Not sure I quite get the definition of the 3 categories, but certainly you indicate 3 interesting exemplars.
First category: maybe a good defitinion of this category is coevolution of problem set with problem solvers. POET does this explicitly; first time I remember seeing this explicitly was in an effort to evolve arithmetic computation: Evolving inductive generalization via genetic self-assembly .
Second category: what is the definition? Simply GAN and its generator discriminator instability as the mechanism for open-endedness? You are right about traditional GAN formulation being driven by data, but a population of GAN agents interacting with each other seems like a very natural context for open-endedness. Starting with a population of only 2 starts to look like a version of Takashi’s Can Mutual Imitation Generate Open-Ended Evolution? .
Third category: various => examples? Maybe simplest version of ‘artificial boredom’ (I like the term!) is wandering in a neutral landscape? Definition seems to need a space of possibilities (latent space is one example? skill-space??) and algorithm to explore the space.