With repeated responses at an identical predictor configuration, pure error is their within-configuration variability. Its residual sum of squares can estimate measurement variance independently of a candidate regression mean function. Without replication, a fitted residual can combine pure error and lack of fit.
New to topics? Read the docs here!