draft: The Race Inside .report_run_id (TOCTOU in hosted reporting)

[abrichr]

Auto-drafted by the daily pipeline on 2026-07-20 and stranded: the run authored the post and pushed the branch, but could not open this PR (ADMIN_TOKEN lacked pull_requests: write). Fixed in #45; opening the backlog by hand.

Merging does not publish this post. The front matter is draft: true and deploy.yml builds with hugo --minify, without --buildDrafts. The file lands on main unpublished. Flip draft: false in a later PR to publish.

The daily scan is paused while any auto-draft/* PR is open, by design, so drafts cannot pile up. Merging or closing all three resumes it. See NEEDS_YOU B10.

[abrichr]

Closing this one, not deleting the branch.

lint_post_substance.py --strict puts it below the floor:

FAIL: thin: 712 words (floor 850)

The core observation is genuinely sharp, and sharper than the draft lets on: O_CREAT | O_EXCL publishes the directory entry before the content, so a racing reader sees a file that exists and is empty, and has to tell "the winner hasn't written yet" apart from "malformed". Fail-closed on that ambiguity and you break a legitimate caller.

That's a lesson about exclusive-create semantics that applies to anyone using a filesystem as a lock, and it would clear the bar written that way. Written as "#163 fixed what #160 shipped", it doesn't.

The branch stays on origin and this closed PR keeps it reachable, so nothing is lost. The source PRs and the angle are recorded in docs/POST_BACKLOG.md.

A closed PR is also the resting state the guard from #45 was built to tolerate, so this branch no longer blocks the daily scan.