[randomparity]

Closes #480.

What this changes

tests/fixtures/quest-log/claim-test.sh's stub gh used mkdir <leaf> as its label-create compare-and-swap. The ubuntu-latest runner image moved from Ubuntu 24.04 to 26.04, where /usr/bin/mkdir is uutils coreutils 0.8.0. That mkdir checks for the leaf and then creates it, so two racing calls both succeed, and the simultaneous-claims rounds failed with exits 0:0.

Reproduction in containers, two mkdir calls racing on one leaf, 2000 trials:

Image	mkdir --version	both succeeded
ubuntu:26.04	uutils coreutils 0.8.0	1991 / 2000
ubuntu:24.04	GNU coreutils 9.4	0 / 2000

The issue's kernel-mkdir(2) hypothesis does not hold, and pre-creating the parent would not have fixed it.

• The stub now claims a label with bash's noclobber redirect (set -C), which opens the description file with O_CREAT|O_EXCL. That open is a shell builtin, so no coreutils implementation decides the race.
• A deliberately racy mkdir shim on the suite's PATH (check, sleep 0.2 s, create; -p passes straight through) makes a regression to a mkdir-based CAS fail the simultaneous-claims rounds on every host, macOS included.
• .claude-plugin/plugin.json 7.12.3 → 7.12.4.

Production tracker.sh / profiles/github.sh are unchanged; they call the GitHub API and never mkdir.

Verification

• Red: with the shim and the old stub, macOS fails claim-test: round 1: exits 0:0, expected one 0 and one 6.
• Green: macOS (bash 3.2) passes; ubuntu:26.04 container passes 5/5 runs.
• Control: main's unmodified suite in the ubuntu:26.04 container fails 3/3 with the CI error.
• just ci passed via the pre-push hook on 915ae4b.

Review

Single-pass adversarial review (routed: no change hazards, size S, one PR, not security-relevant; security pass not triggered) returned approve with three notes:

• medium — the stub comment overclaimed independence from coreutils (mkdir -p still makes the directory): accepted, comment narrowed. The suggested || [[ -d ]] guard was rejected: concurrent mkdir -p tolerates an existing directory, and uutils passed 5/5 container runs.
• low — the shim comment said "deterministically" and implied coverage of the recover rounds: accepted, reworded.
• low — a racing reader can see an empty description and take the malformed-conflict branch: rejected with evidence. The asserted exit codes are unchanged, no test checks that payload, and it replaces main's 404 → exit 4 window.

Deferrals: none.

[randomparity]

Review — PR #481 (issue #480)

• review depth: single-pass (gauntlet, fresh context) · iterations: 1 · exit: none
• verdict: approve — 3 notes (1 medium, 2 low), 0 blocking
• dispositions: medium comment overclaim → accepted-fixed (915ae4b); low shim wording → accepted-fixed (915ae4b); low empty-description window → rejected-with-evidence (no asserted outcome changes)
• security: not triggered (test fixture only; no trigger in the diff)
• deferrals: none · follow-up candidates: none · suppressions: none
• reviewed head: 8d0e5a3; delivered head: 915ae4b (comment-only delta)