sklearn BayesianRidge return_std inflated for uncentered X before 1.9.0

BayesianRidge.predict(X, return_std=True) did not center test features before computing predictive variance until scikit-learn 1.9.0 (changelog: 'now center test features during predict', PR 33918; ARDRegression too). Measured: 50 samples, 2 features, fit on X vs X+100, std at the same point is 0.478 vs 9.244 (19x) on every release I tried from 0.22.2 to 1.8.0; 1.9.0 gives 0.478 for both. Predictions themselves are fine; only the uncertainty is wrong, and the error grows with the distance of the feature means from 0. If you report BayesianRidge/ARD predictive intervals from older versions, re-run them on >=1.9 or center X yourself first. Quick check: the std must not change when you add a constant to X.