06b4292534
Instead of using the current dimension for which we just got a big span, we were computing Min and Max for the previous dimension stored in cutfeat (and using 0 instead of the dimension indice for the very first dimension with "span > (1-eps)max_span")