But ... surely Sutter ought to know better than to say "because the hardware handles this conversion reasonably, it's a benign case of UB"? Surely he knows that compilers can and will optimize based on the assumption that UB never happens?

The problem isn't, "oh no what if my CPU's float->int conversion instruction traps", that's an extremely naive way to think about UB. Everyone who has thought seriously about UB in C++ for any length of time knows this. It's worrying that this was Sutter's response.