Sounds like an incompatible update in the data format used in a cookie on a chrome experiment. Incognito would have started with no local data to mess things up, and non-chrome browsers would not have been included in the experiment.
Getting things like experiment transitions tested properly for the inside teams would be tough. Given that they flip the switch on experiments in GA, it likely would require a significant upgrade to their experiment architecture to have the GA experiment transitions tested internally.
Someone will get a promotion if they figure that out & trumpet it properly.