STDEXEC_ENABLE_EXTRA_TYPE_CHECKING no longer checks senders in connect (since #1809) #2407

With STDEXEC_ENABLE_EXTRA_TYPE_CHECKING, connect used to instantiate completion_signatures_of_t for the sender it connects, so __debug_sender warned when a sender completes with something it did not declare. That came from #1070 (for #1065). The new connect from #1809 no longer does this, so now only code that names completion_signatures_of_t gets the check.

Example: a sender that declares set_value_t(int) but completes with set_value(std::move(rcvr), v), where v is an int lvalue. Compiled with -I<stdexec>/include -DSTDEXEC_ENABLE_EXTRA_TYPE_CHECKING, with GCC 12 and Clang 21:

I tried to put the check back into connect_t::operator(), and found two more things:

• With the flag on, that breaks test/exec/test_any_sender.cpp with GCC 12 and Clang 21. when_any inside an any_sender then gets connected to __debug::__receiver, and static_assert(__nothrow_move_constructible<_Env>) in __fwd (__env.hpp) fails for the type-erased receiver environment. Main compiles that file with the flag on.
• Since ca3797c, the CMake target adds the include directory as SYSTEM. GCC and Clang don't report warnings from system headers, so with -isystem instead of -I none of the cases above warn, also before a new implementation for connect that improves diagnostics on failure #1809. So CMake users don't see these warnings at all.

Is the flag still meant to work? If yes, I'd be happy to make a PR that restores the check in connect, makes it work with any_sender, and adds a build-failure test. For the SYSTEM part, one option is to not mark the include directory SYSTEM when STDEXEC_ENABLE_EXTRA_TYPE_CHECKING is ON. If you'd rather drop the flag, that's fine too.

Claude (Anthropic) was used as assistant.