Test that file descriptor state is clean (Overhaul file descriptor handling)

[andychu]

While investigating a bug that turned out to be related to issue #202, I found two things

• The child process file descriptor state isn't clean. See test/spec.sh command_. bash and dash do one thing, mksh and zsh do another (they have /dev/tty open), but OSH is even worse. There are pipes open for some reason.
• Instead of my own linear search in _GetFreeDescriptor() in core/process.py, I should be using fcntl(..., F_DUPFD, ...) and looking at the return value, like dash does.

[andychu]

NOTE: I just added a spec test case that I'm confused by.

OSH doesn't handle assignment builtins differently than other builtins. But it agrees a lot with zsh, but disagrees with the other shells.

Is this an issue of "change fd state state then fork" vs. "fork then change fd state" ?? Not sure why it would be different.

Maybe check smoosh?

$ test/spec.sh assign --range 37
assign.test.sh
case    line    dash    bash    mksh    zsh     osh
 37     589     ok      ok      ok      pass    pass    redirect after assignment builtin (what's going on with dash/bash/mksh here?)
2 passed, 3 ok, 0 known unimplemented, 0 known bugs, 0 failed, 39 skipped

[andychu]

Yeah there is a bug. Two things to do while fixing this:

• Fix test cases 37 and 38 in spec/assign (related to redirects)
• Test the file descriptor state (where did that go? I don't see it in spec/command_)
  • I guess this is demo/fd-state.sh

[andychu]

#653 mentions several file descriptor constructs not implemented

• <>
• 2>&1-
• {fd}>

[andychu]

OK I did a big refactor that is most of this. We just need some more tests. I started some strace tests in demo/strace-redirects.sh