It seems pretty obvious to me why deterministic code translation is preferable for something like this.

There is zero chance they’d end up with a functional code base after attempting to do this.

Right until LLM hallucinates something that just happens to compile but do the wrong thing (e.g. writing "-" instead of "+" in a computation). Good luck catching that in 10M lines of code.

More importantly, why would you even do that? Feature-wise Java is pretty much a subset of Kotlin, so it's mostly a syntactic transformation, which we know very well how to do reliably with straightforward code.