A long time later… This most recent error is simply caused by there being observed values in the RVs for coupla: the inner workings of prior predictive (and actually posterior predictive) can’t work with this.
Which makes my current life tricky since I really want to sample from my model (which looks a bit like this)
My current solution is to have this backwards-pass model that evidences at the copula, and a very similar forward-pass model that evidences on the marginals… although the forward-pass model is currently incomplete as I figure out pm.Potential 