Are monotone_constraints still restricted to gaussian/bernoulli/tweedie in XGBoost, or is the documentation out of date?

[eyalbarna]

The docs for monotone_constraints state that in GBM and XGBoost the option can
only be used with gaussian, bernoulli, or tweedie distributions (plus quantile in
GBM):
https://docs.h2o.ai/h2o/latest-stable/h2o-docs/data-science/algo-params/monotone_constraints.html

I'd like to confirm whether this restriction is still actively enforced, or
whether the documentation is simply stale.

Native XGBoost imposes no such restriction — its docs describe monotone
constraints purely at the tree-split level, applicable to any feature regardless
of objective (https://xgboost.readthedocs.io/en/stable/tutorials/monotonic.html).
Since constraints act on the raw margin and the log link used by Poisson and Gamma
is monotonic, there's no obvious reason the underlying engine would need those
distributions excluded.

Questions:

1. Is the gaussian/bernoulli/tweedie restriction still enforced in the current
   release, or has it been relaxed while the docs lag behind?
2. If still enforced — is it deliberate, or could it be lifted given native
   XGBoost already supports it?

Context: this matters for actuarial pricing, where Poisson (frequency) and Gamma
(severity) are standard and monotonicity of rating factors is often required.
Being able to use them directly, rather than via a Tweedie proxy, would help.

[hiSandog]

A useful next step would be to pin down whether the break is caused by the dependency matrix or by the docs configuration itself. Capturing the exact Python/package versions and the failing docs command for monotone_constraints would make this easier to turn into a CI regression test.