Add knowledge widening and deepening with checked JiuwenSwarm output

doc/architecture/SPEC/engine/steps/knowledge.md

@@ -29,3 +29,7 @@ PR，且只在 `ALLOW_PUSH=1` 与 `ALLOW_POST=1` 同时成立、非 dry run、
 
 `pr-history` 阶段另接收可选 `pr_count` 与 `budget_usd`，转成整数/浮点后传入 `run_stage`；非法值返回 BLOCKED。
 历史 PR 的完整选择、逐 PR checkpoint、commit 串和 Codex 审阅均由 `kb_service.init_history` 负责。
+
+`knowledge-deepen` 另接收 `budget_usd` 与 `retry_unfinished`，逐功能保存七维实现知识、
+累计预算与草稿。`from_existing` 允许 knowledge/knowledge-deepen 从已合并 KB 开始。
+`kb widen` / `kb deepen` 经同一 playbook/step 指定这两个知识阶段，旧规则 deepen 保持原契约。

doc/architecture/SPEC/kb_service.md

@@ -490,7 +490,7 @@ architecture 页并更新最近索引，已有正文保留。阶段不切换生
 旧路由未覆盖的文件可由清单的显式 feature owner 接管；只有唯一 owner 才补齐本次提炼输入，
 冲突或没有归属的文件继续单列为 unrouted，不改写持久路由。
 
-`--from-existing` 只适用于 knowledge：没有本地 skeleton/modules record 时，显式从已合并的仓库
+`--from-existing` 适用于 knowledge 与 knowledge-deepen：没有本地 skeleton/modules record 时，显式从已合并的仓库
 索引和 owner 路由开始增补；每个路由页须存在且属于本仓库。已有 record 的 review/merge 门禁保持生效。
 不创建或伪造前置阶段记录，输入摘要绑定该模式和 main 的固定 SHA。
 
@@ -519,6 +519,40 @@ token、served model 和调用结果照常记录，judge 的固定记账与预
 可独立复核；上游必须是干净的固定 HEAD。`--write-contracts` 只追加缺少的结构性记录，
 保留过期记录供人工刷新；`--report` 把逐功能/逐文件结果写入 eval 或本地临时文件。
 
+## kb widen / kb deepen：功能广度与实现深度
+
+`kb widen REPO` 经同一 playbook 指定 `knowledge --from-existing`；`kb deepen REPO`
+指定 `knowledge-deepen --from-existing`。支持 `--dry-run`、`--pin`、`--subscription-generator`。
+旧 `--stage deepen` 继续提炼规则和处理 shadow 开关；新知识深读不生成规则、不修改生命周期或旧阶段链。
+
+knowledge-deepen 需要显式 feature/production coverage policy，逐项访问全部功能，包含已有规则的 owner。
+每项提炼 flow、api、configuration、dependencies、failure_modes、tradeoffs、validation 七维，
+只保存有证据的内容到最近 owner 的 `feature-depth-<id>.md`，链接功能概览与 owner 索引。
+只追加索引链接，保留已有 Direct quick map；最终仍检查地图存在且未被截断。
+源码按定义切片，可检索文件前缀之外的实现，并补充可静态解析的第一方被调函数与相关测试。
+flow 至少包含两个不同定义的直接调用；Python 校验限定符号、词法 owner 和导入目标，其它
+语言检查去除注释/字符串后的声明和调用。动态对象分派不能靠同名猜测，语义仍由模型裁判复核。
+
+每维用 `kb:depth` 绑定 feature、facet、完整 pin 与正文哈希，`kb:depth-proof` 绑定精确证据
+区间及其哈希、flow trace。区间必须完整展示给生成器，不允许跨缺口。页面必须是无规则的
+architecture/guide；同一调用独立判定各新增维度的 faithful/non_contradictory/does_not_weaken，
+三项均 yes 的维度才写入。每项 verdict 绑定原段哈希，拒绝或证据不足的草稿保留在本地 checkpoint/trace。
+机器枚举始终用固定英文标识，正文遵循知识语言。单个无证据或无效维度被剔除并记录原因，
+其它有效维度仍须分别通过裁判才保存；flow 不允许用测试调用链替代生产实现。
+来源引用、符号、L1 与知识树校验仍需通过。
+
+`coverage.semantic_depth` 单列完整功能数、七维槽位数和有语义证据的生产文件；
+`coverage.breadth` 保留原功能/文件指标。静态接口卡、旧 owner/facet 标记、规则和路由不能
+增加新深度。七维齐全只说明各有代表性实现知识，不声称全行为或执行测试覆盖。
+`tools/audit_knowledge_depth.py` 在干净的固定上游 HEAD 上独立重算两组指标。
+
+先遍历全部功能，再修复缺口；恢复与显式重试同样优先处理尚无已接受维度的功能，每项每批最多两次尝试。
+重复维度整项剔除，其它有界候选维度继续独立核验。逐项保存预算预留、生成草稿、裁判
+结果与已接受正文；包括第二次尝试在内，中断裁判均可复用完成的生成。恢复固定 KB baseline 与 upstream pin，
+输入变化则拒绝串用。`--budget-usd N` 可提高同批累计 ceiling；`--retry-unfinished` 显式
+重试缺口并保留花费。已发布批次不再修改，合并后用新 state directory。missing facets
+保持明确；订阅生成器 USD 为 unreported，judge 固定记账不等于账单实付。
+
 ## 2026-09-30 kb init：校准集收割（harvest-calibration）
 - 各阶段的 `InitRecord.verdicts[rule_id]` 除 verdict/reasons/model/text_sha/page 外还保存**判定时的规则全文**
   （`section`）与钉点证据（`evidence`），被判 fail 剥离或被 owner 删除的规则也能复原成用例；缺这两项的旧记录使收割

doc/knowledge/SCHEMA.md

@@ -35,6 +35,11 @@ validation 六个维度。它生成的非规则段使用 `kb:knowledge` HTML 注
 不产生规则，也不承诺整文件行为或测试覆盖。功能页和源码记录按最近代码 owner 组织；
 达标报告属于 `eval/`，不能作为产品知识正文保存或通过新增索引增加覆盖率。
 
+`kb deepen`（`knowledge-deepen`）的 `kb:depth` 标记另绑定 feature、七维 facet、完整 pin
+与正文哈希，`kb:depth-proof` 保存精确证据区间及其内容哈希、调用 trace。只用于复核代表性
+实现知识；旧标记、静态接口卡、规则和索引不计入此深度。正文或证据改变、pin 不符或重复
+facet 均失效。页面仍使用既有 architecture/guide 类型，设计推断与未验证测试明确标出。
+
 `check_wiki_lint.py` 强制的只有前五项（`title`、`created`、`updated`、`type`、
 非空 `tags`）；其余为约定字段，写了就必须合法（`confidence` 只能取三值）。
 

eval/knowledge-depth/README.md

@@ -0,0 +1,36 @@
+# JiuwenSwarm implementation-knowledge depth
+
+The pinned run at `f0a69728c96b5961d993449f1a901cbd2f4dac5b` added **173 checked facets across 77 feature-depth pages**. Every accepted block is bound to an actual Codex review, its exact prose, and pinned source spans. The independent audit reports no proof errors.
+
+| Measure | Before | After |
+| --- | ---: | ---: |
+| Feature breadth | 79/79 | 79/79 |
+| Production-file breadth | 2043/2199 (92.91%) | 2043/2199 (92.91%) |
+| Features with checked depth blocks | 0/79 | 77/79 |
+| Checked depth facets | 0/553 | 173/553 (31.28%) |
+| Features with all seven depth facets | 0/79 | 0/79 |
+| Production files participating in depth evidence | 0 | 115 |
+
+Breadth includes checked interface records and feature summaries. The 115 participating files are witnesses for particular claims, not files with exhaustive behavior coverage. Existing owner explanations also remain useful; this conservative depth metric counts only the new bound blocks.
+
+All 79 features received extraction attempts. **Cron and skill evolution have no approved depth facets** after retries; their existing summaries remain. Other missing facets are listed per feature in the [full report](jiuwenswarm-f0a69728c96b.json). A full seven-facet feature is still an open target.
+
+Generation used Zcode GLM-5.3 at low effort; review requested Codex gpt-6-sol at low effort. The completed traces contain 121 generation calls and 112 review calls, including failures. The generator reports served identity when it returns a result; Codex does not report served-model identity. Two interrupted generation attempts have unreported usage and are recorded separately.
+
+Inputs included existing knowledge, 299 distinct source files, and 86 project documents. Of those documents, 24 came from the partial `.doc_project_maintainer` collection, used for 66 features. Historical test ledgers were context rather than evidence of current test passes. Validation facets describe assertions at the source pin; upstream JiuwenSwarm tests were not executed here.
+
+The final batch used $56 of fixed review accounting, plus $4 of earlier diagnostic accounting, within the original $60 ledger ceiling. Native subscription clients report no invoiced USD cost. These amounts are budget reservations, not measured billing or a cap on an actual invoice.
+
+The report contains the frozen knowledge baseline and input digest, code revisions used by the native calls, input paths, before/after audits, knowledge and policy hashes, and per-block native approval bindings. Raw prompts and verdict replies remain in ignored local state.
+
+Recompute the audit against a clean checkout at the exact pin:
+
+```bash
+PYTHONPATH=src python tools/audit_knowledge_depth.py \
+  --repo jiuwenswarm \
+  --upstream /path/to/pinned/jiuwenswarm \
+  --pin f0a69728c96b5961d993449f1a901cbd2f4dac5b \
+  --report /path/to/local/depth-audit.json
+```
+
+`kb widen jiuwenswarm` expands explanatory knowledge. `kb deepen jiuwenswarm` adds implementation facets; `--retry-unfinished` retries gaps with cumulative accounting and priority for features without accepted depth. Use a fresh state directory after publishing and merging a batch. Existing rule-focused `kb init --stage deepen` keeps its original behavior.