What does nested cross-validation estimate?

A researcher performs hyperparameter selection using inner cross-validation and reports the average score from an independent outer cross-validation loop. What does the outer-loop score most directly estimate?

An unbiased estimate of the Bayes error whenever the outer folds are stratified

The expected performance of every hyperparameter configuration considered in the inner loop

The generalization error of the model refitted once on the complete dataset using the globally best hyperparameters

The generalization performance of the full model-selection procedure, including hyperparameter tuning