This isn't the case with SNARKs, they are zero-knowledge no matter what. For a generic zero-knowledge proof, what you point out is completely true. You are not guaranteed zero-knowledge if the CRS is maliciously generated.
For SNARKS, the proofs become zero-knowledge simply by the prover adding in randomness to his proof. This holds even if the verifier generates the CRS maliciously. Indeed, the original use of SNARKS was for verifiable computation where this was exactly what was done. See section 1.2,3.2, and 4.4 of [0]. In particular, note the last paragraph of 4.4 where the verifier is generating the CRS.
(The point you raise about the simulator is a complex one. Zero-knowledge is an existential statement that there can exist such a simulator, therefor all proofs even when there clearly isn't one are zero-knowledge. The verifier can create a CRS and not tell anyone the information needed to simulate, but that information still exists and we could still, in theory, construct a simulator using it. Therefor the thing is zero-knowledge in both theory and practice. This is not always the case with all proof systems.)
Maybe illicit CRS can be a problem for some kinds of checks, but I guess you could feed first Linux kernel commits in the 2016 to some randomness extractor to get something that is harder to manipulate freely.
The CRS issue is that off-chain, a malicious Verifier (buyer) may learn about the witness, i.e., either enc key or solution.
The protocol is ZK because there exists a simulator which can produce an indistinguishable setup, and simulate proofs. This does not mean that if you allow any V* to produce a CRS, the simulator can still simulate.
I cannot think of an explicit attack off-hand, but wonder how you deal with this.