Stop create_mesh flinging a centroid down a hollow

[akaszynski]

Summary

Resolves #70. create_mesh(moveclus=True) could fling a cluster centroid clean through a part, leaving a wedge across the opening of a channel section.

What happens

moveclus projects each cluster centroid onto the surface along the cluster normal. On a hollow section, a cluster lying flat on a wall can have an area-weighted centroid that lands just off the material, over a slot cut in that wall. The ray misses the near wall, runs down the hollow — the side walls are parallel to it, so they fail the |det| < EPSILON test and are skipped — and hits the far wall instead. On the reporter's mesh 3 of 17562 clusters came back at 87 units, the depth of the section, where legitimate projections top out around 4.

There is no nearer intersection to find; brute-forcing all 151888 faces returns the same 87. moveclus and ray_trace arrived in v0.3.0, which is why 0.2 was fine.

The fix

Leave the centroid unprojected when the ray has clearly skimmed off down a hollow, which is what 0.2.x did. Two conditions must both hold: the centroid already sat on the surface before projecting, and the ray then travelled more than three cluster-sizes.

The first condition is what makes this safe. A centroid that starts off the surface is closing a real gap and must keep its projection — a cone apex is the clean example, where ACVD produces ring-shaped clusters whose centroid lands on the axis of revolution and whose ray legitimately exits through the tip. An earlier version of this patch capped on distance alone and broke exactly that case, taking a cone from a surface deviation of 0 to 0.18.

Reporter's mesh: max edge 91.40 to 33.76, against 30.94 for moveclus=False. The extracted component in #70's attachment goes 89.26 to 25.56, matching moveclus=False exactly.

Verified unchanged across 566 remeshes over 43 meshes — cones from height 0.5 to 128, conical countersinks, cylinders, spheres, tori, and the usual example meshes. The only outputs that move are the hollow ones: the two meshes from the issue, the synthetic wall, and a room scan whose max edge improves.

Keeping moveclus=True as the default. Away from this defect it is a real accuracy gain — p99 distance from output points to the input surface is 0.0 on 0.3.x against 2.63 on 0.2.11.

Caveats

Both thresholds are empirical and neither margin is wide. Across the corpus, runaway projections run 3.09 to 9.49 cluster-sizes while the closest legitimate approach from a non-hollow mesh is 2.57; the on-surface ratio separates 0.162–0.462 from 0.517. The constants sit near the geometric midpoint of each pair.

Two of 41 cluster counts on the synthetic wall still fling a point. Loosening either threshold enough to catch them starts suppressing legitimate projections on the cow, human and airplane meshes, so I left them.

A principled version would use the hit face index, which ray_trace already returns and create_mesh currently discards: a legitimate hit lands on a face near the cluster in mesh connectivity, a runaway lands on an unrelated wall. That is a larger change and wants its own PR.

AI Usage

Drafted with Claude Opus 5, reviewed by me before pushing. A second Claude reviewed the first version cold and found it made cones worse, which is what produced the on-surface condition; it also found the original test passed with the cap disabled entirely. Threshold corpus, the 566-config comparison and the mutation check are all its work, spot-checked by me.