Hardened the default get_autoML() stack so the high-level API uses the available safe subset of learners and warns once when optional runtime packages are missing.
Finished the two-stage T-learner update for reduced-modifier settings so first-stage regressions use the confounder set W and the default second stage projects onto the modifier set V.
Tightened task validation for folds, weights, treatment values, and user-supplied nuisance estimates.
Added release checks for the custom website, vignette inventory, and package/docs sync, while excluding deployed site assets from package builds.