I first set up the following cooked 10#:
Kc5 Bf6 Bf7 Pg5 Pg6 Pa3 - Kh8 Sg7 Pa5 (maybe the Pg5 is just
a misprint and must be dropped? Sauce?)
and them played 1.Kd6 a4 2.Bxg7+ Kxg7, then removed the
a file pawns. The result you'll see above. Seems somehow the
program "remembers" the removed piece (as 1.Bb3 loses).
(After site reload and position reenter it works again)
Thanks to Fritz. I somehow thought it only can be won by sacrificing bishop and pg6 to reach wKg6 Pg5 bKg8. No wonder after only three and a half hours of sleep...