Every long human written math paper has cosmetic mistakes in it. I've even read a 9 page paper where they put a 3 where they should have put a 4. It didn't matter at all for their result (they only needed it up to big O). I am not sure how many of these commenters understand the difference.
I do not have remotely enough expertise or time to evaluate the human proof, but an ignorant person running AI on it and reporting mistakes does not really indicate anything substantial.
Also, there are no 250,000 pages to verify because the prompter never released them. There is only the Lean proof (which the prompter certainly would have checked before releasing).