Resolve nightly knowledge backlog and enable gated automatic publication

adapters/afd_plugin/manifest.yaml

@@ -237,11 +237,10 @@ push:
   protected_branches: [main, master]
   allowed: false
 
-# Knowledge service: declared but disabled until afd-plugin has its own
-# calibration set and has run through a shadow period (rollout stage P3b).
+# Owner-authorized automatic publication after its calibration and shadow period.
 knowledge_lifecycle:
   enabled: true
-  mode: shadow
+  mode: auto_merge
   knowledge_dir: repos/afd-plugin
   release:
     trigger: none

adapters/jiuwenswarm/manifest.yaml

@@ -26,10 +26,10 @@ push:
   protected_branches: [develop, main, master]
   allowed: false
 
-# Knowledge service runs in shadow after reviewed initialization.
+# Owner-authorized automatic publication after reviewed initialization.
 knowledge_lifecycle:
   enabled: true
-  mode: shadow
+  mode: auto_merge
   knowledge_dir: repos/jiuwenswarm
   release:
     trigger: none

adapters/vllm_omni/kb-calibration/README.md

@@ -6,11 +6,30 @@ shadow mode only when every `reject` case is caught and at most 20 % of the
 `pass` cases are rejected; re-run it whenever the judge model or version
 changes.
 
-- `good-*`: the eight human-approved rules of InferMatrixCopilot #202, each
-  replayed as an `add` against its page as it was before the merge.
+- `good-*`: the eight human-approved rules of InferMatrixCopilot #202, replayed
+  against their pages before the merge. DIFF-2ag4 explicitly replaces its
+  zero-binding predecessor DIFF-2ag2, retaining the old text as retired and
+  preserving its guards in the stronger successor; the other cases add a rule.
 - `bad-*`: mutations the gate must reject (reversed claim, optional enforcement,
   restated PR description, contradiction of BENCH-1a, unjustified retirement,
   an edit that changes meaning, an invented API) and the historical bot-written
   MMH3-4e that contradicted MMH3-4a (folded by a human in 03b0b62a2).
 
-Evidence is the real upstream PR (title, body, bounded diff excerpt).
+Evidence is the real upstream PR, pinned by merge commit: complete title/body
+and per-file diff, plus exact source-line context where a rule depends on
+unchanged behaviour (runner schema filtering, prefix-cache source planning,
+speech exception handling). Positive and negative mutations of the same PR
+receive identical evidence. The historical MMH3 case retains its original
+evidence. None of the expected labels or scoring thresholds changed.
+
+The 2026-10-09 source audit also corrects two assertions in the good fixtures
+and canonical owner rules: SCHED-1c names `OmniPrefixCacheManager`, and SERV-9d
+preserves engine-exception propagation instead of claiming all engine errors
+are HTTP 400. Its `ValueError` 400 and generic internal-error 500 expectations
+remain intact.
+
+Recorded rerun reasons also correct BENCH-1h to the actual schema predicate
+complement and scope its artifact example to NPU, DIST-1m to exact caller-key
+cleanup without an invented pending-key ownership guard, and QOMNI-1i to the
+real `async_chunk` symbol with honest existing-test coverage. Negative rule
+mutations, labels, source evidence and scoring remain unchanged.

adapters/vllm_omni/manifest.yaml

@@ -401,10 +401,10 @@ push:
   signoff: {name: tzhouam, email: tzhouam@connect.ust.hk}
 
 # Knowledge service (doc/features/knowledge-service.md). Human-only section:
-# switching `mode` to auto_merge is a reviewed code change, never an agent write.
+# Publication policy changes require owner authorization and reviewed code.
 knowledge_lifecycle:
   enabled: true
-  mode: shadow                     # shadow | auto_merge
+  mode: auto_merge                 # owner-authorized; current calibration and both gates still required
   knowledge_dir: repos/vllm-omni
   intake:
     merged_prs: true

doc/architecture/SPEC/kb_service.md

@@ -1,6 +1,6 @@
 # kb_service/ —— 规范
 
-<!-- verified-against: 2026-10-08 -->
+<!-- verified-against: 2026-10-09 -->
 
 `知识服务核心：仓库配置、账本、outbox、CLI · refactor-status: new`
 
@@ -27,6 +27,9 @@
 - `cli`：`infermatrix-copilot kb keygen|status|pause|resume|control`。
 - `reconcile`：负责人已合并的初始化/管理知识历史的显式接收；固定目标与 active、GitHub PR/merger/当前检查、
   完整 first-parent 提交清单、目标树的两个校验器及真实快照验证，签名计划经重新验证后才安装不可变回执。
+  scheduler 仅在签名回执重新验证且提交已进入 active ancestry 后解决对应历史 provenance 告警；
+  外部 PR 告警仅在 live API 证明同仓库、同 PR、同 head 已关闭且未合并时解决。
+  两类解决均由租约栅栏事务保留原告警与证据，不修改原模型判定或授予新知识信任。
 
 ## 不变量
 - 模块不出现任何仓库名；仓库差异只来自 adapter 配置。
@@ -88,6 +91,9 @@ stdlib + PyYAML + `cryptography`（`kb` extra）+ `.adapters` + `.knowledge_serv
   候选改动经 improve 引擎配对实验 + 生产校准集双重把关。
 - `sources`：知识仓库克隆的只读读取（`knowledge_files`、`external_texts`）、只读 GitHub 客户端
   （合并 PR、PR 证据有界摘录、release/tag）、本机 Copilot 运行经验收件目录。
+  PR 证据保留 base/head 分支与 head SHA；intake 区分分支能力与当前默认分支事实，
+  新规则上下文包含完整已占用 ID（含退役与 tombstone），建议按源 PR 和 owner/topic 命名，
+  最终唯一性仍由操作应用和批次合并校验强制保证。
 - `intake`：每个事件由生成器起草类型化操作（只允许 add/edit_same_meaning/replace/retire），
   `apply_operations` 必须接受，最多两轮带精确错误的修复；多事件合并为一个变更集。
   系统提示明确操作字段的字符串类型；未用可选字段省略或为空字符串，`new_page` 是目标路径而非布尔标志。
@@ -96,6 +102,8 @@ stdlib + PyYAML + `cryptography`（`kb` extra）+ `.adapters` + `.knowledge_serv
   外部引用、protected、熔断（按仓库计算）、任何不确定 → human；L1 失败不调用模型。
 - `runtime`：collect → intake → gate → publish；shadow 只记录；变更集文件存 `changesets/<id>.json`。
 - `calibration`：按仓库的校准集评分（坏样例须全部拦下、好样例误拒 ≤ 20%）。
+  一致性检查沿用生产门禁的 changed-rule 归属，既有冲突单独保留；报告保存逐块原因和一致性详情。
+  CLI 校准模型调用写入 replayable trace，按仓库、case 和 judge/consistency 步骤标记。
 - `runner`：经标准 executor 运行 `kb-*` playbook。CLI 新增 `kb run`、`kb calibrate`。
 
 ## 2026-09-28 合并流程、激活、巡检与调度

doc/architecture/SPEC/knowledge_service.md

@@ -1,6 +1,6 @@
 # knowledge_service/ — provider curation components
 
-<!-- verified-against: 2026-10-08 -->
+<!-- verified-against: 2026-10-09 -->
 
 Pinned source identities and depth markers accept full SHA-1 or SHA-256 Git
 commits. Evidence bodies, stored hashes and native approval semantics remain
@@ -52,7 +52,11 @@ outbox item, control record, publisher ack, reconciliation plan, supervised
 reconciliation receipt) never verifies for another. Plans use
 `kb-reconciliation-plan`; only revalidated receipts use the distinct
 `kb-reviewed-reconciliation` purpose, and neither can be replayed as an
-automatic gate verdict or publisher action. It
+automatic gate verdict or publisher action. A separately reviewed historical
+hold uses `kb-reviewed-orphan-hold`, binding its exact queue row, original
+settled sources and active canonical rules without changing model verdicts or
+granting publication trust. Its signature cannot be replayed as an event
+settlement or reconciliation, including by relabelling the envelope. Signing
 needs only `cryptography`.
 
 `containment` implements the opt-in consumer protocol independently of knowledge

knowledge/repos/afd-plugin/_index.md

@@ -1,7 +1,7 @@
 ---
 title: "AFD plugin 仓库知识入口"
 created: 2026-08-05
-updated: 2026-08-06
+updated: 2026-10-09
 type: index
 tags: [afd-plugin]
 sources:
@@ -44,7 +44,7 @@ sources:
 | 核对 connector/graph/DBO 当前支持组合 | [execution-platforms](components/execution-platforms/architecture.md) | v0.26 CUDA/Ascend 平台现状 |
 
 默认 briefing 只加载本入口和仓库规则；深层 owner 页面由 changed-file route 或 `doc_search` / `doc_read` 按需读取。
-- [AFD plugin 仓库规则](rules-init.md)
+- [AFD plugin 初始化与 E2E 门禁规则](rules-init.md)：保留 AFD-I8 的文档/平台边界，GPU MRV2 DBO 按目标版本转入平台规则。
 - [上游 API 漂移与兼容边界（afd-plugin）](rules-seed-vllm-omni-rebase-upstream-api-drift.md)
 - [每次审查 AFD PR 都先套用的通用审查规则。AFD 特有的门禁再看本仓库的规则页。](../../general/review/rules.md)
 - [以下情况先读这页，判断是否应该先改上游：新增或刷新 compat patch、从 vLLM/vLLM-Ascend 复制函数，或在插件里重写上游逻辑（MoE routing、DP metadata、graph capture、EngineC](../../general/review/guides/upstream-first-for-algorithm.md)

knowledge/repos/afd-plugin/components/attention-runtime/_index.md

@@ -1,7 +1,7 @@
 ---
 title: "AFD Attention runtime 入口"
 created: 2026-08-06
-updated: 2026-08-06
+updated: 2026-10-09
 type: index
 tags: [afd-plugin, components, attention-runtime]
 sources:
@@ -36,4 +36,4 @@ sources:
 | 理解初始化、请求和 control metadata 流 | [architecture](architecture.md) | v0.26 当前运行实现 |
 | 调查模型内 handoff | [model-integration](../model-integration/architecture.md) | 角色感知层构造和 metadata consumer |
 | 调查 graph/DBO | [execution-platforms](../execution-platforms/architecture.md) | 平台 wrapper、capture 和 replay |
-- [afd_plugin/v1/worker 审查规则：AFD 事务元数据、控制面发送与 runner 生命周期](rules.md)
+- [afd_plugin/v1/worker 审查规则：AFD 事务元数据、控制面发送与 runner 生命周期](rules.md)：包括 v0.30 GPU MRV2 的实例 prepare seam 与每 stage 控制；不扩大 v0.26 基线支持。

knowledge/repos/afd-plugin/components/attention-runtime/rules.md

@@ -1,10 +1,13 @@
 ---
 title: "afd_plugin/v1/worker 审查规则：AFD 事务元数据、控制面发送与 runner 生命周期"
 created: 2026-09-30
-updated: 2026-09-30
+updated: 2026-10-09
 type: rule
 tags: [afd-plugin]
-sources: []
+sources:
+  - "afd-plugin@4c98691501be37f21735725f7429715ae98b603b:afd_plugin/v1/worker/attention_model_runner_v2.py"
+  - "afd-plugin@4c98691501be37f21735725f7429715ae98b603b:afd_plugin/v1/worker/attention_metadata.py"
+  - "afd-plugin@4c98691501be37f21735725f7429715ae98b603b:tests/unit/v1/worker/test_model_runner_v2.py"
 ---
 
 # afd_plugin/v1/worker 审查规则：AFD 事务元数据、控制面发送与 runner 生命周期
@@ -63,10 +66,34 @@ sources: []
 
 - hook 只在 `cudagraph_mode == CUDAGraphMode.FULL_DECODE_ONLY` 且 `cudagraph_manager` 存在时启用。它用 `MethodType` 覆盖单个 manager 实例的 `run_fullgraph`，用 `_AFD_FULLGRAPH_HOOK_MARKER` 防止嵌套。退出时恢复之前的实例覆盖，或者 `del` 掉实例属性。不要改成修改类属性。
 - replay 的 metadata 先用真实的 `real_tokens`（`scheduler_output.total_num_scheduled_tokens`）构建，再把 `tokens_lens` 改成 padded 的 `desc.num_tokens`；DP 控制同样用 padded 数。`tokens_unpadded_lens` 保持真实值，两者不要合并成一个。
+  - 上述构建和发送描述适用于单 stage。v0.30 GPU MRV2 多 stage 的控制已经由 native prepare seam 发布，FULL replay 必须复用，按 [stage metadata 规则](#gpu-mrv2-dbo-stage-metadata--原生-stage-只附加隔离的-afd-元数据) 检查，不能再发送单 stage payload。
 - hook 自己只恢复 `_afd_is_graph_replaying`。`_afd_pending_metadata`、`_afd_suppress_metadata_send`、`_is_warmup`、`_afd_is_graph_capturing` 都靠外层 `_use_afd_execution_context` 的 `finally` 恢复，所以不能在这个作用域之外单独使用 hook。
 
 <!-- kb:rule status=active since=init-9cae2d2ddaae -->
 
+## gpu-mrv2-instance-seam — GPU MRV2 prepare 包装只覆盖当前实例并恢复方法
+
+- 触发：修改 v0.30 GPU MRV2 的 `_use_afd_ubatch_preparation` 或接入同一 native `UBatchRunner.prepare` 边界；本节适用于包含该升级实现的 checkout，不扩大 v0.26 基线的支持范围。
+- 必须：记录当前实例是否已有 `prepare` 覆盖及原覆盖值，在 `finally` 中恢复原覆盖或删除临时实例属性。委托原绑定方法完成 native 输入切片、Attention metadata 和线程/stream 准备；注释写明缺少回调的原因、上游签名、delegation 例外以及上游提供 prepared-state 回调后的移除条件。
+- 禁止：为此替换 runner 类或模块全局函数、复制重写 native 准备逻辑，或把恢复方法覆盖误写成恢复所有 runner 状态。pending metadata 和执行标志仍由外层 execution/capture 作用域恢复。
+- 验收：无预存覆盖、有预存覆盖和异常退出都恢复 `prepare`；native 准备仍调用一次，外层状态恢复测试覆盖 pending、suppress 和 capture/replay 标志。
+
+^[PR #424]
+
+<!-- kb:rule status=active since=r2026-10-09 -->
+
+## gpu-mrv2-dbo-stage-metadata — 原生 stage 只附加隔离的 AFD 元数据
+
+- 触发：修改 v0.30 GPU MRV2 的 stage 准备、`install_mrv2_ubatch_metadata`、capture 或多 stage FULL replay。
+- 必须：保留 native slices 的传输长度和全 batch 偏移；为每个 stage 克隆 metadata，保留 `stage_idx` 并只附加该 stage 的 `tokens_start_loc`、`requests_start_loc`、`tokens_lens`、`tokens_unpadded_lens`。未填充长度按 `max(0, min(stop, num_tokens_unpadded) - start)` 与真实输入求交。
+- 必须：由调用方在微批线程执行前发布完整控制；多 stage FULL replay 核对 pending metadata 存在且 stage 数等于 descriptor 的 `num_ubatches`，然后复用已发布控制。capture 按 `(num_reqs, num_tokens, num_ubatches)` 匹配事件，不匹配或事件数不完整必须报错。
+- 禁止：重新切片 native input views、Attention metadata 或 worker contexts；从微批线程发送控制；多 stage replay 再发送旧的单 stage payload。
+- 验收：stage metadata 修改互不污染；padding 范围之外的 stage 未填充长度为零；控制发送次数和 descriptor 三个维度的负例都有 focused tests，正常退出和异常退出保留外层状态恢复保证。
+
+^[PR #424]
+
+<!-- kb:rule status=active since=r2026-10-09 -->
+
 ## AFD-I43 — NPU Async CAM stage 构建必须把 builder 0 留给全批次，并保持单 stage 的 AFD 事务
 
 - 全批次 metadata 已由 builder 0 构建，并在 stage 执行期间一直存活。因此 Async CAM stage 构建时用 `ASYNC_MOE_STAGE_METADATA_BUILDER_OFFSET + ubid` 选 builder。native DBO（`is_async_moe_stage_build=False`）的 offset 保持为 0。改动 offset 或 builder 的选择方式，会让 stage metadata 覆盖全批次 metadata 正在使用的可复用缓冲。

knowledge/repos/afd-plugin/components/connectors/_index.md

@@ -1,7 +1,7 @@
 ---
 title: "AFD connectors 入口"
 created: 2026-08-06
-updated: 2026-08-06
+updated: 2026-10-09
 type: index
 tags: [afd-plugin, components, connectors, distributed]
 sources:
@@ -36,4 +36,4 @@ sources:
 | 理解 factory、拓扑、payload 和生命周期 | [architecture](architecture.md) | 三类 connector 当前合同 |
 | 核对 role runtime 如何消费 control/work item | [Attention](../attention-runtime/architecture.md) / [FFN](../ffn-runtime/architecture.md) | caller 生命周期 |
 | 核对平台支持组合 | [execution-platforms](../execution-platforms/architecture.md) | graph/DBO/quantization 矩阵 |
-- [NPU connector（CAM async / CAMP2P）review 规则](rules.md)
+- [NPU connector（CAM async / CAMP2P）review 规则](rules.md)：含 Legacy DBO 物理计数与 opaque op 动态 padding 的分支限定合同。

knowledge/repos/afd-plugin/components/connectors/rules.md

@@ -1,14 +1,34 @@
 ---
 title: "NPU connector（CAM async / CAMP2P）review 规则"
 created: 2026-09-30
-updated: 2026-09-30
+updated: 2026-10-09
 type: rule
 tags: [afd-plugin]
-sources: []
+sources:
+  - "afd-plugin@036640618fc6df6f2c44c67979b91c5c66b0b5a0:afd_plugin/connectors/npu/camp2p.py"
+  - "afd-plugin@036640618fc6df6f2c44c67979b91c5c66b0b5a0:tests/unit/connectors/test_camp2p_connector.py"
 ---
 
 # NPU connector（CAM async / CAMP2P）review 规则
 
+## AFD-CAMP-430-counts — Legacy DBO 的真实 Attention 长度与物理 wire 长度必须分开
+
+- 触发：修改 `codex/v030-main-integration` 分支中 PR #430 引入的 Legacy NPU CAMP2p 多 stage、Attention ranks 大于 FFN ranks 的传输；该规则不声明此分支能力已经进入 main。
+- 强制：control plane 使用独立物理 payload，先将 DP counts 展开到 TP peers，再按 `(ratio, ffn_size)` 的 strided 接收组取每列最大值并重复；本地 connector state 与发送的 control payload 必须使用同一物理计数，原始 Attention payload 保持真实 query 长度。
+- 禁止：原地覆盖 Attention 元数据，或把各 rank 总体最大值当成每个 FFN 接收组的长度；ratio≤1 或单 stage 时保持原 payload 行为。
+- 验收：覆盖 2A1F、4A2F、DP→TP 展开和等量 ranks；逐 stage 核对发送与本地物理计数一致、输入 counts 不变。真实 native A2E/E2A 检查 wire rows 与 control counts 相符，返回前缀仍使用原始长度。 ^[PR #430]
+
+<!-- kb:rule status=active since=pr-430 -->
+
+## AFD-CAMP-430-padding — 动态 CAMP2p padding 必须在 opaque runtime op 内读取当前 stage
+
+- 触发：修改 PR #430 的 Legacy NPU CAMP2p split-stage send 路径；作用域为 `codex/v030-main-integration` 中该修复，GPU 与新 NPU MRV2 DBO 不在此规则范围。
+- 强制：仅在 `attn_size > ffn_size` 且 connector 有多个 stage 时，从当前 forward context 的 `ubatch_idx`、`additional_kwargs["afd_metadata"].connector` 读取本 rank 物理 token 数；在 opaque op 内按差值为 hidden states 补零，只有 expert IDs 存在且需要 padding 时才同时补齐 IDs/scales。进入该多 stage 路径后，即使 padding 为零，也必须把 transfer batch size 更新为物理计数。
+- 禁止：在 op 外按 profile/warmup 元数据预先 padding，让编译图冻结动态分支；禁止用物理长度替换 E2A 接收的原始长度 reference，或合并不同 stage 的 transfer state。
+- 验收：同一 runtime op 在切换实时 stage metadata 后正确处理无 padding 与有 padding，参数化覆盖 IDs 缺失/存在；断言 native send 张量与 batch size 的物理长度、原始前缀与零尾部，以及每 stage 独立状态。native 观测必须核对实际 wire rows，不能只检查 host 计数。 ^[PR #430]
+
+<!-- kb:rule status=active since=pr-430 -->
+
 ## AFD-I58 — async CAM FFN 侧：`token_nums_rankid_layeridx` 必须原样回传，真实层号和 token 数只来自 CAM 元数据
 
 - `afd_async_dispatch_recv` 返回的 `batch_info` 存为 `states.token_nums_rankid_layeridx`，`send_ffn_output` 必须把它原样传给 `afd_async_combine_send`。

knowledge/repos/afd-plugin/components/execution-platforms/_index.md

@@ -1,7 +1,7 @@
 ---
 title: "AFD execution platforms 入口"
 created: 2026-08-06
-updated: 2026-08-06
+updated: 2026-10-09
 type: index
 tags: [afd-plugin, components, execution-platforms]
 sources:
@@ -34,5 +34,6 @@ sources:
 | 遇到什么 | 查看哪里 | 说明 |
 |---|---|---|
 | 理解 CUDA/Ascend、graph/DBO、MLA 和 native op | [architecture](architecture.md) | v0.26 当前平台矩阵 |
+| 核对版本化 MRV2 DBO 校验或 DSV4 本地 NPU E2E 环境 | [rules](rules.md) | v0.30 GPU 升级实现与 v0.26 基线分开；设备和 NIC 按场景校验 |
 | 调查 connector state 来源 | [connectors](../connectors/architecture.md) | control/data plane 和资源 owner |
 | 审查 Ascend global patch | [compatibility](../compatibility/rules.md) | upstream-first 与 non-AFD 回归 |