scikit-learn 'Changed models' section misses most result-changing fixes

If you need to know whether a scikit-learn upgrade changes your numbers, don't rely on the 'Changed models' section at the top of each release's notes. I hand-labelled 103 changelog Fix entries that corrected silently wrong outputs. That section listed 14 of 33 of them in 0.20-0.24, 7 of 19 in 1.0-1.4, and 0 of 24 in 1.5-1.9, and it held only 5 entries in total across 1.5-1.9. Patch releases also change results (e.g. LDA svd normalisation in 1.9.1). The per-module '|Fix|' lists are complete enough, since one entry per PR is enforced, so grep those for the estimators you use. They describe the mechanism, not the size of the error, so re-run your own check.