It's not just about what status you have, but what you actually are. You can view it as analogous to the Newcomb problem, where the predictor/Omega is able to model you accurately enough to predict if you're going to take one or two boxes, and there's no way to fool him into believing you'll take one and then take both. Similarly, your behavior in one situation makes it possible to predict your behavior in other situations, at least with high statistical accuracy, and humans actually have some Omega-like abilities in this regard. If you kill the fat man, this predicts with high probability that you will be non-cooperative and threatening in other situations. This is maybe not necessarily true in the space of all possible minds, but it is true in the space of human minds -- and it's this constraint that gives humans these limited Omega-like abilities for predicting each others' behavior.

(Of course, in real life this is further complicated by all sorts of higher-order strategies that humans employ to outsmart each other, both consciously and unconsciously. But when it comes to the fundamental issues like the conditions under which deadly violence is expected, things are usually simple and clear.)

And while these constraints may seem like evolutionary baggage that we'd best get rid of somehow, it must be recognized that they are essential for human cooperation. When dealing with a typical person, you can be confident that they'll be cooperative and non-threatening only because you know that their mind is somewhere within the human mind-space, which means that as long as there are no red flags, cooperative and non-threatening behavior according to the usual folk-ethics is highly probable. All human social organization rests on this ability, and if humans are to self-modify into something very different, like utility-maximizers of some sort, this is a fundamental problem that must be addressed first.