A cell in the foam is the intersection of half-spaces, one per neighbouring seed — so a cell's faces are its bisectors, and its face normals are its neighbour directions. Put a seed in the middle and one along each face normal of a Platonic solid, and the foam grows that solid. These are not models; there is no mesh. Everything below is points.
Orange dot the cell's own seed · pale dots its neighbours · blue the dual polyhedron they imply · dashed circle where those seeds would sit if the metric had no grain. Drag to orbit.
Push anisotropy up. The seeds pull away from the dashed circle, because the foam measures vertical distance differently from horizontal — grade has to stay a meaningful thing to climb. The solid is still exact: every face is still where it should be, to floating point. It is the space that is stretched, not the shape that is wrong.
Now tick naive placement — seeds at unit directions times a common radius, the obvious way to do it. Watch the verdict. Then switch to a cube and tick it again: the cube is fine. Its face normals are axis-aligned and the metric cannot rotate those, so the first solid anyone tries looks perfect and the bug waits in the second one. That is the trap this code was written to survive, and the error reported is measured live rather than quoted.
This surface is the work product of an agent loop that writes its own tickets and grades itself against gates it also writes. A system like that gets very good at what it can measure and blind to everything else. The correction is people from outside it — which requires there to be something you can open. So: does the squash read as the world having a grain you could learn to use, or as the shapes being broken? There is no test for that, which is exactly why it is being asked.