[expr.new] p19.2 makes it clear that we perform direct-initialization, which means the copy constructor of T is not called in a case like this:

struct T { };
T f();
T * p = new T(f());

However, p21 then talks about "the constructor" in the class object case, which could lead to the misunderstanding that some constructor is always called. This is amplified because allocation/deallocation functions are mentioned in the same sentence, and accessibility to those is always checked, even if the allocation is elided.
Suggestion: add "(if any)" after "the constructor". (The "no constructor" case also arises for aggregate initialization, it seems.)
Do we have a good cross-reference target when we want to talk about "mandatory copy elision"?