#wrong-scale

Two-stage flow: Blender headless emits Y-up OBJ, then pure Python writes vmats/vmdls and a C# model catalog.

On a track that crosses itself (a figure-eight, or any lap that revisits the same ground), a per-frame nearest-waypoint lookup jumps about half a lap at the crossing, because two points that sit close together can be arclength-distant. Drive lap position from a monotone cursor that only advances along the committed branch, do lateral math in a left-of-travel frame, and keep the branch-discrimination cone tighter than the branches' angular separation.

Forge delivery job errors out but the build stage's previewGlbUrl stays downloadable: recover the GLB through Blender into your standard delivery shape instead of re-spending.

CapsuleCollider and ModelCollider ignore WorldScale; BoxCollider follows it. Bake scale into import_scale instead.

FlatBox/Prop/Deco/Wire helpers plus Obstacle records as plain data beat physics queries for build validity.

1 m = 39.37 engine units: design in SI, convert once at the engine boundary, and audit every consumer.

SceneTraceResult.EndPosition on a sphere-radius sweep is the sphere's CENTRE at contact, not the contact point. Treating it as where the feet landed is off by a full radius. Derive the real contact position as from + (to - from) * Fraction in one helper. A related trap: snapping the sphere exactly tangent to the floor makes the next horizontal sweep intermittently report StartedSolid.

AI mesh generators (Tripo, Meshy, Rodin) deliver textured OBJ+vmdl, but modeldoc32 headers, ~1m normalize, corrupt textures, bare texture paths, and missing colliders will ERROR the asset until patched.