feat(shipyard): configure through shipyard-cmake; add the canonical msc.sh

[schmonz]

Do not merge until shipyard has landed bc3082a.

After the cutover the CMake user package registry is neither written nor read; only shipyard-cmake (from the shipyard pkg, at /usr/local/bin/) may configure against shipyard.

What changed

12 cmake/ctest calls in the two workflows → shipyard-cmake / shipyard-ctest. container-tools configures three separate CMake projects — the root (--preset cross, --preset iso), updater/, and menubar/ — each with its own find_package(MavericksShipyard REQUIRED). Those three find_package calls stay; they now resolve from shipyard-cmake's own prefix.

cmake/package_pkg.sh was the repo's only shell-level shipyard locator — $SHIPYARD_SCRIPTS with a --msc-scripts flag and no fallback. It now sources msc.sh, guarded so --msc-scripts still wins and skips the ~2 s probe:

if [ -z "$SHIPYARD" ]; then
  . "$(dirname "$0")/../msc.sh"
fi

msc.sh at the repo root, byte-identical to shipyard's scripts/templates/msc.sh (blob 43c2af0, mode 100644, cmp clean). Root rather than build/: container-tools has no tracked build/, and .gitignore blanket-ignores build*/ (this repo writes build/updater, build/menubar, build-cross, build-iso), so a build/msc.sh would need git add -f and would live inside a directory every build writes into. Check 17 accepts either.

Seven sites no gate reports, converted because they would have shipped wrong:

• three printed (cmake --build <dir>) recipes — cmake/demo_osinit_guard.sh, tests/compat_guard_test.sh, tests/sdk_coverage_test.sh (check 18 deliberately treats a bare ( as not command position)
• four comments handing a reader a plain cmake recipe — CMakeLists.txt ×2, menubar/CMakeLists.txt, updater/CMakeLists.txt

No workflow step was added: install@v1's default mode already puts shipyard-cmake on PATH.

Four tests that could no longer fail — fixed, not just renamed

The repo asserts its own workflows by grepping them:

grep -q 'cmake --preset iso' "$W"  || { echo "missing configure preset" >&2; exit 1; }

shipyard-cmake --preset iso contains cmake --preset iso, so this change is what would have turned four falsifiable assertions into four that pass either way. Renaming the needle to shipyard-cmake … is better only by degree — a comment mentioning the command satisfies it while the run: line says plain cmake, and an existence check structurally cannot see a plain call added beside the converted one.

All four now anchor the command position, and each workflow gains a negative assertion that no line runs a plain cmake/ctest/cpack:

grep -qE '^[[:space:]]*shipyard-cmake --preset iso[[:space:]]*$' "$W" || { echo "missing configure preset" >&2; exit 1; }
…
if grep -qE '^[[:space:]]*(cmake|ctest|cpack)([[:space:]]|$)' "$W"; then
  echo "boot2docker.yml runs a plain cmake/ctest/cpack (only shipyard-cmake configures against shipyard)" >&2
  exit 1
fi

Mutation-checked — six mutants, every one caught by its own message rather than the catch-all. Baseline: both suites exit 0.

#	mutant	test	exit	message	restored
1	boot2docker.yml:34 → cmake --preset iso	boot2docker_ci_test.sh	1	missing configure preset	0
2	boot2docker.yml:35 → cmake --build --preset iso	boot2docker_ci_test.sh	1	missing build preset	0
3	boot2docker.yml:36 → ctest --preset iso …	boot2docker_ci_test.sh	1	missing test preset	0
4	release.yml:158 → cmake -S menubar …	docker_menubar_test.sh	1	release.yml must build the menubar app with shipyard-cmake	0
5	plain cmake --preset iso added beside the converted lines	boot2docker_ci_test.sh	1	boot2docker.yml runs a plain cmake/ctest/cpack …	0
6	plain cmake -S menubar … added beside the converted line	docker_menubar_test.sh	1	release.yml configures menubar with a plain cmake …	0

Mutants 5 and 6 exist because 1–4 alone would leave the new negative assertions themselves untested. All six fail on exit status, not merely on printed output, and the tree was byte-identical to HEAD after each restore.

Verification

check-family-conventions.sh (from shipyard-cmake-collapsed): 12 failures → ok, exit 0. check-shell-portability.sh ok; check-comments.sh exit 0; sh -n on every tracked *.sh and on msc.sh (plus bash -n on the two edited tests); all four workflows parse as YAML. All 16 runnable tests in tests/ exit 0.

Whole-tree sweeps (git ls-files -z | xargs -0 command grep, since .gitignore hides build*/ from a bare grep): no cmake/ctest/cpack left at command position, no .cmake/packages or .local/share/cmake read anywhere, and the one remaining bare-paren hit is (ctest code 77) — prose about an exit code.

msc.sh proven on four paths, sourced from the same if-body shape package_pkg.sh uses: $SHIPYARD_SCRIPTS set → straight through; no shipyard at all → exits 1 with msc.sh's message under both /bin/sh and bash 3.2 (the set -e propagation through a sourced return 1 holds); --msc-scripts given → no probe; and against a stub shipyard-cmake that answers the probe → the helper resolves, which is the tier that replaces the registry.

CI on this PR will fail until shipyard lands bc3082a — shipyard-cmake exists on no runner yet. Expected, not a defect in this PR.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GTxFde2TJMmp5p64DhvDSK