fix(scripts): check-regen-pending --self-test no longer routes its stubs through an ambient package manager

[zhuangjianguo]

Fixes #15990

check:merge-driver runs node scripts/check-regen-pending.mjs --self-test. Its
deferred-merge fixture wrote a stub package.json into a mkdtemp directory under
tmpdir() and ran the stub gate as pnpm -s check:spec-changes. That directory declares
no packageManager and has no parent manifest to inherit one from, so the verdict was
decided by whatever launcher resolved there, not by the tree.

The differentiator, and the reading that settles it

The card asks why origin/main and four unrelated PRs were green while one PR was red
twice. They are not different. The verdict is per-JOB and environmental, and main is
red too — its own push build, run 33981169123 / job 101346594800, 2026-09-05T17:51Z,
fails with the byte-identical signature: the same eleven red assertions, the same three
interleaved greens, the same 1 failure(s) (cases and floor) summary. The greens on the
card were sampled before that crossover, the reds after. No tree-level or config-level
differentiator exists to find.

The mechanism, measured rather than inferred — and it is not "the launcher fails to
resolve":

1. .github/actions/setup-pnpm runs corepack install in the workspace. Measured with a
   scratch COREPACK_HOME: that materialises the pinned pnpm 10.31.0 and writes no
   lastKnownGood.json.
2. The fixture's temp dir pins nothing, so Corepack ignores the repo pin and resolves
   pnpm's latest dist-tag from the registry. Measured: it downloads pnpm 12.3.4
   (latest since 2026-08-26; 12.3.4 published 2026-09-04T14:20Z).
3. pnpm 12's CLI rejects -s: error: unexpected argument '-s' found, exit 2.
   Every stub then collapses to "the gate exited non-zero", which the script correctly
   grades as stale — so exactly the stubs whose expected outcome is not stale go red.
4. Corepack then writes lastKnownGood.json, and it is sticky: the warm re-run costs
   116 ms with no network. That matches the failing job's cadence exactly — every
   runHook in the CI log takes ~130 ms, against 330 ms measured here for the pinned
   pnpm 10.31.0 running the same command in the same directory. The job's launcher
   resolved and ran; it just answered hostilely. Consistent with this, the job log shows a
   Corepack cache hit and contains no download line.
5. actions/cache@v6 saves COREPACK_HOME back on a miss, so a job restores a store
   carrying either a -s-accepting pnpm (green) or a pnpm 12 (red). Which one is not a
   fact about any diff. This reconciles the two observations that a bare "sometimes fails
   to resolve" cannot hold together: flaky across jobs, byte-identical within one.

Not determinable from outside the runner, and stated rather than left silent: which job
first seeded a poisoned store, and on which cache scope. GET /actions/caches answers
403 unauthenticated.

What changed

The gate takes its launcher from OS_REGEN_GATE_LAUNCHER. Production leaves it unset and
every gate still runs as pnpm -s SCRIPTNAME, byte-identical to today. --self-test sets
it — the only thing that ever does — to a launcher the fixture writes for itself, which
reads the named script out of the manifest at the gate directory and runs its body under
/bin/sh. This is direction 1 on the card, and the principle AGENTS.md already states for
check:cross-package-test-inputs: a detector with no dependencies cannot itself fail to
resolve in CI.

Two details worth review:

• The launcher runs /bin/sh by absolute path with argv0 sh. execSync would
  have produced /bin/sh: 1: NAME: not found, which gateCouldNotRun()'s regex does not
  match — the runner name drops out of the diagnosis while exit 127 still classifies. That
  regression showed up as a red assertion during this work and was fixed by making the
  fixture reproduce production's producer shape, not by widening the matcher.
• The launcher lives in a separate temp dir outside the fixture repo, so it cannot be
  swept into the index by the git add -A on side2 and re-open the stat-dirty
  merge --abort crash of [finding] check-regen-pending.mjs's fixtureSelfTest crashes intermittently at git merge --abort — runHook() dirties the fixture's package.json while a merge is in progress #9258.

No case is skipped, disabled, quarantined or relaxed. Two cases are ADDED, pinning the
diagnosis technique permanently: a hostile pnpm first on PATH, refusing the way pnpm 12
refuses, with each reading asserted equal to the same call made without it.

Proof

Reproduced red first, then green under a real, a hostile and an absent launcher. Exit
codes captured by redirect-then-read, never after a pipe. The pre-fix leg restored the
f7db8f4fd blob, verified on disk by git hash-object before measuring; the restore leg
verified back to the HEAD blob with git diff HEAD and git status --porcelain both
empty.

launcher on PATH	pre-fix	post-fix
real pnpm 10.31.0	exit 0	exit 0
refuses, exit 1 (Corepack-style)	exit 1, 11 red assertions	exit 0
pnpm 12 shim: rejects -s, exit 2 — the measured CI shape	exit 1, 11 red assertions	exit 0
no pnpm anywhere on PATH	exit 1, 12 red assertions	exit 0

Gates run locally on the final commit 5a2a2092f, each exit code read before any pipe —
check:merge-driver (the gate this card is about), check:nul-bytes, check:parse-guard,
check:entry-guard, check:cross-package-test-inputs, check:pnpm-filter-targets,
check:agent-test-spelling, check:bash32-floor: all exit 0. The family was derived
with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack (54 families,
nearly all scripts/**-wide); the rest is CI's single pass over the farm.
scripts/measure-self-test-floor.mjs still classifies this file ROSTER, unchanged, and the
battery roster and floor are untouched. pnpm lint was NOT run: this worktree has no
node_modules, so no narrowing is claimed for it — CI's Lint & Repo Gates owns it.

Clause-②: no — the diff is one repo gate script; git diff against the merge base shows
no packages/spec/api-surface/ movement (positive control on the same command and scope:
scripts/ returns the changed file), nothing is published from any package, and the
production gate command is unchanged.

No changeset: scripts/** repo tooling, root manifest is private, nothing ships.

---

🤖 Generated with Claude Code

https://claude.ai/code/session_01ARYe3yQTQCUFm5qPYNgKaJ

---

Generated by Claude Code