Sadly this doesn't scale, as you almost seemed to realise by saying you'd just respond without waiting for the commit step to complete. Next I'll respond to your dismissal of this huge problem in the other thread.
The right thing to do in my mind is to only send requests in 0-rtt if they're specifically flagged as eligible.
Yes, do this. Processing the request isn't bad as long as it doesn't commit.
Most evil attackers won't come back to you complaining that your server gave back an HTTP200 response when it should have returned an HTTP400 because they'd reused a nonce.
I know it was intended flippantly but the attackers do not get the HTTP response code - they're obliged to conduct this attack blind, if they can decrypt the messages then your problem isn't a replay attack but some previously unimagined TLS flaw.
So, suppose your system sees three versions of a transaction to transfer $1M to Dr Evil. You're not worried because two of them "won't commit" but do your end users understand that? (hint: they do not). Let's call these versions A, B and C. The real user's web browser believes transaction A succeeded, and is unaware of B and C.
Dr Evil complains that the money never arrived, and our user checks, alas they see that eventually transaction A failed (it was judged a duplicate of B, and so was C) and this explains the problem. They initiate a new transfer, let's call that Z. This succeeds and Dr Evil is satisfied.
Unfortunately now Dr Evil has $2M instead of $1M.
The server side can easily keep track of which nonces have been used, and not allow any re-used nonces to have any side effects.
This can be implemented in something like spanner by keeping a database of valid nonces, and tying the marking a nonce as used to the other effects of the HTTP request in the same transaction. Spanners global-consistency will ensure only one transaction will commit. In most cases, the data can be sent back to the user without waiting for commit, preventing any cross-ocean latency penalty too.
It's a shame that I don't see a single HTTP/3 server which has any support for a scheme like this.