Solomonoff induction provably converges on reproducing any input function by only observing that function's outputs, and it formalizes Occam's razor as a key principle. I don't think it's fair to call it a heuristic anymore.

Occam's razor is formalized in Solomonoff induction but that doesn't mean that it is not still a useful heuristic in Science more generally.

Sure you can use the simpler version heuristically, but the OP's claim is that it is only a heuristic and not a valid logical inference. I'm saying we now have proof that it is logically valid to claim that one should prefer theories with fewer assumptions, all else being equal.