pm(half-states): name the POPULATION each H22 docblock number was measured on

[claude]

Fixes #17626

H22 pages its closed-card window through closedWindowPagePath — GET /repos/{repo}/issues?state=closed&sort=updated&direction=desc&per_page=100&page=N. Two callers issue the byte-identical request and get different POPULATIONS: the scheduled patrol runner's pages come back card-only, a proxied agent container's come back 49.6% pull requests. Four prose sites in scripts/pm/check-half-states.mjs quoted the container's stream while describing the patrol's, so the measurement recipe they carry — followed correctly — yields the wrong number. This is prose only: no constant moves (MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY stays 139.4, CLOSED_ISSUE_WINDOW_PAGE_CEILING stays 40), no H-row, no self-test change.

The boundary this card asked to survive the fix

The population difference itself is MEASURED, through the sweep's own numbers. The runner's permissions: block as the MECHANISM is NOT claimed — the runner's raw response was never read. Every new sentence below is written that way, and the ⚠️ Cost note says it in those words.

The four sites, before and after

1. The ⚠️ Cost note under CLOSED_ISSUE_WINDOW_DAYS

• before: «46% of the rows this stream returns are PULL REQUESTS, filtered out after paging. The horizon is therefore reached in roughly twice the pages a card-only stream would need.»
• after: the 46% is named as a container reading (49.6%, read 2026-09-11T04:22:36Z, 12 pages / 1,200 rows), the patrol's stream is named card-only, and the runner side is quoted from the sweep's own numbers — the 2026-09-11T01:55:32Z run (34552285557, 2f8ad091) reports 5 pages, 428 in-window closures, reach 2026-09-07; 428 in-window closures cannot come out of 5 PR-inclusive pages, since only ~250 of those 500 rows would be cards. The 2026-09-09 run is the same shape at 6 pages.

2. The same docblock's present-tense divisor sentence

• before: «re-pin (check-half-states: re-measure and re-pin MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY (188.3 of 2026-08-31 — the new alarm reads a factor of 2.11 on objectstack) #16419: the divisor below now reads 415.1)» — present tense against a divisor that now reads 139.4.
• after: restated as history — check-half-states: re-measure and re-pin MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY (188.3 of 2026-08-31 — the new alarm reads a factor of 2.11 on objectstack) #16419 re-pinned 415.1 on 2026-09-06 from the container stream; check-half-states: H22's MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY pin has drifted 2.5x, so every sweep's own coverage sentence is misdescribed by that factor #17254 re-pinned on 2026-09-11 from the card-only slice, and the divisor below now reads 139.4. The ~13 pages this note derives is also marked as a 2026-09-06 container reading, and explicitly not comparable to today's 5 pages on population alone — the two readings are five days AND one population apart, with the decomposition left where PR fix(pm): re-pin H22 closed-issue update rate from a card-only measurement of the stream the patrol actually pages #17627 put it, in MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY.

3. CLOSED_ISSUE_WINDOW_PAGE_CEILING's derivation

• before: «4,000 rows reached 12.8 days on the read above (~312/day averaged over that depth), against a 3-day horizon: ~4.3x headroom.»
• after: the derivation stands, with a ⚠️ POPULATION note beneath it — that read was container-taken, so only ~2,000 of its 4,000 rows were cards. 40 pages return 4,000 rows either way, but on the runner all 4,000 are cards, so the same ceiling reaches about twice as far back: ~25 days rather than 12.8, ~8x headroom rather than ~4.3x. Written as a note, not a change: the direction is more headroom and not less, 40 is unchanged, and re-deriving it properly needs a card-only read at depth that nobody has taken.

4. listRecentlyClosedIssues's rateRows comment

• before: «the pin this rate is checked against was measured over the RAW stream (400 rows / 0.964 days)» — a description of the 188.3 pin, two re-pins ago.
• after: the pin's population as it now stands (500 rows / 3.586 days, read 2026-09-11), plus why the distinction lives at this line: on the patrol runner the stream is card-only, so rateRows holds cards and the observed rate is counted on the same population the pin was; the same request from a container returns 49.6% pull requests and rateRows would count those too — a rate on one population checked against a pin on another. The READ vs ADMITTED argument (check:react-declaration-parity 是唯一没接进任何 workflow 的源码审计门禁,且无 MANIFEST 时静默 skip 退出 0 —— 它现在永远不可能红 #4690 with the numbers the right way up) is unchanged.

The re-measure recipe the file now carries

Stated once, in the ⚠️ Cost note, and pointed at from sites 3 and 4:

A container read is NOT the patrol's stream. Filter pull requests OUT of every page BEFORE counting pages or rating rows, and quote the card-only slice.

On the 2026-09-11T04:22:36Z container read that slice gives 5 pages / 426 rows / ~139.4 per day, against 9 pages / 426 / ~256.2 per day on the whole stream. The slice reproduces the runner's own 5 pages; the raw stream does not. That is the operative half — a seat following the old note from a container reproduced #17254's false RATE PREMISE DRIFTED alarm every time.

Verification

All of it taken on the final commit 034c09ace. Gate families derived with node scripts/pm/dispatch-gates.mjs --commands (no paths passed — it takes its own change set from the merge base), each command run with its exit code captured before any pipe, then reconciled:

✓ dispatch-gates --ran: 40 derived famil(ies) accounted for — 40 run, 0 NOT-MEASURED (a DERIVED zero — all 40 recorded an exit code and none of them is 3).

Every one of the 40 exits 0, including pnpm check:pm-dispatch-gates (detached — it runs past the foreground cap — and waited on with tail --pid). Named readings:

• pnpm check:pm-half-states — self-test 3656 cases pass on the base file (8fa3fe63d) and 3656 cases pass after; the count is unchanged because no pin moved and no self-test assertion names any of the rewritten sentences (grepped 46% of the rows, 415.1, 0.964, 312/day across the tree — the only hits outside this file's own prose are two historical mentions this PR does not touch).
• pnpm check:nul-bytes — check-nul-bytes: OK (scanned 8451 text file(s) ... no raw ASCII control bytes); plus a self-scan of the edited file, grep -naP '[\x00-\x08\x0b\x0c\x0e-\x1f\x7f]', zero hits.
• node scripts/check-comment-mask-corpus.mjs — 6640 files, 0 disagree, 0 unparseable.
• node scripts/check-scripts-symbol-anchors.mjs — 3186 anchors across 248 scripts resolve.
• The remaining 36 derived commands exit 0.

Repository-wide scans (pnpm lint first among them) are CI's run, not this PR's.

Acceptance notes

• noted, not filed: MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY's docblock (landed by PR fix(pm): re-pin H22 closed-issue update rate from a card-only measurement of the stream the patrol actually pages #17627, outside this card's declared surface) states the runner's token scopes as part of its population paragraph. It does label the reading «Measured, not inferred from the workflow file», so it is not a mechanism claim, but it sits closer to one than the ⚠️ Cost note now does. Successor: the next seat re-pinning that divisor.
• noted, not filed: the card names a second possible route — declaring pull-requests: read on half-state-patrol.yml so the runner and a container read the same stream. This PR takes the recipe route the dispatch chose; the permission route is untouched and its cost (roughly doubling H22's page bill) is unmeasured.
• skip-changeset: scripts/pm/** is not published — no package's files[] ships this script.

---

Generated by Claude Code

---

Generated by Claude Code