skills(pm): land the ruled REST-channel prescriptions and closed-inclusive dedup (ceiling 87 → 93) (#14236)

* skills(pm): refuse the issues-list cursor and require a count cross-check

The list-cards row prescribed an endpoint but asserted nothing about the
completeness of what it returns. Measured 2026-08-31/09-01 in two repos, a
walk that follows `Link: rel="next"` reports itself exhausted at 102 items
while the page-number walk on the identical query returns 287 / 448 and
matches `open_issues_count`. The failure is silent and toward the dangerous
side, and a keyword positive control is structurally blind to it: a known-hit
card sitting in the first page answers correctly over a truncated corpus.

Lands both refusals (do not follow the cursor; an enumeration whose total was
never cross-checked is not a reading), the `&page=N` short-page termination
condition, the `open_issues_count`-minus-open-PRs cross-check, the
cause-unknown boundary (GitHub's cursor vs the egress proxy rewriting the
header — the prescription removes the dependency, it does not explain it),
and the keyword-control blind-spot note.

Paid in place: the red-window restatement is retired and its pointer folded
into the section heading at zero line cost, so the red-window rule now has
exactly one home in platform-readings.md's quota section. The ceiling moves
87 -> 93 on the maintainer ruling quoted verbatim beside the row.

* skills(pm): make the dedup prescription state that dedup includes closed cards

Dedup asks whether a subject has been carded before, and that answer lives
disproportionately in closed cards, so an open-scoped scan is blind by
construction rather than by malfunction. Measured 2026-08-31 in both
directions on one day: an open-scoped dedup missed a same-subject card that
had been ruled and closed six hours earlier and a duplicate was filed and
dispatched; a dedup that included closed cards hit a same-shape closed card
and no duplicate was opened.

The obligation lands in the prescription text itself -- zero new quota and no
new mechanism -- so open-scoped dedup becomes an exception that has to be
declared rather than the default. The state and label listing reads keep
`state=open` untouched: that convention governs state reads, where a closed
card is archive, and dedup is not a state read.

---------

Co-authored-by: Claude <noreply@anthropic.com>

Author: os-justin

.claude/skills/pm-dispatch/references/rest-channel.md

@@ -17,6 +17,14 @@ payload → REST → MCP/GraphQL** 的策略住那里,本表是逐操作的通
 
 - ✓ 按标签/状态列卡:`GET /repos/{o}/{r}/issues?state=open&labels=a,b&per_page=N` —— `labels=` 是**真
   AND**(MCP `list_issues` 的 labels 数组是 OR;语义相反那条住 `platform-readings.md`)。
+- ⛔ 完整性自证,**不跟 `Link: rel="next"`**:2026-08-31/09-01 两仓实测,游标自称枯竭停在 102,页号走
+  法得 287/448、与 `open_issues_count` 符 ⇒ 改走 `&page=N` 到短页为止,再用 `GET /repos/{o}/{r}` 的
+  `open_issues_count`(含 PR,减开放 PR)核总数;⛔ 未核总数的枚举不是读数。⚠️ 成因未知(GitHub
+  游标 vs 出口代理改写 `Link`):处方**不依赖该头**,⛔ 不把「已修」写成「已解释」。⚠️
+  关键词阳性对照**结构上**看不见本类(必中卡落在首页照样绿),只有计数核对逮得住。
+- ⛔ **查重必须含 closed**:卡刚关闭时最易被重开(2026-08-31 双向实测:开域查重漏掉 6 小时前
+  刚关闭的同题卡 ⇒ 重复派发;含闭卡那次命中已关闭同形卡 ⇒ 免开重复)。开域查重
+  是**要申报的例外**,不是默认;状态 / 标签列卡读仍 `state=open` —— 那是状态读,不是查重。
 - ✓ 卡 / PR 元数据:`GET .../issues/{n}` · `GET .../pulls/{n}`(assignees、labels、body 齐全 —— MCP
   `list_issues` 永不返回 assignees,这条差别本身就是走 REST 的理由)。
 - ✓ 整条评论线:`GET .../issues/{n}/comments?per_page=100`。
@@ -40,7 +48,7 @@ payload → REST → MCP/GraphQL** 的策略住那里,本表是逐操作的通
 - ✓ 请求复审 `POST .../pulls/{n}/requested_reviewers` · 开 PR `POST .../pulls`(带 `draft=true`;GraphQL 池为 0
   的同一分钟里实测开得出 draft PR ⇒ 交付不必等重置)。
 
-## 不可迁移 —— 只有这几件,围着它们排计划
+## 不可迁移 —— 只有这几件,围着它们排计划(红窗守候规则住 `platform-readings.md` 配额段)
 
 1. **draft → ready 翻转**:GraphQL-only mutation;出口代理只放钉住的 PR-review GraphQL 集(实测拒绝)。
    判据 = REST update-a-pull-request 只收 `title`/`body`/`state`/`base`/`maintainer_can_modify`,**无 `draft`**
@@ -55,8 +63,6 @@ payload → REST → MCP/GraphQL** 的策略住那里,本表是逐操作的通
    GraphQL-only。拿不到时当轮改走「在目的仓重建」配方 —— 纯 REST、配额免疫,配方住
    `platform-readings.md`。
 
-`until remaining > 阈值` 的守候只留给这几件,⛔ 其余一切不为配额空等。
-
 ## 第三桶 —— git 零配额等价物(先问 git,再问 REST)
 
 分支存在性、合并队列分支、按内容判落地、squash 验证:四条 `ls-remote` / `git log` 拼写与各自

scripts/pm/check-skill-line-ratchet.mjs

@@ -311,7 +311,29 @@ export const CEILINGS = new Map([
   // file is the lookup table it points at, so the policy flip did not have to
   // grow the hot file. Set at landed line count (headroom 0, same convention
   // as the entries above).
-  ['.claude/skills/pm-dispatch/references/rest-channel.md', 87],
+  // Raised 87 → 93 by a maintainer ruling (2026-09-01, 总监批 #22) that names
+  // this ceiling and orders its own encoding — the ratchet's own legitimate
+  // exit, and the raising PR quotes the ruling comment. Verbatim and
+  // untranslated (one quotation, wrapped only to fit this comment):
+  //   「**B 获授权**:`rest-channel.md` 天花板 87 → **≤93 行**,提额 PR 引用本裁决
+  //   评论(棘轮自己的合法出口);两条处方按 dev 已测的忠实版落(两个拒绝 +
+  //   `&page=N` 短页终止条件 + `open_issues_count` 减开放 PR 交叉核对 + 成因未知
+  //   边界 + 关键词对照结构性盲区注记)」
+  // and, same batch, the content-ownership call that says what is paid in place:
+  //   「**C 同批**:退役 L58 对红窗常设规则的复述 —— 内容归属裁定:**红窗规则由
+  //   `platform-readings.md` 配额段独家持有**,`rest-channel.md` 只留指路」
+  // Paid in place: that restatement and its adjacent blank are retired (−2), and
+  // the pointer the ruling permits is folded into the 不可迁移 heading at zero
+  // line cost — so the red-window rule now has exactly one home. Spent: +8, the
+  // two enumeration-completeness prescriptions (both refusals, the `&page=N`
+  // short-page termination, the `open_issues_count`-minus-open-PRs cross-check,
+  // the cause-unknown boundary, and the note that a keyword positive control is
+  // structurally blind to this class) plus the closed-inclusive dedup obligation
+  // the same batch ruled onto this file (「提额预算一次用足」). Could not be paid
+  // by re-wrap: the file measures zero reclaimable lines under this gate's own
+  // wrapLine, and re-wrap funding is refused per the 2026-08-17 rule in any
+  // case. Landed count, headroom 0, same convention.
+  ['.claude/skills/pm-dispatch/references/rest-channel.md', 93],
   ['.claude/skills/pm-dispatch/references/review-checklist.md', 84],
   ['.claude/skills/pm-dispatch/references/landing-operations.md', 80],
   // Release-aftercare duties — what a lane PM still owes AFTER a tagged release