On Windows, ActorBase's layout depends on AETHER_NO_THREADING, so a cooperative libaether and the program built against it disagree #3120

What happens

On Windows, ActorBase has two layouts: one in code compiled with -DAETHER_NO_THREADING and another in code compiled without it. The cooperative libaether (make stdlib EXTRA_CFLAGS=-DAETHER_NO_THREADING, the setup make ci-coop uses) and a program built against it by ae (which does not pass the flag) disagree on where every field after thread sits.

With main-thread mode removed (#3112), every actor program built that way dies at exit with heap corruption (0xC0000374), in coop_free_actor -> free(actor->spsc_queue). The runtime reads spsc_queue 8 bytes before where the program's spawn left it. Before #3112 the same mismatch was there, but main_thread_only's 4 bytes happened to absorb it in padding, so the offsets agreed by luck.

offsetof measured with MinGW GCC 16.2, x86_64:

Why

runtime/utils/aether_thread.h takes its no-threads branch first, whatever the platform. That branch says Windows is "handled by the #elif _WIN32 path below (not here)", but with AETHER_NO_THREADING it is not. MinGW defines none of _POSIX_THREADS, __unix__, __linux__ and the others the branch tests, so Windows falls to the bare-metal typedefs: typedef int pthread_t, int mutexes and conditions. The threaded Windows path defines pthread_t as void* (a HANDLE), the mutex as a CRITICAL_SECTION-shaped struct and the condition as a CONDITION_VARIABLE-shaped one.

On Linux and macOS both builds include <pthread.h>, so the layouts agree there, and ci-coop (Linux only) passes.

Fix

On Windows, the no-threads branch uses the same type definitions as the threaded Windows path, so a struct's layout never depends on the flag. Add a test that compiles every runtime struct a program shares with libaether, with and without the flag, and compares their layouts.