I imagine there are technical limitations leading to economic reasons for these kind of mistakes. So much data, sitting in so many massive silos, that they must design their systems on the basis of peoples' access patterns only hitting the most recent subset.
That, and the opposite economic reason - that you can charge for the older/richer/more complete dataset.
(EDIT: I know it's odd that I was whinging about the same thing a few posts up. I don't really know what the resolution is for that.)
I think the true reason is really just pervasive ignorance — everybody royally fucks this up and doesn't question it for a second.
If you have N items with M on each page, and the same request for 'page 2' always returns items N-M through N-2M, with the contents shuffling off the end as N increases, you're an abject failure. SELECT … LIMIT M OFFSET N-(P-1)•M is in almost every single web app and totally bullshit. It's incredibly depressing, but we'll probably be stuck with it for at least the rest of our lifetimes.
Nearly everyone makes the fundamental pagination mistake (Blogger being the sole exception), but calendar-based archives are a basic assumption that everybody implements. Facebook doesn't expose it in their interface but it's possible through the API. For Twitter it's completely fucked — the only people with access to your old tweets are the Library of Congress.