CI failure looks unrelated

Could you post on the forum? Either for this change or for adopting SPEC7 project-wide

Would be nice to have it spelled out explicitly somewhere. Backwards compat break procedures and all that.

It is spelled out by the SPEC except the duration of SciPy's deprecation cycle being two releases long. For us, that means:

• for two releases, no warnings
• then we turn on warnings in the decorator and wait two more releases
• then we remove the decorators and all old random_state/seed validation. Whatever the user passed for rng gets passed into rng = np.random.default_rng(rng) for validation/standardization.

Interesting. Applying the decorator to the class looks like it almost worked, but the decorator appears not to preserve a __code__ attribute. (IIUC functools.__wraps__ doesn't get everything right.) If we find a solution for this, it would be good to update SPEC 7 accordingly.

The other failures are gh-21831 and gh-21828.

gh-21833 shows a few other classes with initializers that accept seed or random_state, but they have other issues that will require more work than simply applying the decorator.

• BootstrapMethod and PermutationMethod expose an attribute random_state, so applying the decorator is not enough.
• The QMC classes accept seed but they have their own check_random_state function that behaves differently from everything else; it passes RandomStates through, but it uses default_rng on everything else. Rather than further generalize the decorator, I think we'll just copy-paste.

With that in mind, it might not be worth the time to generalize the decorator to replace the documentation here. If you edit the documentation manually, would putting the decorator back on the __init__ method fix the remaining error, do you think?

gh-21833 shows a few other classes with initializers that accept seed or random_state, but they have other issues that will require more work than simply applying the decorator.

• BootstrapMethod and PermutationMethod expose an attribute random_state, so applying the decorator is not enough.
• The QMC classes accept seed but they have their own check_random_state function that behaves differently from everything else; it passes RandomStates through, but it uses default_rng on everything else. Rather than further generalize the decorator, I think we'll just copy-paste.

With that in mind, it might not be worth the time to generalize the decorator to replace the documentation here. If you edit the documentation manually, would putting the decorator back on the __init__ method fix the remaining error, do you think?

Sounds reasonable, done

Failures are unrelated; saw those in gh-21867.

@j-bowhay This looks pretty good, but there are no uses of rng in the tests - were the tests not seeded before? I guess that hasn't bitten us with random failures, but do you agree it would be good to include it (at least to make sure it's working)?

The permutation is to prevent underflow/overflow during the computation, the seeding shouldn't impact the results as basically any permutation will do but I will add it to a test as I agree it would be good to have for coverage