docs(qa,pm): qa-run joins the sweep exclusion list; run records owe defect extraction at close-out

.claude/skills/checklist-test/SKILL.md

@@ -88,7 +88,7 @@ node scripts/checklist-select.mjs <selector> --json
 运行发现该项的 `steps` 写错了(路由搬家、键改名、过期路径还需要清 localStorage)——
 这是清单在起作用。那是一次清单**编辑**:在 **worktree** 里做(PD#11):修订该项、递
 增 `revision`、追加一条 `history`、保持 `node scripts/check-platform-checklist.mjs`
-绿,落在任务分支上。运行中发现的产品缺陷进 `FOLLOW-UPS.md`(或立成 issue),作为
+绿,落在任务分支上。运行中发现的产品缺陷按 §4 收束抽取成独立卡;条款侧可记
 expected-fail 探针 —— 永不在真实缺陷上把条款打成绿。
 
 ## 4. 结果 issue —— 一次运行一张 GitHub issue,纯文本
@@ -103,8 +103,8 @@ expected-fail 探针 —— 永不在真实缺陷上把条款打成绿。
 用 `issue_write`(github MCP)立单:
 
 - **标题** —— `QA run · <selector> · <framework-sha[:8]> · <date>`
-- **标签** —— 恒带 `qa-run`;任何条款失败就加 `bug`(P0/P1 再加 `regression`),让
-真实缺陷直接从运行 issue 可分诊,不必再立第二张。
+- **标签** —— 只带 `qa-run`,⛔ 不挂 `bug`/priority 等工作标签:run 记录是协议载体、
+不入分诊 sweep,⛔ 不是可派发单元;工作标签随抽取出的缺陷卡走(RUNNER 抽取义务条款)。
 - **正文**,按此顺序:
  - **环境指纹** —— framework sha、`.objectui-sha`、端口、db、seed、时间戳。
  - **范围** —— 选择器 + 每项所对的 `revision`。
@@ -114,6 +114,10 @@ expected-fail 探针 —— 永不在真实缺陷上把条款打成绿。
  body)/ ref 定位的 selector 路径,足以在全新启动上重新命中,外加 oracle 的
  expected-vs-actual。足够让人或全新 agent 不靠你的截图复现。
  - 派生的整项判定 + fixture 缺口清单(如有)。
+- **收束抽取(RUNNER 抽取义务条款)** —— 运行中发现的每个产品缺陷,收束报告时逐个抽
+ 取为独立 issue:标题自含、复现与机制条目化、指回本 run issue 取全量证据;缺陷卡不
+ 挂 `qa-run`,正常进分诊首触。清单准确性发现与 fixture 缺口归波次锚卡(sweep 跟踪
+ issue)收口,⛔ 不抽取;环境阻塞记录在案即可。
 
 同一份逐条款表 + 环境准备与测试的耗时之比,回报给维护者(chat),并链接已立的 issue。
 

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

@@ -255,7 +255,7 @@ of record)。**读侧闭环(维护者 2026-08-12 批准):接管与巡检只读
 仓没有车道标签,这形状是它们队列卡的常态);③ 有 `domain:*` 无 pm-state。②③ 只取 `updated_at`
 早于 ~2 分钟的卡且不是可选项 —— 半标注卡是协议自己按设计生产的(一次标签写入即把老卡打成半标
 注),只带路由或状态机其一的卡对两个视图同时不可见。排除:`tracking`、`status:parked`(其正常形
-状恰是「带域标签无 pm-state」)、全部 `pm:seat` 贴;存量大时每轮限量、优先最新。**紧急卡直接分诊**
+状恰是「带域标签无 pm-state」)、全部 `pm:seat` 贴、`qa-run` 记录(协议载体非工作);存量大时每轮限量、优先最新。**紧急卡直接分诊**
 (维护者 2026-08-13):维护者点名或 p0 嫌疑 ⇒ 立即起 `claude-fable-5` 分诊子代理,不等 Routine 班
 次;授权面 = 分诊本身(定级/路由/标签/既有评论格式),⛔ 不写码不认领;产出落卡,与 Routine 分诊同格式同效力(细则见 `references/dispatch-runbook.md`)。
 **跨仓 pin 链的窗口级兜底也在本 sweep**(联动单第一产者仍是接受座位;⛔ 只立单不执行 bump):

docs/qa/platform-checklist/README.md

@@ -281,7 +281,8 @@ smoke) ∪ any item whose `source` cites a PR in the release. The tracking issue
 sweep links here and hosts discussion; results stay OUT of the repo — every run files
 one `qa-run` GitHub issue as its record (text only: the verdict table + a reproduction
 rule per failure, never screenshots; `runs/` is git-ignored), and a run that finds a real
-regression carries the `bug` label so it triages straight from that issue. Item text,
+regression extracts it into its own standalone card at close-out (RUNNER.md, extraction
+obligation — the run issue itself stays `qa-run`-only, excluded from triage). Item text,
 fixtures learned, and new traps discovered flow **back into the ledger** as revisions —
 that is the accumulation the one-off checklists never had.
 

docs/qa/platform-checklist/RUNNER.md

@@ -52,9 +52,9 @@ test-run output the clause's `evidence` field names.
  — the vendored `/_console` bundle may be stale (skill §2);
  - then capture the **reproduction rule** — ordered steps / API calls (method · path ·
  body) / the ref-targeted selector path + expected-vs-actual — into the run's result
- issue, which is labeled `bug`. A `fail` with no reproduction rule in its issue is not
-a completed verdict. (The screenshot that convinced you is a live judgment aid, not
-report content — describe what it showed in one line; never attach it.)
+ issue. A `fail` with no reproduction rule in its issue is not a completed verdict.
+ (The screenshot that convinced you is a live judgment aid, not report content —
+ describe what it showed in one line; never attach it.)
 3. **Classify blockers honestly.** Missing seed/persona/fixture → `blocked(fixture)`,
  and *record the gap on the item* (`fixtures.knownGaps` or `blocked`) so the next
  sweep doesn't rediscover it. A defect in the fixture itself (seed silently failing,
@@ -114,8 +114,8 @@ contradicts it, and correct it here when it does.
 ## Run records — the GitHub issue is the report
 
 Every completed run — **pass or fail alike** — files **one GitHub issue** as its durable
-record, labeled `qa-run` (plus `bug` when any clause failed). **Nothing lands in the
-repo** — not the JSON, not screenshots; `runs/` is git-ignored except its README.
+record, labeled `qa-run` and nothing else (extraction obligation below). **Nothing lands
+in the repo** — not the JSON, not screenshots; `runs/` is git-ignored except its README.
 
 **The issue is text only.** Screenshots and DOM dumps are oracles you consult *live* to
 reach a verdict — never report artifacts. What the report carries for a defect is the
@@ -158,3 +158,21 @@ per-clause verdict table (text oracle evidence) · a reproduction rule per `fail
 derived item verdicts + fixture gaps. The durable, version-controlled truth is still the
 checklist under `areas/`; a run is a dated assertion about one build, and it lives in its
 issue, not the tree.
+
+### Extraction obligation — the run record is a protocol carrier, not work
+
+Run issues are excluded from the PM's backlog sweep, same class as the `pm:seat` post
+(maintainer ruling, 2026-08-17: 「把 qa-run 加进 sweep 排除清单 —— run 记录和 pm:seat
+贴同类:协议载体,不是工作」). Nobody fishes a run record for dispatchable work — so
+closing out the report includes the extraction, owed by the runner:
+
+- **The run issue carries run evidence only.** It is ⛔ not a dispatchable unit, and it
+ ⛔ never carries work labels (`bug`, priority, …) — work labels ride the extracted
+ cards.
+- **Product defects found during the run**: at close-out, extract each one into its own
+ standalone issue — self-contained title, reproduction and mechanism itemized in the
+ card, a pointer back to the run record for the full evidence chain. A defect card does
+ not carry `qa-run`; it enters triage first-touch normally.
+- **Checklist-accuracy findings and fixture gaps** close out through the wave's anchor
+ card (the sweep's tracking issue) — ⛔ not extracted.
+- **Environment blockers**: recorded in the run record is enough.

docs/qa/platform-checklist/runs/README.md

@@ -14,9 +14,10 @@ durable, **text-only** report (pass or fail alike):
 - the issue **body** hosts the per-clause verdict table (pass / partial / fail / blocked)
  and the scope — the selector that chose the items + the `revision` each ran against;
 - **every `fail` carries a reproduction rule** in that same issue — ordered steps / API
- calls (method · path · body) / the ref-targeted selector path + expected-vs-actual —
- and the issue gains the `bug` label (a real regression may additionally get its own
- linked issue) so it triages straight from the run;
+ calls (method · path · body) / the ref-targeted selector path + expected-vs-actual;
+ the run issue itself stays `qa-run`-only (protocol carrier, excluded from the triage
+ sweep) — at close-out each product defect is extracted into its own standalone card
+ (RUNNER.md, extraction obligation), and that card is what triages;
 - **screenshots are never part of the report** — they are live judgment aids that die
  with the run environment, described in one line of text, never attached or linked.