In #4858, I implemented close_on_exec logic matching MRI, but because we do not always use direct native IO streams there are lingering issues. Toward the end of that issue, I discovered that ServerSocketChannel can be unwrapped by our fileno logic, but changing the CLOEXEC bit does not appear to affect it. The fd remains available and open in an exec'ed subprocess.

This needs further investigation.

Note that the properly-unwrapped fileno from JDK socket streams can't be directly used in a subprocess, making this a bit trickier to test. See #5008.