We switched to Quickhull in 2018.3 that is a much more robust approach to convex hulls than we had before, in general. However, in this particular case we're hitting the limits of it, as the input mesh is arguably tiny. Notice all of its vertices contained within a 0.03 cube even after being scaled x100 on the GO. Quickhull takes the mesh before this scaling, so has to deal with even smaller values, at which scale it doesn't seem to be able to correctly expand the initial tetrahedron to the actual hull. I think after it finds the initial tetrahedrons, all the remaining vertices are discarded as being located on the surface of the tetrahedron. I recommend applying scale to the mesh before passing it to the mesh collider, either in a 3D tool or in the mesh importer settings.