I'm not sure where you see a disagreement with section 5.2.2/4 of the standard:
"It is implementation-defined whether the lifetime of a parameter ends when the function in which it is defined returns or at the end of the enclosing full-expression."
Regardless of where exactly the lifetime of the temporary ends (end of function or end of expression), your example would be correct and not UB.
I don't think it is in dispute whether or not the return expression inside the function is evaluated before the lifetime of the parameter ends, but that is when the copy of a into r is made.
Where do you see a disagreement with the posted article?
If the lifetime of the temporary ended with the function body, the assignment (which reads the temporary via the returned reference) happens afterwards and that's UB. Essentially,
int const &r = f(3);
and
int r = f(3);
would both be UB under that implementation-defined case.
The posted article, second row of "common mistakes", describes returning a const reference parameter, when called with a prvalue, as blanket UB.
"6.1 A temporary object bound to a reference parameter in a function call persists until the completion of the full-expression containing the call."
So, while the parameter's (the reference "int const &a") lifetime is implementation-defined, the temporary to which that parameter refers to is the full expression of the call. Sanity is restored.
const int& f(int parameter) { return parameter; }
const int i = f(42);
Is strictly undefined behavior under the simpler C++03 rules/wording of 5.2.2/4: "The lifetime of a parameter ends when the function in which it is defined returns.")
P0135R0 would loosen the requirements on parameter's lifetime as a means of enabling guaranteed copy elision, in such a way that an implementation could technically decide to make this to have defined behavior (not that I suggest relying on it!)
"It is implementation-defined whether the lifetime of a parameter ends when the function in which it is defined returns or at the end of the enclosing full-expression."
Regardless of where exactly the lifetime of the temporary ends (end of function or end of expression), your example would be correct and not UB.
I don't think it is in dispute whether or not the return expression inside the function is evaluated before the lifetime of the parameter ends, but that is when the copy of a into r is made.
Where do you see a disagreement with the posted article?