statsmodels Logit: complete separation gives only ConvergenceWarning

Do not rely on PerfectSeparationWarning to detect separation in statsmodels Logit. In a 40-point simulation (x ~ U(4,8), sm.Logit(y, add_constant(x)).fit(disp=0)), statsmodels 0.14.5 and 0.15.0 gave: y = (x > 6), complete separation -> slope 247 / 251, converged=False, ConvergenceWarning only. One 1 at the largest x (quasi-separation) -> slope 426 / 510, ConvergenceWarning + RuntimeWarning. All-zero y -> PerfectSeparationWarning appears, but the returned slope is -4.19 on 0.14.5 and -1.51 on 0.15.0 for the same data, se 8.6e8 / 3.6e5. All results are returned, none raise. Check r.mle_retvals['converged'] and huge bse, or count 0/1 outcomes per group before fitting; for a finite estimate use a penalized fit: sklearn LogisticRegression (C=1) gave 2.45 and 0.90 for the first two cases and raised ValueError on the all-zero one; fit_regularized(alpha=1.0) gave 1.17 and -0.55 (L1 that also penalizes the intercept, so choose alpha and the penalized columns deliberately).