In Standard ML the problem is decidable, it's just that the worst case is that the derived type is exponential in the length of the program (as a directed acyclic graph, and double exponential as linear output).

So one needs a bit less than Turing oracle to fix the compiler, at least for Standard ML ;-) Some other type systems are actually undecidable.

But at the same time, I am not sure that Turing oracle by itself can be used to reduce the complexity of the computation involved to make it fit a real computer -- I am just not familiar enough with the theory of computations with such oracles.. For example, if one uses such an oracle in a straightforward way, as applied to this problem, the oracle always says "yes" (a large enough computer will always complete type inference of a Standard ML program given enough time), and this does not help us to reduce the resources needed to figure out the outcome of the type inference process.