The withdrawal was marked sent. The chain had no record of it. Both were right: they were reading different nodes.
On my machine the custodial wallet talked to a single node. Broadcast, then ask for the account nonce, and the answer included what I had just sent, every time. Sequential, consistent, free.
Production talked to an endpoint sitting in front of a pool of nodes. The write landed on one, and the next read was served by another that had not seen the pending transaction yet. So the allocator handed out the same nonce twice. The second transaction replaced the first. A payout silently stopped existing.
The expensive part was not the fix. It was working out which withdrawals had actually landed: reconciling my own records against the chain, one by one, because that state had been derived from a read I did not control.
Now I do not derive a nonce from a remote read at send time. Nonces are allocated from my own database, one allocator per wallet, inside a transaction. The chain confirms in-flight state, it is not the source of it.
The gap between my machine and production was never config. It was a consistency guarantee a laptop gives you for free and no distributed system ever promised.
If you run custodial wallets or on-chain payouts, I would go looking for the place where your system still trusts a read-after-write it cannot guarantee...