> remains O(n^2), even in real life.

This is something of a grandiose claim. How are you thinking of a physical insertion sort working? If you don't have a model, you can't say anything about the time requirements.

But note that if we conceptualize insertion sort like this:

    loop:
      pick up a book
      find the place within the sorted books where this book belongs
      open a space for the book
      insert the book

the four steps in that model are O(1), O(log n), O(1), and O(1), and the loop repeats n times, so we have an upper bound of O(n log n).

The reason insertion sort is O(n^2) when operating on an array is that step 3, "open a space for the book in our hand" is O(n) in that case, because we can only move one book at a time.