OK, this is my second attempt to play a language lawer. It seems like
there is a bug in the set<T> specifications.

Something like this should be illegal:

set<int> s;
s.insert(1);
s.begin() = 0;

But the current DWP seems to mandate that it's legal.

1. 23.1 table 66 says that for any non-const value a of the
   type "container X<T>", a.begin() returns iterator
   (not const_iterator).

2. 23.3.3 says that the type of set<T>::iterator is
   implementation defined.

3. 23.1.2 table 70 says that assosiative container is a container.

4. 23.1 table 66 says that for any container X<T>, X<T>::iterator
   is an iterator type, pointing to T, where iterator is any
   iterator category, except output iterator. 23.1.2.6 further
   says that an iterator for an associative container is of a
   bidirectional iterator category.

5. 24.1.3 table 75 says that if a is forward iterator, then *a
   has a type T& (not const T&). Same is true for bidirectional
   and random access iterators, because they are also forward
   iterators (tables 76, 77)

6. 23.1 table 66 says that T must be assignable.