… aw dude, I am sorry for taking up so much time for this non-error. (so embarressing…)
I was wrong, but this does reveal a misconception I had. I was implicitly assuming that if I sampled points from the model under the seed handler while the model was conditioned I would expect inference, which is incorrect. I guess I was confusing Ancestral sampling [1, 2, 3] as a form of inference.
To sample the model’s posterior you need wrap the model in an inference algorithm. =P