The main idea of an HSM is that the private key cannot be exported out of it. The way you describe your scheme means that you will have to get the signature out, then create the private key from that outside the HSM. Obviously if your server is compromised then an attacker can get to that private key by extracting it from memory, or by forcing the application to perform a signature.
Another advantage of HSMs is that the operation and algorithms are generally validated to work correctly and not be applicable to e.g. side channel analysis. You cannot rely on that kind of reliability by performing the calculations outside of the HSM.
There is an advantage to derive a single key pair from the existing one: you would not have to store the private key in memory while you are not using it. It is possible to keep it in memory and try to destroy it once you've used it.
That said, if the memory contents can be retrieved, if side channels are possible, if the signature generation can be achieved by an adversary then you cannot be sure that the private key gets copied and security gets lost.
Initially it seemed that the key pair was created specifically for the message. In that case it might be hard to establish trust in the public key that is part of the key pair. Without trust in the public key verification of signatures is meaningless.
