For quadratic datafit, the step includes the stepsize, because the datafit is smooth. For Poisson and Gamma we take a stepsize of 1, because the datafit is not smooth so there is no good rule to tune it. See https://github.com/scikit-learn-contrib/skglm/pull/326#discussion_r2245118947