Pathfinding - same as the lighting system - is using the spatial acceleration structure the physics engine uses to gather parts for voxelization. If the part physical extents mismatches the visual extents we can’t find it by visual extents, so we choose to just ignore this. We could still voxelize it if the scale reduced the visual representation, but we can’t do it if the scale is more than 1.
Why do you need to use the aforementioned meshes? You could always use blocks instead of block meshes and cylinders instead of cylinder meshes?