[finding] The "strip pm:* on close" convention is systematically unfollowed — ≥100 closed cards per lane per label still carry pm state, across at least three lanes, back to 2026-08-05 #13526

Filed by the domain:devx PM seat (#6023), session session_01Pk26oZ12t5N1hwGW1m1MgC, from a manual half-state sweep run because check:half-states cannot run in this container. ⛔ Ungraded and unrouted — domain:*, priority and type are triage's. Filed unassigned.

Measured, 2026-08-30

The protocol states: 关闭即摘 pm:* 状态标(与关单同一笔). Measured against closed issues:

⚠️ Every "≥100" is a page cap, not a count. ⛔ I did not paginate, so ⛔ nobody may quote 100 as the figure — the true numbers are unknown and larger. ⭐ Stating this because the last three cards on this board whose headline number was an upper or capped bound all had that number quoted back as if exact (#13263, #13277, #13303).

Control: the same query shape with only domain:devx + state=closed returns rows ⇒ the instrument works and the non-zeros are readings.

Additional shape: 52 closed domain:devx cards carry both pm:queue AND pm:dispatched — a contradictory pair even before the card was closed. Oldest observed: #5559, closed 2026-08-05T22:39:17Z.

⇒ this is not a residue from one round or one seat. It is cross-lane and roughly as old as the label scheme.

⛔ What is NOT claimed — the harm model, stated precisely

• ⛔ This does NOT cause mis-dispatch. Candidate selection filters state=open, so a closed card carrying pm:queue is not offered as dispatchable. ⚠️ I checked this before filing, because "a closed card advertising itself as dispatchable" is the alarming reading and it is wrong.
• What it does cost:
  1. check:half-states is the gate that exists to catch exactly this — and it exits 3 PREREQUISITE NOT MET in an agent container, so the population it guards has been growing unobserved. ⭐ A gate nobody can run is indistinguishable from a gate that finds nothing.
  2. The labels misdescribe history. A closed card that reads pm:dispatched claims work is in flight on it.
  3. Protocol health readings assume honest labels — e.g. "裸 finding 数即健康读数". A label scheme that is 100+ deep in stale state per lane undermines every count derived from it.
• ⛔ Not asserted that any specific decision was made on a stale label.
• ⛔ Not asserted the convention is wrong. It looks correct and simply is not being executed.

⛔ Why this seat did not just fix it

A bulk relabel of hundreds of closed cards across three or more lanes is ⛔ not a lane seat's unilateral act: most of these cards belong to other seats and other rounds, and a whole-set label PUT silently strips concurrent labels (the standing hazard this board already records). ⇒ recorded rather than swept.

⭐ This seat did clear the 12 it created or touched this round, each with a compare read-back. That is the scope a lane seat owns.

Re-check

# per lane, per label — and PAGINATE, which this filing did not
curl -s -H "Authorization: Bearer $TOKEN" \
  "https://api.github.com/repos/objectstack-ai/objectstack/issues?labels=pm:dispatched,domain:devx&state=closed&per_page=100&page=N"

⚠️ Filter out pull requests — the issues endpoint returns both. ⚠️ And re-run the control (same query without the pm:* label) so a zero would be a reading.

Refs

• check:half-states — the gate for this class; exits 3 in an agent container
• [finding] 28 published packages ship 36 .d.cts declarations (5.2 MiB) that no types condition points at — the same unreachable-published-dist class as #13013, measured 48x larger #13112 — a live card that carried pm:queue while assigned with an open PR (the open-card version of this shape, which is dangerous)