A collision world's segment list is not a display representation, and drawing it one primitive per segment shows the joins

Fact

Physics stores walls as independent segments, because that is what a collision test wants: each one is checked on its own and neighbours never interact. A renderer handed the same list will happily draw one rounded box, capsule or extrusion per segment — and every rounded end then protrudes through the neighbour it was supposed to continue into.

On a straight run nothing shows. On a curve, each segment's cap pokes out at a slightly different angle, and a smooth wall renders as a scalloped, ribbed one. The tell is that the artefact tracks curvature: worst where the polyline turns fastest, invisible where it is straight.

Why it matters

The geometry is right, the collision is right, and the mesh builder is right — so the ribbing reads as a bug in the primitive, and gets chased in its cap tessellation or its radius. The defect is upstream of all of that: the display needs an object the physics representation does not contain.

The same shape appears wherever a simulation's data structure is drawn directly — a navmesh drawn per triangle, a path drawn per waypoint, a heightmap drawn per cell.

How to apply

• Recover the polylines first: chain segments whose endpoints coincide within a tolerance, in either direction, then extrude each chain as one mitred rail with a shared cross-section. Cap the miter length so a sharp corner does not produce a spike.
• A pass over a few dozen segments is O(n²) and runs once at load; do not complicate it.
• Assert the chaining rather than eyeballing it: every segment must land in exactly one chain, and the total should be segments x per-segment + chains x per-cap. A silent failure to chain just puts the ribbing back.
• Round cross-sections are the exception — overlapping spheres along a tube read as continuous, so wire guides can stay per-segment.