The change is pure operational plumbing: the scheduled re-judging workflow swaps a maintenance PAT for the workflow's own GITHUB_TOKEN (which carries the `actions: write` the job already declares), and `rejudge-parked.mjs` now wraps each `gh run rerun` in try/catch so one failure does not strand the remaining picks — while still `process.exit(1)` at the end so the fault stays visible rather than being swallowed. No content under content/ is touched, no citations are added (verification section confirms none), and no run records are involved. Nothing here loosens a §3.15 check: the panel itself still runs, parking on the budget window is unchanged, and re-running is only triggered when the window has room, not to dodge a dispute. The DECISIONS entry states the failure, the date, and stamps 'Supervised-by: founder (direction in session, 2026-09-18)', which §3.15's founder-direction clause expressly permits to be recorded in the decisions log with date and channel. It also discloses honestly that the operator re-ran #353's panel by hand, rather than hiding that manual action. AGENTS.md is untouched.

The change does not edit AGENTS.md or weaken the gated review process; it repairs the scheduled mechanism that re-runs parked arbiter workflows using the job token with its declared permission. The script now attempts all selected PRs while preserving fail-visible behavior by exiting nonzero after any rerun failure. The decision entry records the operational change and its supervision with a date and channel, and no citations, content assessments, confidential material, or fabricated locators are introduced.

This pull request contains strictly operational and tooling fixes for the budget-window rejudging workflow (`budget-window.yml` and `rejudge-parked.mjs`), along with the corresponding log entry in `docs/DECISIONS.md`. It switches the token to the workflow's `GITHUB_TOKEN` (which possesses `actions: write`) and makes the rejudging script iterate through all candidates before exiting with an error if any fail. No constitutional rules, CI gates, or substantive content files are modified or weakened.

The diff is an ops fix only: budget-window.yml uses the job’s declared GITHUB_TOKEN (actions: write) instead of the maintenance PAT that could not re-run workflows; rejudge-parked.mjs tries every picked PR before exiting non-zero; DECISIONS.md records the failure mode, fix, and founder supervision dated 2026-09-18. No AGENTS.md edit, no content/citations, no provenance stamps, and no loosening of §3.15 concurrence, fabrication, or confidence checks.

This is a pure infrastructure fix: the scheduled re-judging step switches from the maintenance PAT to the workflow's own GITHUB_TOKEN so the already-declared actions:write permission actually works, and the script now tries every parked PR before failing instead of aborting on the first refusal. It does not edit AGENTS.md, does not touch content, citations, or assessments, and does not weaken any §3.15 check — the arbiter re-run still triggers the same multi-model ratification gate; only the broken token path is repaired, and the run still fails visibly if any re-run cannot be issued. The DECISIONS.md entry is appended (append-only in spirit), accurately describes the mechanical change, and records the founder direction with date and channel as the constitution requires. The operator's manual re-run of #353's panel is a disclosed remediation, not a bypass: it routes through the same panel process. Nothing in the diff fabricates provenance, presents AI records as human ones, or dodges a gate; the discrepancy between 'five scheduled runs failed' in the workflow comment and the PR body's framing is immaterial to any rule.