claims on how this works

Apple's libc has a pretty wild feature (guarded by _DARWIN_UNLIMITED_SELECT, on by default) which allows fds above FD_SETSIZE. It works by checking the address of the fd_set: if it's within the current thread's stack, then the call will fail, under the assumption that it's a stack-allocated fd_set.
But if the address is NOT within the current thread's stack, select() assumes you know what you're doing and will allow the call, trusting you have allocated sufficient memory for the high fds in the fd_set.

so there may be some condition to fulfill to make it work.

which I think is not satisfied for qemu main loop fds?

complete macos g_poll:

gint
g_poll (GPollFD *fds,
	guint    nfds,
	gint     timeout)
{
  struct timeval tv;
  fd_set rset, wset, xset;
  GPollFD *f;
  int ready;
  int maxfd = 0;

  FD_ZERO (&rset);
  FD_ZERO (&wset);
  FD_ZERO (&xset);

  for (f = fds; f < &fds[nfds]; ++f)
    if (f->fd >= 0)
      {
	if (f->events & G_IO_IN)
	  FD_SET (f->fd, &rset);
	if (f->events & G_IO_OUT)
	  FD_SET (f->fd, &wset);
	if (f->events & G_IO_PRI)
	  FD_SET (f->fd, &xset);
	if (f->fd > maxfd && (f->events & (G_IO_IN|G_IO_OUT|G_IO_PRI)))
	  maxfd = f->fd;
      }

  tv.tv_sec = timeout / 1000;
  tv.tv_usec = (timeout % 1000) * 1000;

  ready = select (maxfd + 1, &rset, &wset, &xset,
		  timeout == -1 ? NULL : &tv);
  if (ready > 0)
    for (f = fds; f < &fds[nfds]; ++f)
      {
	f->revents = 0;
	if (f->fd >= 0)
	  {
	    if (FD_ISSET (f->fd, &rset))
	      f->revents |= G_IO_IN;
	    if (FD_ISSET (f->fd, &wset))
	      f->revents |= G_IO_OUT;
	    if (FD_ISSET (f->fd, &xset))
	      f->revents |= G_IO_PRI;
	  }
      }

  return ready;
}