The gotcha here is that if we get ahead of ourselves, and read to
the end of the stream, we detect bad chains too early, so instead
incrementally build the page chain cache, which is also quicker
and behaves more similarly to the previous code.