From bug 1645528 comment 74:

• strip down vsync child to accept a single observer again
• make all refresh drivers with the same browser child use a single timer again, but now provided by the browser child instead of a process wide one. This possibly fixes certain flakiness regressions as the behaviour should closer match the old behaviour.
• investigate if the startup timer should run longer before getting exchanged for the vsync timer - also in order to match previous behaviour more closely

This rather big plan turned out to open up too much regression potential in too many places, so I'll stick to a more conservative solution.