FWIW Linux software RAID doesn't do that. I think the argument was that differences like this were mostly related to power loss where some disks have the new da

FWIW Linux software RAID doesn't do that. I think the argument was that differences like this were mostly related to power loss where some disks have the new data and others the old data. And at the block layer, it's impossible to tell which is which and so the code just picks a winner basically at random. I'm not 100% convinced myself that a 'majority wins' strategy like you described wouldn't be superior, but I can see why they decided otherwise.