Please excuse the brainstorming . . . but could it be that the B-adjustment screw is set with the pulley just a little too close to the cluster? Maybe the contours of the bigger cassette when using the WiFli rd have a spot where the cassette is closer to the pulley than it is at the extremes of the cluster -- but just BARELY too close, so that it sometimes works and sometimes doesn't. The condition described seems to be what would (logically) happen if the rd encountered some resistance when trying to make the shift.

Shifting the fd would change the length of the chain / distance of the pulley from the cassette, thereby freeing the rd. And shifting just the rd would register a shift electronically, but the rd would physically back away from the obstruction. Are all your chainrings, cassette, and chain length within recommended specs?

I don't know if this makes sense -- just trying to help. It's a theory.