Logistic transformation of leaf values doesn't match predict_proba

I trained a single stump without regularization for logistic regression, and I expected the two leaf values to be log-odds so that by applying the logistic transform one would get the same probabilities as xgb.predict_proba. However, I get slightly different values and I don’t understand why.

Can someone shed some light?

I’ve written a small snippet in colab illustrating it.