kb: 支持保留未知项的基础知识发布与深化交接

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

@@ -1,6 +1,6 @@
 # engine/steps/knowledge.py —— 规范
 
-<!-- verified-against: 2026-10-04 -->
+<!-- verified-against: 2026-10-06 -->
 
 `知识服务的 knowledge.* 步骤 · refactor-status: new`
 
@@ -34,3 +34,22 @@ PR，且只在 `ALLOW_PUSH=1` 与 `ALLOW_POST=1` 同时成立、非 dry run、
 累计预算与草稿。`from_existing` 允许 knowledge/knowledge-deepen 从已合并 KB 开始。
 `unlimited_subscription` 显式传入两角色订阅无限模式；预算和语义完成门槛由同一 stage owner 处理。
 `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。
+
+## 2026-10-06 显式基础知识部分发布
+
+`foundation_mode` 缺省 `strict`，未显式选择时不添加部分发布参数。
+`partial` 只允许订阅无限的 `knowledge`，或带原始 `foundation_record` 的
+`knowledge-deepen`；CLI 的 `--foundation-record` 经步骤转换为 `Path`，以
+`foundation_record_path` 传给 `run_stage`。不兼容的阶段、缺失记录和非法模式由
+stage 拒绝，步骤返回 BLOCKED，不能以参数缺失自动降级为部分发布。
+
+`knowledge` 的显式部分发布只重放原批次的真实任务和独立评审，不追加模型调用，
+不重置修正次数，也不改变生成身份。新深度批次将模式及原发布记录/收据哈希纳入
+身份；七维认可门槛和完整目录分母保持不变。`published` 只表示 PR 已发布，
+仍须单独保留 `init_complete=false`、六维基础知识缺口与结构覆盖结果。
+原始 `blocked`/`partial` 状态仍返回 BLOCKED；不能把这些状态当作已完成。
+
+证明与发布收据契约由 [`foundation_publication`](../../kb_service/foundation_publication.md)
+负责，使用入口见[部分发布说明](../../../../knowledge/partial-foundation.md)。
+接口测试：`test_kb_foundation_partial_interfaces.py`；原生证明和门槛测试：
+`test_kb_foundation_explicit_partial.py`。

doc/architecture/SPEC/kb_service.md

@@ -599,7 +599,7 @@ proof `basis` 缺省 supported，旧区块不改写。verified_absent 当前只
 可选政策 `semantic_depth: {per_facet_gt: 0.90}` 对每个维度严格执行大于目标，并要求
 每个功能至少一个认可区块；79 功能时每维至少 72，553 分母不缩减。target_met 与
 批次遍历完成分别记录。提取前复核结构覆盖；结构门槛不足先阻断，不消耗模型调用。
-结构覆盖与深度门槛均须满足才能发布。深度未达标 dry run 为 partial，保存预览但步骤不成功；发布阻断。
+默认 strict 流程中，结构覆盖与深度门槛均须满足才能发布。深度未达标 dry run 为 partial，保存预览但步骤不成功；发布阻断。
 执行 API 的 feature_ids 只限定调度范围；审计仍包含完整政策中的所有功能和维度。
 显式列表保留执行顺序并纳入输入摘要；零深度功能仍优先，之后按列表顺序开展首轮。
 并行批次各自保留检查点，合并认可页面后须重新通过全量验收，不能发布分组的部分结果。
@@ -618,6 +618,14 @@ modules/knowledge 的公共阶段预算仅在显式启用时改为 None；单独
 仍保留原记账上限。该模式绑定检查点身份，无此选项时的旧摘要保持不变；
 feature-discovery 与 knowledge-deepen 继续使用各自已有的检查点及身份实现。
 
+`--foundation-mode partial` 是显式基础知识部分发布及其新深度批次交接选项，
+不降低七维认可门槛。基础发布只重放已经结束的订阅批次，保留原生成身份、任务、
+修正次数、真实未知项与 `init_complete=false`；源码、原生证明、格式和结构覆盖
+仍须通过。新深度批次要求原始已合并发布记录，绑定模式、记录和收据哈希，
+同一完整功能目录和七维分母不变。收据、紧凑 PR 正文和 prepared 恢复契约见
+[`foundation_publication`](kb_service/foundation_publication.md)，操作入口见
+[部分发布说明](../../knowledge/partial-foundation.md)。
+
 ### 2026-10-02 轻量认可、13 功能任务与轨迹留存
 
 政策可指定 `semantic_depth.acceptance_mode: lightweight`，CLI 可显式选择

doc/architecture/SPEC/kb_service/foundation_publication.md

@@ -0,0 +1,42 @@
+# kb_service/foundation_publication.py —— 规范
+
+<!-- verified-against: 2026-10-06 -->
+
+## 职责
+
+为显式 `--foundation-mode partial` 保存和复核基础知识发布收据，并为新深度批次
+验证原发布记录。调用方的默认完整门槛保持不变；使用方式见
+[部分发布说明](../../../knowledge/partial-foundation.md)。
+
+## 不变量
+
+- 只发布已经完成初轮扫描、保留真实原生任务的固定源码批次。已有基线未覆盖的
+  功能和 owner 缺少初轮任务时拒绝发布；已处理范围不能称为已补齐全部知识。
+- 生产文件结构覆盖达到原政策目标；源码、目录与政策身份、引用、格式及原生
+  认可仍须有效。不能修改功能集合、分母、语义认可条件或失败记录来满足门槛。
+- `freeze_receipt` 重放每个任务的原生证明，再冻结生成身份、KB/source SHA、
+  全部终态任务的输入/结果摘要、实际进入页面的认可区块、输出文件哈希及未知项。
+  收据按内容哈希在原始 Git 外阶段目录独占创建；既有字节或路径损坏必须拒绝。
+- `verify_prepared_receipt` 核对原始 partial 模式、收据范围和 prepared 文件，
+  从固定源码及实际准备发布的字节重新计算覆盖。不能只信检查点里的计数。
+  旧 strict prepared 发布保持原行为，不能偷偷转换为 partial。
+- `foundation_handoff` 只接受原始 published、非 dry-run 的 partial 记录。
+  记录、收据、任务、prepared 文件、当前合并基线及政策必须一致，且真实 PR
+  状态为 MERGED。返回原记录/收据路径与哈希、固定 pin、目录绑定和输出文件哈希。
+  下游以这些字节建立新批次身份，发生变化即拒绝复用。
+- 发布状态只表示 PR 传输。收据保留 `init_complete=false`、原六维目标结果和
+  所有未知项；结构覆盖、六维基础知识、七维深度认可及实际测试执行分别表述。
+
+## PR 正文与恢复
+
+`render_partial_body` 只渲染汇总数量、固定身份、真实未知维度和收据引用；完整
+模型输入、输出、评审及历史仍在 Git 外。费用未报告时为未知。
+若原 prepared 发布因正文过长而未成功，官方恢复路径须先通过全部原生、源码和
+收据校验，再仅刷新正文元数据：commit、文件字节、任务和修正次数不得改变，
+不得重新生成知识或伪造发布状态。
+
+## 验证
+
+`test_kb_foundation_explicit_partial.py` 覆盖原生证明、范围、收据、恢复与门槛；
+`test_kb_foundation_partial_interfaces.py` 覆盖 CLI、campaign、worker 与 assembly
+身份传递；`test_kb_foundation_partial_prbody.py` 覆盖紧凑正文和官方正文恢复。

doc/knowledge/partial-foundation.md

@@ -0,0 +1,34 @@
+# Explicit partial foundation publication
+
+The default initialization workflow requires all six foundation facets for every
+catalog feature. `--foundation-mode partial` is an explicit alternative for an
+existing, source-pinned subscription batch whose genuine native tasks have
+already finished. It reconstructs approved sections and inspects the declared
+scope without calling either model again. Unknown sections remain unknown;
+local publication exclusions retain their original native evidence archives.
+
+Publication still requires the configured production-file threshold, valid
+catalog and source inventories, genuine native proofs, valid knowledge files,
+and a complete terminal scope. Missing initial feature tasks block publication
+unless that feature was already covered in the merged baseline. The external
+immutable receipt records the original generation identity, source/catalog and
+policy, terminal tasks, active approvals, exact output hashes and unknowns.
+Prepared publication retries recheck these bindings and the actual pinned
+coverage before writing. Publication status describes PR transport: the
+foundation target remains false when six-facet gaps remain, and
+`init_complete` remains false.
+
+After this foundation PR merges, start a **new depth batch** with
+`--foundation-mode partial --foundation-record /path/to/original/knowledge.json`.
+The original published record and receipt, merged output bytes and fixed source
+pin become immutable depth inputs. The campaign and standalone depth audit use
+the same structural prerequisite; six-facet foundation completion is reported
+separately. The audit also requires native approval reports in this mode.
+
+This option preserves the catalog and all semantic acceptance criteria. Each
+feature must have recognized depth knowledge, each of the seven depth facets
+must exceed the policy threshold, and production coverage must meet the policy
+target. The denominator remains the complete frozen catalog times seven.
+Partition workers report their actual execution scope separately from global
+semantic completion; completing a partition never implies the whole catalog
+has passed. Retrieval, independent review and CI remain delivery gates.

eval/knowledge-depth/assemble_depth_campaign.py

@@ -172,7 +172,8 @@ def _filtered(text, page, retired):
     return text
 
 
-def assemble(root: Path, state: Path, source: Path, *, retirement_reports=()):
+def assemble(root: Path, state: Path, source: Path, *, retirement_reports=(),
+             foundation_mode="strict", foundation_record_path: Path | None = None):
     """Return metadata and validated writes without modifying any input."""
     root, state, source = root.resolve(), state.resolve(), source.resolve()
     metadata_raw = (state / "campaign.json").read_bytes()
@@ -181,6 +182,25 @@ def assemble(root: Path, state: Path, source: Path, *, retirement_reports=()):
     if not isinstance(repo, str) or not re.fullmatch(r"[a-zA-Z0-9_-]+", repo) \
             or not all(isinstance(sha, str) and re.fullmatch(r"[0-9a-f]{40}", sha) for sha in (baseline, pin)):
         raise ValueError("campaign needs a safe repo and immutable source/knowledge SHAs")
+    if foundation_mode not in ("strict", "partial") or (
+            foundation_mode == "partial") != bool(foundation_record_path):
+        raise ValueError("partial assembly requires its foundation record; strict assembly cannot use one")
+    if campaign.get("foundation_mode", "strict") != foundation_mode:
+        raise ValueError("assembly foundation mode differs from campaign")
+    foundation = None
+    if foundation_mode == "partial":
+        from infermatrix_copilot.kb_service.foundation_publication import foundation_handoff
+        from infermatrix_copilot.kb_service.init_support import InitError
+        from infermatrix_copilot.kb_service.sources import KnowledgeRepo
+        try:
+            foundation = foundation_handoff(foundation_record_path, knowledge=KnowledgeRepo(root),
+                                            baseline=baseline, repo=repo, pin=pin)
+        except InitError as exc:
+            raise ValueError(str(exc)) from exc
+        if campaign.get("foundation") != foundation:
+            raise ValueError("assembly foundation record differs from campaign")
+    elif campaign.get("foundation") is not None:
+        raise ValueError("strict assembly cannot reuse partial foundation metadata")
     if _git(root, "rev-parse", baseline + "^{commit}").decode().strip() != baseline \
             or _git(source, "rev-parse", "HEAD").decode().strip() != pin \
             or _git(source, "status", "--porcelain", "--untracked-files=no").strip():
@@ -236,6 +256,12 @@ def assemble(root: Path, state: Path, source: Path, *, retirement_reports=()):
                (("stage", "knowledge-deepen"), ("repo", repo), ("pin", pin), ("kb_base_sha", baseline))):
             raise ValueError("worker checkpoint identity differs from campaign: " + worker)
         depth = record.get("depth", {})
+        if foundation is not None and (depth.get("foundation_mode") != "partial"
+                                        or depth.get("foundation") != foundation):
+            raise ValueError("worker foundation identity differs from campaign: " + worker)
+        if foundation is None and (depth.get("foundation_mode", "strict") != "strict"
+                                   or depth.get("foundation") is not None):
+            raise ValueError("strict campaign cannot reuse partial worker state: " + worker)
         entries, accepted = depth.get("features", {}), depth.get("accepted", {})
         if not isinstance(entries, dict) or not isinstance(accepted, dict) or set(entries) - set(features):
             raise ValueError("worker checkpoint has invalid feature state: " + worker)
@@ -267,6 +293,8 @@ def assemble(root: Path, state: Path, source: Path, *, retirement_reports=()):
                 eligible_features.add(feature.id)
             old = baseline_pages.get(page)
             related = f"[功能概览]({PurePosixPath(feature.page).name}) · [owner 入口](_index.md)"
+            if foundation_mode == "partial" and _baseline(root, baseline, "knowledge/" + feature.page) is None:
+                related = "[owner 入口](_index.md)"
             expected_body = _outside_blocks(old) if old is not None else ["# " + _one_line(feature.title) + "：实现深读", related]
             if _outside_blocks(candidate) != expected_body:
                 raise ValueError("accepted page alters prose outside approved blocks: " + feature.id)
@@ -342,6 +370,8 @@ def assemble(root: Path, state: Path, source: Path, *, retirement_reports=()):
               "eligible_feature_count": len(eligible_features),
               "pages": changes, "index_links": new_links, "planned_files": sorted(writes),
               "receipts": receipts, "native_archive_audit_required": True, "problems": []}
+    if foundation is not None:
+        report.update(foundation_mode="partial", foundation=foundation)
     return report, writes
 
 
@@ -351,6 +381,8 @@ def main():
     parser.add_argument("--state", type=Path, required=True)
     parser.add_argument("--source-tree", type=Path, required=True)
     parser.add_argument("--retirement-report", type=Path, action="append", default=[])
+    parser.add_argument("--foundation-mode", choices=("strict", "partial"), default="strict")
+    parser.add_argument("--foundation-record", type=Path)
     parser.add_argument("--report", type=Path)
     parser.add_argument("--apply", action="store_true")
     args = parser.parse_args()
@@ -364,7 +396,8 @@ def main():
                              for n in range(campaign["workers"]))]:
                     handle = stack.enter_context(lock.open("r"))
                     fcntl.flock(handle, fcntl.LOCK_EX | fcntl.LOCK_NB)
-            report, writes = assemble(args.root, args.state, args.source_tree, retirement_reports=args.retirement_report)
+            report, writes = assemble(args.root, args.state, args.source_tree, retirement_reports=args.retirement_report,
+                                     foundation_mode=args.foundation_mode, foundation_record_path=args.foundation_record)
             report["applied"] = args.apply
             if args.apply:
                 for page, text in writes.items():