Nope, no plans to go in that direction. We are trying to limit our focus to models that are tractable. None of these are because of the combinatorial multimodality in the posterior (not just index switching, which is manageable). In models that are tractable, you can almost always marginalize out any discrete parameters for much greater efficiency than you’d get from sampling them.
We do support Gaussian processes, which are essentially non-parametric curve fitters that work sort of like a smooth version of K-nearest neighbors.
I think what @Max_Mantei was getting at in the second paragraph is that you can approximate a Dirichlet process with a large Dirichlet. But you won’t be able to get good convergence in multiple chains because of multimodality. You see the same thing with a simple Latent Dirichlet allocation model (i.e., multinomial mixed membership mixture model).