The traditional hard fork sequence is

1. Rules change on the chain at block X
2. Sometime later somebody produces a block larger than 1M, and fork occurs. Perhaps block X+0, perhaps X+1, perhaps X+100 or X+10000.

Tighten this to require a larger-than-1M block when the rule changes, on block X+0. Doing so creates a more predictable behavior with the network.

Requested by Bitmain and BU to implement an anti-wipeout feature; myself and other WG members concurred via the wider rationale of creating a more predictable network upgrade.