Can interactive zero-knowledge proof systems be implemented using secure two-party computation?

I am defining multi-party computation using the real-ideal paradigm (see A Pragmatic Introduction to Secure Multi-Party Computation). That is, for any successful attack on an MPC protocol in the real world, there exists a simulator that carries out this attack successfully in the ideal world. It follows that security in the real world must be equivalent to security in the ideal world.

I am defining interactive zero-knowledge proof systems for a language $L$ using the original definition from The Knowledge Complexity of Interactive Proof Systems. That is, a pair $(A, B)$ of interactive Turing machines must fulfill

1. Completeness: given $x \in L$, $B$ accepts with very high probability;
2. Soundness: given any prover $A'$ and $x \not\in L$ passed into $(A', B)$, $B$ accepts with very low probability;
3. Zero-Knowledge: there exists a probabilistic polynomial-time simulator that can simulate the entire exchange of messages between $A$ and $B$ for any input $x \in L$.

Now, the paper Zero-Knowledge from Secure Multiparty Computation mentions the following:

Zero-knowledge protocols can be viewed as a special case of secure two-party computation, where the function verifies the validity of a witness held by the prover.

That is, given $L \in \mathcal{NP}$, there exists an algorithm $A$ such that $x \in L \iff \exists w\colon A(x, w) = 1$ (definition of $\mathcal{NP}$). One party $P_1$ acts as the prover, another $P_2$ as the verifier. $P_1$ knows $x$ and $w$, $P_2$ knows only $x$. They execute $A(x, w)$ together to determine whether $x \in L$ or not.

Clearly, $w$ is not revealed to the verifier $P_2$ due to the MPC protocol. However, is the definition of zero-knowledge not more general? If the prover $P_1$ sent, for some reason, the solution to some instance of an $\mathcal{NP}$-complete problem¹, no polynomial-time simulator could simulate this assuming $\mathcal{P} \neq \mathcal{NP}$. The created proof system would not be zero-knowledge.

So, given that an MPC protocol could exchange non-simulatable messages, an MPC protocol cannot actually be used to implement a zero-knowledge proof system for some language $L \in \mathcal{NP}$, can it?

¹ The solution can be made dependent on $x$ such that it is not constant and thus easily simulatable.