AtlasIntel released a poll which predicted the actual lead: https://projects.fivethirtyeight.com/polls/20241102_SwingSta...
So, “muuuch farther than the projections of ANY polls” is wrong.
Correlated error between states is expected, and not relevant (ie. if you think Mangione poll is affected by this few % correlated error, again it doesn’t change the interpretation).
Overall, election results tend to come pretty close to the polls. They are often wrong by a little bit, occasionally by a lot, which is what you’d expect based on the theory of polling and statistics. The conclusion is asking a large random sample of people what they think, using the best methodologies the polling industry has developed, is a pretty good way of finding out what most people think (within a few % margin of error), certainly more objective than a vibe check of your personal echo chamber. If your argument depends on completely dismissing polling results, you’re probably wrong.
To be clear, that's not the MARGIN of error in those statistics, but rather the difference between predicted and actual. In a five-tier scale, separated between groups, there is already statistical manipulation by having fewer respondents per-category, and therefore a higher margin of error between each. A margin of error even in low single-digit percentages can completely compromise the outcomes of data presented in this fashion.
He ended up winning every single swing state, muuuch farther than the projections of ANY polls that was getting time-of-day.
Even if you want to die on that purposely misinterpreting hill, it still shows they were inaccurate and unreliable.