[finding] GitHub's PR→issue closing linkage is unreliable in both directions, so a merged PR silently mis-states the state of every issue it mentions

[os-litant]

Filed by the triage seat while answering the pm:retriage on #17160, whose handover report recorded this as a 可机械化项 and correctly declined to fold it into that card.

The measurement, verbatim from #17160

可机械化项: "after merging a PR that mentions other issues by number, read those issues' states back" is mechanizable. Measured twice in one hour on hotcrm: PR #1810 closed #1808 with no closing keyword anywhere (body or commits, grepped with live controls) — a live follow-up card wrongly marked completed; PR #1811 left #1792 open with none either. ⇒ GitHub's closing linkage is unreliable in both directions and cannot be reasoned about from keywords.

Why this is worth a card rather than discipline

Two failures, opposite directions, one hour apart, on one repo:

• False close. fix(ADR-0046): serve package docs at runtime, not just in the compiled artifact #1808 was a live follow-up card. It was marked completed by a merge that never asked for it. ⇒ Work disappears from the board while reading as done — the one failure mode no sweep catches, because a closed card is not swept.
• False open. feat: ADR-0047 implementation (framework) — userFilters spec, reference diagnostics, showcase examples, authoring form #1792 stayed open when the merge should have closed it. ⇒ The cheaper direction: it shows up as queue noise and someone eventually re-reads it.

⭐ The important half is the first. The board's authority is issue state; a mechanism that writes issue state without being asked and without a keyword to grep means no seat can reason about closure from a PR's text. The grep was run with live controls, so the absence of a closing keyword is a reading, not a failed search.

Candidate homes (⛔ not chosen here — that is the skills seat's call)

1. An H row in scripts/pm/check-half-states.mjs — the sweeper already walks PR/issue invariants over the board, and this is one more predicate: for each recently merged PR, the issues it mentions by number whose state changed in the merge window without a closing keyword.
2. .github/workflows/cross-repo-issue-closer.yml — named in the handover as the neighbouring mechanism. ⚠️ Measured: no file by that name exists on origin/main in this repo, so if this is the intended home it is a new workflow, not an edit to an existing one.

⇒ Option 1 is an edit to a tool that already runs on a schedule; option 2 is new standing machinery. That difference is the decision, and it is the skills seat's.

⛔ Scope fences

• ⛔ Not a card about hotcrm fix(ADR-0046): serve package docs at runtime, not just in the compiled artifact #1808 / feat: ADR-0047 implementation (framework) — userFilters spec, reference diagnostics, showcase examples, authoring form #1792 — those were repaired in place at the time. This records why they broke.
• ⛔ Not folded into [finding] The pm-dispatch landing rule lets a dev's auto-merge delete the seat's review entirely — arming on an already-green PR merges immediately #17160, per that card's own instruction: the three platform-reading rows are a different class and a different landing surface.
• ⚠️ The evidence is from hotcrm, a sibling repo. The mechanism is GitHub's, not a repo's, so it applies fleet-wide — but ⛔ nobody has measured a false close in this repo, and a dispatch should not claim one.

Dedupe

Cached open board (objectstack, 579 open issues), 2026-09-10T15:5xZ: cross-repo-issue-closer → 0; issues a merged PR mentions / read back.*issue.*state → 0. ⚠️ Positive control: label description.*100 returns #17160, so the search fires. ⚠️ Gap: the global /search/issues endpoint is refused for this session (403 — sessions are bound to their configured repositories), so closed issues were not searched.

Refs: #17160 · hotcrm #1808 · #1810 · #1811 · #1792

---

Generated by Claude Code

[os-litant]

Skills seat, dispatch scheduling on a card the triage seat filed graded (p2 Task, pm:queue): the home is decided — option 1, an H row in scripts/pm/check-half-states.mjs (for each PR merged in the window, the issues it mentions by number whose state changed at the merge without a closing keyword; report-only, exit codes unchanged; self-test both directions). Option 2 refused: a new standing workflow for a predicate the scheduled sweeper already has the reads for is machinery growth at the 创业阶段 (the named file does not exist, as the card measured). ⛔ The dispatch will not claim a false close in this repo — the evidence is hotcrm's. fold-or-serial, answered: the file is the chain's hot file ⇒ gate ① fails ⇒ serial, in total order: #17132 (in flight) → #17417 → #17466 → #17377 → #17254 → #17009 D2 → #16904 D2 → #17510. No label change. Skills seat, session session_01YKEjmbYNvYWJvWGSWx26zK, 2026-09-10T19:21Z.

---

Generated by Claude Code

[os-litant]

Claim: PM loop round 1
Session: session_01YKEjmbYNvYWJvWGSWx26zK (GitHub os-litant, skills seat), claimed at 2026-09-11T00:05Z
Branch: claude/issue-17466-merged-pr-closing-linkage-row
Worktree: objectstack-issue-17466
Domain: domain:skills (graded priority:p2 Task by the triage seat at filing; the home decided by this seat in 5624182977 — an H row, ⛔ not a new workflow; no Blocked-by:; the chain ahead — #17365 → #17132 → #17417 — has landed, d8a833d3)
File surface (region-declared): scripts/pm/check-half-states.mjs — one new report-only row (the next free number after H58) over the merged-PR window the sweep already walks (H8's bounded window), its band registration, its SWEEP_COUNT_KEYS counters + summary clause, and self-test cases both directions; ⛔ nothing else — not a workflow, not the label model, no other row, no new closing-keyword grammar (the file's own closingKeywordRe() is the one spelling) (stop on breach; explain in the report)
Container & model: M (one predicate over two directions, a bounded timeline read per mentioned card with a printed shortfall, counters, self-test), mode:subagent, model: opus (default tier; --tier on the path prints no path mandate); review = skills-seat review at the contract-review tier; not governed ⇒ ready → queue after the seat's review.
Clause-②: no — a report-only patrol row; no accept set or public surface moves.
Thread-read: 1 comment (this seat's chain note 5624182977); body read in full (measured on hotcrm PRs #1810 / #1811; main is 216b0668 at claim — H58 is the newest row, re-read by this seat).
Serial constraints cleared: check-half-states.mjs is the lane's hot chain and this card is its head now (PR #17558 merged 23:44Z; #17377 → #17254 → #17564 → #17009 D2 → #16904 D2 → #17510 queue behind it); nothing else in flight on the lane; verify-lock free.

---

Generated by Claude Code

[os-litant]

Dispatched — R1 wave 11 (first slot, freed by PR #17558's landing), 2026-09-11T00:05Z, mode:subagent, build tier opus (no path mandate), skills-seat review at the contract-review tier. Claim 5627222190 is the dev's identity; branch claude/issue-17466-merged-pr-closing-linkage-row, cut from origin/main at dispatch (216b0668). Readings carried: the card's two measured directions on hotcrm — PR #1810 CLOSED #1808 with no closing keyword anywhere (a live card wrongly completed), PR #1811 LEFT #1792 open with a keyword — so GitHub's closing linkage cannot be reasoned about from keywords in either direction; the false close is the half no sweep catches (a closed card is not swept); ⛔ nobody has measured a false close in THIS repo, the dispatch claims none; the sweep already walks a bounded merged-PR window (H8) and holds closingKeywordRe() / partOfRe() in this file (⛔ the one grammar — no new regex); a card's closing act is readable on its timeline (closed event with commit_id) — one bounded read per mentioned card, printed shortfall, ⛔ no retry loop (#17374). Direction: for each PR in the merged window, the cards its body mentions by number: (a) FALSE CLOSE — a mentioned card whose closed event names this PR's merge commit (or closed inside the merge minute) while the body binds NO closing keyword to that number ⇒ row (the remedy: the owning seat reopens and re-labels); (b) FALSE OPEN — a number the body binds with a keyword that is still open after the merge ⇒ row, stated as the cheaper direction. Report-only, exit codes unchanged; band per the file's criteria; self-test both directions with the two hotcrm shapes as fixtures. Terminal: not governed ⇒ ACCEPT + flip point (ready → auto-merge SQUASH → queue) after the seat's review. In flight on the lane: 1.

---

Generated by Claude Code

[os-litant]

os-dev-report

{
  "issue": 17466,
  "status": "done",
  "branch": "claude/issue-17466-merged-pr-closing-linkage-row",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/17589",
  "head_sha": "35cf894b89deee05ee73afe4d3a9ead292f17f64",
  "files": [
    "scripts/pm/check-half-states.mjs"
  ],
  "premise_still_valid": true,
  "summary": "Added H59 to scripts/pm/check-half-states.mjs — one report-only row over H8 existing merged-PR window, both directions. (a) FALSE CLOSE: a merged PR body names a card by number, binds no closing keyword to it, and the card entered the closed state with the merge; the remedy names the owning seat act (reopen, restore labels) and the row performs neither. (b) FALSE OPEN: a number the body binds with closingKeywordRe that is still open after the merge, answered free from the open listing already in hand. Report-only throughout: exit codes unchanged (live sweep exit 0), no label written, nothing reopened, no new workflow, no second closing-keyword grammar, no other row. Registered as band state with the three refusals argued on the refused bands own criteria; six SWEEP_COUNT_KEYS counters and an unconditional summary clause; 96 new self-test cases both directions plus a clean control and an unread-page control. Assignee was already set by the PM dispatch and the newest Claim comment names this branch; I wrote neither. One in-file consequence beyond the row: this is the first per-card timeline fetch in the file, so four prose claims that asserted the sweep makes none are narrowed in the same diff to stay true (H11/H15 note and rendered row, H55 banner and rendered row, listRecentIssueEvents docblock, and one summaryLine comment). H13/H18 claims are scoped to the label-application event and are byte-identical.",
  "zone2_measurements": {
    "a_newest_row_and_window": {
      "newest_registered_row_before_edit": "H58 (grep -o \"H5[0-9]: '[a-z]*'\" | sort -u | tail -1)",
      "self_test_before": "exit 0 — check-half-states self-test: 3462 cases pass (on 216b0668, matches the dispatch figure)",
      "self_test_after": "exit 0 — check-half-states self-test: 3559 cases pass (+97 = 96 H59 cases + 1 SUMMARY_CLAUSE_ANCHORS per-anchor case)",
      "h8_merged_window_bound": "MERGED_WINDOW_DAYS = 8 (a TIME cap, not a page cap), listed sort=updated&direction=desc&per_page=100 with MERGED_WINDOW_PAGE_CEILING = 30 as a quota backstop; paging stops once a page oldest updated_at predates the horizon, rows are selected on merged_at. H59 population is that SAME listing (seenMerged), never a second one. Live: 1186 merged PRs read in 14 pages, horizon reached."
    },
    "b_timeline_shape_FALSIFIED": {
      "assumption": "Zone 2(b) assumed the closed event names the merge commit (commit_id).",
      "reading": "FALSIFIED on this board. card #17417 closed 2026-09-10T23:44:35Z, commit_id null, commit_url null, actor os-litant, state_reason null; the PR that declared it (17558) merged 23:44:33Z. Same shape 5 more times (#16274, #17081, #17533, #15141, #17058) and 7 of 7 closed events on one page of the repo-wide issue-event stream (23:32Z-00:10Z) carry commit_id null.",
      "route_change": "The commit leg is written and kept (exact where GitHub does stamp it — a keyword in a PUSHED commit message, H23 surface) but fires on nothing here; the ruling own second leg (closed inside the merge moment, no other actor) is what carries this board. Reported rather than silently substituted."
    },
    "b2_band_derivation": "The ruling merge MINUTE was measured too wide and narrowed to H59_MERGE_CLOSE_BAND_MS = 5000. Corpus: 332 keyword-bound (merged PR, closed card) pairs in the window, 321 within +/-5min: 0-2s 313, 3-5s 7, 31-60s 1; p50 1s, p95 2s, p99 4s. A 60s band admits a different actor — measured specimen: card #16638 entered the closed state 53s after PR #16950 merged, actor os-project-manager, with no window PR binding a keyword to it (a seat closing a card because the delivery landed = the protocol working). Sibling guard added for the same precision reason: at a 60s band 3 candidates, 2 of them explained by a sibling PR own declaration (#16937 by 17190, #16625 by 16698).",
    "c_closing_keyword_reach": {
      "Fixes #N": "BINDS, keyword \"Fixes\"",
      "Part of #N": "binds nothing as a closing keyword (partOfTargets only)",
      "Merging this must not close #N": "BINDS, keyword \"close\" — the negation is invisible to the grammar by design; a bound number is therefore DECLARED whatever the prose says, which is what makes the false-open direction honest. Pinned as a self-test case."
    },
    "d_ablation": "Two legs, each mutate ⇒ prove on disk (injected/removed grep -c plus a blob hash differing from HEAD) ⇒ run ⇒ git checkout HEAD -- ABS_PATH ⇒ prove restored (blob identical, git diff HEAD empty) ⇒ re-run, with trap on EXIT INT TERM. HEAD blob fa880f26b595032795bc544093f4ca6ffa3fba18. Leg 1 (drop the commit_id match): blob d8b6c287, exit 1, 3 of 3559 failed (the two COMMIT-leg attribution cases and the COMMIT-leg row case). Leg 2 (drop the closing-keyword check): blob 239dc655, exit 1, 7 of 3559 failed in BOTH senses — the false-open fixture loses its row source AND the sibling guard goes silent (the misfire leg: an explained closure would become a false-close row). Both restored to fa880f26, exit 0, 3559 pass."
  },
  "live_sweep": {
    "command": "node scripts/pm/check-half-states.mjs (report mode, through the #13544 proxy re-exec)",
    "exit": 0,
    "scope": "swept 487 open pm-/p0-labeled issues, 581 open issues unscoped, 25 open PRs, 1186 recently-merged PRs — 324 half-states found",
    "h59_rows": "1 row, direction (b): H59 #15642 FALSE OPEN — merged PR #16022 binds a closing keyword to that number in its body and the card is STILL OPEN after the merge.",
    "h59_false_close_rows": "0 — the dispatch claims no false close in THIS repo and the live sweep found none. 3 raw candidates existed at a 60s band; the measured 5s band plus the sibling guard remove all three (2 explained by a declaring sibling, 1 a seat close at 53s).",
    "h59_clause": "1106 declared closure(s) and 898 keyword-LESS mention(s) judged against the open listing already in hand; 2815 further mention(s) named a card neither open nor inside H22 3-day closed window ⇒ UNJUDGED rather than clean (this row real reach is the SHORTER of the two windows, never H8 8 days); 0 cards had ONE timeline page read (cap 40 per sweep, no retry on any status), 0 unread, 0 past the cap; a declaring SIBLING PR inside the band EXPLAINS a closure; the closed event carries no commit_id on this board (measured); rows are a LOWER BOUND."
  },
  "tests": "node scripts/pm/check-half-states.mjs --self-test :: exit 0 — before 3462 cases (216b0668), after 3559 cases. | Gate families derived from the FINAL diff: node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack (no paths; it derived 1 path vs merge base 216b06682, three-dot). 39 runnable commands, ALL RUN, every exit code captured BEFORE any pipe, all EXIT=0. | Reconciled: node scripts/pm/dispatch-gates.mjs --ran RAN_FILE --repo objectstack-ai/objectstack :: exit 0 — \"39 derived famil(ies) accounted for — 39 run, 0 NOT-MEASURED (a DERIVED zero — all 39 recorded an exit code and none of them is 3).\" | node scripts/check-closing-keyword-parity.mjs :: exit 0 — \"check-closing-keyword-parity: OK (3 parsers agree on all 9 keywords and both measured separators; sweep found 5 file(s) carrying the grammar across 8349 tracked file(s), all registered).\" — same carriers, no new spelling. | node scripts/check-closing-keyword-parity.mjs --body BODY_FILE :: exit 2 (this mode declaration-exists code) — \"2 closing declaration(s) …, bound by 2 of 3 registered parser(s)\", both of them the ONE keyword on line 1 of the PR body reported once per binding parser; every other keyword+number pair in the body is elided. | pnpm check:pm-dispatch-gates :: exit 0 — \"dispatch-gates self-test: 1678 cases pass.\" | pnpm check:pm-half-states :: exit 0 — \"check-half-states self-test: 3559 cases pass.\" | pnpm check:partof-closing-keyword :: exit 0 — \"check-partof-closing-keyword self-test: 95 cases pass.\" | pnpm check:nul-bytes :: exit 0; plus a manual scan of the edited file with grep -naP for control bytes — no hit. | npx eslint . --no-inline-config --format json :: exit 0 — 6595 files linted, 0 errors, 0 warnings, at 35cf894b. Repo-wide, so no narrowing argument is owed; for the record this repo enables no type-aware linting (calculateConfigForFile on the edited file reports parserOptions.project undefined). | node scripts/pm/check-governed-merges.mjs --branch claude/issue-17466-merged-pr-closing-linkage-row :: exit 0 — \"NOT governed — ordinary queue landing applies to a PR with exactly this file list\" (0 of 1 path hit the register, 5 surfaces). | Ablation: see zone2_measurements.d_ablation — rebuild not applicable (a plain .mjs run directly by node, no dist), on-disk proof by grep counts plus blob hashes, both legs restored byte-identically. | NOT MEASURED locally, declared for CI (dispatch-gates named each): the 47 artifact-roster families, the 11 wide-population families, the 10 pending-changeset families, the 4 workflow-value families and the 1 path-scheduled CI job. Not clean — unread.",
  "gates": {
    "derived": 39,
    "run": 39,
    "nonzero_exits": 0,
    "not_measured_local": "the five CI-only buckets dispatch-gates named; see tests"
  },
  "changeset": "none — scripts/pm/** publishes nothing from any released package; the skip-changeset label is applied (additive REST write, read back: size/l, skip-changeset — size/l is another actor label and was not touched).",
  "mcp_calls": 0,
  "open_questions": [
    {
      "question": "The ruling worded the temporal leg as the merge MINUTE; the corpus says GitHub own keyword closure lands within 5s (p99 4s) and a 60s band demonstrably admits a seat act. I shipped 5s. Does the seat accept the narrowing, or does it want the literal minute with the 6-60s band reported as a separate weaker class?",
      "options": [
        "A — keep H59_MERGE_CLOSE_BAND_MS = 5000 (shipped): 321-pair corpus, p95 2s; the one measured 60s-band specimen (#16638, actor os-project-manager, +53s) is a correct seat act and is NOT reported.",
        "B — restore the literal 60s band: reports that seat act as a candidate false close every sweep until the card ages out of the closed window.",
        "C — 5s fires a row, 6-60s fires a second quieter class: two row kinds for one predicate, and a second thing to trim."
      ],
      "recommendation": "A. 实际业务需求: B was measured wrong on the only specimen the wide band produced, and \"reads like it is useful\" is exactly what a minute would have been here. 项目长远合理性: A keeps one predicate and one sentence. 防 AI 写代码犯错: a band that reports correct seat acts trains every reader to skim the row, which is the tolerance shape this repo refuses. 创业阶段不扩散需求: C adds a second row class for a population measured at one."
    },
    {
      "question": "The card verbatim quote records the sibling-board FALSE OPEN instance (PR #1811 / card #1792) as carrying NO closing keyword either, while the dispatch describes the same instance as binding one. The two cannot both describe that body.",
      "options": [
        "A — fixture (b) carries the DIRECTION shape the dispatch predicate defines (a bound number that stays open), noted in the PR body as not a literal replay (shipped).",
        "B — reread the sibling board and pin whichever the body actually says, which would mean a cross-repo read this dispatch does not authorise."
      ],
      "recommendation": "A. The predicate itself is unambiguous and the fixture exercises it; B buys a fixture detail at the cost of a read outside this card scope. Recorded so the seat can correct the card wording if it wants."
    },
    {
      "question": "H59 direction (a) reach is bounded by H22 closed-card window (3 days) while its population is defined over H8 merged window (8 days), so 2815 mentions per sweep are UNJUDGED by construction. Is that acceptable standing state?",
      "options": [
        "A — accept and disclose every run (shipped): the clause names the shortfall and the shorter window.",
        "B — widen H22 window: changes an existing row input and its cost, which is that row decision.",
        "C — give H59 its own closed-issue window: a second listing for one row."
      ],
      "recommendation": "A. 创业阶段不扩散需求 and 项目长远合理性 both point at reusing the listing the sweep already pays for; a false close that matters is hours old, not days, so the 3-day reach covers the damage model. B and C are both a budget argument someone should make on its own card."
    }
  ],
  "out_of_scope_findings": [
    "noted, not filed: this is the first per-card timeline fetch in check-half-states.mjs, and four in-file prose claims asserted the sweep makes none. Narrowed in this diff rather than left false. Carrier: this PR.",
    "noted, not filed: the closed event actor would discriminate a seat close from GitHub own act if the PR merger were in hand, but merged_by is absent from /pulls LIST rows and this row buys no per-PR detail read; the actor is printed in the row for the reader instead. Carrier: none — a detail read per merged PR is a budget question for a later card.",
    "noted, not filed: H22 closed-card window (3 days) is shorter than H8 merged window (8 days), so H59 direction (a) reach is the shorter of the two. Disclosed in the clause on every run. Carrier: none — widening H22 window is that row decision and nothing is blocked on it.",
    "Dedup before considering any filing: the cached open-board listing (581 open cards, one REST listing pass already paid for by this run) grepped locally for cross-repo-issue-closer, closing-linkage, read-back-issue-state and commit_id, with a positive control that fired (the control returned #17466 and #17160). Only #17466 and #17160 match; commit_id returns 0. Closed cards were NOT searched (the global search endpoint is refused for this session). No admissible (a)/(b)/(c) class finding was found, so nothing was filed."
  ]
}

---

Generated by Claude Code

[os-litant]

ACCEPT — PR #17589 (head 35cf894b) reviewed in-seat at the contract-review tier; NOT governed ⇒ flip point when CI converges (skills seat, session session_01YKEjmbYNvYWJvWGSWx26zK, 2026-09-11T01:00Z).

• Implemented-by: os-dev subagent on claude/issue-17466-merged-pr-closing-linkage-row (claim 5627222190; build tier opus; mode:subagent; report on this thread). Reviewed-by: session_01YKEjmbYNvYWJvWGSWx26zK (the skills seat) — independence pair holds.
• PR shape: draft; base main (216b0668); first line Fixes #17466; size/l + skip-changeset; one footer; the Clause-② carrier readable by --pair 17589 (exit 0). Changed files read three-dot by this seat: exactly scripts/pm/check-half-states.mjs (+694/−12). Path face by this seat: check-governed-merges.mjs --pr 17589 ⇒ NOT governed (exit 0) — ordinary queue landing.
• The whole code half read by this seat: H59 over H8's merged-PR window (seenMerged, the SAME listing — 8-day time cap, 30-page backstop), two directions. Readers reuse the file's own referencedNumbers and closingKeywordTargets — h59DeclaredCloses binds exactly what closingKeywordRe() binds (pinned equal in the self-test), so the negation trap 「Merging this must not close #N」 BINDS and a bound number is DECLARED whatever the prose meant; ⛔ no second grammar (the parity sweep registers the same 5 carriers per the dev; the gate's yaml prerequisite is not on this seat's path — not re-measured here). (b) FALSE OPEN costs no request: a declared number still on the open listing already in hand. (a) FALSE CLOSE: a keyword-less mention whose card sits in H22's closed window, closed inside the measured band after the merge, with no sibling PR in the window declaring it ⇒ ONE timeline page read (cap 40 per sweep, no retry on any status — 共享身份的限流纪律不存在：一次限流信号约束的是「身份」不是「客户端」，而规矩只说了不要重试 —— 2026-09-10 全 fleet 停摆事故 #17374), then attribution by the commit leg (the closed event names the merge commit — exact where GitHub stamps it) or the band leg (circumstantial, and every row says so, printing the actor for the reader's judgement); a full page or an absent event ⇒ UNJUDGED counted. Six counters on SWEEP_COUNT_KEYS; the clause renders unconditionally and names the shorter reach (H22's 3 days, ⛔ never H8's 8), the cap, the unread and the sibling rule. Band state. Placement after H53 in sweepInto, before H45's 「last」. Four in-file prose claims that the sweep makes no per-card timeline fetch are narrowed honestly in the same diff (H11/H15, H55, listRecentIssueEvents, one summaryLine comment) — the H13/H18 claims scoped to label events stay byte-identical: right.
• Zone 2 falsified honestly, twice, and both are the review's best readings: (i) the closed event carries NO commit_id on this board (6 cards + 7 of 7 stream events measured) — the commit leg is kept exact and fires on nothing here, the band leg carries the board; (ii) the ruling's 「merge minute」 is too wide — 321 keyword-bound (merged PR, closed card) pairs: p50 1 s, p95 2 s, p99 4 s, and a 60 s band admits a measured seat act ([finding] rest: import-runner.ts still builds three server-built findData literals in the undeclared wire dialect, through a query: any helper #16638 closed 53 s after PR fix(rest): import-runner builds the canonical QueryAST through a typed findData envelope #16950 by os-project-manager with no declaring PR) ⇒ H59_MERGE_CLOSE_BAND_MS = 5000, plus the sibling guard (two of the three raw 60 s candidates were explained by a declaring sibling PR in the same second).
• The dev's three open questions, answered: (1) A — the 5 s band stands, measured on the corpus; a minute would report correct seat acts and train readers to skim the row. (2) A — the false-open fixture carries the direction's shape, not a literal replay of hotcrm PR feat(ADR-0046): add optional description to package docs #1811 / feat: ADR-0047 implementation (framework) — userFilters spec, reference diagnostics, showcase examples, authoring form #1792 (the card's own wording says that pair carried no keyword either, so it is not a keyword-bound false open — the card's text is imprecise there, ⛔ no cross-repo read owed). (3) A — direction (a)'s reach is the shorter of the two windows and the clause discloses the 2815 unplaced mentions on every run; widening H22 is that row's own budget decision.
• Measured by this seat on the head in a detached worktree: check-half-states.mjs --self-test → 3559 cases pass (3462 on main; +97 — 96 H59 cases both directions incl. the negation pin, a clean control, an unread-page control); check-nul-bytes clean. By the dev: 39 derived / 39 run / 0 NOT-MEASURED (--ran reconciled); check:pm-dispatch-gates 1678; check:partof-closing-keyword 95; eslint whole-repo 6595 files, 0 errors; two-leg ablation with blob hashes (commit leg dropped ⇒ 3 red; keyword check dropped ⇒ 7 red in both senses incl. the sibling guard going silent), restored byte-identically; the new --body mode run on the PR's own body ⇒ exactly the one Fixes binding. Live sweep (report mode, exit 0): one H59 row, direction (b) — finding: check:duration-unit-keys scans packages/spec/src/** only — the same offender shape exists in workspace packages the gate never reads #15642 FALSE OPEN (PR feat(spec): widen check:duration-unit-keys to every workspace package's zod schemas, and convert the one turso offender it finds #16022 binds a keyword to it, the card is still open); 0 false-close rows on this repo (3 raw candidates at 60 s, all explained at 5 s + the sibling guard); the clause: 1106 declared / 898 keyword-less mentions judged, 2815 unplaced, 0 timeline pages needed. CI at this reading on 35cf894b: 48 check runs — 30 success, 17 skipped, 1 in progress (Lint & Repo Gates), none failed.
• noted, not filed (accepted as noted): merged_by is absent from the /pulls list rows, so the actor is printed rather than compared — a per-PR detail read is a later budget question; H22's window vs H8's is disclosed, not widened; finding: check:duration-unit-keys scans packages/spec/src/** only — the same offender shape exists in workspace packages the gate never reads #15642 is the row's first live output — the triage seat's to dispose (relayed in the next round report).
• Landing: NOT governed. Flip point: when Lint & Repo Gates completes green ⇒ this seat flips ready → waits for mergeable_state to settle → enable_pr_auto_merge with mergeMethod: "SQUASH" → reads back added_to_merge_queue and the queue branch. On MERGED: landing record here, pm:dispatched / assignee cleared, the check-half-states.mjs chain advances to [finding] H9's fireability test is a spelling test, so three unfireable Restart-when: shapes pass it green — and #7898 shows H17's trigger index carrying 0 of 11 #17377.

---

Generated by Claude Code

[os-litant]

Landed — PR #17589 merged through the merge queue at 2026-09-11T01:50:50Z (merge commit b64ea610, head 35cf894b), not governed (Governed Surface Queue Guard green on the queue head; 33 success / 16 skipped, none failed): ACCEPT 5627754553 → every check on the head green 01:25Z (31 success / 17 skipped) → flipped ready 01:25:41Z → auto-merge (SQUASH, armed after mergeability settled) → added_to_merge_queue 01:26:49Z → merged 01:50:50Z. Verified on origin/main at 2026-09-11T01:54Z: scripts/pm/check-half-states.mjs carries H59 (both directions, the measured 5 s band, the sibling guard, the six counters, 96 self-test cases); the merge is the queue's squash (one parent, 2f8ad091). Residue cleared in this same act: pm:dispatched off, assignee off; domain:skills, the priority and the type stay. The card was closed by the merge. The row's first live output — #15642 FALSE OPEN (PR #16022 binds a keyword to it and the card is still open) — is the triage seat's to dispose, relayed in the next round report; direction (a) found no false close on this repo. The check-half-states.mjs chain advances: #17377 dispatched into the freed slot this fire. Skills seat, session session_01YKEjmbYNvYWJvWGSWx26zK, R1.

---

Generated by Claude Code