Frankly, if you have to ask, just use a (system) CSPRNG.
People massively overblow the impact of the speed difference for most scenarios.
Yes, A simulation running on a large scale etc. will probably need a statistically good RNG w/o any security properties but (say) a game generating a seed occasionally will not be bottlenecked by using a CSPRNG. I would say it is worth it just to not have any mental load and slightest change of misuse.
Also, if the salt part is for salting and hashing passwords, forget about the whole idea and use a proper password hash be it Argon2id, scrypt, PBKDF2, whatever. It doesn't really matter which one and ideally a library should have chosen one of the algorithms with good parameters and nonce generation.
(I know Argon2 calls its nonce a salt too but that is irrelevant. It should come from a CSPRNG)
People massively overblow the impact of the speed difference for most scenarios. Yes, A simulation running on a large scale etc. will probably need a statistically good RNG w/o any security properties but (say) a game generating a seed occasionally will not be bottlenecked by using a CSPRNG. I would say it is worth it just to not have any mental load and slightest change of misuse.
Also, if the salt part is for salting and hashing passwords, forget about the whole idea and use a proper password hash be it Argon2id, scrypt, PBKDF2, whatever. It doesn't really matter which one and ideally a library should have chosen one of the algorithms with good parameters and nonce generation.
(I know Argon2 calls its nonce a salt too but that is irrelevant. It should come from a CSPRNG)