seed steps are killed by SIGPIPE on their first byte of output

[luthermonson]

run_seed spawns each seed: command with both stdout and stderr set to Stdio::piped() and then calls .status():

// src/deployer.rs
let status = Command::new("sh")
    .args(["-c", cmd])
    .current_dir(&workdir)
    .env("PREVIEW_URL", preview_url)
    .env("PREVIEW_HOST", hostname)
    .env("PR", pr_number.to_string())
    .stdout(Stdio::piped())
    .stderr(Stdio::piped())
    .status()
    .await;

tokio's Command::status() drops the parent's stdin/stdout/stderr handles before awaiting the child, so both pipes' read ends close immediately. The first byte the seed step writes to fd 1 or 2 therefore raises SIGPIPE and kills it.

The failure is invisible: the step dies in single-digit milliseconds with exit 141, no output is captured anywhere (the pipes were dropped), and the only trace is

WARN switchboard::deployer: seed step failed — continuing step=2 command=...

Reproduced on a preview node against ephpm/wordpress-sample:

$ cd /srv/ephpm/sites/ephpm-wordpress-sample-pr-4
$ sh -c "DOCROOT=. HOST=x bash seed/install.sh" | true ; echo ${PIPESTATUS[0]}
141
$ sh -c "DOCROOT=. HOST=x bash seed/install.sh" >/dev/null 2>&1 ; echo $?
2      # runs normally, fails later for its own unrelated reason

That step has failed on every deploy of PRs #1, #3 and #4 of wordpress-sample, always in 5-11 ms, always before it reaches the network — it dies on its first echo. It looked like an application bug for weeks.

Build steps are only half affected: run_build uses stdout(Stdio::null()) (safe) but stderr(Stdio::piped()) with the same .status(), so a build step that writes to stderr dies the same way.

Suggested fix

Use .output().await and log the captured stdout/stderr on failure — which is what the piping presumably intended in the first place, and would have made this self-diagnosing. Alternatively Stdio::null() for both if the output genuinely is not wanted.

The captured output would be valuable on its own: today a failing seed step gives an operator nothing but the command string.

Workaround

Apps must redirect every seed step's output, e.g. - "bash seed/foo.sh >> .seed.log 2>&1". Done in ephpm/wordpress-sample#5.