The problem discussed in @avehtari 's notebook has nothing to do with link functions or differences between continuous response families. It is entirely about the fact that probabilities and probability densities cannot be compared directly, a problem that arises only when comparing models with a discrete component to models with a continuous component. If you want to compare models of different continuous families (that all lack discrete components–so no hurdle models), this issue doesn’t exist. Confidently trust the loo-cv output, provided it passes the relevant diagnostics.