That may indeed have been an assumption of the original architecture, but it's orthogonal to the end-to-end argument in Internet design, which is about moving logic out of the network entirely and into applications (more precisely, about recognizing that the boundary between network and application is productively debatable, and had, up to the point where Saltzer and Clark and Reed wrote the paper, been defaulting too much towards the network). An end-to-end-architected networking application can be oblivious to its addressing, or even the network layer below it.
If anything, my intuition is that the unreasonable effectiveness of CGNAT --- which is exactly what Huston is writing about --- is strong evidence that the end-to-end paper was deeply correct.