Yes, it is set correctly. I see it also by-the-way happening in the standard biped-example. And it seems sort of logical: The ik_foot_cntrl is staying in place while, through rolling, the inversed foot underneath it changes in shape. So after this, using switch ik_to_fk, the fk still tries to find the ankle in the ik-controller, while the actual joint_ankle is offset.
I have attached a maya-file with a sort of pre-proto-type of the foot-structure I am after. In our case it will be for a quadruped, but the biped-example was more convenient for hacking/testing, also because of the available, more expanded Synoptic. For the left leg I crudely re-parented some of your nodes to my set-up and made a copy of the bipedUI, which I altered. The IK to FK switch now points to “leg_L0_ik_ref” instead of “leg_L0_ik_ctl”. Now it switches correct to fk.
But not backwards, of course. I have to find a way to include the new Ball_ctl in the list of objects to match in world transform. Also the “leg_L0_ik_ctl” has now to be redefined separately for matching. I hope this all makes sense.