In your first examples, cutpoints must be ordered. If you use the ordered transform this will happen during mcmc sampling and should be fine although you need to provide an initval.
However if you call .eval you will not see ordered values because the transform only plays a role during mcmc sampling, it’s not part of the generative random graph that you are accessing when you call .eval().
These pages give more details on transforms: Transformations — PyMC dev documentation
Linked issue about a proposed alternative to ordered transforms: Implement `Ordered` distribution factory · Issue #7297 · pymc-devs/pymc · GitHub
For why the constrained uniform or other attempts are failing I’ll have to take a look. Just wanted to clear perhaps a source of confusion