Skip to content

fix(pm): re-pin H22 closed-issue update rate from a card-only measurement of the stream the patrol actually pages #17627

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
os-litant merged 1 commit into from
Sep 11, 2026
+61 −35
Merged

fix(pm): re-pin H22 closed-issue update rate from a card-only measurement of the stream the patrol actually pages #17627

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
96 changes: 61 additions & 35 deletions scripts/pm/check-half-states.mjs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16471,18 +16471,40 @@ export const CLOSED_ISSUE_WINDOW_PAGE_CEILING = 40;
* disagree with.
*
* ⛔ This is closed-issue UPDATE EVENTS per day, NOT closures per day. The two
* differ by ~17x on this board and confusing them is precisely the defect this
* differ by ~13x on this board and confusing them is precisely the defect this
* window carried (see the docblock above).
*
* read 2026-09-06T21:59Z, `GET /repos/{repo}/issues?state=closed&sort=updated`
* window 400 rows spanning 0.964 days
* rate 400 / 0.964 = ~415 closed-issue update events/day
* (re-read at depth: 800 rows / 2.051d = ~390/day, 1200 rows / 3.092d = ~388/day)
* ⚠️ AND IT IS MEASURED OVER THE STREAM THE PATROL PAGES, which is NOT the one
* an agent container gets from the same URL. The scheduled runner's token
* declares `contents: read` + `issues: write` and no pull-request scope, and its
* `/issues` pages come back CARD-ONLY; the byte-identical request from a proxied
* container returns 49.6% pull requests. Measured, not inferred from the
* workflow file: the 2026-09-11T01:55Z run reports 5 pages, 428 in-window
* closures and ~141.3/day, and replaying THIS pager over a container read of the
* same hour gives 5 pages / 426 / ~139.4 on the CARD-ONLY slice against 9 pages /
* 426 / ~256.2 on the whole one — and 428 in-window closures cannot come out of 5
* PR-inclusive pages at all, since only ~250 of those 500 rows are cards.
*
* ⛔ So a re-pin read from a container rates a different population than the
* sweep it is checked against, and the 0.34 factor the DRIFTED clause has been
* printing is TWO things multiplied: a real ~0.64 slowdown of the stream since
* 2026-09-06, and a ~0.53 population difference that is not tempo at all.
*
* read 2026-09-11T04:22Z, `GET /repos/{repo}/issues?state=closed&sort=updated`
* — card-only rows, the population the runner pages
* window 500 rows spanning 3.586 days — the 5 pages a 3-day horizon costs
* rate 500 / 3.586 = ~139 closed-issue update events/day
* (re-read at depth: 200 rows / 1.352d = ~148/day, 400 rows / 2.769d = ~144/day,
* 605 rows / 3.969d = ~152/day)
*/
export const MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY = 415.1;
export const MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY = 139.4;

/** WHEN `MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY` was measured. */
export const MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY_AT = '2026-09-06';
/**
* WHEN `MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY` was measured — and therefore when
* it EXPIRES on the calendar whatever a sweep observes: `RATE_PREMISE_STALE_DAYS`
* past this date is 2026-10-11.
*/
export const MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY_AT = '2026-09-11';

/**
* H22's premise, in the shape `classifyRatePremise` judges (#16393).
Expand Down Expand Up @@ -22702,12 +22724,12 @@ async function selfTest() {
t('windows: …and 2.73d at the rate measured 8 days later — same cap, 25% more window', Number(windowCoverageDays(300, MEASURED_COMMITS_PER_DAY).toFixed(2)), 2.73);
t('windows: …which is why H23\'s cap became a TIME cap', COMMIT_WINDOW_DAYS, 3);
// ⛔ H22's is the one whose DIVISOR was the surprise: its rows are consumed by
// closed-issue UPDATE activity, not by closures, and the two differ ~39x at
// the divisor re-pinned 2026-09-06, against the ~10.7 closures/day measured
// closed-issue UPDATE activity, not by closures, and the two differ ~13x at
// the divisor re-pinned 2026-09-11, against the ~10.7 closures/day measured
// when the defect was found. The old 4-page cap read as a closure window and
// was nothing of the kind.
t('windows: H22\'s old 4-page cap bought only 0.96d of UPDATE recency at the 2026-09-06 rate', Number(windowCoverageDays(400, MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY).toFixed(2)), 0.96);
t('windows: …and the honest divisor is ~39x the closure rate it read as', Number((MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY / 10.7).toFixed(1)), 38.8);
t('windows: H22\'s old 4-page cap bought 2.87d of UPDATE recency at the 2026-09-11 rate', Number(windowCoverageDays(400, MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY).toFixed(2)), 2.87);
t('windows: …and the honest divisor is ~13x the closure rate it read as', Number((MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY / 10.7).toFixed(1)), 13);
t('windows: …which is why H22\'s cap became a CLOSURE-time cap', CLOSED_ISSUE_WINDOW_DAYS, 3);
t('windows: …and its horizon is seen by 12 consecutive runs', sweepOverlap(CLOSED_ISSUE_WINDOW_DAYS), 12);
// #13499 — H8's window is no longer a page cap at all, so the arithmetic that
Expand Down Expand Up @@ -22878,10 +22900,13 @@ async function selfTest() {
t('H22 window: a 1-day-old closure is inside', issueClosedWithinWindow(closed13606(1, 1), CH), true);
// Direction 2 — the row the OLD 4-page cap structurally could not reach. The
// cap covered ~0.96d of UPDATE recency at the divisor pinned 2026-09-06
// (~2.12d at the one pinned when the repair landed), so a card closed 2.5
// days ago and untouched since fell out of it: not late, GONE. Measured cost
// at the time of the repair: 19 residue carriers closed within 3 days missed.
t('H22 window: the old 4-page cap reached only ~1.0 days of update-recency', windowCoverageDays(400, MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY) < 2.5, true);
// (~2.12d at the one pinned when the repair landed, ~2.87d at the one pinned
// 2026-09-11), so a card closed 2.5 days ago and untouched since fell out of
// it: not late, GONE. Measured cost at the time of the repair: 19 residue
// carriers closed within 3 days missed. ⚖️ The days this cap buys move with
// every re-pin — which is the whole argument against a page cap — so what is
// pinned below is its SHORTFALL against the horizon, not a fixed count of days.
t('H22 window: the old 4-page cap reached ~2.9 days of update-recency, still short of the 3-day horizon', windowCoverageDays(400, MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY) < 3, true);
t('H22 window: …but the 3-day closure horizon admits that card', issueClosedWithinWindow(closed13606(2, 2.5), CH), true);
// …and the window still HAS an edge; a boundary that admits everything is none.
t('H22 window: a 4-day-old closure is outside', issueClosedWithinWindow(closed13606(3, 4), CH), false);
Expand Down Expand Up @@ -23103,28 +23128,29 @@ async function selfTest() {
const rateStream16393 = (step) =>
closedStream(Array.from({ length: 30 }, (_, i) => ratePage16393(step * (i + 1))));

// 2x the pin: 0.11-day steps, so the horizon falls on page 28 — 2800 rows
// read across a 2.97-day span is ~942.8/day against a pin of 415.1.
// Well past the band: 0.11-day steps, so the horizon falls on page 28 — 2800
// rows read across a 2.97-day span is ~942.8/day against a pin of 139.4.
const driftedStats16393 = {};
const driftedRows16393 = await listRecentlyClosedIssues(driftedStats16393, NOW13606, rateStream16393(0.11));
t('#16393 stream: the pass stops on the first page past the horizon', driftedStats16393.closedPages, 28);
t('#16393 stream: …and the rate is over the rows READ', Math.round(driftedStats16393.closedRateObserved * 100) / 100, 942.76);
t('#16393 stream: ⛔ …not over the rows ADMITTED, which is a different number', Math.round(observedRatePerDay(driftedRows16393, 'updated_at') * 100) / 100 === Math.round(driftedStats16393.closedRateObserved * 100) / 100, false);
t('#16393 stream: …a board running 2x the pin FIRES the alarm', h22Premise(driftedStats16393.closedRateObserved).state, 'drifted');
t('#16393 stream: …and the factor is stated so it can be acted on', h22Premise(driftedStats16393.closedRateObserved).message.includes('a factor of 2.27'), true);
t('#16393 stream: …with both rates beside it', h22Premise(driftedStats16393.closedRateObserved).message.includes('observed ~942.8 closed-issue updates/day') && h22Premise(driftedStats16393.closedRateObserved).message.includes('= 415.1/day'), true);
t('#16393 stream: …a board running past the band FIRES the alarm', h22Premise(driftedStats16393.closedRateObserved).state, 'drifted');
t('#16393 stream: …and the factor is stated so it can be acted on', h22Premise(driftedStats16393.closedRateObserved).message.includes('a factor of 6.76'), true);
t('#16393 stream: …with both rates beside it', h22Premise(driftedStats16393.closedRateObserved).message.includes('observed ~942.8 closed-issue updates/day') && h22Premise(driftedStats16393.closedRateObserved).message.includes('= 139.4/day'), true);
// ⛔ The rule that makes this an alarm rather than a self-healing constant.
t('#16393 stream: ⛔ a drifted sweep does NOT rewrite the pin', MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY, 415.1);
t('#16393 stream: …nor its date', MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY_AT, '2026-09-06');

// 1x the pin: 0.25-day steps, horizon on page 13 — 1300 rows over 3.0 days is
// ~433.3/day, a factor of 1.04, and the alarm stays silent. ⚖️ It is the very
// stream that ran 2x the pin until the 2026-09-06 re-measure: the board moved
// by that much in six days, and the fixture did not move at all.
t('#16393 stream: ⛔ a drifted sweep does NOT rewrite the pin', MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY, 139.4);
t('#16393 stream: …nor its date', MEASURED_CLOSED_ISSUE_UPDATES_PER_DAY_AT, '2026-09-11');

// 1x the pin: 0.75-day steps, horizon on page 5 — 500 rows over 3.0 days is
// ~166.7/day, a factor of 1.20, and the alarm stays silent. ⚖️ The step is
// stated RELATIVE to the hand-pinned rate, so it moves when the pin does — it
// was 0.25 against the 415.1 pinned 2026-09-06 — because a fixture asserting
// "in band" against a rate nobody re-measured passes for a board nobody read.
const okStats16393 = {};
await listRecentlyClosedIssues(okStats16393, NOW13606, rateStream16393(0.25));
t('#16393 stream: an in-band board reaches its horizon in thirteen pages', okStats16393.closedPages, 13);
t('#16393 stream: …and observes a rate at the pin', Math.round(okStats16393.closedRateObserved * 100) / 100, 433.33);
await listRecentlyClosedIssues(okStats16393, NOW13606, rateStream16393(0.75));
t('#16393 stream: an in-band board reaches its horizon in five pages', okStats16393.closedPages, 5);
t('#16393 stream: …and observes a rate at the pin', Math.round(okStats16393.closedRateObserved * 100) / 100, 166.67);
t('#16393 stream: …which raises NO alarm', h22Premise(okStats16393.closedRateObserved).state, 'ok');
t('#16393 stream: a board with nothing dateable observes no rate at all', (await (async () => { const s = {}; await listRecentlyClosedIssues(s, NOW13606, closedStream([[{ number: 1, updated_at: 'nope' }]])); return s.closedRateObserved; })()), null);
t('#16393 stream: …and an empty board likewise', (await (async () => { const s = {}; await listRecentlyClosedIssues(s, NOW13606, closedStream([])); return s.closedRateObserved; })()), null);
Expand All @@ -23137,13 +23163,13 @@ async function selfTest() {
const clause16393 = (observed, extra = {}) => h22RatePremiseClause(observed, { nowMs: NOW13606, ...extra });
t('#16393 clause: a drifted divisor is marked loud', clause16393(942.76).startsWith('⚠️ RATE PREMISE DRIFTED'), true);
t('#16393 clause: …and names the re-measure act', clause16393(942.76).includes('re-pinned by hand, from a fresh measurement'), true);
t('#16393 clause: an expired pin is marked loud too', clause16393(415.1, { nowMs: Date.parse('2026-11-01T00:00:00Z') }).startsWith('⚠️ RATE PREMISE EXPIRED'), true);
t('#16393 clause: an expired pin is marked loud too', clause16393(139.4, { nowMs: Date.parse('2026-11-01T00:00:00Z') }).startsWith('⚠️ RATE PREMISE EXPIRED'), true);
t('#16393 clause: an unobserved rate is marked loud, never silent', clause16393(null).startsWith('⚠️ RATE PREMISE UNOBSERVED'), true);
t('#16393 clause: an in-band divisor is NOT marked loud', clause16393(433.33).includes('⚠️'), false);
t('#16393 clause: an in-band divisor is NOT marked loud', clause16393(166.67).includes('⚠️'), false);
// ⛔ …but it still SPEAKS. A check that is only visible when it fires cannot
// be told apart from a check somebody deleted — `renderRatePremise`'s rule.
t('#16393 clause: …and still speaks, so a deleted check cannot pass for a healthy board', clause16393(433.33).startsWith('rate premise OK'), true);
t('#16393 clause: …naming both rates even when it agrees', clause16393(433.33).includes('observed ~433.3/day') && clause16393(433.33).includes('= 415.1/day'), true);
t('#16393 clause: …and still speaks, so a deleted check cannot pass for a healthy board', clause16393(166.67).startsWith('rate premise OK'), true);
t('#16393 clause: …naming both rates even when it agrees', clause16393(166.67).includes('observed ~166.7/day') && clause16393(166.67).includes('= 139.4/day'), true);

// ---- The clause on the summary line. Only time-INDEPENDENT properties are
// asserted here, for the reason above; the verdict wording is pinned by the
Expand Down
Loading