Here's the problem you're not getting here.
Lets say tomorrow the society (lets presume all of them are WASPs) decides to divide themselves between people whose name start with a J (the most common letter for names in US, like James, John, Jennifer, Jessica, Jesse etc) and non-Js. And this J group declares themselves as the 'upper caste' with Js only marrying with Js and non-Js only among non-Js. Also they even live in physical isolation from each other.
Two thousand years later, which group will have higher IQ?
The answer is, you can't tell. Whatever is the starting group IQ of these two groups is, that's exactly what it would be for the latter.
Same goes with your theory about upper castes having higher IQ. Now the place where you might be right is (not that you realize that), if there is an IQ test available only to the J's, and they prefer to marry to the highest IQ person willing to marry them, but the other caste does not have this test available. Now, over generations, one group will be selecting for the IQ but the other won't be able to.
The 'IQ test' could be education (and jobs which require education like priesthood) which was generally only available or restricted to the Brahmins (so not 'all' higher castes). The more 'Vedas' you memorized (which weren't written down for a long period), the more esteemed scholar you were, this is why the have last names like Dvivedi or Dubey ('a scholar who memorized 2 vedas') or Trivedi (a scholar who memorized 3 vedas) or Chaturvedi or Chobey (a scholar who memorized all 4 vedas).
Here, you COULD see that if Brahmins are only allowed to be educated, and IFF they are choosing for those who show a 'higher score' in terms of IQ, then there can be an optimization for IQ.
A control group would be a higher caste which isn't optimizing for education or any other proxy for IQ (like being in successful business is a kind of IQ test, not as strong of an IQ test as being able to memorize vedas, but still, it is).