Imputation
Replacing a missing value with an estimate, so a row with a gap can still be used.
The value you write in is fabricated, and the model cannot tell it apart from a measurement. A filled 4.0 and a measured 4.0 are the same float, with no flag and no weight, which is why the honest question is what damage a method does rather than which one is best.
Mean imputation shows that concretely. A value placed at the mean adds nothing to the total squared distance from the mean, so variance falls while the row count rises — and a spike of identical values can't co-move with anything, so correlations drop toward zero too.
Two habits worth keeping. Add a boolean column recording where you filled, since missingness is often predictive in its own right. And remember the statistic is fitted, so compute it on training rows only and refit inside every cross-validation fold.