Very good point! Thanks!
With the correction, it interestingly enough produces the good behavior also at size=2. It also delays SIMD until size=5. But then it bizarrely stops doing SIMD again after size=64.
https://godbolt.org/z/P979nY4nf
The iterator version stays SIMD-y also after size=64, but stops at some point. What?! I don't know enough to understand what's going on. Anyone?