Suggested resolution:

Change in 7.7 [expr.const] bullet 10.18.2 as follows:

An expression E is a core constant expression unless the evaluation of E, following the rules of the abstract machine (6.9.1 [intro.execution]), would evaluate one of the following:

• ...
• a new-expression (7.6.2.8 [expr.new]), unless either
  • the selected allocation function is a replaceable global allocation function (17.6.3.2 [new.delete.single], 17.6.3.3 [new.delete.array]) and the allocated storage is deallocated within the evaluation of E, or
  • the selected allocation function is a non-allocating form (17.6.3.4 [new.delete.placement]) with an allocated type T, where
    • the placement argument to the new-expression points to an object whose type is similar to T (7.3.6 [conv.qual]) that is transparently replaceable (6.7.4 [basic.life]) by the object created by the new-expression or, if T is an array type, to the first element of such an object of a type similar to T, and
    • the placement argument points to storage whose duration began within the evaluation of E;
• ...

Consider:

  struct A { int n; };
  struct B : A { int m; };

  constexpr int f() {
   B b = {{0}, 0};
   A *p = &b;
   new (p) A{1};     // does not transparently replace *p
   return p->n;      // UB, p refers to out-of-lifetime object
  }
  constexpr int k = f();

It seems unreasonable to expect implementations to track situations where a placement-new evaluations that don't transparently replace the original object.

The status quo wording in 7.7 [expr.const] bullet 10.18.2 involving "similar" was added by issue 2922.