set -e is suppressed inside if/&&/|| conditions and everything they call,
including the whole body of update_direct_target, which both entry points
invoke as a condition. So most failures there fell through silently: a
failed git checkout let the merges and the final git reset --hard run on
the previously checked-out branch (corrupting it, in main's loop over
several child PRs), and a failed git commit-tree let the function return
success, after which main retargeted the PR onto a base its head does not
contain. A failed conflict comment still added the label, leaving a resume
with no state marker.

Every log_cmd call site now states its failure policy: run (log, execute,
die on failure; die's explicit exit aborts from any context) or try (log,
execute, hand the status to the caller) for the merges whose failure is an
expected outcome. set -e stays as a backstop only.

https://claude.ai/code/session_01STkeSJ7cLrmrNn4aTDYkwH

Two helpers swallowed gh errors with `|| echo ""` and returned the negative
answer instead (verified against the previous script version with the new
test scenarios):

- has_sibling_conflicts: a failed `gh pr list` read as "no siblings", so the
  resume deleted the kept parent branch while another conflicted PR may
  still need it for its resolution.
- pr_has_conflict_label: a failed `gh pr view` read as "no label", ending
  the run green without resuming anything.

Both now die on command failure, so the branch and label survive and the
next push retries. read_state_marker gets the same explicit handling; there
the gh failure already aborted the run (pipefail propagated it to an
errexit-active caller), but silently, with the error message hidden by
2>/dev/null.

Also reorders main() to push the updated heads before retargeting their
PRs, matching the conflict-resolved path: if the run fails between the two
steps, a retargeted PR with a stale head sits conflicting with its new
base, GitHub suppresses its CI, and no label exists to resume it.

https://claude.ai/code/session_01STkeSJ7cLrmrNn4aTDYkwH

Re-applied onto the restructured script: the merge --abort guard and the
origin/OLD_BASE existence check (now via the PR-number abandon_resume).
The regression test becomes scenario F, adapted to the PR_BASE payload.

https://claude.ai/code/session_01STkeSJ7cLrmrNn4aTDYkwH

See https://github.com/scortexio/autorestack-action/actions/runs/27349317276

Trunk meanwhile absorbed most of this PR (#42 reordered main(), #50
hardened the label and comment reads), so what remains is the
has_sibling_conflicts fix plus a new one of the same kind: list_child_prs
failures were swallowed by the process substitutions consuming it, so a
failed listing read as 'no children' and let main() delete the merged
branch under the children it never saw. Callers now capture the output and
die on failure.

https://claude.ai/code/session_01STkeSJ7cLrmrNn4aTDYkwH

https://claude.ai/code/session_01STkeSJ7cLrmrNn4aTDYkwH

…appers

# Conflicts:
#	tests/test_conflict_resolution_resume.sh
#	update-pr-stack.sh

…flation

# Conflicts:
#	tests/mock_gh.sh
#	tests/test_conflict_resolution_resume.sh
#	update-pr-stack.sh

…appers

…flation

See https://github.com/scortexio/autorestack-action/actions/runs/28656754137