This is a tricky one. For some context / history, see here: GLTF Parser Hooks.
Basically, the issue is that Babylon’s multi-material / submeshes is not a 1-to-1 mapping to glTF mesh primitives. This makes it difficult to get correct round-tripping behavior. The loader doesn’t attempt to load the glTF mesh primitives into Babylon.js submeshes as there are many edge cases.
One possible workaround is to manually merge the meshes using Mesh.MergeMeshes after the load if you know this will work for your scenario.