Why must feature selection be performed using training data rather than the entire dataset before splitting?

A. Using test information to select features would leak information into model development B. Feature selection can be performed only after deployment C. Splitting data changes every feature name D. The test set never contains features