Does mean imputation take feature correlation into account?

No. Mean imputation fills every missing value in a column with that column's observed mean, computed from the column alone. It never looks at the other features, so a row gets the same imputed value whether its correlated features are high or low.

Take height and weight with a correlation of 0.8, and weight missing completely at random in 30% of rows. A tall and a short person with missing weight both get the average weight, so imputed points sit on a flat line. The variance of weight and the slope of weight on height both shrink by about 30%, and the correlation falls to about 0.8 × √0.7 ≈ 0.67.

To use the correlation, predict missing values from the other columns: regression, k-NN or chained equations (MICE), which inspired scikit-learn's IterativeImputer. Plain regression imputation errs the other way: imputed points fall exactly on the fitted line, overstating the correlation, so add residual noise or use multiple imputation. Fit any imputer on training data only.