What is the main statistical role of cross-fitting?

In double machine learning, nuisance functions are trained on K−1 folds and evaluated through an orthogonal score on the remaining fold. What is the main purpose of this cross-fitting procedure?

To separate nuisance fitting from score evaluation, reducing overfitting bias and restrictive empirical-process requirements

To guarantee that every nuisance learner is unbiased on each fold

To ensure that all observations receive identical nuisance predictions

To select the nuisance learner with the minimum cross-validation error