I'm pretty sure it's easier to sort by -timestamp than to run a complex ranking model across all recent tweets (which would have to be windowed by timestamp anyway), and then sort by the output of that.

Seems to me like they are going out of their way to force their "top tweets" algorithm on users.

It's a global platform with billions of users, with many users follow hundreds of people, some users followed by millions, huge volumes of media tweets, broad distribution of posting frequency... even with those basic scaling concerns, I don't see how consistent reverse chronology could be implemented as naively as "yo just sort by -timestamp".

I would bet "top tweets" improves engagement with the passive consumption audience while saving money by abandoning consistency and improving caching.