This bug doesn’t actually affect the selection results.
The issue is caused by missing normalization of the target values y. However, in minibatch, the selection process treats each target (i.e., each column of y) independently.
As a result, during each batch selection, only one target is used, and the SSC scores of all candidates are scaled by a constant factor. This scaling doesn’t change their relative order, so the selection outcome remains the same.
Normalization would only affect the results when multiple targets are used together in a single batch selection — but that’s not the case in the minibatch.