Sweep runner: wait no longer reports a just-finished job as not running (flaky required check)

[seathatflowsinourveins]

Scope

• What this PR changes, in one or two sentences: fixes a check-then-act race in the landscape sweep runner's wait (a job that finished between its done check and its lock check was reported as "not running"), adds a deterministic test and failure diagnostics, and records two proven mistakes in the anti-pattern log.
• Base commit: b72c588f
• Lane: lane:foundation
• Owned paths touched: tools/sota-convergence/landscape-sweep/codex_job.py, tests/test_landscape_sweep_harness.py, docs/harness-defaults.md, manifests/evidence.json (hot-file commit, last).

SOTA sources

• flock(2), Linux man-pages (https://man7.org/linux/man-pages/man2/flock.2.html): a lock is released when its last descriptor is closed, so a runner that writes its completion marker before it closes the lock lets a reader decide from the marker once the lock is free.
• This repository's runner: tools/sota-convergence/landscape-sweep/codex_job.py (finish writes done before run closes the job lock; running probes the lock). No upstream tool is replaced; the change is three lines in our own wrapper around codex exec.
• The anti-pattern log's contract: docs/harness-defaults.md, "Upstream verification and compounding learning" (AGENTS.md: record each proven mistake).

Evidence-class table

Claim	Evidence class	Command / receipt
A job that finishes during the lock check is reported as done	synthetic (fixture: the lock check itself completes the job)	python3 -m unittest tests.test_landscape_sweep_harness.WaitFinishRaceTests; failing first on the unfixed runner: 'done exit=none (the job is not running)' != 'done exit=0'
A job that never ran is still reported as not running	synthetic	the same class, second test
The hosted failure on PR #360 (run 36836458360) was this race	inference: the run kept no output of wait; a single-core stress loop did not reproduce it before the fix	the test now prints wait's output and the runner's files on failure
The harness's other tests still pass	local_integration	python3 -m unittest tests.test_landscape_sweep_harness tests.test_adoption_docs_consistency

Local commands run

$ python3 -m unittest tests.test_landscape_sweep_harness.WaitFinishRaceTests     (test added, runner unfixed)
FAILED (failures=1)   AssertionError: 'done exit=none (the job is not running)' != 'done exit=0'
$ python3 -m unittest tests.test_landscape_sweep_harness tests.test_adoption_docs_consistency
Ran 209 tests in 48.440s   OK (skipped=4)   (exit 0)
$ python3 scripts/validate.py
{"components": 69, "hashed_files": 8720, "profiles": 4, "receipts": 180, "status": "passed"}   (exit 0)
$ git diff --check origin/main HEAD
(no output, exit 0)

Decision record

None: a defect repair. The two anti-pattern rows in docs/harness-defaults.md record what happened and the rule that prevents it (this race; and the evidence-manifest rows that a hot-file re-run of PR #548 dropped twice on 2026-10-01).

Host evidence

Not applicable.

Checklist

• New/changed GitHub Actions are pinned to a full commit SHA with a version comment (none changed).
• New/changed workflows declare top-level permissions: contents: read (none changed).
• No secrets are printed, logged or committed; no new required secret was added without a documented owner.
• No new paid hosting, subscription or billing surface was introduced.
• Peer-owned untracked files and worktrees were preserved (not deleted, moved or overwritten).

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-01T09:19:19.893251Z	2414ee8	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.