You can never assume 100% scaling with FLOP changes only, because there are other equally important factors, like fillrate and memory bandwidth.
I didn't. The scaling factors I used throughout are less than 100% precisely because of other things not scaling. In other words, the bottlenecks you mentioned are already built into my calculations.
6EU GT1 is only 37.5% FLOPs of the 16EU GT2, but much better 50% of fillrate.