Next Gen STEP AP242 Native-ish BREP Geometric Modeling Kernel
Editable V8 geometry reference
geometry-reference.expected.json retains the original implementation's 171
binding names and solved world matrices, together with 21 qualified display-mesh
fingerprints. Its source hash and implementation commit identify the design's
pre-refactor provenance; meshReferenceRevision identifies its display projection.
geometry-reference.meshes.zip contains the numeric data behind those qualified
fingerprints. Each entry is named <SHA-256>.bin and contains
little-endian position and normal arrays (Int32 length, then Float64 elements),
followed by indices (Int32 length and elements). Signed zero is normalized.
The test verifies each entry against its declared fingerprint.
The numeric golden is necessary because a byte hash cannot distinguish geometry
changes from platform math roundoff. Windows/Linux comparison on 2026-10-16 found
identical array lengths and indices, maximum position delta 1.013e-13 mm, and
maximum normal-component delta 2.267e-10 (spring normals). The live test allows
1e-9 mm per position component and 1e-9 per normal component, requires finite
values, and compares indices exactly. World matrices retain their 1e-9 bound.
The analytic-edge export correction on 2026-10-24 changed display sampling on four
definitions. Symmetric quad diagonals also gained a rounding-scale deterministic
tie-break. Before refreshing these display fingerprints, all 21 definitions'
surfaces, curves, trim bindings, vertices and topology were compared with the prior
qualified kernel within 1e-9; all 171 world matrices were preserved. The prior
kernel's meshes first passed the original frozen reference. The new Windows and
Linux meshes have identical indices and array lengths, with maximum position
and normal differences within the existing component bounds. No design dimensions
or exact construction geometry changed.
On 2026-11-06 the STEP writer began deriving loop direction (each edge used once each
way, counter-clockwise about the outward normal), and the coil sweep's tube faces were
bound opposed to their B-spline nets, whose normals point into the wire. The parts
round-trip through STEP, so 19 of the 21 definitions re-mesh from different loop start
points. Before refreshing, every definition's triangle area, signed volume and bounding
box were compared with the prior archive and agreed within 1e-6 mm², 1e-6 mm³ and
1e-9 mm; one definition has 8 fewer triangles for the same area. The archive was
frozen from the Linux projection, which earlier comparisons place within the component
bounds of the Windows one.
To recover the archive, run the demo at the qualified current revision with
default V8Spec (see the JSON's reproduction field), then from the current
repository root:
python scripts/freeze-editable-v8-reference.py path/to/baseline.mesh.json
The command writes under ignored artifacts/local/v8-ci/. It refuses any input
whose meshes do not exactly match all declared fingerprints; it cannot rebaseline
changed geometry. The checked-in archive uses the qualified Windows projection
with all 21 hashes verified. It is a compressed durable regression golden,
not a per-run mesh export. Its full numeric arrays permit tolerance comparisons
without the boundary failures introduced by rounding and hashing floats.