I feel your pain,

I once tried to use graphframes to flatten a complex tree, ended up using graphX (which is even worse to use but at least it is more flexible).

So maybe take a look at graphX? Beware, it is terrible to use.

I wonder what happened to making cypher (or tinkerpop or whatever) available on spark/databricks.

That being said:

Maybe you can try to put the type of object into your edges. The vertices are typeless and contain only values. The edges have a different value depending on what vertices are connected.