chore: saved Claude workflows stay LF; budgeted-autonomy close-out cleanup

[IanFrelinger]

The budgeted-autonomy close-out's one cleanup PR (owner decision of 2026-10-13, decisions log row "Workstream", landed by #793).

What changed

• Saved workflows stay LF. .gitattributes pins .claude/workflows/*.js to text eol=lf. With core.autocrlf=true, as on this Windows host, they were checked out CRLF. The Workflow tool then refused project-audit by name: "script contains control characters that would be hidden in the approval dialog". So the phase F handoff had to run an LF copy through scriptPath.
• The checker refuses control characters. scripts/phase-end/check_workflows.js now refuses a workflow that holds a control character other than tab and newline. It names the code point and the line, and prints the fix: rm <file> && git checkout -- <file>. Before, the checker printed ok for a CRLF workflow that the tool refuses.
• The test. tests/scripts/claude-workflows.test.sh:
  • asserts every .claude/workflows/*.js is pinned LF (git check-attr eol);
  • adds a sixth broken fixture, a valid script with CRLF line ends, which must be refused for that reason.
• The registry row. ci/cert-gate-assertions.md row 82 names both new rules.
• The controller's docs. In tools/autonomy-controller/README.md:
  • The intro says the controller takes tasks the operator adds with task add, optionally naming a card, and that nothing reads cards from GitHub. The package docstring in ashlar_controller/__init__.py says the same.
  • The outbox.py row says the operator enqueues and flushes the posts, and the tick posts nothing itself.
  • "Not in PR 3" says what PR 5 added: the Codex and Cursor adapters, and the transfer driven by task transfer.
• The pilot file. tools/autonomy-controller/PILOT-PR4.md, the PR 4 pilot's throwaway file (Add the file tools/autonomy-controller/PILOT-PR4.md containing one short paragraph saying it was written by the budgeted autonomy controller's pilot of PR 4 on 2026-10-11 in the workspace it was given. Create no other file and change nothing else. #777), is removed. Nothing references it.
• The changelog. CHANGELOG.md gets an [Unreleased] "Fixed" entry.

After merge

Merging does not rewrite existing Windows checkouts: git rewrites no file whose blob is unchanged, and the .js blobs did not change. So each active Windows clone needs the fix once: rm <file> && git checkout -- <file> for both workflows. Until then claude-workflows.test.sh fails there, with that command in its message. I will run it in the session checkout.

Verified

• Observed red, then green, by hand on this host.

  • Before the checker change, the test's first byte check, which used grep $'\r', passed on CRLF copies. Git for Windows' grep cannot see a CR before a line end. That check was replaced, first by tr/cmp and then by the checker rule.
  • The tr/cmp version went red on the CRLF working copies and green after the re-checkout.
  • The final version goes red on a CRLF copy of project-audit.js, with holds a control character (U+000D at line 1) and the fix. It goes green on the LF copy.

• scripts/mutation-check.sh, with ASHLAR_MUTATION_RUNNER="bash scripts/phase-end/mutation-runner.sh" (the test_validate.py suite plus this test), at 248c0886c:

  mutation close-m1-attribute-dropped: KILLED red=failed:1/35 green=passed:35/35 ref=248c0886cd5aaf4b29de6d6725f7251baee505cd runner=override
  mutation close-m2-attribute-crlf: KILLED red=failed:1/35 green=passed:35/35 ref=248c0886cd5aaf4b29de6d6725f7251baee505cd runner=override
  mutation close-m3-control-check-skipped: KILLED red=failed:1/35 green=passed:35/35 ref=248c0886cd5aaf4b29de6d6725f7251baee505cd runner=override
  mutation close-m4-cr-allowed: KILLED red=failed:1/35 green=passed:35/35 ref=248c0886cd5aaf4b29de6d6725f7251baee505cd runner=override
  

  • m1 and m2 failed on the attribute check: eol: unspecified and eol: crlf.
  • m3 and m4 failed on "the checker accepted the broken fixture crlf.js".

• An adversarial review (portability, records, breakage) found these, and all are fixed here:

  • the registry row was not updated;
  • the checker did not enforce the rule its header claims;
  • the README's card wording was wrong.

  The reviewer also ran the attribute check on Linux (Ubuntu 24.04, autocrlf unset), checked every output form of git check-attr, ran the test from another cwd and from a worktree, and confirmed the .js blobs are unchanged, so no renormalize is needed.

• Other checks: scripts/check-docs-phantom-paths.sh reports 0 missing. The knowledge graph is current; a rebuild changed only line endings in this checkout.

🤖 Generated with Claude Code