The first work, I linked, arguably gets around this issue completely. There we train a neural network to approximate Bayesian prediction directly. It accepts the dataset as input. Thus, there is no optimisation procedure per dataset, just a forward pass. So far, this seems to be pretty stable.
Just because you optimize network parameters to fit a surrogate model of the data doesn't mean you are not doing local optimization without guarantees on robustness or global optimality w.r.t. network parameters to fit the surrogate. Am i missing something?