I hadn't thought about that, but I'd actually like to see if it's better to do the opposite. Insert half a leap second twice as frequently as we currently do with a leap second. If that's successful, switch to inserting a quarter leap second and so on, maybe until we get to a state where we adjust a tiny fraction of a leap second on a daily basis. The point is that even if you have a system screw up time adjustment 1 day, it's such a small time error that shouldn't affect anything, and people can immediately work on fixing it.

And actually, what you explained about waiting until some significant threshold like 5 minutes is already something we do today in a way. It's just that we wait years and our threshold is 1 second before we actually make a change.