The specific attack you describe doesn't seem possible, because, at the point where the fake client would need to commit to a nonce, the server doesn't yet know the other clients' nonces, but only their hashes. Provided that the nonces have sufficient entropy to resist brute force attacks, it should not be possible force the server to learn them just from the hashes. This is precisely the point of a commitment scheme.
(Some methods of combining the nonces can be vulnerable to replay attacks, where the misbehaving client simply echoes another client's hash and nonce. But this attack is generic and can be carried out by any client that can see another client's messages before sending their own, and in any case can be prevented either by including a client ID in the hash input, or by aborting the protocol if any two hashes match.)
The critical part here, which is not fully explicit in your description of the protocol, is that no client should reveal its nonce before it has received hashes from all other clients (and from the server, if the server contributes input to the random number generation). This ensures that all participants must commit to their nonces before they learn any other client's nonce.