Six-decimal incremental PEP programs round arc centers, so an arc's end
tangent misses its neighbor's by about gap / radius (1.6e-5 rad on the
offset 5/32" fillet in the reported nest). Two parts resting one spacing
apart then gave the slide contact a Minkowski cone a hair wider than a
half-plane, which the classifier read as the whole plane: every direction
blocked, so the top part could not be pushed sideways along the part below.
SlideContact now reads a cone up to TangentSlack (1e-4 rad) past a
half-plane as a noisy half-plane and widens its boundary and concave-ray
comparisons by the excess, so near-tangential slides read as tangential.
A direction admitted that way can enter material by at most travel x
excess at that contact, and every other contact is still classified on its
own; wider cones remain real overlap and block. Raising JointTolerance was
not needed for this case and is left at 1e-6.
Whole-job benchmarks (synthetic nests plus the PEP job at 3 and 60 parts,
--parallel 1) are identical before and after for every engine.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A part placed exactly one part spacing from its neighbor touches the
neighbor's offset contour. PEP drawings round arc centers to six decimals,
so their offset contours carry hairline kinks and joint gaps where the parts
touch; the slide kernel then classifies that contact as blocking in every
direction, and Push moved the part 0 even straight away from the neighbor.
An obstacle whose bounds end at or behind the moving part's trailing side
can only fall further behind, so Push and PushBoundingBox now skip it
(within Tolerance.Epsilon) instead of only when strictly behind.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ajandClaude Opus 5.5