docs(pm-dispatch): three maintainer-ruled standing rules + the Part-of bookkeeping-sentence warning

.claude/agents/os-dev.md

@@ -412,9 +412,9 @@ not decoration:**
   2026-08-04:这是创业项目,核心能力优先):能力扩张默认从紧,无拉动的声明面按
   implementation-first 处置,已发布零消费的能力不因沉没成本获得豁免。
 
-Your recommendation must be justified on all four axes;四轴冲突时如实呈现权衡,交维护
-者拍板。同样,`main` 在你脚下碎了、依赖未合并、CI 基础设施故障时,返回 `blocked`(附证
-据)—— 先重试到足以确认不是你的改动。
+Your recommendation must be justified on all four axes;四轴冲突时推荐按拉动排序:实测拉动→长远,
+零拉动→不扩散,防错破平,安全与难逆恒人工(只排推荐);权衡如实呈现交维护者拍板。`main`
+在你脚下碎了、依赖未合并、CI 基础设施故障⇒`blocked`(附证据),重试到足以排除你的改动。
 
 ## 终报消息 —— 恰好这段 JSON,不带任何环绕散文
 

.claude/skills/pm-dispatch/SKILL.md

@@ -868,9 +868,17 @@ finding 定级换结果态,永不留挂)/ `Blocked-by:` 活性现验(合并一
 - **创业阶段不扩散需求** — **创业阶段聚焦原则**(维护者 2026-08-04 指示:
   「我们是一个创业项目,应该先专注于核心能力」):能力扩张默认从紧,
   无拉动的声明面按 implementation-first 处置,已发布零消费的能力不因沉没成本获得豁免。
+  **过渡也从紧 —— 创业阶段不渐进**(维护者 2026-08-27 裁,逐字:「项目在创业阶段,
+  用户也很少,短期不考虑渐进。」):废弃别名/拼写与能力退役默认**立即退休**,
+  不设分阶段窗口、不留双拼写宽限;staged 选项仅凭具名外部用户证据才可呈报为推荐。
 
 推荐意见必须基于这四条轴给出理由;四轴冲突时如实呈现权衡,
-交维护者拍板。**标准四棱卡面块是落卡与升级的必备件**(四棱维护者 2026-08-11 接受;
+交维护者拍板。**分歧推荐序按拉动定向**(维护者 2026-08-27 裁「tong y 4」):
+有实测拉动 ⇒ 荐长远形态一次付清;零拉动 ⇒ 荐不扩散(defer/remove、记录在案、重开自由);
+防错轴破余下平局向响亮/结构性;安全/权限边界与破坏性难逆动作恒在人工地板。
+⛔ 此序只排推荐:分歧块照旧升级,四棱同向置信门与代裁面不变,
+细则见 `references/decision-analysis.md`。
+**标准四棱卡面块是落卡与升级的必备件**(四棱维护者 2026-08-11 接受;
 标准块 2026-08-18 裁定「同意」;同裁定的摘要视图/批量决裁通道已退役,
 墓碑与原话见 `references/lanes/director.md`;收件箱由维护者定期与 AI 讨论消化,⛔ 不 assign
 推送 —— 维护者 2026-08-19 裁定:「我感觉决策卡推给我太麻烦了,我需要和ai讨论才能判断,
@@ -904,6 +912,11 @@ skills和决裁。」的归属半句,skills 面与其车道照旧归 skills 席)
   发布动作的完整清单与「围绕发布的工作照旧归座位」细则以 AGENTS.md「release act」
   段为权威,含**合并 Version Packages PR**)。发现未经人工的发布痕迹按事故立案,
   ⛔ 不代跑补救性发布 —— 机械通道的存在不构成授权,遇到那类通道当缺陷上报。
+- **发版提醒纪律**(维护者 2026-08-27 拍板,逐字:「要发版本我会自己点。
+  除非重要修复要发版本你可以提醒我。」):Version Packages PR 默认永不催 ——
+  挂着的发布 PR 是常态,⛔ 不进维护者摘要;唯一例外 = 重要修复已合 main 在等发布,
+  且提醒必须点名该修复与不发版的用户可见后果,让维护者仅凭提醒即可判缓急;
+  与上条相组合 —— 本条管**何时开口**,上条管**谁动手**。
 - **Governed 面由维护者人工合并,合并即审核记录**(三裁一脉:2026-08-08
   「adr 只能由维护者自己确认,人工合并,ai 不得擅自合并」;2026-08-11「所有 skills 的更新和adr
   类似,需要人工审核」;2026-08-18 对「人工合并即人工审核,事后审计代替事前门」

.claude/skills/pm-dispatch/references/decision-analysis.md

@@ -42,3 +42,5 @@
 ④ 创业阶段不扩散(remove 优于 declare-and-maintain,每个已声明的键都是永久义务);
 一行推荐 + 字母选项(A/B/…);**一行强制置信缺口(「本分析看不见什么」)**。
 四棱行同受六项约束:每行论据从业务立场写,机制名词只作括号补充。
+**四棱分歧推荐序**(2026-08-27「tong y 4」):②实测拉动⇒荐①长远终态;零拉动⇒荐④不扩散;
+③破余下平局向响亮/结构;安全与难逆恒人工;⛔ 只排推荐,分歧照旧升级、代裁面不扩。

.claude/skills/pm-dispatch/references/lanes/director.md

@@ -25,6 +25,8 @@ issue 编号;机器判据恒英文,引文逐字不译。
   策」四轴(`check:skill-frame-sync` 管的四份拷贝不动,⛔ 本文不复制框架本体)。
 - 本原则重锚 2026-08-26 流程裁定(「决裁流程今天这样很好，决策卡5张一组，从业务的角度给我具体解
   释，说明可选项，并进行四维分析。应该写入技能。」):流程未变,2026-08-26 定流程、2026-08-27 升格定归属。
+- **创业阶段不渐进**(维护者 2026-08-27,逐字:「创业阶段不渐进 应该写入项目总监skills」):
+  废弃别名与退役默认立即退休,无窗口无双拼写宽限;staged 仅凭具名外部用户证据可荐。
 
 ## 职责一:`needs:contract-review` 复审链
 

.claude/skills/pm-dispatch/references/review-checklist.md

@@ -5,11 +5,11 @@
 
 - **PR 形态**:存在、是 draft、目标 `main`、正文首行引用卡片
   —— **`Fixes #<n>` 仅当合并应当关卡**;只落地了可实施一半(另一半在决策箱或按范围排除)
-  ⇒ 必须 `Part of #<n>`,否则合并会静默关掉一张正躺在决策箱里的卡,
-  而 `needs-user-decision` 的收件箱过滤只看 open issue。翻 ready 之前亲核首行,别只信报告。
-- **闭合关键词两读**:
-  翻 ready 前亲扫正文 —— 合并前唯一的人工关口(⛔ 关键词永不挨着另一张 open 卡编号;
-  解析器不理会否定句);合并后那一读见落地细则,解析事实见平台读数。
+  ⇒ 必须 `Part of #<n>`,否则合并静默关掉决策箱里的卡(收件箱过滤只看 open)。⚠️ 切
+  `Part of` 的正文最易再武装自动关闭:解释「为何不关」的记账句(谁在合并后手动关卡)、
+  否定句、引号/反引号包裹,解析器一律照关;安全拼法 = 卡号旁零动词,关闭安排写进评论。
+- **闭合关键词两读**:翻 ready 前亲核首行、亲扫全正文(⛔ 关键词永不挨另一张 open 卡编号)
+  —— 合并前唯一的人工关口,别只信报告;合并后那一读见落地细则,解析事实见平台读数。
 - **`Part of` 收口的卡不会自动关,`pm:dispatched` 必须手工摘**(动作要件见落地细则 B):
   ACCEPT 那一刻记进落地待办,⛔ 不留给下次巡检 —— 漏摘让在飞视图多算一张无主开卡。
 - **范围检查**(取 changed files,⛔ 不看报告自述):无 `content/docs/releases/` 改动、

scripts/check-skills-token-ratchet.mjs

@@ -263,9 +263,14 @@ export const CEILINGS = new Map([
   ['skills/objectstack-formula/SKILL.md', 6002], //     -53 (was 6055)
   ['skills/objectstack-i18n/SKILL.md', 6338], //        -11 (was 6349)
   ['skills/objectstack-platform/SKILL.md', 12705], //   -11 (was 12716)
-  // Unchanged: this skill's only id-shaped tokens are a CLI usage line and a
-  // JSON shape example. The one edit there (`#457` -> `#<n>`) is byte-neutral.
-  ['skills/objectstack-pm-dispatch/SKILL.md', 14239],
+  // 14239 -> 14391: the pull-directed split-resolution order joined the decision
+  // frame (maintainer ruling 2026-08-27, verbatim and untranslated: 「tong y 4」 —
+  // accepting the four-rule set), and this file carries TWO enforced frame copies
+  // (check:skill-frame-sync COPIES), so the rule ships to third-party installers
+  // with the frame it amends — the #5130 drift is exactly a frame-semantics change
+  // that skipped this mirror. +152 tokens across both copies, compressed to the
+  // minimal anchor form; the raising PR's body carries the arithmetic.
+  ['skills/objectstack-pm-dispatch/SKILL.md', 14391],
   ['skills/objectstack-query/SKILL.md', 5552], //       -17 (was 5569)
   ['skills/objectstack-ui/SKILL.md', 25125], //         -29 (was 25154)
   ['skills/objectstack-upgrade/SKILL.md', 8333], //      -2 (was 8335)
@@ -713,8 +718,11 @@ function selfTest() {
     // ⚠️ Scoped to the SKILL.md rows since #12392 widened the map: summing ALL
     // ceilings would compare this re-measure's subtotal against a total that
     // now includes 27 rows it never touched, which is a different claim.
+    // 117943 -> 118095: shifted by exactly the +152 ruling-authorized raise on
+    // the pm-dispatch row (2026-08-27 「tong y 4」, see that row), so the strip's
+    // lowering claim keeps its original slack instead of being silently eaten.
     ['the re-measure lowered the SKILL.md subtotal',
-      skillMdCeilings.reduce((a, [, n]) => a + n, 0) < 117943, true],
+      skillMdCeilings.reduce((a, [, n]) => a + n, 0) < 118095, true],
 
     // ── the #12392 extension basis ───────────────────────────────────────
     // Same reasoning as the pins above: the gate never reads this sha, so it

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

@@ -312,7 +312,12 @@ export const CEILINGS = new Map([
   // the maintainer-action ledger. A lane charter is read per seat session like
   // every entry above. Set at the landed line count read from this ratchet's own
   // run (headroom 0, same convention as the entries above).
-  ['.claude/skills/pm-dispatch/references/lanes/director.md', 75],
+  // 75 → 77: maintainer-directed placement (2026-08-27, verbatim: 「创业阶段不渐进
+  // 应该写入项目总监skills」) — the no-gradualism standing rule lands here by that
+  // directive; the file sat at headroom 0 with no losslessly compressible slack in
+  // the touched sections (arithmetic in the raising PR), so the directive is the
+  // ruling that pays for exactly the two lines it dictates.
+  ['.claude/skills/pm-dispatch/references/lanes/director.md', 77],
   // 399 → 405 (#11126): maintainer-ruled (2026-08-23, option B, quoted in that
   // PR) — the +6-line cross-repo dispatch-gates caveat, sized so the queued
   // #11137 (395→399 on main) and this PR's +6 compose to exactly 405.

skills/objectstack-pm-dispatch/SKILL.md

@@ -692,7 +692,12 @@ forming, more permissive once it is stable — and this axis reads it from there
 like every other project-specific rule.
 
 Your recommendation must be justified on **all four** axes. If they conflict,
-present the trade-off honestly and let the maintainer decide.
+present the trade-off honestly and let the maintainer decide. When they split,
+rank the recommendation **pull-first**: real measured pull → the long-term
+shape; zero pull → non-proliferation (defer/remove, record, reopen freely);
+the AI-safety axis breaks remaining ties toward the loud option; security and
+destructive actions stay with the human. The order ranks recommendations
+only — a split still escalates.
 
 ### 9. Round report, then next round
 
@@ -841,7 +846,10 @@ Analyze every option on four fixed axes:
   declaration. How tight the default is comes from the project's conventions
   file, not from this template.
 Justify your recommendation on all four axes; if they conflict, present the
-trade-off and let the maintainer decide.
+trade-off and let the maintainer decide, ranking the recommendation pull-first
+(real pull → long-term shape; zero pull → non-proliferation; the AI-safety
+axis breaks ties; security/destructive → always the maintainer); the ranking
+orders recommendations only — a split still escalates.
 
 Return "blocked" (with evidence) when the default branch is broken under you, a
 dependency issue is unmerged, or CI infrastructure fails — after retrying