Hard to reason about all the permutations, but sounds like you over-parametrize when you do both. How do the pair plots of the unconstrained parameters look like?
You can keep them by passing idata_kwargs={"include_transformed": True} to pm.sample (IIRC)