I think I'll go with the simple answer that the crosstabs are wacky until we actually get any voting results.
Also I question the utility of aggregating unreliable crosstabs. Better than nothing, but probably better to just rely on actual polls of those subgroups instead and average those. Like we do have specific polls of young people, black people, etc that are appropriately weighted and sampled.
If it were inflation I think we'd see more topline damage. But toplines haven't been that bad. And we'd see broader change. This is just in a few subgroups. In others Biden is doing better.
I think directionally the changes are real, the magnitudes are too large to be credible. And aren't really evident in robust subsample polls like the Harvard Youth Poll.