Moravec's paradox doesn't actually tell us very much. All it tells us is that our intuitions about the relative difficulty of specific problems were incorrect. This is not very surprising - we know extremely little about how intelligence works, and we have a long history of underestimating / overestimating the difficulty of various problems which goes far back into the history of mathematics. Our brains evolved to solve some pretty specific tasks, which made some tasks (like playing Chess) seem very difficult although they could easily be solved by a computer program, because we simply had no need for Chess playing abilities in order to survive. This doesn't mean that the low-level tasks aren't difficult in an absolute sense, it just means we shouldn't extrapolate this observation to other tasks we currently consider very difficult. Note that computer-vision was usually thought to be one of the low-level tasks AI researchers thought was too difficult to be solved very soon, and then deep-learning changed that pretty quickly.

The way that I read your argument (and I may be reading it wrong) is that we shouldn't expect all blue-collar jobs to be taken by robots before all white-color jobs. This assertion I think is probably true in the specific way it is stated, but I think you imply that we won't need to worry about human jobs eventually only being available to a cognitive elite, where people who have lower cognitive ability find themselves unemployed and their jobs being automated out. This claim does not follow from the argument you gave, except in the specific case in which AI tends to replace jobs in an evenly-distributed way across the intelligence spectrum, or in the case in which we find other economically useful jobs for humans as quickly as AI replaces them. Both of those situations I think are unlikely.