`<thread>`: `promise::set_value_at_thread_exit` sets the value too early

Describe the bug
from https://en.cppreference.com/w/cpp/thread/promise/set_value_at_thread_exit

Stores the value into the shared state without making the state ready immediately. The state is made ready when the current thread exits, after all variables with thread-local storage duration have been destroyed.

but this is not true for our implementation.

This is probably a known issue because I found a comment in the implementation // TRANSITION, ABI but I didn't find a github issue about it so I decided to create it.

Command-line test case

C:\Temp>type main.cpp
#include<iostream>
#include<future>
using namespace std;

struct foo {
	~foo() { cout << "~foo() "; }
};

int main() {
	promise<void> prom;
	future<void> f = prom.get_future();
	thread t([&prom] {
		thread_local foo foo;
		prom.set_value_at_thread_exit();
	});
	f.get();
	cout << "gotten ";
	t.join();
}

C:\Temp>cl /EHsc /W4 /WX .\main.cpp
Microsoft (R) C/C++ Optimizing Compiler Version 19.36.32502 for x64
Copyright (C) Microsoft Corporation.  All rights reserved.

main.cpp
Microsoft (R) Incremental Linker Version 14.36.32502.0
Copyright (C) Microsoft Corporation.  All rights reserved.

/out:main.exe
main.obj

C:\Temp>main.exe
gotten ~foo()

Expected behavior
The output should be “~foo() gotten”

STL version

    Microsoft Visual Studio Community 2022 Preview
    Version 17.6.0 Preview 2.0

Additional context
DevCom-10322768 and internal VSO-1779407 / AB#1779407

I played with the example a bit and I think this line is wrong:

https://github.com/microsoft/STL/blob/9231abe46d466a0f262db234fd3bd4de1170ee45/stl/inc/thread#L56

I added more details to Devcom-10322768 as my comment

If we could use thread_local in STL we could add a thread_local struct and call _Cnd_do_broadcast_at_thread_exit in its destructor...

vNext note: Resolving this issue will require breaking binary compatibility. We won't be able to accept pull requests for this issue until the vNext branch is available. See https://github.com/microsoft/STL/issues/169 for more information.

Start with stl/inc/thread at the referenced line around 56, then run the command-line test case from the issue to observe the ordering of the output. The fix is complete when set_value_at_thread_exit makes the future ready only after thread-local destruction, while accounting for the stated vNext ABI-compatibility constraint.