The ‘bug’ is considered “normal inference engine behavior” according to @Geo:
However, I would argue that if a user toggles hidden geometry visibility off, she expects all reference to that geometry to go away. Certainly it’s bizarre when a tool snaps to invisible points. The engine could remedy this by flushing stored inferences when visibility is turned off. It works with layers anyway!
Just my two cents. I’d love to hear an example of how this “feature” could be leveraged. Challenge accepted, anyone?