The point is that if you try to build an agent based on a theory of general intelligence whose problem space includes (with significant probability mass) intractable problems, then any agent which has quality-of-solution performance guarantees on that problem space will have impractical resource bounds.

I analyze the problem in terms of continuous intelligence metrics, not in terms of binary guarantees. Agents with worse performance rate lower but it doesn't mean you have any guarantees in the picture. As a simplified analogy, think of an exam with a collection of problems and a time-limit for solving each. The overall score is a weighted sum over the scores of individual problems. What happens if we have intractable problems in the set (problems which cannot be solved in the given time limit)? Obviously, no examinee will get non-zero score for those. However this doesn't in any way impair the relative comparison of examinees based on the other problems.

In general I think that what you call "creativity" is some not sort of "magic wand" or "oracle" that came from the sky, bestowed unto us by the god of infinite random trials, rather it is the combined effect of all the heuristic modules in our brain, whose mechanics we can't well access by introspection and verbalize.

I think you have a false dichotomy in there :) My claim is that these heuristic modules are so heuristic that there is no practical way to invent them without copying the answer from the actual brain.

In general, my confidence that there is a hard step is significantly higher than my confidence than it is between chimp and human.