If perfect is not possible let's do nothing? Your approach is likely the reason for the current bad global rule. You want to replace it with another bad global rule.
But in the real world I'd use proxy conditions to arrive at a rate of correct guesses that's satisfactory for me just like with everything else with computers (even basic mouse clicks), not use your total "safe" ban (safety has a trade-off, you can't resolve that universally)
I (and presumably a lot of other people) don't believe that it's possible to build a system that can dynamically determine when it's acceptable to steal focus with low enough error rate to make it better than a total ban.
Seemingly you disagree and believe that such a system is possible. Is there anything preventing you from implementing it to prove the doubters wrong?
But in the real world I'd use proxy conditions to arrive at a rate of correct guesses that's satisfactory for me just like with everything else with computers (even basic mouse clicks), not use your total "safe" ban (safety has a trade-off, you can't resolve that universally)