I do downvote people who appear to be completely mind-killed

I think your mindkill detection algorithms need some tuning; they have both false positives and false negatives.

Rather [...] with reasonable running time

I know of no credible way to do it with unreasonable running time either. (Unless you count saying "AIXI can solve any solvable problem, in principle, so use AIXI", but I see no reason to think that this leads you to a solution with low Kolmogorov complexity.)

I don't think your argument from superintelligent AI works; exactly where it fails depends on some details you haven't specified, but the trouble is some combination of the following.

• For your first premise to be uncontroversial around here, I think you need to either take it as applying only to the form of the laws of physics and not to initial conditions, arbitrary constants, etc. (in which case you can't identify "this universe" and still have it be of low complexity) or adopt something like Tegmark's MUH that amounts to running every version of the universe (all boundary conditions, all values for the constants, etc.) in parallel (in which case what gets taken over by a superintelligent AI is no longer the whole thing but a possibly-tiny part, and specifying that part costs a lot of complexity).
• You need to say where in the universe the AGI is, which imposes a large complexity cost -- unless ...
• ... unless you are depending on it taking over the whole universe so that you can just point at the whole caboodle and say "that thing" -- but then presumably its agent-detection facilities are a tiny part of the whole (not necessarily a spatially localized part, of course), and singling those out so you can say "agents are things that that identifies as agents" again has a large complexity cost from locating them.