A CI gate

Orientation

Two audits (18 and 19 Sep 2026) found the same gap: the only tracked workflow publishes the wiki, so nothing but the lead’s own runs stands between a change and main, and an outside reader cannot reproduce a completed full suite. The auditor’s own attempt timed out at 180 s. Measured today: the full suite takes 11 to 12 minutes on the Mac (14 cores) and 18 to 35 minutes on a 4-core x86_64 Linux machine, because the corpus builds one program at a time; and it writes generated files into the source tree (examples/**/zig-out, toolchain/.zig-cache), so two runs cannot share a tree and an orphan from one run has deleted another’s outputs. This step makes the suite reproducible by anyone and gates pull requests on it. Worker: a fresh Claude Opus session, bypass permissions, own worktree and Herdr tab, launched after step 42 lands (both edit toolchain/src/corpus.zig). The lead owns the wiki, audit and acceptance. Nothing here changes the lead’s acceptance checklist: CI is a floor, not a replacement.

Write scope

.github/workflows/ci.yml (new; site.yml untouched), toolchain/build.zig, toolchain/src/corpus.zig and the test helpers it uses, toolchain/README.md (a “Running the suite” section), .gitignore. No language or runtime change, no harness, no wiki, no audit/, HANDOFF.md. No push: the lead pushes the branch and opens the pull request that proves the gate. Never use tr; ls is aliased, use /bin/ls.

Parts

  1. Generated files leave the source tree. Every file the suite makes (mo build outputs, brick caches, temporary corpora) goes under one directory chosen by an environment variable (say MO_TEST_OUT), defaulting to a fresh temporary directory per run, never under examples/. Prove it: run the suite from a read-only copy of examples/; and run two filtered suites at once in one tree, both green. RED first: a test that fails today because a run leaves files under examples/.
  2. The corpus builds in parallel. Today the corpus step is serial (load average 1.0 on 4 cores for half an hour). Build and run corpus programs with a bounded worker pool sized from the CPU count, output still deterministic and failures still attributed to their program. Numbers: wall time before and after on this Mac; the lead measures Linux.
  3. Two gates in one workflow. On every pull request and push to main: fast (Zig 0.16.0 pinned by exact version, zig build, the unit tests by filter, mo fmt --check and mo check over examples/, the rejects’ first diagnostics), which must finish inside 10 minutes; and full (zig build test --summary all, unfiltered) on ubuntu-latest and macos-latest, with a timeout above the measured time, the summary line and exit code in the job summary, and the log uploaded as an artifact. Every step’s real exit status propagates; no || true, no continue-on-error. Least privilege: permissions: contents: read; actions pinned by commit SHA; no secrets; nothing from a fork’s pull request runs with write access.
  4. The known flake is named, not hidden. The TLS corpus test “a fatal alert where a hello belongs” fails about 1 in 5 alone on Darwin. Do not retry it silently and do not skip it: report its rate from 20 runs on each platform in the report, and if it makes the gate useless, propose a quarantine list that prints every quarantined test in the job summary; the lead decides.
  5. Negative controls, described for the lead to run on the proving pull request: a planted failing unit test, a planted corpus mismatch, and a planted formatter difference each turn the gate red with the failure named; removing them turns it green.

Done when

Every local process under toolchain/bench/step36/guard.py, each quoted run teed into a filed log with an exit file beside it (under toolchain/bench/ci-gate/). zig build exit 0; the focused tests and both proofs of part 1 with real summary lines and exit codes; actionlint clean on the workflow (install it with Homebrew). The unfiltered full suite, Linux and the pull request are the lead’s. Small commits as yourself with a Co-Authored-By line naming your model. Write your final report to toolchain/CI-GATE-REPORT.md and commit it last, with a clean worktree. While anything runs, wait in the foreground.