In addition, if I understand it correctly, the goal of ppc is to resample observed or new variables given as input the parents points in a trace (e.g, the likelihood, new input x or new/missing groups). You don’t usually resample unobserved variables as you seem to be trying in your example. That’s because the posterior for unobserved variables may not look like any known distribution and therefore we cannot just use random number generation routines (you would need mcmc sampling again).
Yes, my little example is a bit pathological. I distilled it from a much larger model that is more sensible, but that also exhibits this strange behavior.
Note that in my larger model, I am using the super-useful (and under-appreciated) model factory idiom described by @lucianopaz here and here.
But perhaps I am misunderstanding your point?