For non-observed GRW it may be better to define in the non-centered way, which is NOT what pm.GaussianRandomWalk does as you said. For observed, only centered works, but it wouldn’t matter as we are not sampling those.
I have thought about it and considered using different forms for observed and unobserved distributions.
Would be good to confirm NUTS indeed prefers the non-centered case. I would guess so as it can do proposals closer to the unit range.
This might also make sense for Censored distributions as NUTS can’t sample the direct logp, but it may be happy to sample a latent uncensored distribution that is then clipped (or perhaps not, but that’s what we suggest in the docs anyway)