skills(pm): encode the axis-weighting and decision-batch presentation rulings (#14142)

* skills(pm): encode the axis-weighting and decision-batch presentation rulings

Two maintainer rulings from the 2026-09-01 PM chat, both of which explicitly
order their own encoding into the skills.

Axis weighting: long-term soundness carries the highest weight, at least 50%,
when the axes conflict. Recommendations lead with the long-term reading and the
other three axes together cannot outvote it. Landed in all four
check:skill-frame-sync copies plus the card-face serialization, with two guards
the ruling implies: the weight is read through that axis's own definition
(shrink special-cases and contract accretion, never grow them), so it cannot
underwrite speculative expansion; and it ranks RECOMMENDATIONS only, never
authority, so a 50% weight cannot make a human-floor matter dispatchable.

Decision-batch presentation: exactly 5 cards per batch, one batch per report,
stop and wait for the reply. "Detailed explanation" is prose per card; a table
row, a one-line verdict or a changed/unchanged grid does not discharge the duty
— compressing the batch is a violation of it. Applicability is bounded to
presenting items for a RULING, never to already-decided result statements.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Msg17tAHJ3jVTYFgHydCm2

* Keep the quoted ruling on one line in the token-ratchet ledger

#11106: a governed quotation that soft-breaks stops being findable by the
things that grep it, and the raise's authorization is exactly such a quote.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Msg17tAHJ3jVTYFgHydCm2

---------

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

Author: claude[bot]

.claude/agents/os-dev.md

@@ -412,9 +412,12 @@ 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。
+Your recommendation must be justified on all four axes;长远合理性权重恒 ≥50%(2026-09-01
+裁「四维分析中，长期合理应该权重最高，至少50%」):推荐以它领起,另三轴合起来投不翻
+它,按本义读(缩小而非扩大特例/契约增生),⛔ 不为投机扩张背书;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(权重与序都只排推荐,⛔
+不授权、不移人工地板);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

@@ -882,6 +882,12 @@ finding 定级换结果态,永不留挂)/ `Blocked-by:` 活性现验(合并一
 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`。
+**长远合理性权重恒 ≥50%**(维护者 2026-09-01 裁,逐
+字:「四维分析中，长期合理应该权重最高，至少50%」):四轴冲突时推荐以长远的读数领
+起,另三轴合起来投不翻它;权重按该轴本义读 —— 缩小而非扩大特例与契约增生 —— ⛔
+不据它为投机扩张背书。与分歧推荐序 compose 不取代:后者按拉动定指向,本条定下限。⛔ 权
+重是推荐规则不是授权规则:50% 不把人工地板的事变成可派发的事,安全/权限边界与破坏性
+难逆动作照旧恒交维护者。
 **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

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

@@ -10,9 +10,14 @@
 
 ## 常设决裁批流程(2026-08-26 裁;裁决原话与归席见主文件「升级与决策」)
 
-- 每批 ≤5 张,批序:在飞被阻塞 > 运营阻塞 > 用户可见 > 结构性;每卡 = 六项写法 + 四棱块;
-  回批一行式「1A 2B …」,「X, 但…」= 附带条件随执行落地;裁后四件原子执行不再请示,
-  一口气跑完;批间新入箱的卡滚入下一批,收件箱读数为零才收工。
+- 每批**恰好 5 张,呈完即停等回批**(2026-09-01
+  裁「每批5张详细解释，不是这样一堆列给我，这个也要写入skills」),⛔ 永不一条消息投多
+  批;批序:在飞被阻塞 > 运营阻塞 > 用户可见 > 结构性;每卡 = 六项写法 + 四棱块;回批一行
+  式「1A 2B …」,「X, 但…」= 附带条件随执行落地;裁后四件原子执行不再请示,一口气跑
+  完;批间新入箱的卡滚入下一批,收件箱读数为零才收工。
+- **「详细解释」= 逐卡散文**:业务问题 + 每选项真实代价 + 四维 + 一个推荐;⛔ 表格行、一
+  行判词、「变/未变」栅格不构成详解 —— 压缩批次是违规不是摘要,重分析轮同纪律。适
+  用面 = 任何席位向维护者**呈递需其裁决的决策项**,⛔ 不及结果陈述(轮报表格照旧)。
 
 ## 落卡分析模板(主文件「落卡/升级流程」②的 references 细则;中文,2026-08-19 裁)
 
@@ -42,5 +47,8 @@
 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。
+**①长远权重下限**(2026-09-01 裁「四维分析中，长期合理应该权重最高，至少50%」):①恒
+≥50%,推荐以①领起,另三轴合起来投不翻①;①按本义读(缩小而非扩大特例/契约增生),⛔ 不
+据它为投机扩张背书;⛔ 只排推荐不授权 —— 50% 不把人工地板的事变成可派发的事。
 **应用仓推荐序特例**(2026-08-31 裁「既然是平台缺陷，就应该等待平台处理」):阻塞源是平台
 缺陷 ⇒ 恒荐等待,⛔ 不荐绕行(形状容错/复刻平台规则)、不荐劈半落地。

scripts/check-skills-token-ratchet.mjs

@@ -332,7 +332,25 @@ export const CEILINGS = new Map([
   // 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],
+  // 14391 -> 14549: the axis WEIGHTING joined the decision frame — long-term
+  // soundness carries the highest weight, at least 50%. Maintainer ruling
+  // 2026-09-01, verbatim and untranslated (kept on ONE line, #11106: a governed
+  // quotation that soft-breaks stops being findable by the things that grep it):
+  // 「四维分析中，长期合理应该权重最高，至少50%」
+  // Same shape and same reason as the +152 row above it: this file carries TWO
+  // enforced frame copies (check:skill-frame-sync COPIES), and a rule that
+  // changes WHICH RECOMMENDATION the frame yields is exactly the #5130 drift
+  // class if it ships to third-party installers with only the old tie-break —
+  // the customer's agent would weigh the axes co-equally while this repo weighs
+  // long-term ≥50%, and no report would reveal it. +158 tokens across both
+  // copies, against +152 for the comparable 2026-08-27 rule. Compressed twice
+  // before landing: the first cut measured +242, and the provenance was then
+  // removed on the #5451 route-B convention — the published copy carries the
+  // RULE without this repo's dates or quoted rulings (verified: the file holds
+  // zero of each), so the verbatim ruling lives in the internal copies and in
+  // the raising PR's body. No genuine deletion was available: the rule makes no
+  // existing sentence redundant, and a re-wrap moves no tokens and pays nothing.
+  ['skills/objectstack-pm-dispatch/SKILL.md', 14549],
   ['skills/objectstack-query/SKILL.md', 5552], //       -17 (was 5569)
   // 25125 -> 25143: the CRM UI Blueprint — the catalog's module-completeness
   // list, and the only place an agent is told what a finished module contains —
@@ -826,8 +844,12 @@ function selfTest() {
     // and correctly does not shift this pin. Subtotal after: 118830, i.e. the
     // same 80 tokens of slack the strip's claim had before either raise — which
     // is the point of shifting rather than widening.
+    // 118910 -> 119068: same operation for the 2026-09-01 axis-weighting raise
+    // on the pm-dispatch row (+158, see that row) — one SKILL.md row, so the
+    // shift equals the raise exactly, and the strip's claim keeps the same 80
+    // tokens of slack it has carried since before any of the three raises.
     ['the re-measure lowered the SKILL.md subtotal',
-      skillMdCeilings.reduce((a, [, n]) => a + n, 0) < 118910, true],
+      skillMdCeilings.reduce((a, [, n]) => a + n, 0) < 119068, 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

@@ -273,7 +273,27 @@ export const CEILINGS = new Map([
   // Lowered 48 → 46 by the #12081 soft-break closure (lowering is always
   // legitimate): three of this file's measured Han+ASCII-punct breaks merged
   // back into their paragraphs, and two of the three paid for themselves.
-  ['.claude/skills/pm-dispatch/references/decision-analysis.md', 46],
+  // Raised 46 → 54 by TWO same-day maintainer rulings on the decision frame,
+  // both of which explicitly order their own encoding into the skills. Verbatim
+  // and untranslated — the decision-batch presentation rule:
+  //   「每批5张详细解释，不是这样一堆列给我，这个也要写入skills」
+  // and the axis-weighting rule, whose card-face half lands here (its full
+  // statement is in the frame's single source, SKILL.md 「升级与决策」, paid from
+  // that file's existing headroom):
+  //   「四维分析中，长期合理应该权重最高，至少50%」
+  // +5 for the presentation teeth (exactly-5 / stop-and-wait / ⛔ multi-batch,
+  // and the prose-per-card duty with its applicability boundary — presenting
+  // items for a RULING, never already-decided result statements), +3 for the
+  // weight floor with its two guards (read through axis ①'s own definition, so
+  // it cannot underwrite speculative expansion; recommendation-only, so it
+  // cannot move the human floor). Could not be paid in place: all 46 lines are
+  // already packed to their own minimum at the 120-byte cap (measured per line
+  // — the file was densified at phrase boundaries by #13389 and the widest
+  // reflow saving available is zero), and re-wrap funding is refused per the
+  // 2026-08-17 rule in any case; the only other in-place payment was deleting a
+  // ruled clause, refused on the state-machine precedent. Landed count, headroom
+  // 0 again, same convention.
+  ['.claude/skills/pm-dispatch/references/decision-analysis.md', 54],
   // 134 → 133: whole-text restructuring round, PR-2 (maintainer ruling
   // 2026-08-23) — the three write-side sanitizer rows consolidated to one
   // author rule + one measured-behaviour row per surface (body / comment).
@@ -368,7 +388,21 @@ export const CEILINGS = new Map([
   // and 4 of the 8 paid for themselves. Zero content change — the diff is
   // byte-identical after whitespace normalization (39,313 B both sides), so the
   // 4 lines are re-flow slack, never a cut. Headroom 0 again, same convention.
-  ['.claude/agents/os-dev.md', 466],
+  // Raised 466 → 469 by the axis-weighting ruling (maintainer 2026-09-01, PM
+  // chat, verbatim and untranslated): 「四维分析中，长期合理应该权重最高，至少50%」.
+  // This file is one of the four `check:skill-frame-sync` COPIES, so a rule that
+  // changes which recommendation the frame yields has to reach it — the #5130
+  // drift is exactly a frame-semantics change that skipped a mirror. +3 lines,
+  // folded into the existing binding sentence rather than added as a new
+  // paragraph. ⚠️ Re-wrap funding was AVAILABLE here and was REFUSED: three
+  // paragraphs nearby carry wrap artifacts (two orphan lines of 6 and 8 bytes)
+  // worth exactly the 3 lines needed, and merging them would have landed this
+  // file at 466/466 with no raise — that is 筹行, banned by the 2026-08-17
+  // ruling, since the ratchet governs content volume and lines are only its
+  // machine-readable proxy. Measured both ways before reverting to the raise.
+  // The density fix is legitimate on its own and belongs in a net-reducing PR
+  // (2026-08-29 ruling), not this one. Landed count, headroom 0, same convention.
+  ['.claude/agents/os-dev.md', 469],
   // #9473: the other four `.claude/skills/` are read in full by the sessions
   // that use them too — the erosion mechanism the ratchet exists to stop
   // isn't specific to the pm-dispatch surface. Set at current counts on

skills/objectstack-pm-dispatch/SKILL.md

@@ -698,6 +698,12 @@ 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.
 
+**Axis ② carries the highest weight — at least 50%** when the axes conflict:
+lead with the long-term reading, and the other three together cannot outvote
+it. Read that weight through axis ②'s own definition — shrink special-cases
+and contract accretion, never grow them — so it never underwrites speculative
+expansion. It ranks recommendations only, never authority.
+
 ### 9. Round report, then next round
 
 Print a round report to the maintainer: a table of issue → verdict → PR link →
@@ -848,7 +854,11 @@ Justify your recommendation on all four axes; if they conflict, present the
 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.
+orders recommendations only — a split still escalates. Long-term soundness
+carries the highest weight, at least 50%: lead with it, the other three
+together cannot outvote it, and read it as shrinking special-cases rather than
+as licence to expand speculatively; weight ranks recommendations only, never
+authority.
 
 Return "blocked" (with evidence) when the default branch is broken under you, a
 dependency issue is unmerged, or CI infrastructure fails — after retrying