Like pretty much every unsupervised learning problem, there is no data which can tell you what N should be. If there was, you could trivially use that measure to convert it to a supervised learning problem.

More generally, all of those metrics are pretty fluffy (the scree plot is probably the best, as it doesn't give you a p-value).

In order to actually do this right, you need multiple studies and CFA, but the standard in the field is to refit your exploratory model as a CFA on the original dataset, which is almost certainly causing overfitting in most cases.