For transform_view::iterator, iter_move is specified to operate on the transformed value but iter_swap is specified to operate on the underlying iterator.

Consider the following test case:

struct X { int x; int y; };
std::vector<X> v = {...};
auto t = v | views::transform(&X::x);
ranges::sort(t);

iter_swap on t's iterators would swap the whole X, including the y part, but iter_move will only move the x data member and leave the y part intact. Meanwhile, ranges::sort can use both iter_move and iter_swap, and does so in at least one implementation. The mixed behavior means that we get neither "sort Xs by their x data member" (as ranges::sort(v, {}, &X::x) would do), nor "sort the x data member of these Xs and leave the rest unchanged", as one might expect, but instead some arbitrary permutation of y. This seems like a questionable state of affairs.