pm-dispatch references: platform-facts sweep — five measured deliverables across four member cards

[hotlong]

Filed by the skills seat (session session_018WuTtyckQa1VcXwgd52JpN) under this shift's maintainer grant (2026-08-13, verbatim, untranslated: 「集中处理skills车道中的所有任务，包括分诊和集中开卡你都可以处理。」). Sweep-promotion of four graded findings plus one absorbed item, all hitting the same criterion: a measured platform/process fact whose deliverable is a self-contained row/line in .claude/skills/pm-dispatch/references/** — no SKILL.md lines, no code. Per the sweep rule: this card is the claim target; member cards keep their labels but are not separately dispatchable while this card is open.

Members and itemized deliverables (落点 | content)

1. PM gate-reading does not pin the run to the PR's CURRENT head — a cancelled run on a superseded head gets rerun, burning a heavyweight cycle and producing a false red #8434 — references/review-checklist.md: two clauses — (a) pin every gate reading to the PR's CURRENT head: read pulls/{n}.head.sha first; a run whose head_sha differs is a reading about a dead commit — neither its green nor its red counts, in either direction; (b) never rerun a run on a superseded head: a cancelled conclusion on a non-current head needs no action at all — the newer push already has its own runs. Plus one references/platform-readings.md row: cheap sibling workflows success + heavyweight carrier cancelled on the same head is the expected signature of ordinary supersession (siblings finish in seconds, carriers take 10–15 min), not evidence of selective failure.
2. [finding] list_issues returns no assignee field, so a lane inventory cannot tell a claimed card from a free one — add to the platform-readings trap table #8374 — references/platform-readings.md row: list_issues never returns assignees (no such fields member; absent from the response either way), so a lane listing cannot distinguish a claimed card from a free one — a listing is a shortlist, and every entry needs a full issue_read before it can be claimed.
3. [finding] GitHub MCP issue_read returns HTML-entity-escaped bodies (' / "), so a body round-trip (read → edit → write) from an MCP-backed seat risks corrupting the original text #8272 — references/platform-readings.md row: MCP issue_read bodies are HTML-entity-escaped while comments come back raw; a body round-trip via issue_write is unsafe from MCP-backed seats — machine-greppable lines (e.g. the blocked-by reverse index) may live in comments instead, so unlock scans must read comments too, and any body rewrite should be verified against the raw body via REST first.
4. finding: a dev session is invisible across GitHub accounts, so archive_session / probe / poke are all unavailable to a successor seat — the hand-off checklist asks for something the successor cannot do #8187 — references/dispatch-runbook.md, extending the existing "SendMessage 直投 not-reachable 是设计而非故障" line: session handles are account-scoped — get_session / archive_session / the trigger-poke route all answer not found for another account's session, and that answer is never a death signal. Plus references/seat-post-protocol.md 退场收尾清单: the outgoing PM archives its own dev sessions before standing down; if a session must be left running, the seat post says so and marks the archive obligation non-transferable.
5. [finding] Three mechanizable items from a 16-hour spec-lane shift: os-dev's background-wait stall (6 instances), the verify-lock convention devs invented, and enable_pr_auto_merge's silent no-op #8294 item 3 only — references/platform-readings.md rows: (a) enable_pr_auto_merge without an explicit mergeMethod: "SQUASH" is a silent no-op on this repo — it falls back to the forbidden merge-commit method; the echo's empty method: field is the failure signal; (b) measured data points worth a line each: single-container heavy-verify concurrency sweet spot ≈3; merge-queue landing latency ~15–30 min per PR, serialized. Items 1–2 of [finding] Three mechanizable items from a 16-hour spec-lane shift: os-dev's background-wait stall (6 instances), the verify-lock convention devs invented, and enable_pr_auto_merge's silent no-op #8294 are absorbed elsewhere (the os-dev role-file card) — ⛔ this PR must NOT close [finding] Three mechanizable items from a 16-hour spec-lane shift: os-dev's background-wait stall (6 instances), the verify-lock convention devs invented, and enable_pr_auto_merge's silent no-op #8294; reference it in prose only, never in a closing position, and do not create a development link to it.

Constraints

• File surface: .claude/skills/pm-dispatch/references/** ONLY. ⛔ No SKILL.md edits (line ratchet untouched). Rows land in the existing table/rule style, self-contained, ⛔ no issue-number references in the operational text (check:pm-skill-id-lint).
• Serial: references/dispatch-runbook.md is held by the dispatch-strategy card's in-flight work — this sweep dispatches after that PR lands and rebases over it.
• Model: claude-fable-5 mandatory (pm-dispatch skill root); the 2026-08-13 quota-exhaustion exemption applies (floor opus, downgrade + reason recorded in the claim comment).
• Merge path: ADR-class three-piece — draft PR, ⛔ no auto-merge / queue / ready-flip by any AI seat; the maintainer lands it by hand.
• PR first line: Fixes #8434, Fixes #8374, Fixes #8272, Fixes #8187 (all four fully delivered here). The PR body carries the per-member 落点|before|after list (sweep five-condition rule).

[hotlong]

Evidence refinement for member item 5's enable_pr_auto_merge row, measured live this shift (2026-08-13 ~14:29–14:30Z, two attempts): passing an explicit mergeMethod: "SQUASH" on this repo still echoes (method: MERGE, …) — the parameter is coerced/overridden, most plausibly by the merge queue governing main (whose landed history is one commit per PR regardless). The row's wording should therefore anchor on the failure signal, not the method name: an empty method: in the echo is the silent-no-op signal; a non-empty echo (even MERGE) has landed PRs on this repo through the queue. Keep the "always pass an explicit mergeMethod" instruction, but do not assert the echo will read back SQUASH — the implementing dev should fold this into the same row.

---

Generated by Claude Code

[hotlong]

Scope amendment (member item 6 added) — from the Part-of/auto-close investigation card's grading: the protocol fallback half of that card (make "re-pull the lane inventory after every merge and diff it against the expected state" an explicit post-merge step where the landing text lives, i.e. references/landing-operations.md) moves INTO this sweep. Reason: that one line is .claude/skills/pm-dispatch/** text and therefore under the fable-mandatory clause — it rides this sweep's fable dispatch instead of upgrading a whole investigation card's model for one sentence. The step is justified by the measured incident alone (it is what caught the wrongly-closed card), independent of what the mechanism investigation concludes. The investigation card itself stays scoped to the timeline-mechanism read + the scripts/pm/check-half-states.mjs contradiction guard, on opus.

---

Generated by Claude Code

[hotlong]

Scope amendment (member item 7 added) — from the Part-of/auto-close investigation's ACCEPT: the author-side keyword-hygiene rule, one deliverable in references/review-checklist.md's PR-shape section (beside the existing Fixes-vs-Part of first-line rule): GitHub's closing-keyword parser binds close/fix/resolve (+inflections, but NOT gerunds) to a #N on the same line and ignores all surrounding prose — negations, modals, and warnings included; the sentence written to prevent an auto-close is exactly what performed one. Therefore: ⛔ never put a closing keyword next to another open card's number in a PR body; write "card N is not addressed here" / "out of scope: card N", or put the keyword in backticks (measured: code spans do not fire). Reviewers check the PR body for this shape before ready-flip on any Part of PR; a report-only detector for the contradiction also patrols it.

Evidence and phrasing source: the mechanism record now in scripts/pm/check-half-states.mjs's header (four live readings) — the implementing dev should keep the reference row self-contained per the id-lint rule.

---

Generated by Claude Code