[deanbot]

What changed and why

The auditor's "Measured, no finding" section listed every waiter state file in the state folder on the machine it ran on, by file name, ending in -wait.json. That had three problems:

• the report is posted as an issue, so it put local file names on GitHub;
• most of the files were leftovers from PRs merged long ago;
• the numbers looked like counts over the period, though they are streaks at the time of the run.

scripts/audit.mjs now reads each state file only to learn its waiter's kind and PR, and names no file:

• <slug>-pr-poll-<n>-wait.json: the executor's waiter on PR #n, counted only while that PR is open. The open PRs come from one gh pr list --state open read, inside the same gather step as every other read, so a failure exits 2 like they do.
• <slug>-pr-state[-<clone>]-wait.json: the reviewer's waiter. The clone name, a local name, is dropped.
• <slug>-pr-qa-state-wait.json: QA's waiter.
• Any other name is skipped: another repository whose slug starts the same way, or an old state format.

The markdown line now reads, for example: Wake-ups that found nothing: 4 waiters on this machine, for open PRs and the roles, with their idle streaks at the time of the run (not counts over the period): the executor's on PR #3 2, the executor's on PR #12 4, the reviewer's 1, QA's 5. What a wake-up catches isn't recorded, so there is no finding.

--format json gives idle: [{ "waiter": "pr-poll", "pr": 3, "idle": 2 }, …, { "waiter": "pr-state", "idle": 1 }, { "waiter": "pr-qa", "idle": 5 }]. It no longer has a file key: the JSON named files too.

SPEC's Auditor section describes the new form. The test fake gh gains an openPrs key, used only for pr list --state open. Other callers are unchanged.

Done, and the checks that show it

All tests below are in scripts/test/run.mjs, in the audit section, named audit idle (#108): ….

• The markdown names no state file and no local path. A seeded state folder gives a report with no -wait.json, no .agent-loop, no home path and no clone name ("the markdown names no state file, folder or clone").
• Waiters for closed PRs are left out. With PRs Kiro adapter #3 and Remove spec-path config field from all adapters #12 open, feat(start-reviewer): sentinel file stop mechanism #11 merged, a lookalike slug and an old-format name in the folder, only Kiro adapter #3 and Remove spec-path config field from all adapters #12 count, plus the role waiters. The test checks the exact line. With no PR open, only the role waiters count.
• The line says the numbers are current streaks, not period counts. The exact-line test includes "(not counts over the period)".
• A missing state folder and an empty one both still give "no waiter state for this repository on this machine".
• SPEC's Auditor section describes the summary form.
• JSON: the exact idle array is checked, and the JSON text contains no file or clone name.

Mutation-checked:

• Dropping the open-PR filter fails the exact-line, JSON and no-open-PR tests.
• Dropping the "(not counts over the period)" wording fails the exact-line test.

Calls a reviewer might push back on

• The line lists every counted waiter by role and PR, rather than a count with only the highest streak named. PR numbers are public, and a project has few open PRs, so each streak stays readable.
• Several reviewer clones each appear as "the reviewer's". The clone name is local, so it isn't shown.
• If the folder holds only leftovers or unknown names, the line is the same "no waiter state for this repository on this machine": nothing left counts.
• An executor waiter is identified by the PR number in its file name. A waiter for a PR that was reopened counts again.

Known limits

• The streaks are still those of the machine the audit runs on. A waiter on another machine isn't seen, as before.

Restart note

None: no loop text changed, and audit.mjs is read when it runs. Setup doesn't need re-running.

Closes #108