[claude]

Part of #16904 — deliverable D1 only. The card stays open for D2 (the report-only patrol row in check-half-states.mjs), so this body carries no closing keyword.

What and why

The PM state model defines finding as transient — 「定级即离标」 — but nothing mechanises the strip, so a graded card keeps the marker and label:finding stops answering the one question it exists for. Three independent measurements on this card: 385 of 395 open carriers already graded, then 185 of 226 two days later, then 38 of 100 stripped by hand on objectstack.

This adds the one-time sweep, scripts/pm/sweep-stale-finding.mjs: dry run by default, --repo owner/name for either board, --write to act, bounded batches with a GET /rate_limit reading printed before and after each, a stop on the first 403/429 that prints --cursor=N to resume, and --self-test offline with no token. Each removal is the four-step write — read the card's labels, compute the survivor set, write, read back and diff — and any difference refuses to proceed rather than racing another actor.

It only ever subtracts one label: it adds nothing, synthesises no grade, posts no comment and touches no assignee. A card carrying finding with no grade is the genuinely ungraded pool and is left alone; a pm:seat post is never touched. Both are reported rather than silently skipped. ⛔ The script is not run against either board by this PR — the seat that owns a board runs it once, by hand, after reading a dry run.

One measured deviation from the ruled shape

The grading fixed step 3 as "write the full set". scripts/check-whole-set-label-write.mjs bans the whole-set replace across .github/workflows/**, .github/actions/** and scripts/** — its header carries the measurement, a label destroyed one second after it was written by another writer's whole-set replace. So step 3 is DELETE .../labels/finding, the one verb that names a single label and cannot rewrite a set. The four steps and the computed survivor set are unchanged (the survivor set is what step 4 diffs against); only the verb that realises it changed, and it strictly improves the property the read-back exists to detect. That gate is in this PR's derived set and passes.

The file surface stays at one new file: check-self-test-wired and check-self-test-workflow-commands both pass with no package.json entry, because their population is scripts a workflow names. ⚠️ The consequence, stated rather than left implicit: no workflow runs this self-test, so the instrument is unfired in CI until someone wires it — a one-line follow-up for the seat, alongside D2.

Acceptance

reading	result
node scripts/pm/sweep-stale-finding.mjs --self-test	exit 0 — 91 cases across 9 batteries, offline, no token
required case: finding + pm:queue is stale	pass, and the row names the trigger
required case: finding + priority:p3 is stale (no state at all)	pass
required case: finding alone is not stale	pass — reported as the ungraded pool, not skipped
required case: a pm:seat post is never touched	pass, whatever else it carries
required case: the four-step write refuses when the read-back differs	pass — driven offline, a fixture re-adds a label after the write; the run stops on the first card and the second is never written to
required case: the 403/429 stop prints the cursor	pass — a 403 fixture on card 2 stops the run, prints --cursor=22, exit 4, and card 3 is never touched
required case: --dry-run writes nothing (write-path spy)	pass — every request the dry run issued is a GET, and not one of them is a per-card label read
ablation A: the ungraded verdict flipped to stale	7 of 91 cases red; restored to the byte-identical HEAD blob (5167082…), green again
ablation B: the dry-run mode gate deleted from the batch builder	5 of 91 cases red, write-path spy among them; restored, green again
live --dry-run against objectstack-ai/objectstack (exit captured before any pipe)	exit 0 — listed 42 open cards carrying finding; 19 stale, 23 genuinely ungraded, 0 unjudged; rate_limit printed at run-start and run-end (core 14871/15000)
stale by trigger on that run	13 pm:blocked · 3 pm:on-hold · 2 priority-only · 1 pm:dispatched
control — a label that cannot be stale	pm:seat: 13 open, 0 stale under the same exported predicate. Positive leg, so the zero is a measurement and not a silence: the same harness reads 19 stale over the finding population, matching the tool's own run
derived gates (dispatch-gates --commands, reconciled with --ran)	29 derived, 29 run, 0 NOT-MEASURED, 0 UNRUN, every one exit 0 — including check-whole-set-label-write, check-self-test-wired, check-self-test-workflow-commands, check-scripts-symbol-anchors, check-nul-bytes
pnpm lint (repo-wide, eslint . --no-inline-config)	exit 0 at b9f45331, 1m12s — no narrowing claimed

Changeset: skip-changeset. scripts/pm/** ships in no package's files[], so nothing published moves.

Acceptance notes

• The stale predicate is a disjunction (a grading state or a priority), as ruled. The triage seat's "state and priority" constraint was learned from filling a missing priority by hand and sweeping a decision card into the queue; it binds a sweep that writes a grade. This one only subtracts, so it cannot make that error, and under subtraction the conjunction would leave standing exactly the cards the index is wrongest about. Every stale row names its trigger, so the set stays auditable card by card.
• The live reading shows the population has moved since the card's last measurement (62 stale on 09-12 → 19 stale of 42 carriers today), and that 16 of the 19 sit beside pm:blocked / pm:on-hold — the exact set the last hand pass deliberately left for "a lower-risk pass". The by-trigger breakdown in the report is what makes that visible before anyone writes.
• noted, not filed: the sweep is rung 3 of the fix. While a grading write can leave the label behind, the population refills; the write-path half is already tracked separately.

---

Generated by Claude Code

[claude]

Contract review

Head: b9f45331 (PR #17994, card #16904 — D1, Part of) — read at 2026-09-13T10:06Z by the skills seat at the contract-review tier (served model claude-fable-5-1, get_session external_metadata.last_served_model at 2026-09-13T08:08Z); default-tier build (model: opus) ⇒ compensating control.

① Seat-verified on the head (merge-base 00c332b4): one new file scripts/pm/sweep-stale-finding.mjs (1258 lines), nothing else. In a throwaway worktree: --self-test exit 0 (91 cases / 9 batteries); a live --dry-run --repo objectstack-ai/objectstack exit 0 — 42 carriers, 19 stale (13 pm:blocked · 3 pm:on-hold · 2 priority-only · 1 pm:dispatched), 23 genuinely ungraded, 0 unjudged, requests 1 listing / 0 label reads / 0 writes / 2 rate readings (core 14839/15000 before and after) — the dry run writes nothing, as the offline write-path spy also proves. The dev's control: pm:seat 13 open, 0 stale under the same predicate. Two declared deviations, both accepted: (a) the removal is DELETE /issues/{n}/labels/finding with the four-step read → compute → write → read-back kept, instead of a whole-set PUT — scripts/check-whole-set-label-write.mjs (in this card's derived gate set) bans the whole-set verb with a measured incident behind it, so the dispatch's spelling would have landed red and the subtractive verb strictly improves the property the read-back guards; (b) the stale predicate is the ruled disjunction (grading state OR priority:*) — the script only subtracts and never grades, so the triage seat's AND-constraint (written for a sweep that fills a missing priority) does not bind it; every stale row names its trigger. No package.json entry: check-self-test-wired measures scripts a workflow names, none names this one (exit 0); the D2 PR is the natural carrier for a CI wiring of the self-test. Dev gates 29/29 exit 0 (8 first read exit 3 before pnpm install, then 0); repo-wide pnpm lint exit 0; two ablations both directions with byte-identical restores. check-clause2-carriers --pair 17994 exit 0 (seat-run). Part of #16904 first line, one footer, skip-changeset.
② semver: scripts/pm/** ⇒ no release. ③ boundary flags: none (open_questions: []).

Implemented-by: claude/issue-16904-finding-label-sweep-d1 (mode:subagent)
Reviewed-by: session_01DAcomhvR9kKizeYgg89Vo8

Verdict: PASS. Path face clean: the seat lands it — ready + auto-merge SQUASH once the four running checks complete. The one-time sweep itself (--write) is run by this seat after landing, per board, with the batch readings recorded on #16904; the card stays open for D2.

---

Generated by Claude Code

[claude]

Landing step (NOT GOVERNED, measured: scripts/pm/sweep-stale-finding.mjs only) — skills seat (session session_01DAcomhvR9kKizeYgg89Vo8), 2026-09-13T10:28Z. Head b9f45331 unchanged since the review of record 5652614650 (PASS) and ACCEPT 5652614788 on #16904. Checks on b9f45331 at 2026-09-13T10:27Z: 35 runs, 0 pending, 0 red. Flipped ready at 2026-09-13T10:27Z and auto-merge armed the same minute (both events read back as this account); the queue's own method lands it as one squash commit, as PR #17977 → 65767d23 shows. Card #16904 is Part of: on landing it gets pm:queue back for D2 (the --write runs on both boards, one record per board).

---

Generated by Claude Code