test(NO-TASK): Filter the bash 5 job-control warning, and preflight the CI matrix locally

[aaronware]

main is red, and #71 is only the messenger — the failure arrived with #70 and reproduces on
every PR opened since.

Why it failed

test/shell-init.test.js sources snippets in bash -i. With no controlling terminal, bash
complains about job control on stderr, and the helper filtered that noise out:

.filter((line) => !line.includes('no job control'))

That is the entire message bash 3.2 prints — the version macOS ships. Bash 5, on every CI
runner, prints a second line first:

bash: cannot set terminal process group (2031): Inappropriate ioctl for device
bash: no job control in this shell

The filter dropped the second line and let the first through, so two assert.equal(stderr, '')
checks failed on Node 22.12 and 24. Agent-readiness never ran at all — it needs: test.

The fix matches the family rather than one member of it.

Why it got past us, and what now stops it

npm test passing on a workstation is evidence about macOS, not about ubuntu-latest, and this
CLI shells out — to bash, to git, to the filesystem — where the two differ in behaviour and in
wording. Nothing we had could have caught this before CI did.

• npm run preflight — typecheck, build, test. Also runs on git push via a new husky
  pre-push hook.
• npm run preflight:linux — stages the working tree into a node:<version> container per
  matrix entry, runs git clean -Xdf inside so the container sees a clean checkout plus your
  uncommitted edits, then install → typecheck → build → test → the agent-readiness floor.

The matrix, the linter version and the score floor are read out of ci.yml rather than restated,
and the scorer that was inlined in the workflow now lives in scripts/agent-lint-report.mjs so
the workflow and the preflight compute the score with the same code. A preflight that could
quietly disagree with the gate it previews would be worse than no preflight.

CLAUDE.md is new: which changes make preflight:linux non-optional, and how to write a test
that survives the crossing — never assert on the exact text a system tool emits, prefer a
positive match over an empty-string equality, filter noise by family rather than by the phrasing
your laptop happens to print.

Verification

Run locally before pushing, both directions:

	macOS	Linux 22.12	Linux 24
fix reverted	151 pass	—	149 pass / 2 fail — CI's exact numbers
fix applied	151 pass	151 pass	151 pass

Green on macOS and red on Linux in a single command, which is the failure mode this PR exists to
remove.

One thing to decide separately

Agent-readiness measures 86.4% in the container against the recorded 84% floor. Ignored
files are stripped inside it, so that should be a CI-faithful number rather than the inflated
local reading ci.yml warns about — if this PR's CI run agrees, the floor is 2.4 points behind
the score and worth raising deliberately.

🤖 Generated with Claude Code