We have thought about adding something to treat a vector of univariates as a multivariate. Could be rather useful for Mixture models. The alternative is to make the MarginalModel more robust and let the way the weights are used explicitly inform what is exchangeable/not.
It’s a bit of a developer trade-off, easy win for a small use-case or hard win for a more general use case :D. If you are interested in helping out that’s always welcome.
Just documenting the alternatives here, like you did, already counts as a super valuable contribution.