I'm not sure if this is all correct. What you're describing doesn't exactly sound like Solomonoff induction, but you do seem to have a grasp of the principles involved.

Solomonoff induction does not discard any program that is consistent with the observation so far. But for any observation string there are an infinite number of programs that produce that string. There is a shortest one, an infinite class of that program prefixed by some whole number of no-operations (computations that undo themselves). And compilers implementing that same program in encodings of other programming languages. And interpreters implementing that same program in encodings of other programming languages. And entire universes containing people who happen to be simulating one of these (which may be considered an unreliable type of interpreter). And arbitrary nestings of any of the above any number of times. None of this is discarded. But again, the set is infinite, so no matter what distribution you choose, the probabilities after some point must decrease for it to converge.

The point about the witch isn't that "witch" is a complex cost to encode in a program (although it is), but that "she did it" fails to compress the data at all, because you still have to encode what the pronoun "it" is referring to. Because a "witch" can be blamed for literally anything, adding a "witch" to the uncompressed hypothesis "it" adds no predictive power whatsoever. (If you can compress "it" some other way, then you can make predictions without the witch and she is useless to your model.)

God, who can likewise be credited for anything (even what looks like evil--"all part of God's plan", or "God works in mysterious ways", right?) is the same as the witch: no predictive power over "it" whatsoever. And worse, God's complexity cost is not just relatively big like any intelligent mind (such as the witch) would be, but literally infinite if we say that God is omniscient: If God is a "halting oracle", then God is not even contained in the set of all computer programs, because He is not computable: He can't even be a hypothesis, only approximated. And to get a better approximation, you must use a longer computer program that encodes more of Chaitin's constant, which is provably not compressible by any halting program. Better approximations of God get bigger without limit. The approximate God hypothesis has literally infinitesimal probability--you can't escape it: The better the approximation gets, the less likely it is.

And the true God hypothesis is not even in the running. It literally cannot be proved by induction at all. Nor can you take God as an axiom. (I will dismiss it as the fallacy of special pleading: applying this privilege lets us prove anything, even false gods.) The only hope then is proving deductively from some logical necessity, or giving up on omniscience as defined, which of course, opens the possibility of there being beings greater than whichever God you choose.