Fix test-driver output parsing crash on Ubuntu

[am-on]

When running tests on Ubuntu VMs, the test-driver can crash if the executed command calls sudo and writes to stderr. For example:

vm.execute("sudo bash -c \"echo 'Created foo → bar.\n' >&2 && echo 'foo' \"")

sudo spawns a new TTY for stderr on Ubuntu that gets read by the test-driver's output parsing. However, the test-driver only expects base64-encoded stdout and fails when stderr data is read.

• Normal command (note stderr usees /dev/ttyS0)

> bash -c 'echo $$; ls -l /proc/$$/fd'
lr-x------ 1 root root 64 Feb 18 06:10 0 -> /dev/hvc0
l-wx------ 1 root root 64 Feb 18 06:10 1 -> pipe:[20942]
l-wx------ 1 root root 64 Feb 18 06:10 2 -> /dev/ttyS0
lr-x------ 1 root root 64 Feb 18 06:10 3 -> /proc/664/fd

• Using sudo (note stderr uses /dev/pts/0)

> sudo bash -c 'echo $$; ls -l /proc/$$/fd
lrwx------ 1 root root 64 Feb 18 06:10 0 -> /dev/pts/0
l-wx------ 1 root root 64 Feb 18 06:10 1 -> pipe:[20943]
lrwx------ 1 root root 64 Feb 18 06:10 2 -> /dev/pts/0
lr-x------ 1 root root 64 Feb 18 06:10 3 -> /proc/670/fd

Because stderr is now read along with the stdout, the test-driver's base64 decoding code crashes on unexpected data:

• encoding step,
• receiving the data,
• decoding step.

A minimal reproducible example is provided in
am-on/nix-test-driver-ubuntu-bug.

Related issues:

• Test driver crashes if special characters are present #84
• apt-get install breaks the test runner #5

(Worked on this during Thaiger sprint: Thaigersprint/thaigersprint-2025#1)

[am-on]

There's a PR that could fix the issue on the test-driver side - NixOS/nixpkgs#382260 but it has a downside of hiding the stderr of the executing command from the machine logs.

Any opinions on which side should the issue be fixed?

[picnoir]

Thanks!

I think this is an improvement over the previous behavior.

Let's merge that as it is and improve upon it if we need to.

[picnoir]

Any opinions on which side should the issue be fixed?

Forgot to respond about that. Ideally, we'd fix that upstream I think.

We'll revert this PR if you manage to fix that upstream.