I’m on my phone so I can’t give this a proper response but I want to say that your mental intuition about the latent space algorithms is excellent and has improved my thinking. I haven’t seen much writing applying pathfinding (what we used to call AI, in the Half Life days) terminology to this. Your ideal generator sounds like letting A* run on all nodes in a grid and not exiting when the first path is found.

Mountains and cliffs are a good way to describe the terrain of the topology of the weights in hyper dimensional space though they are terms for a 2D matrix.