[BUG] set_params silently ignores nested parameters, breaking GridSearchCV

[KaranSinghDev]

Description

Currently, pyGAM estimators fail to properly route nested parameters. Because set_params() does not recursively handle scikit-learn's double-underscore (__) routing syntax, nested parameters are silently ignored.

This results in a critical silent failure: users running GridSearchCV over a range of term-specific parameters (e.g., varying n_splines) are unknowingly evaluating the exact same default model repeatedly, as the terms never actually update.

Steps to Reproduce

from pygam import LinearGAM, s

gam = LinearGAM(s(0, n_splines=10))

# Scikit-Learn meta-estimators tune nested parameters using this syntax:
gam.set_params(terms__0__n_splines=30)

# The parameter is silently ignored. The term remains unchanged.
assert gam.terms[0].n_splines == 10

So any hyperparameter tuning on specific terms via Pipeline or GridSearchCV is currently a placebo.

I think we can override set_params in GAM to intercept kwargs containing __. Split the key, resolve the target object within the TermList, and apply setattr directly to the nested term.

[ishan372or]

Hi, just checking in on this issue. Are you currently working on a fix for it? If not, I’d be interested in trying to implement it and opening a PR.

[KaranSinghDev]

@ishan372or I was working on it but it is taking longer than usual as I am working on other proejcts as well, you can try it, if you want any assistance let me know.

[ishan372or]

Thanks for the update. Since you’re okay with it, I’ll take a shot at implementing the fix and open a PR.