Those questions are out of the realm of science at this point. Just because some people would rather not have to face such questions doesn't mean scientists shouldn't follow the evidence wherever it leads.
It uses a fitness function - a means of determining how well the solution fits the target goal. In biological terms, that would be how well the individual survives and thrives. In mathematical terms, it's an overall fitness score based upon what you want.
So if the target is mobility and you get useless mutations that will only be useful when the final protein is in place many generations in the future, how does the algorithm know to save these mutations? In irreducibly complex systems, they don't function if all the pieces are not in place. In order for the algorithm to be able to produce an irreducibly complex system, it needs foresight to save the pieces that could be useful in the future.
No, you don’t need foreknowledge; the GA doesn’t retain that information. Mutations happen, and often the same mutation will pop up every few dozen generations. The key is the sample size you test - literally millions (or in the case of biological evolution, trillions of trillions) of solutions.