[jonathan-taylor]

Follows #224, whose new test test_max_dense_design fails on main.

The guard in #224 works (the covariances are None above max_dense_design, and the standard errors raise). The test fails at the next line, dir(G):

• scikit-learn's BaseEstimator.__dir__ keeps each attribute for which hasattr() is true, and hasattr catches only AttributeError.
• Its HTML display (a notebook's) calls inspect.getmembers, which reads every attribute dir() lists.

So dir() and displaying a fitted GAM evaluated all seven computed properties (design_, posterior_covariance_, frequentist_covariance_, edf_influence_, penalty_matrix_, fixed_effects_, fixed_effects_covariance_):

• below the max_dense_* limits, they formed the dense design and the P x P matrices;
• with a categorical() term, at any size, fixed_effects_ raises NotImplementedError, so dir() and the display failed;
• above max_dense_design with an s() term, fixed_effects_ raises ValueError from _check_covariance, so they failed too.

GAM.__dir__ now leaves these properties out, so neither dir() nor the display computes them. They're still there when asked for (G.design_). It keeps scikit-learn's hasattr filter (for available_if) for the remaining names.

Tests:

• test_max_dense_design adds the s() case over the limit;
• a new test_display_forms_nothing checks that dir() and estimator_html_repr form no covariance, list none of the computed properties, and that they still evaluate when asked for.

Run locally (glmstar 0.7.0 wheel with this branch's estimator.py, whose C++ is unchanged since): tests/gam 446 passed, 160 skipped; tests/gam/test_gam_sandwich.py with R, 16 passed.

Changelog bullet under Unreleased / Bug fixes.

🤖 Generated with Claude Code