T/PMI-heavy NIST parts; in the embedded run it is 26 of 57. The robustness page tracks per-operator pass rates on a real vendor-file corpus.
• Anchor coverage in §1. The anchor re-tessellation pass is skipped, with the reason recorded per row, when a shipped mesh is too large or the corpus run has spent its wall-clock budget; the report's notes field carries the thresholds. Triangle count and watertightness are reported for every solid regardless.
• cascadio volume units are unverified. §3's volume figures carry the same note tools/cascadio_convert.py attaches to every row; a raw ratio between §1/§2's mm volumes and §3's is not a real-world discrepancy until the units are checked.
• cascadio is file-granularity, NativeCAD is solid-granularity. A STEP file with several solids is one cascadio row and several NativeCAD rows; §3 reports cascadio's side as measured, per file.
• Mayo and Datasmith carry no numbers. §4/§5 are reproducible methodology sourced from each project's public documentation, so a figure only ever sits next to a name that was measured here.
• The torus primitive in §2 meshes watertight to under 1% of its own analytic volume by construction (ops/primitives.js, CADSB-26), so a row near that ceiling is expected.