> Kolmogorov complexity of a given program.
You seem fundamentally confused about the objects of study of information theory. They're not programs, they're e.g. strings of symbols. We measure by the information content of those strings based on likelihood / programs. Information theory asks "given some bit sequence A, how much information is in it?" not "is it an optimal program?" - instead we measure the information in it by constructing or otherwise proving facts about programs that generate or predict it. We talk about Kolmogorov complexity of strings (/ signals / states / whatever) as measured by programs, not Kolmogorov complexity of programs themselves.
Obviously programs are also themselves representable strings of symbols, and this is why we find the usual suspects of self-reference paradoxes in IT. But that doesn't mean the measure does not exist, or that it's not possible to find in lots of interesting, easily-computable cases. It's a bit like handing me a ruler and asking me how long it is - sure, if I don't trust any ruler I'll have a hard time measuring it. But I don't have to trust that specific ruler to do so, and the fact it's a device used for measuring itself is completely incidental to my measuring of it.