fix: complete sklearn 1.9 support — drop self.alphas overwrite (closes #1032)

[immu4989]

Closes #1032.

What's happening

#1032 (filed by @jakevdp) reports that import econml.orf errors on scikit-learn 1.9 because n_alphas was removed from LassoCV.__init__.

The recent commit e546416 added a >= 1.7 version dispatch in WeightedLassoCV.__init__ / WeightedMultiTaskLassoCV.__init__ that translates n_alphas=<int> into alphas=<int> on the super().__init__() call. That fixes the import-time TypeError jakevdp reported.

But e546416 also added self.alphas = alphas at the end of the dispatch, which clobbers the value sklearn's __init__ had correctly recorded back to the constructor's alphas kwarg (None by default):

if parse(sklearn.__version__) >= parse("1.7"):
    super().__init__(
        eps=eps, alphas=alphas if alphas is not None else n_alphas,  # alphas=100 ✓
        ...)
    self.n_alphas = n_alphas
    self.alphas = alphas    # ← overwrites sklearn's alphas=100 back to None

On sklearn 1.7–1.8 the loose param-validation tolerated the resulting self.alphas = None. On sklearn 1.9 the stricter _param_validation rejects it, so SparseLinearDML.fit (via _DebiasedLasso.fit → WeightedLassoCV) and DebiasedLasso.fit raise:

sklearn.utils._param_validation.InvalidParameterError:
The 'alphas' parameter of WeightedLassoCV must be an int in the range [1, inf)
or an array-like. Got None instead.

Fix

Drop the overwrite. super().__init__(...) already records self.alphas from the translated value. self.n_alphas is still set so callers can introspect the original wrapper kwarg.

Same edit applied to both WeightedLassoCV.__init__ and WeightedMultiTaskLassoCV.__init__.

Also bumps pyproject.toml scikit-learn >= 1.0, < 1.9 → < 1.10 so users can actually install sklearn 1.9.

Verification (locally, sklearn 1.9.0 + narwhals)

• import econml.orf ✓ (jakevdp's exact repro)
• Smoke fits all succeed on sklearn 1.9: LinearDML, SparseLinearDML, CausalForestDML, DMLOrthoForest, LinearDRLearner, SparseLinearDRLearner
• New test_default_alphas_fits_on_strict_sklearn regression test in TestLassoExtensions:
  • Asserts WeightedLassoCV().alphas is not None and WeightedMultiTaskLassoCV().alphas is not None (the dispatched value must survive)
  • Calls .fit(...) on each, verifying sklearn 1.9 strict validation accepts the result
  • Pre-fix on sklearn 1.9: fails with AssertionError: WeightedLassoCV.alphas was clobbered to None by __init__ (#1032)
  • Post-fix: passes
• Broader sweep: 48 passed, 0 failed across test_linear_model.py + test_dml.py + test_treatment_featurization.py on sklearn 1.9.

[kbattocchi]

Thanks for the clean root-cause analysis and the regression test — this is the smallest correct fix for #1032, and I appreciate that you traced the failure back to the self.alphas = alphas overwrite introduced by #1031 rather than just papering over the InvalidParameterError.

Two follow-ups I'm planning as separate commits (on top of this) so I'm not blocking the merge here:

1. Factor out the shared kwargs dict so both __init__ methods stop duplicating the whole super().__init__(...) body across the version dispatch.

2. Add a get_params() override on both wrappers to drop n_alphas on sklearn >= 1.7. Without this, self.get_params() still returns n_alphas (because sklearn's _get_param_names inspects our wrapper's __init__ signature, not the parent's), which leaks into lasso_path(**path_params) inside LinearModelCV.fit. On 1.7-1.10 that emits a FutureWarning storm (once per CV fold); on 1.11 n_alphas will be removed and it will become a TypeError. Proactively silencing shields us from both.

Approving.

[kbattocchi]

Reopening to trigger a fresh CI run — the previous run is >30 days old, so GitHub's "Re-run failed jobs" is unavailable and manual workflow_dispatch runs don't update the PR's check status.