docs(north-star): the road grades a wrong-result run by Priority rule 2; rule 3 does not hold the p2 pm:blocking card that unblocks the lane's P0/P1

.claude/skills/pm-dispatch/references/execution-duties.md

@@ -47,7 +47,7 @@
 - 分诊预见的后续工作是一张卡,触发条件落地时立卡;⛔ 分诊一句话永不授权无卡编辑。
 - 阻塞项无主 ⇒ 被挡席认领做掉,不限大小;在该卡走完整认领、尊重其热文件串行队。
 - 阻塞项在飞 ⇒ 等;卡 PR 叠未合 PR 或 `Blocked-by:` 未合的,合前 ⛔ 不开线程(维护者令除外)。
-- 取卡前置 = `docs/NORTH-STAR.md`「Priority」第 3 条:本车道队列开放 P0/P1 每次取卡现读。
+- 取卡前置 = `docs/NORTH-STAR.md`「Priority」第 3 条(含豁免):本车道队列开放 P0/P1 取卡时现读。
 - 开放 = 未派发(无 `pm:dispatched`、无 assignee);在飞的 P0/P1 不挡任何候选,只余取卡全序。
 - 取卡全序:维护者直派插队卡(出处三件)> 契约面卡(判据见 `references/lanes/spec.md`)> 标签序。
 - 标签序:`priority:p0` > `pm:blocking` > 功能点位次 > p1 > p2 > p3 > 无级。

.claude/skills/pm-dispatch/references/triage-duties.md

@@ -60,7 +60,7 @@
 - 定级按北极星「Priority」1–3 条;首行 `Path: <定义项> | <清单项 id/缺项> | P<路步> | none`。
 - 定义项取北极星「What gets built」或仪器卡的车队决定;无则关;`none` 非分诊缺口。
 - 「无则关」= 首触即关 not_planned,带理由与入队三门,⛔ 不定 p3、不 hold;下行同此。
-- 所入车道队列 P0/P1 开着(= 未派发,在飞不算)时,三门皆无的 p2/p3 `tooling` 卡同样关。
+- 所入车道队列 P0/P1 未派时,三门皆无的 p2/p3 `tooling` 卡同样关;「Priority」第 3 条豁免除外。
 - 路由属分诊技术判断,⛔ 不升级哪个仓的问题;父单永不派发,`pm:epic` 卡 `pm:retriage` 即答。
 - 算涉 cloud:亦含托管面(objectos-runtime)、修复待其改动的 `Blocked-by:` cloud 卡、病因在 cloud。
 - 不算涉 cloud:修复与验收全在本仓,只在 cloud 测得或只缺其事实读数;仅引 cloud 卡作历史。

docs/NORTH-STAR.md

@@ -8,7 +8,7 @@
 
 **What counts as done**: an application built on the protocol is done only once it has passed verification — the checklist items its metadata maps to pass in one run, and `objectstack verify` is green on it. A delivery that has not passed verification does not count as done.
 
-**The one measure: the road**: a new AI developer (a human accepting alongside) starts from `npm create objectstack`: write metadata → run it locally and see it → verification loudly rejects what is wrong and lets through what is right → publish and install into an environment → connect an agent (MCP in the community edition) and have it complete one real business operation in the application → iterate once on a one-sentence customer request. Walk it for real every week, and the result lands as a run record. The first step that breaks is the current P0; one that runs through but gives a wrong result is P2. The four promises are each kept by a step: ① AI can write it · ②③ it executes · ④ you own it · ⑤ agents can operate it; the sixth step, ⑥, closes the loop.
+**The one measure: the road**: a new AI developer (a human accepting alongside) starts from `npm create objectstack`: write metadata → run it locally and see it → verification loudly rejects what is wrong and lets through what is right → publish and install into an environment → connect an agent (MCP in the community edition) and have it complete one real business operation in the application → iterate once on a one-sentence customer request. Walk it for real every week, and the result lands as a run record. The first step that breaks is the current P0; one that runs through but gives a wrong result is graded by "Priority" rule 2. The four promises are each kept by a step: ① AI can write it · ②③ it executes · ④ you own it · ⑤ agents can operate it; the sixth step, ⑥, closes the loop.
 
 **Features on the road**: the features the platform has to deliver, one per line, arranged by the steps of that road; top to bottom is the order — this table is the roadmap, and editing it is editing priority. Each row reads `axis · feature (in business language) · checklist area · item ids`, and one feature is a whole strip from protocol through runtime to the frontend; every item in `docs/qa/platform-checklist/` belongs to exactly one row here (an item not yet written is marked "checklist item to be written"). An id without a prefix belongs to the checklist area before it; one from another area is written in full.
 
@@ -113,7 +113,7 @@
 **Priority**:
 1. Security and data integrity always come first and do not wait for the road.
 2. The road broken, or a capability on the checklist broken ⇒ P0/P1; data silently wrong or lost ⇒ P1, subject to the grading rules' reach discount; any other run that goes wrong ⇒ P2; neither on the road nor on the checklist ⇒ p3 or not taken on. Grading reads "Features on the road": change that table to change priority, ⛔ never re-grade card by card — except for one re-grade pass when the grading rules themselves change, and none otherwise.
-3. While a lane's queue still holds an open P0/P1, that lane dispatches no p2/p3 tooling card or contract-hygiene card; a process or tooling card that blocks other work is P2 and carries `pm:blocking`; it does not inherit priority up the chain.
+3. While a lane's queue still holds an open P0/P1, that lane dispatches no p2/p3 tooling card or contract-hygiene card, except a p2 `pm:blocking` card that such a P0/P1 names on its `Blocked-by:` line, which is not held because it unblocks that P0/P1; a process or tooling card that blocks other work is P2 and carries `pm:blocking`; it does not inherit priority up the chain.
 4. Because AI writes it and AI uses it: the ontology must be writable by AI — what is wrong must be loudly refused with a prescription, never silently stored; and usable by people and agents — declared objects, actions and agent / tool / skill metadata are honoured at runtime, what the MCP surface exposes is what the application really can do, and an agent that oversteps is stopped by permissions and leaves a trace. The half people use has a standard too: an author who writes not one line of UI code gets a UI that looks as if a designer had handled it — one design system, one layout; looking wrong is as much a defect as failing to render, and is refused at the "see it" step. One wrong sentence in the documentation and skills written for AI is a product defect.
 
 **Stage posture**: speed over compatibility. Few people use it: a retirement takes effect at once, with no transition window and no alias double-spelling; compatibility layers, migration windows and the like default to p3; keys declared but not honoured are retired in batches per release, not one card per key.