By letting the code formatter exceed the limit if it allows for more readable formatting in certain cases. Going for a 100% hard limit means sometimes it'll shuffle chunks of code around because of 1-2 characters and that just doesn't make a lot of sense.

Or in other words, by formatting the code more like a human than dumping the source tree with a blind set of rules. If Copilot is possible then so is an AI model able to consider how code actually looks on the screen.

How do you quantify the tipping point where a longer but continuous line is prettier than a broken line? (broken where?)

They already answered the question: you use statistical methods (aka AI).