Poisson and readnoise variances are added to the frame at the start of the pipeline. These are propagated all the way through the pipeline if Kappa 1.0 or newer is used. If an older version of Kappa is used then the variance will be lost when the data is formed into a datacube. A warning will be shown if this happens.
The values of adjacent pixels in the output frame are correlated due to resampling in the y and λ directions, so strictly speaking the variance of the final frame should be represented by an enormous covariance array. This would not be practical, so the variance array is propagated by resampling it in the same way as the data array, which while not rigorously correct should provide a useful estimate of the variance.