I went looking through the source code and found that the adjustment window is indeed defined in block time as PastBlocksMax = 12 while nProofOfWorkTargetSpacin

I went looking through the source code and found that the adjustment window is indeed defined in block time as PastBlocksMax = 12 while nProofOfWorkTargetSpacing = 150 (half a minute for each of 5 different pow algorithms), giving an expected adjustment window of 12 * 150 = 1800 seconds. Apparently, the adjustment window, which needs to be way longer than the allowed timestamp drift, is in fact way shorter...