Mathematically yes; in practice one is still sampling a density using HMC, so the prior sampling process can break in all the same ways as the posterior sampling process. The following issues seem related to your model/prior:
- Meaning of divergences in prior predictive checks
- Weird output for a very simple model with sample_prior = "only"
- [Clarification] Divergences during sample_prior = "only" model fitting
I’m surprised that the sampler isn’t throwing more warning messages here. It is usually fairly warning-happy with these kinds of things, so I’ve probably missed something somewhere in the model.