The mixture model is wrong, as it assumes the zeros could have come from either component which is not true. It’s mixing densities and masses.
That’s what the epsilon truncation tries to avoid, although it’s fundamentally just a hack to reuse the mixture implementation.
As to why your model fails to sample I’m not sure. Perhaps the priors are too diffuser or you don’t have enough data.
You can break the observations into a gamma likelihood for the non zeros and a binomial for the number of zeros which should be equivalent to the hurdle model, without the epsilon truncation hack. Does that also fail to sample?