I think you are underestimating the spiral effect. Working with a documented and structured code makes it easier to understand and reason with even for LLMs. At the end of the day its working with patterns, grammar, trying to deduce and implement intent. Its possible where we enter the era of blackbox applications where no one knows the internal, and yet for the required inputs it provides the expected outputs. Every failure would re-trigger the evaluation of the whole application to accomodate that one failure in requirement.