This one involves some language-lawyering, which is not my forte.

While developing a constexpr, array-based vector that supports non-default constructible types, I have run into the following inconsistency between clang and gcc. 

https://godbolt.org/z/PYdq5P

```
#include <memory>

struct Foo {
    mutable int n = 0;
};

union U {
    Foo foo;
    constexpr U() {}
};

struct cvector {
    U storage[1];
    constexpr cvector() {
        std::construct_at(&storage[0].foo);
    }
    constexpr ~cvector() {
        std::destroy_at(&storage[0].foo);
    }
};

constexpr cvector v;
```

After some slack discussion that I didn't entirely follow, it seems that http://eel.is/c++draft/expr.const#7.2 combined with http://eel.is/c++draft/expr.const#6.2 may exclude the constexpr use of destroy_at of an instance of a class with mutable members. I am not confident with this analysis, nor am I confident that this is desirable behavior, but am reporting it for consideration.