Let a host that needs no launcher run its own smoke test

[frangarcj]

Every macOS leg of the executor has been dying since the Rosetta gate landed, after a complete and valid build. From the arm64 job of vitasdk/autobuilds#36:

21:55:13  self-contained headers OK (73 checked)
21:55:18  Vita GCC/binutils contract OK
21:55:18  [100%] Built target check-toolchain-contract
21:55:18  scripts/ci/build-host.sh: line 160: run[@]: unbound variable
21:55:19  ##[error]No files were found with the provided path: out/*. No artifacts will be uploaded.

The SDK is finalized, the static validation passes, the toolchain contract passes, and then the script exits and stages nothing. The Intel host fails behind it for the reason it should — no build-machine SDK found for x86_64-apple-darwin (needs arm64-apple-darwin) — because the arm64 artifact it cross-builds from was never uploaded. arm64-apple-darwin is a required host, so no snapshot can publish at all; the last one that did, sdk-snapshot-20260825.612.1, was built from 79a92784, the commit before this landed.

smoke_test_bootstrap's launcher prefix is an array that is empty for every host that runs its own binaries — the comment above it says so — and macOS's /bin/bash is 3.2:

$ /bin/bash --version | head -1
GNU bash, version 3.2.57(1)-release (arm64-apple-darwin25)
$ /bin/bash -c 'set -u; run=(); "${run[@]}" echo hola'
/bin/bash: run[@]: unbound variable
$ /bin/bash -c 'set -u; run=(); ${run[@]+"${run[@]}"} echo hola'
hola
$ /bin/bash -c 'set -u; run=(arch -x86_64); ${run[@]+"${run[@]}"} uname -m'
x86_64

build_and_stage() already carries the tolerant form, and the reason for it, three lines below the break.

The second commit is separable — take only the first if you would rather. This is the second time the same class has hit this file: fixed once in 9233c494a, reintroduced by the launcher. Deciding case by case which array can be empty is the judgement that failed, so every expansion in the scripts that run on a published host takes the tolerant form, and a test refuses any that do not. Run against the parent of this branch it names lines 161–165 of build-host.sh — the ones that broke the nightly:

$ tests/ci/test-empty-array-expansion.sh
scripts/ci/build-host.sh:161:  "${run[@]}" build/vitasdk/share/vdpm/bootstrap-vitasdk.sh --install-dir "$install_root"
scripts/ci/build-host.sh:162:  VITASDK="$install_root" "${run[@]}" "$install_root/bin/vdpm" --help >/dev/null
...

The test is static because it has to be: BASH_COMPAT=32 does not restore the 3.2 behaviour, so nothing on a Linux runner reproduces this by executing. The rewrite is mechanical and means the same thing whenever the array has elements; ${#name[@]} and ${!name[@]} are fine in 3.2 and are left alone. tests/ is outside the rule — it runs on ubuntu only.

Every suite in the checks job passes locally, and both bashes parse every rewritten script (bash -n, /bin/bash -n). tests/package/test-vdpm-bundle.sh fails here with and without this change: it wants a stage-1 SDK, and it is not in checks.

---

AI tools were used in preparing this PR (Claude Opus 5, Anthropic).