Assuming Alice is paying Bob with an old style time-locked tx. Under your proposal, after the hardfork, Bob is still able to confirm the time-locked tx on both networks. To fulfil your new rules he just needs to send the outputs to himself again (with different tx format). But as Bob gets all the money on both forks, it is already a successful replay

On 25 Jan 2017, at 15:15, Natanael <natanael.l@gmail.com> wrote:

Den 25 jan. 2017 08:06 skrev "Johnson Lau" <jl2012@xbt.hk>:
What you describe is not a fix of replay attack. By confirming the same tx in both network, the tx has been already replayed. Their child txs do not matter.

Read it again.

The validation algorithm would be extended so that the transaction can't be replayed, because replaying it in the other network REQUIRES a child transaction in the same block that is valid, a child transaction the is unique to the network. By doing this policy change simultaneously in both networks, old pre-signed transactions *can not be replayed by anybody but the owner* of the coins (as he must spend them immediately in the child transaction).

It means that as soon as spent, the UTXO sets immediately and irrevocably diverges across the two networks. Which is the entire point, isn't it?