As crazy as it may end up being, setting up my engine thing to use an Entity-Component System (ECS) -based approach might actually be more sane, particularly if I use multiple threads/fibers for assembling GPU data and command buffers.

The reasoning is simply that it means never having t walk the entire scene or worry about "virtual" nodes to compose stuff, but rather querying exactly what I need ("give me every instance of That Thing").

Which is really useful when I need to do things like assembling a buffer of 4x4 matrices to render thousands of identical objects in a single pass (which also requires knowing the number of instances and how many vertices to draw).