实现: 建立正文双证据上下文基础设施

.claude/skills/read-context/SKILL.md

@@ -11,6 +11,16 @@ SoT 对齐:包结构=专题-03 §4.2 **四层上下文**;字段级 aiContext 裁
 
 ## 步骤
 
+正文 `continuation` 的 v1 流程在通用四层裁剪之前先执行以下可信读取链，禁止跳步：
+
+1. **固定检索计划**：从已确认细纲的人物、关系、物品、地点、力量体系和硬事件生成 `RetrievalPlan`；计划身份排除 `runId`，查询范围在写手启动前冻结。
+2. **卡索引**：生产只调用 search skill 的 `work` 授权面，只读 active Canonical entity 和有效 binding；参考作品回放只读预注册 `upgrade_book` 评测卡，固定 `productionRetrievalEligible=false`。
+3. **原文回读**：抽取卡只保留 `sourceVersion/sourceRefs/stateAsOf` 索引。每条历史事实必须顺着块级来源指针，在 `REPEATABLE READ READ ONLY` 快照内读取 `chapter <= asOf` 的原文；缺少指针的卡只记 `unverifiedIndexHint`，不能单独进入事实证据。
+4. **双证据组装**：连续前四章全文先进入 `proseEvidence` 基线，再按卡指针补充历史原文并去重；`factEvidence` 只接收冻结历史原文、正式设定、Canonical 状态和细纲声明的新事实。正式设定、Canonical 状态、细纲新事实必须保留各自不可变权威引用，不强造历史原文。
+5. **冻结与回显**：按 `score DESC, sourceVersion ASC, sourceId ASC, sourceOffset ASC` 稳定排序，生成 JSON `WriterContext v1` 和 Markdown `RetrievalManifest`。清单与上下文身份排除运行 ID、时间戳和执行节点；预算裁剪必须回显来源与 `token_budget` 原因。
+
+对应实现：`scripts/writer_contract.py`、`scripts/retrieve_writer_sources.py`、`scripts/assemble_writer_context.py`。写手只能接收最终冻结的 `WriterContext v1`，不得直接接收卡片全文，也不得自行搜索或回读文件。
+
 1. 读 `works/<书>/装配.yaml`:拿槽位绑定与知识绑定。**未绑定的 `knowledge/` 内容一律不读**(=Layer 3 的授权门)。
 2. 读 `meta/schemas/` 相关型的 `aiContext`,得出本用途的可见字段集:`true` 可见;`false` 不可见;`[用途…]` 仅列出的用途可见。
 3. 按**四层**取数并逐字段裁剪——层是分组骨架,取数规则按**来源**逐条执行:
@@ -20,9 +30,9 @@ SoT 对齐:包结构=专题-03 §4.2 **四层上下文**;字段级 aiContext 裁
      - 本章细纲:出自大纲当前章条目;
      - 输出合同:字数区间/frontmatter 要求/新设定申报块等(按 agent 的产出合同);
      - L0 内部同按稳定度排:细纲在前(同章多轮重生成不变),本回合修改意见最后。
-   - **Layer 1 近邻正文**(续写不可省略):
+   - **Layer 1 近邻正文**(正文 continuation v1 不可省略):
-     - 上一章末尾 1–2 个场景**原文**(衔接锚,不得摘要化);
+     - 目标章之前连续四章**全文**作为已验证基线；作品不足四章时从第一章开始，不重复；
-     - 再前一章只给 frontmatter 摘要(一句话概要+出场+伏笔动作);
+     - 卡索引命中的来源原文作为补充，按不可变来源引用去重并稳定排序；
      - `状态.md` **全文**(伏笔台账/时间线/角色最新状态)——"叙事现在时"快照,归近邻层。
    - **Layer 2 作品事实**(一致性/规划不可省略;只读已确认):
      - **设定**:`设定.md` 四节;`generation` 用途裁掉「谜底与真相」「结局方向」「弃案记录」等非 generation 字段;

.claude/skills/read-context/scripts/assemble_writer_context.py

@@ -0,0 +1,459 @@
+#!/usr/bin/env python3
+"""确定性组装 WriterContext v1 与可审阅 RetrievalManifest。
+
+组装器不访问数据库、不调用模型。调用方必须先完成检索计划、卡索引和
+sourceRefs 原文回读，再把冻结结果交给本模块。
+"""
+
+from __future__ import annotations
+
+import copy
+import hashlib
+import json
+from typing import Any, Mapping, Sequence
+
+from writer_contract import (
+    MANIFEST_VERSION,
+    canonical_json,
+    normalize_text,
+    retrieval_identity,
+    validate_writer_context,
+)
+
+
+class AssemblyError(ValueError):
+    """上下文不连续、预算不足或证据合同非法时抛出。"""
+
+
+def _hash_text(text: str) -> str:
+    """对已经归一化的证据文本计算带算法前缀的哈希。"""
+
+    return "sha256:" + hashlib.sha256(text.encode("utf-8")).hexdigest()
+
+
+def _source_key(source_ref: Mapping[str, Any]) -> tuple[str, str, int]:
+    """来源版本、来源 ID 与偏移共同定义片段身份。"""
+
+    return (
+        str(source_ref.get("sourceVersion") or ""),
+        str(source_ref.get("sourceId") or ""),
+        int(source_ref.get("startCodePoint") or 0),
+    )
+
+
+def _whole_source_key(source_ref: Mapping[str, Any]) -> tuple[str, str]:
+    """同一块全文已存在时，卡片子区间不再重复注入。"""
+
+    return (str(source_ref.get("sourceVersion") or ""), str(source_ref.get("sourceId") or ""))
+
+
+def _normalize_prose(raw: Mapping[str, Any], *, recent: bool, purpose: str | None = None) -> dict[str, Any]:
+    """规范化单条原文证据，并机械复核内容哈希和字符区间。"""
+
+    required = {"chapter", "sourceRef", "text"}
+    if not isinstance(raw, Mapping) or not required.issubset(raw):
+        raise AssemblyError("原文证据缺少 chapter/sourceRef/text")
+    chapter = raw["chapter"]
+    if isinstance(chapter, bool) or not isinstance(chapter, int) or chapter <= 0:
+        raise AssemblyError("原文证据 chapter 必须是正整数")
+    source_ref = copy.deepcopy(dict(raw["sourceRef"])) if isinstance(raw["sourceRef"], Mapping) else None
+    if source_ref is None or not source_ref.get("sourceId") or not source_ref.get("sourceVersion"):
+        raise AssemblyError("原文证据缺少不可变来源引用")
+    text = normalize_text(str(raw["text"]))
+    if not text:
+        raise AssemblyError("原文证据不能为空")
+    source_ref["chapter"] = chapter
+    source_ref.setdefault("startCodePoint", 0)
+    source_ref.setdefault("endCodePoint", len(text))
+    if source_ref["endCodePoint"] > len(text) and source_ref["startCodePoint"] == 0:
+        raise AssemblyError("原文来源字符区间超过文本长度")
+    return {
+        "evidenceId": str(raw.get("evidenceId") or f"prose:{source_ref['sourceId']}:{source_ref['startCodePoint']}"),
+        "chapter": chapter,
+        "sourceRef": source_ref,
+        "contentSha256": _hash_text(text),
+        "purpose": purpose or str(raw.get("purpose") or "card_source"),
+        "text": text,
+        "isRecentBaseline": recent,
+    }
+
+
+def _normalize_fact(raw: Mapping[str, Any]) -> dict[str, Any]:
+    """规范化事实证据，并保留来源类型与风险优先级。"""
+
+    required = {"evidenceId", "fact", "sourceType", "sourceRef"}
+    if not isinstance(raw, Mapping) or not required.issubset(raw):
+        raise AssemblyError("事实证据缺少 evidenceId/fact/sourceType/sourceRef")
+    source_type = str(raw["sourceType"])
+    if source_type not in {"historical_prose", "formal_setting", "canonical_state", "fine_outline_declared_new"}:
+        raise AssemblyError("事实证据 sourceType 非法")
+    source_ref = copy.deepcopy(dict(raw["sourceRef"])) if isinstance(raw["sourceRef"], Mapping) else None
+    if source_ref is None or not source_ref.get("sourceId") or not source_ref.get("sourceVersion"):
+        raise AssemblyError("事实证据缺少不可变来源引用")
+    fact = normalize_text(str(raw["fact"]))
+    if not fact:
+        raise AssemblyError("事实证据内容不能为空")
+    risk = str(raw.get("riskLevel") or "medium")
+    if risk not in {"low", "medium", "high"}:
+        raise AssemblyError("事实证据 riskLevel 非法")
+    return {
+        "evidenceId": str(raw["evidenceId"]),
+        "fact": fact,
+        "sourceType": source_type,
+        "sourceRef": source_ref,
+        "contentSha256": _hash_text(fact),
+        "riskLevel": risk,
+    }
+
+
+def _select_recent_baseline(recent_chapters: Sequence[Mapping[str, Any]], *, as_of: int) -> list[dict[str, Any]]:
+    """选择冻结章起连续四章全文；作品不足四章时从第一章开始。"""
+
+    by_chapter: dict[int, dict[str, Any]] = {}
+    for raw in recent_chapters:
+        normalized = _normalize_prose(raw, recent=True, purpose="recent_full_chapter")
+        chapter = normalized["chapter"]
+        if chapter > as_of:
+            raise AssemblyError("近章原文包含目标章或未来章")
+        if chapter in by_chapter:
+            raise AssemblyError(f"第 {chapter} 章完整原文重复")
+        by_chapter[chapter] = normalized
+    first = max(1, as_of - 3)
+    expected = list(range(first, as_of + 1))
+    missing = [chapter for chapter in expected if chapter not in by_chapter]
+    if missing:
+        raise AssemblyError(f"连续四章基线缺章: {','.join(map(str, missing))}")
+    return [by_chapter[chapter] for chapter in expected]
+
+
+def _outline_contract(fine_outline: Mapping[str, Any]) -> dict[str, Any]:
+    """只保留 WriterContext 合同字段，检索用实体列表不会被倾倒给写手。"""
+
+    required = {"sourceRef", "hardConstraints", "adjustableBeats", "declaredNewFacts"}
+    if not isinstance(fine_outline, Mapping) or not required.issubset(fine_outline):
+        raise AssemblyError("fineOutline 缺少严格合同字段")
+    return {key: copy.deepcopy(fine_outline[key]) for key in ("sourceRef", "hardConstraints", "adjustableBeats", "declaredNewFacts")}
+
+
+def _coverage_elements(fine_outline: Mapping[str, Any]) -> list[dict[str, str]]:
+    """从细纲提取人物、关系、物品、地点和力量体系覆盖目标。"""
+
+    groups = (
+        ("entities", None),
+        ("relations", "character_relation"),
+        ("items", "item"),
+        ("locations", "location"),
+        ("powerSystems", "power_system"),
+    )
+    result: list[dict[str, str]] = []
+    seen: set[str] = set()
+    for field, fallback_type in groups:
+        values = fine_outline.get(field, [])
+        if not isinstance(values, list):
+            raise AssemblyError(f"fineOutline.{field} 必须是数组")
+        for index, raw in enumerate(values):
+            if isinstance(raw, str):
+                item = {"id": f"{field}:{index}", "type": fallback_type or "unknown", "name": raw}
+            elif isinstance(raw, Mapping):
+                item = {
+                    "id": str(raw.get("id") or f"{field}:{index}"),
+                    "type": str(raw.get("type") or fallback_type or "unknown"),
+                    "name": str(raw.get("name") or ""),
+                }
+            else:
+                raise AssemblyError(f"fineOutline.{field}[{index}] 类型非法")
+            if not item["name"]:
+                raise AssemblyError(f"fineOutline.{field}[{index}] 缺少名称")
+            if item["id"] not in seen:
+                seen.add(item["id"])
+                result.append(item)
+    return result
+
+
+def _build_coverage(
+    fine_outline: Mapping[str, Any],
+    facts: Sequence[Mapping[str, Any]],
+    prose: Sequence[Mapping[str, Any]],
+    cards: Sequence[Mapping[str, Any]],
+) -> list[dict[str, Any]]:
+    """把五类细纲要素映射到最终保留的事实与原文证据。"""
+
+    prose_by_source = {
+        str(item["sourceRef"]["sourceId"]): item["evidenceId"]
+        for item in prose
+    }
+    declared = fine_outline.get("declaredNewFacts", [])
+    result: list[dict[str, Any]] = []
+    for element in _coverage_elements(fine_outline):
+        fact_ids = [item["evidenceId"] for item in facts if element["name"] in item["fact"]]
+        prose_ids: list[str] = []
+        for card in cards:
+            if str(card.get("name")) != element["name"] and str(card.get("type")) != element["type"]:
+                continue
+            for ref in card.get("sourceRefs") or []:
+                evidence_id = prose_by_source.get(str(ref.get("sourceId")))
+                if evidence_id and evidence_id not in prose_ids:
+                    prose_ids.append(evidence_id)
+        is_declared = any(
+            isinstance(item, Mapping)
+            and (str(item.get("factId")) == element["id"] or element["name"] in str(item.get("text") or ""))
+            for item in declared
+        )
+        if is_declared:
+            status, reason = "declared_new", ""
+        elif fact_ids and prose_ids:
+            status, reason = "supported", ""
+        elif fact_ids:
+            status, reason = "style_gap", "缺少历史表现原文"
+        elif prose_ids:
+            status, reason = "card_gap", "原文可证但缺少冻结事实索引"
+        else:
+            status, reason = "unsupported", "没有可信事实证据"
+        result.append(
+            {
+                "elementId": element["id"],
+                "elementType": element["type"],
+                "name": element["name"],
+                "status": status,
+                "factEvidenceIds": sorted(fact_ids),
+                "proseEvidenceIds": sorted(prose_ids),
+                "gapReason": reason,
+            }
+        )
+    return result
+
+
+def _manifest(
+    *,
+    plan_id: str,
+    facts: Sequence[Mapping[str, Any]],
+    prose: Sequence[Mapping[str, Any]],
+    pattern_references: Sequence[Mapping[str, Any]],
+    omitted: Sequence[Mapping[str, str]],
+) -> dict[str, Any]:
+    """由最终入包来源集合计算稳定 manifest，不含 runId 或时间戳。"""
+
+    unique: dict[tuple[str, str, int], dict[str, Any]] = {}
+    for item in [*facts, *prose]:
+        ref = copy.deepcopy(dict(item["sourceRef"]))
+        unique[_source_key(ref)] = ref
+    for raw_ref in pattern_references:
+        ref = copy.deepcopy(dict(raw_ref))
+        unique[_source_key(ref)] = ref
+    sources = [unique[key] for key in sorted(unique)]
+    omitted_rows = sorted(
+        [copy.deepcopy(dict(item)) for item in omitted],
+        key=lambda item: (str(item.get("reason")), str(item.get("sourceId"))),
+    )
+    payload = {
+        "manifestVersion": MANIFEST_VERSION,
+        "planId": plan_id,
+        "sources": sources,
+        "omittedSources": omitted_rows,
+    }
+    return {**payload, "manifestId": retrieval_identity(payload)}
+
+
+def _markdown_manifest(manifest: Mapping[str, Any]) -> str:
+    """渲染不含运行时元数据的人类审阅清单，保证字节稳定。"""
+
+    lines = [
+        "# 正文检索清单",
+        "",
+        f"- 计划：`{manifest['planId']}`",
+        f"- 清单：`{manifest['manifestId']}`",
+        f"- 纳入来源：{len(manifest['sources'])}",
+        f"- 排除来源：{len(manifest['omittedSources'])}",
+        "",
+        "## 纳入来源",
+        "",
+    ]
+    if manifest["sources"]:
+        for source in manifest["sources"]:
+            location = f"第{source['chapter']}章" if "chapter" in source else "权威版本"
+            lines.append(f"- `{source['sourceId']}` @ `{source['sourceVersion']}`，{location}")
+    else:
+        lines.append("- 无")
+    lines.extend(["", "## 排除来源", ""])
+    if manifest["omittedSources"]:
+        for source in manifest["omittedSources"]:
+            lines.append(f"- `{source['sourceId']}`：{source['reason']}")
+    else:
+        lines.append("- 无")
+    return "\n".join(lines) + "\n"
+
+
+def _context_size(context: Mapping[str, Any]) -> int:
+    """用最终规范 JSON 的 Unicode code point 数作为确定性预算单位。"""
+
+    return len(canonical_json(context))
+
+
+def assemble_context(
+    *,
+    run_id: str,
+    attempt: int,
+    mode: str,
+    purpose: str,
+    quality_policy_version: str,
+    work_id: int,
+    target_chapter: int,
+    as_of: int,
+    source_version: str,
+    authorization_snapshot: Mapping[str, Any],
+    source_status: str,
+    retrieval_plan: Mapping[str, Any],
+    retrieval_result: Mapping[str, Any],
+    fine_outline: Mapping[str, Any],
+    narrative_state: Mapping[str, Any],
+    recent_chapters: Sequence[Mapping[str, Any]],
+    output_contract: Mapping[str, Any],
+    token_budget: Mapping[str, int],
+    pattern_references: Sequence[Mapping[str, Any]] = (),
+    generated_at: str,
+) -> dict[str, Any]:
+    """组装稳定 WriterContext，并返回规范 JSON 与 Markdown manifest。"""
+
+    if retrieval_plan.get("runId") != run_id or retrieval_plan.get("asOf") != as_of:
+        raise AssemblyError("retrievalPlan 与当前 runId/asOf 不一致")
+    if retrieval_plan.get("filters", {}).get("workId") != work_id:
+        raise AssemblyError("retrievalPlan 与当前 workId 不一致")
+    max_chars = token_budget.get("maxContextChars")
+    if isinstance(max_chars, bool) or not isinstance(max_chars, int) or max_chars <= 0:
+        raise AssemblyError("tokenBudget.maxContextChars 必须是正整数")
+
+    baseline = _select_recent_baseline(recent_chapters, as_of=as_of)
+    baseline_blocks = {_whole_source_key(item["sourceRef"]) for item in baseline}
+    supplemental: list[dict[str, Any]] = []
+    seen_fragments = {_source_key(item["sourceRef"]) for item in baseline}
+    for raw in retrieval_result.get("proseEvidence", []):
+        item = _normalize_prose(raw, recent=False)
+        if item["chapter"] > as_of:
+            raise AssemblyError("卡来源原文包含目标章或未来章")
+        if _whole_source_key(item["sourceRef"]) in baseline_blocks:
+            continue
+        key = _source_key(item["sourceRef"])
+        if key not in seen_fragments:
+            seen_fragments.add(key)
+            supplemental.append(item)
+    supplemental.sort(key=lambda item: _source_key(item["sourceRef"]))
+    facts = sorted(
+        [_normalize_fact(item) for item in retrieval_result.get("factEvidence", [])],
+        key=lambda item: ({"high": 0, "medium": 1, "low": 2}[item["riskLevel"]], item["evidenceId"]),
+    )
+    cards = [copy.deepcopy(dict(item)) for item in retrieval_result.get("cards", [])]
+    inherited_omitted = [
+        {"sourceId": str(item.get("sourceId") or "unknown"), "reason": str(item.get("reason") or "not_relevant")}
+        for item in retrieval_result.get("manifest", {}).get("omittedSources", [])
+        if isinstance(item, Mapping)
+    ]
+
+    selected_facts: list[dict[str, Any]] = []
+    # 连续前四章全文是正文实验 v1 的不可裁剪基线。预算容不下时必须失败关闭，
+    # 不能静默退化成只保留最近一两章。
+    selected_prose: list[dict[str, Any]] = list(baseline)
+    omitted = list(inherited_omitted)
+    candidates: list[tuple[str, dict[str, Any]]] = []
+    candidates.extend(("fact", item) for item in facts if item["riskLevel"] == "high")
+    candidates.extend(("fact", item) for item in facts if item["riskLevel"] == "medium")
+    candidates.extend(("fact", item) for item in facts if item["riskLevel"] == "low")
+    candidates.extend(("prose", item) for item in supplemental)
+
+    def make_context() -> dict[str, Any]:
+        """用当前选择集生成完整上下文，供预算试算与最终冻结。"""
+
+        ordered_prose = sorted(
+            selected_prose,
+            key=lambda item: (
+                0 if item["isRecentBaseline"] else 1,
+                item["chapter"] if item["isRecentBaseline"] else 0,
+                _source_key(item["sourceRef"]),
+            ),
+        )
+        ordered_facts = sorted(selected_facts, key=lambda item: item["evidenceId"])
+        manifest = _manifest(
+            plan_id=str(retrieval_plan["planId"]),
+            facts=ordered_facts,
+            prose=ordered_prose,
+            pattern_references=pattern_references,
+            omitted=omitted,
+        )
+        context = {
+            "schemaVersion": "writer-context-v1",
+            "runId": run_id,
+            "attempt": attempt,
+            "mode": mode,
+            "purpose": purpose,
+            "qualityPolicyVersion": quality_policy_version,
+            "workId": work_id,
+            "targetChapter": target_chapter,
+            "asOf": as_of,
+            "contextSnapshot": {
+                "manifestId": manifest["manifestId"],
+                "contextSha256": "sha256:" + "0" * 64,
+                "generatedAt": normalize_text(generated_at),
+            },
+            "sourceVersion": source_version,
+            "authorizationSnapshot": copy.deepcopy(dict(authorization_snapshot)),
+            "sourceStatus": source_status,
+            "retrievalPlan": copy.deepcopy(dict(retrieval_plan)),
+            "retrievalManifest": manifest,
+            "fineOutline": _outline_contract(fine_outline),
+            "narrativeState": copy.deepcopy(dict(narrative_state)),
+            "factEvidence": ordered_facts,
+            "proseEvidence": ordered_prose,
+            "patternReferences": [copy.deepcopy(dict(item)) for item in pattern_references],
+            "evidenceCoverage": _build_coverage(fine_outline, ordered_facts, ordered_prose, cards),
+            "outputContract": copy.deepcopy(dict(output_contract)),
+            "tokenBudget": {"maxContextChars": max_chars, "usedContextChars": 0},
+            "omittedSources": sorted(copy.deepcopy(omitted), key=lambda item: (item["reason"], item["sourceId"])),
+            "acceptanceEligible": mode == "production" and purpose == "production",
+        }
+        # usedContextChars 自身位数会影响 JSON 长度，迭代到数值稳定。
+        for _ in range(8):
+            used = _context_size(context)
+            if context["tokenBudget"]["usedContextChars"] == used:
+                break
+            context["tokenBudget"]["usedContextChars"] = used
+        context["contextSnapshot"]["contextSha256"] = retrieval_identity(context)
+        return context
+
+    baseline_context = make_context()
+    if _context_size(baseline_context) > max_chars:
+        raise AssemblyError("上下文预算不足以容纳细纲硬约束与连续前四章全文基线")
+    for kind, item in candidates:
+        target = selected_facts if kind == "fact" else selected_prose
+        target.append(item)
+        trial = make_context()
+        if _context_size(trial) > max_chars:
+            target.pop()
+            omitted.append({"sourceId": str(item["sourceRef"]["sourceId"]), "reason": "token_budget"})
+
+    context = make_context()
+    # 添加裁剪回显可能占用少量预算；若越界，按最低优先级继续移除。
+    while _context_size(context) > max_chars and (selected_prose or selected_facts):
+        removable_prose = [item for item in selected_prose if not item["isRecentBaseline"]]
+        if removable_prose:
+            removed = removable_prose[-1]
+            selected_prose.remove(removed)
+        elif selected_facts and selected_facts[-1]["riskLevel"] != "high":
+            removed = selected_facts.pop()
+        elif selected_facts:
+            removed = selected_facts.pop()
+        else:
+            raise AssemblyError("上下文预算不足以保留连续前四章全文基线")
+        omitted.append({"sourceId": str(removed["sourceRef"]["sourceId"]), "reason": "token_budget"})
+        context = make_context()
+    if _context_size(context) > max_chars:
+        raise AssemblyError("上下文预算不足以保留可用证据")
+    context["tokenBudget"]["usedContextChars"] = _context_size(context)
+    context["contextSnapshot"]["contextSha256"] = retrieval_identity(context)
+    normalized_context = validate_writer_context(context)
+    return {
+        "context": normalized_context,
+        "contextJson": canonical_json(normalized_context),
+        "manifestMarkdown": _markdown_manifest(normalized_context["retrievalManifest"]),
+    }
+
+
+__all__ = ["AssemblyError", "assemble_context"]

.claude/skills/read-context/scripts/retrieve_writer_sources.py

@@ -0,0 +1,518 @@
+#!/usr/bin/env python3
+"""根据固定卡索引计划回读冻结历史原文。
+
+抽取卡只提供定位与历史状态索引。任何由卡承载的历史事实都必须展开
+sourceRefs 并成功读取 asOf 以内的原文；正式设定、Canonical 状态和细纲
+新事实通过独立不可变引用进入事实证据。
+"""
+
+from __future__ import annotations
+
+import copy
+import hashlib
+import pathlib
+import sys
+from typing import Any, Callable, Mapping, Protocol, Sequence
+
+from writer_contract import PLAN_VERSION, TIE_BREAK, canonical_json, normalize_text, retrieval_identity
+
+
+class RetrievalError(ValueError):
+    """检索计划、授权、冻结或来源引用不能证明安全时抛出。"""
+
+
+class CardIndexRepository(Protocol):
+    """卡索引仓储只接收固定计划，不允许写手自行追加查询。"""
+
+    def search(self, plan: Mapping[str, Any]) -> list[dict[str, Any]]:
+        """返回带稳定排序字段、历史里程碑和来源指针的卡索引。"""
+
+
+class ProseRepository(Protocol):
+    """原文仓储只能读取冻结线内的不可变来源引用。"""
+
+    def read_source_refs(
+        self,
+        *,
+        work_id: int,
+        as_of: int,
+        source_refs: Sequence[Mapping[str, Any]],
+    ) -> list[dict[str, Any]]:
+        """展开已校验的来源指针。"""
+
+
+def _load_search_cards() -> Callable[..., list[dict[str, Any]]]:
+    """延迟导入 search skill，保持纯数据测试不触发嵌入依赖。"""
+
+    search_scripts = pathlib.Path(__file__).resolve().parents[2] / "search" / "scripts"
+    sys.path.insert(0, str(search_scripts))
+    from search import search_cards
+
+    return search_cards
+
+
+def _chapter(value: Any, path: str) -> int:
+    """严格解析正整数章号，拒绝 bool 和猜测性字符串。"""
+
+    if isinstance(value, bool):
+        raise RetrievalError(f"{path} 必须是正整数章号")
+    if isinstance(value, str) and value.isdigit():
+        value = int(value)
+    if not isinstance(value, int) or value <= 0:
+        raise RetrievalError(f"{path} 必须是正整数章号")
+    return value
+
+
+def _milestone_chapter(value: Mapping[str, Any]) -> int:
+    """兼容实验库中英文里程碑章号字段，但不解析模糊文本。"""
+
+    for key in ("chapter", "章", "chapterNo", "order_no"):
+        if key in value:
+            return _chapter(value[key], f"milestone.{key}")
+    raise RetrievalError("卡里程碑缺少可证明的绝对章号")
+
+
+def stable_sort_cards(cards: Sequence[Mapping[str, Any]]) -> list[dict[str, Any]]:
+    """按合同固定卡片顺序，同分召回不会因数据库执行计划漂移。"""
+
+    copied = [copy.deepcopy(dict(card)) for card in cards]
+    try:
+        return sorted(
+            copied,
+            key=lambda item: (
+                -float(item["score"]),
+                str(item["sourceVersion"]),
+                str(item["sourceId"]),
+                int(item.get("sourceOffset") or 0),
+            ),
+        )
+    except (KeyError, TypeError, ValueError) as error:
+        raise RetrievalError("卡缺少稳定排序字段") from error
+
+
+def _outline_elements(fine_outline: Mapping[str, Any]) -> list[dict[str, str]]:
+    """按合同顺序提取需要检索的人物、关系、物品、地点和力量体系。"""
+
+    groups = (
+        ("entities", None),
+        ("relations", "character_relation"),
+        ("items", "item"),
+        ("locations", "location"),
+        ("powerSystems", "power_system"),
+        ("stateNeeds", "unknown"),
+    )
+    result: list[dict[str, str]] = []
+    seen: set[tuple[str, str]] = set()
+    for field, fallback_type in groups:
+        values = fine_outline.get(field, [])
+        if not isinstance(values, list):
+            raise RetrievalError(f"fineOutline.{field} 必须是数组")
+        for index, raw in enumerate(values):
+            if isinstance(raw, str):
+                item = {"id": f"{field}:{index}", "type": fallback_type or "unknown", "name": raw}
+            elif isinstance(raw, Mapping):
+                item = {
+                    "id": str(raw.get("id") or f"{field}:{index}"),
+                    "type": str(raw.get("type") or fallback_type or "unknown"),
+                    "name": str(raw.get("name") or ""),
+                }
+            else:
+                raise RetrievalError(f"fineOutline.{field}[{index}] 类型非法")
+            if not item["name"]:
+                raise RetrievalError(f"fineOutline.{field}[{index}] 缺少 name")
+            identity = (item["type"], item["name"])
+            if identity not in seen:
+                seen.add(identity)
+                result.append(item)
+    return result
+
+
+def build_retrieval_plan(
+    *,
+    run_id: str,
+    work_id: int,
+    target_chapter: int,
+    as_of: int,
+    fine_outline: Mapping[str, Any],
+    card_index_version: str,
+    prose_index_version: str,
+    token_budget: Mapping[str, int],
+    top_k: int = 5,
+) -> dict[str, Any]:
+    """从已确认细纲确定查询集合，并在执行前冻结计划身份。"""
+
+    target = _chapter(target_chapter, "targetChapter")
+    freeze = _chapter(as_of, "asOf")
+    if freeze >= target:
+        raise RetrievalError("asOf 必须早于 targetChapter")
+    if not isinstance(work_id, int) or isinstance(work_id, bool) or work_id <= 0:
+        raise RetrievalError("workId 必须是正整数")
+    if not isinstance(top_k, int) or isinstance(top_k, bool) or top_k <= 0:
+        raise RetrievalError("topK 必须是正整数")
+    hard_constraints = fine_outline.get("hardConstraints", [])
+    if not isinstance(hard_constraints, list) or any(not isinstance(item, str) for item in hard_constraints):
+        raise RetrievalError("fineOutline.hardConstraints 必须是字符串数组")
+    suffix = " ".join(hard_constraints)
+    queries = [
+        {
+            "queryId": f"query-{index:03d}-{item['id']}",
+            "text": normalize_text(f"{item['name']} {suffix}".strip()),
+            "entityTypes": [item["type"]],
+            "purpose": "fact_and_prose_evidence",
+            "topK": top_k,
+        }
+        for index, item in enumerate(_outline_elements(fine_outline), 1)
+    ]
+    plan = {
+        "planVersion": PLAN_VERSION,
+        "runId": run_id,
+        "asOf": freeze,
+        "queries": queries,
+        "cardIndexVersion": str(card_index_version),
+        "proseIndexVersion": str(prose_index_version),
+        "filters": {
+            "workId": work_id,
+            "asOfChapter": freeze,
+            "sourceStatus": "active",
+            "authorizationRequired": True,
+        },
+        "tieBreak": TIE_BREAK,
+        "tokenBudget": copy.deepcopy(dict(token_budget)),
+    }
+    plan["planId"] = retrieval_identity(plan)
+    return plan
+
+
+def _validate_source_ref(ref: Mapping[str, Any], *, as_of: int, path: str) -> dict[str, Any]:
+    """校验块级原文引用完整且不越过冻结线。"""
+
+    required = {"sourceId", "sourceVersion", "chapter", "blockId", "startCodePoint", "endCodePoint"}
+    if not isinstance(ref, Mapping) or not required.issubset(ref):
+        raise RetrievalError(f"{path} 缺少块级来源定位")
+    chapter = _chapter(ref["chapter"], f"{path}.chapter")
+    if chapter > as_of:
+        raise RetrievalError(f"{path} 包含目标章或未来章")
+    start = ref["startCodePoint"]
+    end = ref["endCodePoint"]
+    if any(isinstance(value, bool) or not isinstance(value, int) for value in (ref["blockId"], start, end)):
+        raise RetrievalError(f"{path} 的块或字符区间非法")
+    if ref["blockId"] <= 0 or start < 0 or end <= start:
+        raise RetrievalError(f"{path} 的块或字符区间非法")
+    if not str(ref["sourceId"]).strip() or not str(ref["sourceVersion"]).strip():
+        raise RetrievalError(f"{path} 缺少来源 ID 或版本")
+    return copy.deepcopy(dict(ref))
+
+
+def freeze_card(card: Mapping[str, Any], *, as_of: int) -> dict[str, Any]:
+    """仅用冻结线内里程碑重建卡状态，不透传终态摘要和未来字段。"""
+
+    freeze = _chapter(as_of, "asOf")
+    milestones = card.get("milestones")
+    if not isinstance(milestones, list):
+        raise RetrievalError(f"卡 {card.get('cardId')} 缺少历史里程碑")
+    kept: list[dict[str, Any]] = []
+    for index, milestone in enumerate(milestones):
+        if not isinstance(milestone, Mapping):
+            raise RetrievalError(f"卡里程碑[{index}] 必须是对象")
+        chapter = _milestone_chapter(milestone)
+        if chapter <= freeze:
+            normalized = copy.deepcopy(dict(milestone))
+            normalized["chapter"] = chapter
+            for alias in ("章", "chapterNo", "order_no"):
+                normalized.pop(alias, None)
+            kept.append(normalized)
+    kept.sort(key=lambda item: (item["chapter"], str(item.get("id") or "")))
+    if not kept:
+        raise RetrievalError(f"卡 {card.get('cardId')} 在冻结线内没有可证明状态")
+    refs = [
+        _validate_source_ref(ref, as_of=freeze, path=f"card.sourceRefs[{index}]")
+        for index, ref in enumerate(card.get("sourceRefs") or [])
+    ]
+    return {
+        "cardId": str(card.get("cardId") or ""),
+        "type": str(card.get("type") or ""),
+        "name": str(card.get("name") or ""),
+        "score": float(card.get("score") or 0),
+        "sourceId": str(card.get("sourceId") or ""),
+        "sourceVersion": str(card.get("sourceVersion") or ""),
+        "sourceOffset": int(card.get("sourceOffset") or 0),
+        "sourceRefs": refs,
+        "stateAsOf": kept,
+        "sourceKind": str(card.get("sourceKind") or ""),
+        "productionRetrievalEligible": bool(card.get("productionRetrievalEligible")),
+    }
+
+
+class ProductionCardIndexRepository:
+    """生产卡仓储固定调用 search.py 的 work 授权面。"""
+
+    def __init__(self, *, search_function: Callable[..., list[dict[str, Any]]] | None = None):
+        self._search = search_function or _load_search_cards()
+
+    def search(self, plan: Mapping[str, Any]) -> list[dict[str, Any]]:
+        """按计划逐查询召回，并再次失败关闭验证生产资格。"""
+
+        cards: list[dict[str, Any]] = []
+        for query in plan.get("queries", []):
+            entity_types = query.get("entityTypes") or []
+            card_type = entity_types[0] if len(entity_types) == 1 and entity_types[0] != "unknown" else None
+            cards.extend(
+                self._search(
+                    intent=query["text"],
+                    scope="work",
+                    work_id=plan["filters"]["workId"],
+                    ttype=card_type,
+                    purpose="generation",
+                    top=query["topK"],
+                )
+            )
+        unique: dict[str, dict[str, Any]] = {}
+        for card in stable_sort_cards(cards):
+            if (
+                card.get("sourceKind") != "canonical_entity"
+                or card.get("sourceStatus") not in {"active", "authorized"}
+                or card.get("bindingStatus") != "active"
+                or card.get("productionRetrievalEligible") is not True
+            ):
+                raise RetrievalError("生产检索命中非 active Canonical entity 或无效 binding")
+            unique.setdefault(str(card.get("cardId")), copy.deepcopy(card))
+        return stable_sort_cards(list(unique.values()))
+
+
+class ReplayCardIndexRepository:
+    """回放卡仓储只接受预注册 upgrade_book 评测投影。"""
+
+    def __init__(
+        self,
+        cards: Sequence[Mapping[str, Any]],
+        *,
+        preregistered_card_ids: Sequence[str],
+        _validated_replay: bool = False,
+    ):
+        if not _validated_replay:
+            raise RetrievalError("回放仓储必须通过 from_replay_config 完成授权与泄露门禁")
+        expected = {str(item) for item in preregistered_card_ids}
+        actual = {str(item.get("cardId")) for item in cards}
+        if not expected or expected != actual:
+            raise RetrievalError("回放卡与预注册 ID 不一致")
+        self._cards = []
+        for card in cards:
+            if (
+                card.get("sourceType") != "upgrade_book"
+                or card.get("evaluationStatus") != "eval_draft"
+                or card.get("productionRetrievalEligible") is not False
+            ):
+                raise RetrievalError("回放只允许不可生产检索的 upgrade_book 评测卡")
+            self._cards.append(copy.deepcopy(dict(card)))
+
+    @classmethod
+    def from_replay_config(
+        cls,
+        config: Mapping[str, Any],
+        *,
+        cards: Sequence[Mapping[str, Any]],
+        preregistered_card_ids: Sequence[str],
+    ) -> "ReplayCardIndexRepository":
+        """复用 replay-eval 的授权、冻结来源和内容泄露审计后构造仓储。"""
+
+        scripts = pathlib.Path(__file__).resolve().parents[2] / "replay-eval" / "scripts"
+        sys.path.insert(0, str(scripts))
+        from audit_leakage import audit_snapshot
+        from check_snapshot import check_authorization, check_target_sources
+
+        if not isinstance(config, Mapping):
+            raise RetrievalError("回放配置必须是对象")
+        target = _chapter(config.get("targetChapter"), "config.targetChapter")
+        snapshot = config.get("snapshot")
+        if not isinstance(snapshot, Mapping):
+            raise RetrievalError("回放配置缺少冻结快照")
+        as_of = _chapter(snapshot.get("asOfChapter"), "config.snapshot.asOfChapter")
+        if target != as_of + 1:
+            raise RetrievalError("回放 targetChapter 必须等于 asOfChapter+1")
+        authorization_result = check_authorization(config.get("authorization"))
+        if not authorization_result.get("ok"):
+            raise RetrievalError("回放授权快照未通过既有门禁")
+        source_result = check_target_sources(target, config.get("sources", []))
+        if not source_result.get("ok"):
+            raise RetrievalError("回放来源包含目标章、未来章或不可证明区间")
+        leakage = config.get("leakageAudit")
+        target_facts = leakage.get("targetFacts") if isinstance(leakage, Mapping) else None
+        audit_result = audit_snapshot(snapshot, target_facts, as_of=as_of, target=target)
+        if not audit_result.get("ok"):
+            raise RetrievalError("回放冻结快照未通过既有内容泄露审计")
+        return cls(
+            cards,
+            preregistered_card_ids=preregistered_card_ids,
+            _validated_replay=True,
+        )
+
+    def search(self, plan: Mapping[str, Any]) -> list[dict[str, Any]]:
+        """返回预注册集合；计划不能把评测卡转换为生产卡。"""
+
+        del plan
+        return stable_sort_cards(self._cards)
+
+
+class FrozenProseRepository:
+    """复用 load_reference_work 的唯一冻结原文 SQL 入口。"""
+
+    def __init__(self, *, dsn: str, tenant_id: int):
+        self.dsn = dsn
+        self.tenant_id = tenant_id
+
+    def read_source_refs(self, *, work_id: int, as_of: int, source_refs: Sequence[Mapping[str, Any]]) -> list[dict[str, Any]]:
+        """延迟导入 replay-eval，避免复制 SQL 或建立第二套权限语义。"""
+
+        scripts = pathlib.Path(__file__).resolve().parents[2] / "replay-eval" / "scripts"
+        sys.path.insert(0, str(scripts))
+        from load_reference_work import load_frozen_prose_rows
+
+        return load_frozen_prose_rows(
+            dsn=self.dsn,
+            tenant_id=self.tenant_id,
+            work_id=work_id,
+            as_of=as_of,
+            source_refs=source_refs,
+        )
+
+
+def _content_hash(text: str) -> str:
+    """对规范化文本计算带算法前缀的内容哈希。"""
+
+    normalized = normalize_text(text)
+    return "sha256:" + hashlib.sha256(normalized.encode("utf-8")).hexdigest()
+
+
+def _authoritative_evidence(facts: Sequence[Mapping[str, Any]]) -> list[dict[str, Any]]:
+    """把正式设定、Canonical 状态和细纲新事实转换为独立事实证据。"""
+
+    allowed = {"formal_setting", "canonical_state", "fine_outline_declared_new"}
+    result: list[dict[str, Any]] = []
+    for index, raw in enumerate(facts):
+        source_type = raw.get("sourceType")
+        source_ref = raw.get("sourceRef")
+        fact = normalize_text(str(raw.get("fact") or ""))
+        if source_type not in allowed or not fact or not isinstance(source_ref, Mapping):
+            raise RetrievalError(f"authoritativeFacts[{index}] 合同非法")
+        if not source_ref.get("sourceId") or not source_ref.get("sourceVersion"):
+            raise RetrievalError(f"authoritativeFacts[{index}] 缺少不可变来源引用")
+        result.append(
+            {
+                "evidenceId": str(raw.get("factId") or f"authoritative:{index}"),
+                "fact": fact,
+                "sourceType": source_type,
+                "sourceRef": copy.deepcopy(dict(source_ref)),
+                "contentSha256": _content_hash(fact),
+                "riskLevel": str(raw.get("riskLevel") or "medium"),
+            }
+        )
+    return result
+
+
+def retrieve_writer_sources(
+    *,
+    plan: Mapping[str, Any],
+    card_repository: CardIndexRepository,
+    prose_repository: ProseRepository,
+    authoritative_facts: Sequence[Mapping[str, Any]] = (),
+) -> dict[str, Any]:
+    """执行固定计划，展开卡来源并形成事实/原文两条证据线。"""
+
+    as_of = _chapter(plan.get("asOf"), "plan.asOf")
+    work_id = plan.get("filters", {}).get("workId")
+    if not isinstance(work_id, int) or isinstance(work_id, bool) or work_id <= 0:
+        raise RetrievalError("plan.filters.workId 非法")
+    cards = [freeze_card(card, as_of=as_of) for card in stable_sort_cards(card_repository.search(plan))]
+    source_refs: list[dict[str, Any]] = []
+    unverified: list[dict[str, Any]] = []
+    for card_item in cards:
+        if card_item["sourceRefs"]:
+            source_refs.extend(card_item["sourceRefs"])
+        else:
+            unverified.append(
+                {
+                    "cardId": card_item["cardId"],
+                    "reason": "missing_source_refs",
+                    "classification": "unverifiedIndexHint",
+                }
+            )
+    deduplicated = {
+        (str(ref["sourceVersion"]), str(ref["sourceId"]), int(ref["startCodePoint"])): ref
+        for ref in source_refs
+    }
+    ordered_refs = [deduplicated[key] for key in sorted(deduplicated)]
+    prose_rows = prose_repository.read_source_refs(work_id=work_id, as_of=as_of, source_refs=ordered_refs) if ordered_refs else []
+    def ref_key(ref: Mapping[str, Any]) -> tuple[str, str, int]:
+        """同一块可有多个字符区间，映射键必须包含偏移。"""
+
+        return (str(ref.get("sourceVersion")), str(ref.get("sourceId")), int(ref.get("startCodePoint") or 0))
+
+    rows_by_source = {ref_key(row.get("sourceRef", {})): row for row in prose_rows}
+    missing = [ref["sourceId"] for ref in ordered_refs if ref_key(ref) not in rows_by_source]
+    if missing:
+        raise RetrievalError(f"卡来源未能回读原文: {','.join(map(str, missing))}")
+
+    prose_evidence: list[dict[str, Any]] = []
+    for index, ref in enumerate(ordered_refs):
+        row = rows_by_source[ref_key(ref)]
+        text = normalize_text(str(row.get("text") or ""))
+        if not text:
+            raise RetrievalError(f"来源 {ref['sourceId']} 原文为空")
+        prose_evidence.append(
+            {
+                "evidenceId": f"prose:card:{index:04d}",
+                "chapter": _chapter(row.get("chapter"), "prose.chapter"),
+                "sourceRef": copy.deepcopy(ref),
+                "contentSha256": _content_hash(text),
+                "purpose": str(row.get("purpose") or "card_source"),
+                "text": text,
+                "isRecentBaseline": False,
+            }
+        )
+
+    fact_evidence = _authoritative_evidence(authoritative_facts)
+    for card_item in cards:
+        if not card_item["sourceRefs"]:
+            continue
+        latest = card_item["stateAsOf"][-1]
+        fact_text = normalize_text(str(latest.get("fact") or latest.get("台阶") or canonical_json(latest)))
+        fact_evidence.append(
+            {
+                "evidenceId": f"fact:card:{card_item['cardId']}",
+                "fact": fact_text,
+                "sourceType": "historical_prose",
+                "sourceRef": copy.deepcopy(card_item["sourceRefs"][0]),
+                "contentSha256": _content_hash(fact_text),
+                "riskLevel": "medium",
+            }
+        )
+    fact_evidence.sort(key=lambda item: item["evidenceId"])
+    sources = sorted(
+        [copy.deepcopy(ref) for ref in ordered_refs]
+        + [copy.deepcopy(item["sourceRef"]) for item in fact_evidence if item["sourceType"] != "historical_prose"],
+        key=lambda item: (str(item["sourceVersion"]), str(item["sourceId"]), int(item.get("startCodePoint") or 0)),
+    )
+    manifest_payload = {
+        "manifestVersion": "writer-retrieval-manifest-v1",
+        "planId": plan["planId"],
+        "sources": sources,
+        "omittedSources": [
+            {"sourceId": f"card:{item['cardId']}", "reason": item["reason"]}
+            for item in unverified
+        ],
+    }
+    manifest = {**manifest_payload, "manifestId": retrieval_identity(manifest_payload)}
+    return {
+        "cards": cards,
+        "factEvidence": fact_evidence,
+        "proseEvidence": prose_evidence,
+        "unverifiedIndexHints": unverified,
+        "manifest": manifest,
+    }
+
+
+__all__ = [
+    "RetrievalError", "CardIndexRepository", "ProseRepository", "ProductionCardIndexRepository",
+    "ReplayCardIndexRepository", "FrozenProseRepository", "stable_sort_cards", "freeze_card",
+    "build_retrieval_plan", "retrieve_writer_sources",
+]

.claude/skills/read-context/scripts/test_assemble_writer_context.py

@@ -0,0 +1,264 @@
+#!/usr/bin/env python3
+"""事实证据与原文证据双线组装测试。"""
+
+from __future__ import annotations
+
+import copy
+import hashlib
+import pathlib
+import sys
+import unittest
+
+sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+
+from assemble_writer_context import AssemblyError, assemble_context  # noqa: E402
+from writer_contract import retrieval_identity, validate_writer_context  # noqa: E402
+
+
+def digest(text: str) -> str:
+    """生成测试证据的规范 SHA-256。"""
+
+    return "sha256:" + hashlib.sha256(text.encode("utf-8")).hexdigest()
+
+
+def prose(chapter: int, block_id: int, text: str, *, recent: bool = False, purpose: str = "card_source") -> dict:
+    """构造块级可引用原文。"""
+
+    return {
+        "evidenceId": f"prose:{chapter}:{block_id}",
+        "chapter": chapter,
+        "sourceRef": {
+            "sourceId": f"chapter:{chapter}:block:{block_id}",
+            "sourceVersion": f"chapter-{chapter}-block-{block_id}-v1",
+            "chapter": chapter,
+            "blockId": block_id,
+            "startCodePoint": 0,
+            "endCodePoint": len(text),
+        },
+        "contentSha256": digest(text),
+        "purpose": purpose,
+        "text": text,
+        "isRecentBaseline": recent,
+    }
+
+
+def fact(fact_id: str, text: str, source_type: str = "formal_setting", risk: str = "medium") -> dict:
+    """构造独立权威事实证据。"""
+
+    return {
+        "evidenceId": fact_id,
+        "fact": text,
+        "sourceType": source_type,
+        "sourceRef": {
+            "sourceId": f"{source_type}:{fact_id}",
+            "sourceVersion": f"{source_type}-v1",
+        },
+        "contentSha256": digest(text),
+        "riskLevel": risk,
+    }
+
+
+def plan(run_id: str = "run-a") -> dict:
+    """构造合法固定检索计划。"""
+
+    value = {
+        "planVersion": "writer-retrieval-plan-v1",
+        "runId": run_id,
+        "asOf": 6,
+        "queries": [],
+        "cardIndexVersion": "cards-v1",
+        "proseIndexVersion": "prose-v1",
+        "filters": {
+            "workId": 8,
+            "asOfChapter": 6,
+            "sourceStatus": "active",
+            "authorizationRequired": True,
+        },
+        "tieBreak": "score DESC, sourceVersion ASC, sourceId ASC, sourceOffset ASC",
+        "tokenBudget": {"maxContextChars": 50000},
+    }
+    value["planId"] = retrieval_identity(value)
+    return value
+
+
+def fine_outline() -> dict:
+    """构造含五类覆盖目标的细纲输入。"""
+
+    return {
+        "sourceRef": {"sourceId": "fine-outline:7", "sourceVersion": "outline-v3", "chapter": 7},
+        "hardConstraints": ["甲携剑抵达城门"],
+        "adjustableBeats": ["可调整过渡方式"],
+        "declaredNewFacts": [],
+        "entities": [{"id": "character:甲", "type": "character", "name": "甲"}],
+        "relations": [{"id": "relation:甲乙", "type": "character_relation", "name": "甲乙关系"}],
+        "items": [{"id": "item:剑", "type": "item", "name": "剑"}],
+        "locations": [{"id": "location:城门", "type": "location", "name": "城门"}],
+        "powerSystems": [{"id": "power:灵力", "type": "power_system", "name": "灵力"}],
+    }
+
+
+def recent_chapters(last: int = 6) -> list[dict]:
+    """构造第一章到冻结章的完整章节原文。"""
+
+    return [prose(chapter, 100 + chapter, f"第{chapter}章完整原文。") for chapter in range(1, last + 1)]
+
+
+def base_kwargs() -> dict:
+    """构造 assemble_context 的公共合法输入。"""
+
+    return {
+        "run_id": "run-a",
+        "attempt": 1,
+        "mode": "production",
+        "purpose": "production",
+        "quality_policy_version": "writer-production-v1",
+        "work_id": 8,
+        "target_chapter": 7,
+        "as_of": 6,
+        "source_version": "raw-file-v1:sha256:" + "a" * 64,
+        "authorization_snapshot": {
+            "snapshotId": "auth-1",
+            "allowedPurpose": "generation",
+            "verifiedAt": "2026-07-20T00:00:00Z",
+        },
+        "source_status": "active",
+        "retrieval_plan": plan(),
+        "retrieval_result": {
+            "cards": [],
+            "factEvidence": [],
+            "proseEvidence": [],
+            "unverifiedIndexHints": [],
+            "manifest": {
+                "manifestVersion": "writer-retrieval-manifest-v1",
+                "manifestId": "sha256:" + "b" * 64,
+                "planId": plan()["planId"],
+                "sources": [],
+                "omittedSources": [],
+            },
+        },
+        "fine_outline": fine_outline(),
+        "narrative_state": {
+            "time": "当日",
+            "location": "城门",
+            "characterPositions": {"甲": "城外"},
+            "immediateSituation": "即将交战",
+        },
+        "recent_chapters": recent_chapters(),
+        "output_contract": {
+            "targetChars": 4000,
+            "minChars": 3600,
+            "maxChars": 4400,
+            "frontmatterRequired": False,
+            "newSettingDeclarationRequired": True,
+        },
+        "token_budget": {"maxContextChars": 50000},
+        "generated_at": "2026-07-20T00:00:00Z",
+    }
+
+
+class AssembleWriterContextTest(unittest.TestCase):
+    def test_previous_four_full_chapters_are_baseline(self):
+        result = assemble_context(**base_kwargs())
+        baseline = [item for item in result["context"]["proseEvidence"] if item["isRecentBaseline"]]
+        self.assertEqual([item["chapter"] for item in baseline], [3, 4, 5, 6])
+        self.assertTrue(all(item["purpose"] == "recent_full_chapter" for item in baseline))
+        validate_writer_context(result["context"])
+
+    def test_when_fewer_than_four_exist_start_at_chapter_one_without_duplicates(self):
+        kwargs = base_kwargs()
+        kwargs.update(
+            {
+                "target_chapter": 3,
+                "as_of": 2,
+                "recent_chapters": recent_chapters(2),
+                "retrieval_plan": {**plan(), "asOf": 2, "filters": {**plan()["filters"], "asOfChapter": 2}},
+            }
+        )
+        kwargs["retrieval_plan"]["planId"] = retrieval_identity({key: value for key, value in kwargs["retrieval_plan"].items() if key != "planId"})
+        kwargs["retrieval_result"]["manifest"]["planId"] = kwargs["retrieval_plan"]["planId"]
+        result = assemble_context(**kwargs)
+        chapters = [item["chapter"] for item in result["context"]["proseEvidence"]]
+        self.assertEqual(chapters, [1, 2])
+
+    def test_missing_chapter_in_required_continuous_baseline_fails_closed(self):
+        kwargs = base_kwargs()
+        kwargs["recent_chapters"] = [item for item in recent_chapters() if item["chapter"] != 4]
+        with self.assertRaises(AssemblyError):
+            assemble_context(**kwargs)
+
+    def test_card_prose_is_deduplicated_and_stably_sorted_after_baseline(self):
+        kwargs = base_kwargs()
+        duplicate = prose(6, 106, "第6章完整原文。")
+        supplemental_b = prose(2, 202, "补充乙。")
+        supplemental_a = prose(1, 201, "补充甲。")
+        kwargs["retrieval_result"]["proseEvidence"] = [supplemental_b, duplicate, supplemental_a]
+        result = assemble_context(**kwargs)
+        evidence = result["context"]["proseEvidence"]
+        self.assertEqual(sum(item["sourceRef"]["sourceId"] == duplicate["sourceRef"]["sourceId"] for item in evidence), 1)
+        supplemental = [item for item in evidence if not item["isRecentBaseline"]]
+        self.assertEqual([item["sourceRef"]["sourceId"] for item in supplemental], ["chapter:1:block:201", "chapter:2:block:202"])
+
+    def test_fact_and_prose_evidence_are_separate_and_five_types_report_coverage(self):
+        kwargs = base_kwargs()
+        kwargs["retrieval_result"]["factEvidence"] = [
+            fact("fact:character", "甲保持警惕", risk="high"),
+            fact("fact:relation", "甲乙关系稳定"),
+            fact("fact:item", "剑仍在甲手中"),
+            fact("fact:location", "城门已经关闭"),
+            fact("fact:power", "灵力消耗受限"),
+        ]
+        result = assemble_context(**kwargs)
+        context = result["context"]
+        self.assertTrue(all("fact" in item and "text" not in item for item in context["factEvidence"]))
+        self.assertTrue(all("text" in item and "fact" not in item for item in context["proseEvidence"]))
+        self.assertEqual(
+            {item["elementType"] for item in context["evidenceCoverage"]},
+            {"character", "character_relation", "item", "location", "power_system"},
+        )
+        self.assertTrue(all(item["status"] in {"supported", "style_gap"} for item in context["evidenceCoverage"]))
+
+    def test_same_snapshot_and_inputs_have_same_manifest_and_context_hash(self):
+        first = assemble_context(**base_kwargs())
+        second_kwargs = base_kwargs()
+        second_kwargs["run_id"] = "run-b"
+        second_kwargs["retrieval_plan"]["runId"] = "run-b"
+        second = assemble_context(**second_kwargs)
+        self.assertEqual(first["context"]["retrievalManifest"]["manifestId"], second["context"]["retrievalManifest"]["manifestId"])
+        self.assertEqual(first["context"]["contextSnapshot"]["contextSha256"], second["context"]["contextSnapshot"]["contextSha256"])
+        self.assertEqual(first["manifestMarkdown"], second["manifestMarkdown"])
+
+    def test_budget_keeps_hard_outline_recent_prose_and_high_risk_fact_with_trace(self):
+        kwargs = base_kwargs()
+        kwargs["retrieval_result"]["factEvidence"] = [
+            fact("fact:high", "甲的高风险身份事实" + "甲" * 300, risk="high"),
+            fact("fact:low", "低风险背景" + "乙" * 300, risk="low"),
+        ]
+        kwargs["retrieval_result"]["proseEvidence"] = [prose(1, 900, "很早的补充原文" + "旧" * 500)]
+        kwargs["recent_chapters"] = [
+            prose(chapter, 100 + chapter, f"第{chapter}章" + "近" * 350)
+            for chapter in range(1, 7)
+        ]
+        kwargs["token_budget"] = {"maxContextChars": 7500}
+        result = assemble_context(**kwargs)
+        context = result["context"]
+        self.assertIn("甲携剑抵达城门", context["fineOutline"]["hardConstraints"])
+        self.assertIn("fact:high", {item["evidenceId"] for item in context["factEvidence"]})
+        baseline = [item["chapter"] for item in context["proseEvidence"] if item["isRecentBaseline"]]
+        self.assertEqual(baseline, [3, 4, 5, 6])
+        self.assertLessEqual(context["tokenBudget"]["usedContextChars"], 7500)
+        self.assertTrue(context["omittedSources"])
+        self.assertIn("token_budget", {item["reason"] for item in context["omittedSources"]})
+
+    def test_budget_that_cannot_hold_four_full_chapters_fails_closed(self):
+        kwargs = base_kwargs()
+        kwargs["recent_chapters"] = [
+            prose(chapter, 100 + chapter, f"第{chapter}章" + "近" * 500)
+            for chapter in range(1, 7)
+        ]
+        kwargs["token_budget"] = {"maxContextChars": 3000}
+        with self.assertRaisesRegex(AssemblyError, "连续前四章全文基线"):
+            assemble_context(**kwargs)
+
+
+if __name__ == "__main__":
+    unittest.main()

.claude/skills/read-context/scripts/test_retrieve_writer_sources.py

@@ -0,0 +1,399 @@
+#!/usr/bin/env python3
+"""卡索引驱动的冻结原文检索测试。"""
+
+from __future__ import annotations
+
+import copy
+import pathlib
+import sys
+import unittest
+
+SCRIPT_DIR = pathlib.Path(__file__).resolve().parent
+sys.path.insert(0, str(SCRIPT_DIR))
+sys.path.insert(0, str(SCRIPT_DIR.parents[1] / "replay-eval" / "scripts"))
+sys.path.insert(0, str(SCRIPT_DIR.parents[1] / "search" / "scripts"))
+
+from load_reference_work import begin_read_snapshot  # noqa: E402
+from search import search_cards  # noqa: E402
+from retrieve_writer_sources import (  # noqa: E402
+    ProductionCardIndexRepository,
+    ReplayCardIndexRepository,
+    RetrievalError,
+    build_retrieval_plan,
+    freeze_card,
+    retrieve_writer_sources,
+    stable_sort_cards,
+)
+
+
+SOURCE_VERSION = "raw-file-v1:sha256:" + "a" * 64
+
+
+def source_ref(chapter: int, block_id: int, offset: int = 0) -> dict:
+    """构造可回读的冻结原文引用。"""
+
+    return {
+        "sourceId": f"chapter:{chapter}:block:{block_id}",
+        "sourceVersion": SOURCE_VERSION,
+        "chapter": chapter,
+        "blockId": block_id,
+        "startCodePoint": offset,
+        "endCodePoint": offset + 4,
+    }
+
+
+def card(card_id: str, score: float, chapter: int = 3, offset: int = 0) -> dict:
+    """构造含历史里程碑和原文指针的抽取卡索引。"""
+
+    return {
+        "cardId": card_id,
+        "type": "character",
+        "name": f"角色{card_id}",
+        "score": score,
+        "sourceVersion": SOURCE_VERSION,
+        "sourceId": f"canonical-entity:{card_id}",
+        "sourceOffset": offset,
+        "sourceRefs": [source_ref(chapter, int(card_id), offset)],
+        "milestones": [
+            {"chapter": 1, "fact": "初始状态"},
+            {"chapter": chapter, "fact": "冻结线内状态"},
+            {"chapter": 9, "fact": "未来终态"},
+        ],
+        "sourceKind": "canonical_entity",
+        "sourceStatus": "active",
+        "bindingStatus": "active",
+        "productionRetrievalEligible": True,
+    }
+
+
+class FakeCardRepository:
+    """只返回预置卡片并记录固定计划查询。"""
+
+    def __init__(self, cards):
+        self.cards = cards
+        self.plans = []
+
+    def search(self, plan):
+        self.plans.append(copy.deepcopy(plan))
+        return copy.deepcopy(self.cards)
+
+
+class FakeProseRepository:
+    """按来源引用返回冻结片段，不访问数据库。"""
+
+    def __init__(self):
+        self.calls = []
+
+    def read_source_refs(self, *, work_id, as_of, source_refs):
+        self.calls.append((work_id, as_of, copy.deepcopy(source_refs)))
+        rows = []
+        for ref in source_refs:
+            rows.append(
+                {
+                    "sourceRef": copy.deepcopy(ref),
+                    "chapter": ref["chapter"],
+                    "text": "历史原文",
+                    "purpose": "card_source",
+                }
+            )
+        return rows
+
+
+class FakeConnection:
+    """记录事务声明，证明读取器先锁定只读可重复读快照。"""
+
+    def __init__(self):
+        self.statements = []
+
+    def execute(self, statement, params=None):
+        self.statements.append((" ".join(statement.split()), params))
+        return self
+
+
+class FakeQueryResult:
+    """模拟 psycopg 查询结果。"""
+
+    def __init__(self, rows):
+        self.rows = rows
+
+    def fetchall(self):
+        return self.rows
+
+
+class FakeSearchConnection:
+    """为 search_cards 提供字段策略与单条 Canonical 卡结果。"""
+
+    def __init__(self):
+        self.statements = []
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc, traceback):
+        return False
+
+    def execute(self, statement, params=None):
+        normalized = " ".join(statement.split())
+        self.statements.append((normalized, params))
+        if "muse_meta_schema" in normalized:
+            return FakeQueryResult([("character", {"秘密": False})])
+        return FakeQueryResult(
+            [
+                (
+                    "entity",
+                    11,
+                    {
+                        "型": "character",
+                        "名称": "甲",
+                        "一句话摘要": "历史角色",
+                        "字段": {
+                            "秘密": "不可见",
+                            "sourceRefs": [source_ref(3, 11)],
+                            "milestones": [{"chapter": 3, "fact": "历史事实"}],
+                        },
+                    },
+                    "active",
+                    0.9,
+                    2,
+                    {"sourceRefs": [source_ref(3, 11)]},
+                    "active",
+                    "active",
+                )
+            ]
+        )
+
+
+class RetrieveWriterSourcesTest(unittest.TestCase):
+    def setUp(self):
+        self.plan = build_retrieval_plan(
+            run_id="run-a",
+            work_id=8,
+            target_chapter=6,
+            as_of=5,
+            fine_outline={
+                "entities": [
+                    {"id": "character:甲", "type": "character", "name": "甲"},
+                    {"id": "item:剑", "type": "item", "name": "剑"},
+                ],
+                "relations": [],
+                "locations": [],
+                "powerSystems": [],
+                "hardConstraints": ["甲使用剑"],
+            },
+            card_index_version="cards-v1",
+            prose_index_version="prose-v1",
+            token_budget={"maxContextChars": 20000},
+        )
+
+    def test_cards_use_documented_stable_order_and_ties_repeat(self):
+        cards = [card("3", 0.8, offset=2), card("2", 0.8, offset=1), card("1", 0.9)]
+        expected = ["1", "2", "3"]
+        for _ in range(3):
+            self.assertEqual([item["cardId"] for item in stable_sort_cards(cards)], expected)
+
+    def test_plan_identity_excludes_run_id(self):
+        other = copy.deepcopy(self.plan)
+        other["runId"] = "run-b"
+        rebuilt = build_retrieval_plan(
+            run_id="run-b",
+            work_id=8,
+            target_chapter=6,
+            as_of=5,
+            fine_outline={
+                "entities": [
+                    {"id": "character:甲", "type": "character", "name": "甲"},
+                    {"id": "item:剑", "type": "item", "name": "剑"},
+                ],
+                "relations": [],
+                "locations": [],
+                "powerSystems": [],
+                "hardConstraints": ["甲使用剑"],
+            },
+            card_index_version="cards-v1",
+            prose_index_version="prose-v1",
+            token_budget={"maxContextChars": 20000},
+        )
+        self.assertEqual(self.plan["planId"], rebuilt["planId"])
+
+    def test_state_as_of_uses_only_milestones_at_or_before_freeze(self):
+        frozen = freeze_card(card("1", 1.0), as_of=5)
+        self.assertEqual([item["chapter"] for item in frozen["stateAsOf"]], [1, 3])
+        self.assertNotIn("未来终态", str(frozen))
+
+    def test_target_future_and_unprovable_sources_fail_closed(self):
+        for bad_ref in (
+            source_ref(6, 1),
+            source_ref(7, 1),
+            {"sourceId": "unknown", "sourceVersion": SOURCE_VERSION},
+        ):
+            unsafe = card("1", 1.0)
+            unsafe["sourceRefs"] = [bad_ref]
+            with self.subTest(bad_ref=bad_ref), self.assertRaises(RetrievalError):
+                retrieve_writer_sources(
+                    plan=self.plan,
+                    card_repository=FakeCardRepository([unsafe]),
+                    prose_repository=FakeProseRepository(),
+                )
+
+    def test_cards_expand_to_prose_and_authoritative_facts_keep_independent_refs(self):
+        prose = FakeProseRepository()
+        no_ref_card = card("2", 0.7)
+        no_ref_card["sourceRefs"] = []
+        result = retrieve_writer_sources(
+            plan=self.plan,
+            card_repository=FakeCardRepository([card("1", 0.9), no_ref_card]),
+            prose_repository=prose,
+            authoritative_facts=[
+                {
+                    "factId": "setting:1",
+                    "fact": "正式设定事实",
+                    "sourceType": "formal_setting",
+                    "sourceRef": {
+                        "sourceId": "setting:8",
+                        "sourceVersion": "setting-v2",
+                    },
+                    "riskLevel": "high",
+                },
+                {
+                    "factId": "outline:new:1",
+                    "fact": "细纲声明的新事实",
+                    "sourceType": "fine_outline_declared_new",
+                    "sourceRef": {
+                        "sourceId": "fine-outline:6",
+                        "sourceVersion": "outline-v3",
+                        "chapter": 6,
+                    },
+                    "riskLevel": "medium",
+                },
+            ],
+        )
+        self.assertEqual(len(prose.calls[0][2]), 1)
+        self.assertEqual(result["proseEvidence"][0]["text"], "历史原文")
+        self.assertEqual(
+            {item["sourceType"] for item in result["factEvidence"]},
+            {"historical_prose", "formal_setting", "fine_outline_declared_new"},
+        )
+        self.assertEqual(result["unverifiedIndexHints"][0]["cardId"], "2")
+
+    def test_snapshot_transaction_is_repeatable_read_and_read_only(self):
+        connection = FakeConnection()
+        begin_read_snapshot(connection)
+        self.assertEqual(
+            connection.statements[0][0],
+            "SET TRANSACTION ISOLATION LEVEL REPEATABLE READ READ ONLY",
+        )
+
+    def test_production_repository_only_accepts_active_canonical_binding(self):
+        calls = []
+
+        def search_function(**kwargs):
+            calls.append(kwargs)
+            return [card("1", 0.9)]
+
+        repository = ProductionCardIndexRepository(search_function=search_function)
+        result = repository.search(self.plan)
+        self.assertEqual(result[0]["cardId"], "1")
+        self.assertEqual(calls[0]["scope"], "work")
+
+        for field, value in (
+            ("sourceKind", "eval_draft"),
+            ("sourceStatus", "draft"),
+            ("bindingStatus", "inactive"),
+            ("productionRetrievalEligible", False),
+        ):
+            unsafe = card("1", 0.9)
+            unsafe[field] = value
+            repository = ProductionCardIndexRepository(search_function=lambda **_: [unsafe])
+            with self.subTest(field=field), self.assertRaises(RetrievalError):
+                repository.search(self.plan)
+
+    def test_search_cards_reuses_active_entity_and_binding_sql(self):
+        connection = FakeSearchConnection()
+        result = search_cards(
+            "甲的历史状态",
+            scope="work",
+            work_id=8,
+            ttype="character",
+            purpose="generation",
+            top=5,
+            connection_factory=lambda _: connection,
+            embedder=lambda _: [0.1, 0.2],
+        )
+        sql = connection.statements[1][0]
+        self.assertIn("en.status='active'", sql)
+        self.assertIn("b.binding_status='active'", sql)
+        self.assertIn("en.source_action_policy='allowed'", sql)
+        self.assertEqual(result[0]["sourceKind"], "canonical_entity")
+        self.assertTrue(result[0]["productionRetrievalEligible"])
+        self.assertNotIn("秘密", result[0]["visibleFields"])
+
+    def test_replay_repository_is_preregistered_upgrade_book_and_never_production_eligible(self):
+        replay_card = card("11", 0.9)
+        replay_card.update(
+            {
+                "sourceKind": "eval_draft",
+                "evaluationStatus": "eval_draft",
+                "sourceType": "upgrade_book",
+                "productionRetrievalEligible": False,
+            }
+        )
+        replay_config = {
+            "targetChapter": 6,
+            "snapshot": {"asOfChapter": 5, "data": {"chapters": [{"chapter": 5, "text": "安全历史"}]}},
+            "sources": [
+                {"sourceId": "chapter:5", "sourceVersion": "chapter-v1", "chapter": 5}
+            ],
+            "authorization": {
+                "sourceStatus": "active",
+                "copyrightStatus": "licensed",
+                "sourceVersion": SOURCE_VERSION,
+                "allowedPurpose": ["offline_evaluation"],
+                "authorizationSnapshot": {
+                    "id": "auth-1",
+                    "version": "v1",
+                    "immutable": True,
+                    "sourceVersion": SOURCE_VERSION,
+                    "sourceStatus": "active",
+                    "allowedPurpose": ["offline_evaluation"],
+                    "checkedAt": "2026-07-20T00:00:00Z",
+                    "revalidationAt": "2026-07-21T00:00:00Z",
+                },
+            },
+            "leakageAudit": {
+                "targetFacts": {"targetChapter": 6, "forbiddenFacts": []}
+            },
+        }
+        repository = ReplayCardIndexRepository.from_replay_config(
+            replay_config,
+            cards=[replay_card],
+            preregistered_card_ids=["11"],
+        )
+        self.assertFalse(repository.search(self.plan)[0]["productionRetrievalEligible"])
+
+        denied = copy.deepcopy(replay_config)
+        denied["authorization"] = {}
+        with self.assertRaises(RetrievalError):
+            ReplayCardIndexRepository.from_replay_config(
+                denied,
+                cards=[replay_card],
+                preregistered_card_ids=["11"],
+            )
+
+        wrong = copy.deepcopy(replay_card)
+        wrong["sourceType"] = "extract_chapter"
+        with self.assertRaises(RetrievalError):
+            ReplayCardIndexRepository.from_replay_config(
+                replay_config,
+                cards=[wrong],
+                preregistered_card_ids=["11"],
+            )
+        with self.assertRaises(RetrievalError):
+            ReplayCardIndexRepository.from_replay_config(
+                replay_config,
+                cards=[replay_card],
+                preregistered_card_ids=["12"],
+            )
+
+
+if __name__ == "__main__":
+    unittest.main()

.claude/skills/read-context/scripts/test_writer_contract.py

@@ -0,0 +1,239 @@
+#!/usr/bin/env python3
+"""正文写手输入输出合同的确定性与失败关闭测试。"""
+
+from __future__ import annotations
+
+import copy
+import hashlib
+import pathlib
+import sys
+import unittest
+
+sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
+
+from writer_contract import (  # noqa: E402
+    ContractError,
+    calculate_target_chars,
+    canonical_json,
+    han_count,
+    normalize_text,
+    retrieval_identity,
+    validate_writer_context,
+    validate_writer_output,
+)
+
+
+def valid_context() -> dict:
+    """构造覆盖全部必填字段的最小合法上下文。"""
+
+    context = {
+        "schemaVersion": "writer-context-v1",
+        "runId": "writer-run-001",
+        "attempt": 1,
+        "mode": "production",
+        "purpose": "production",
+        "qualityPolicyVersion": "writer-production-v1",
+        "workId": 8,
+        "targetChapter": 489,
+        "asOf": 488,
+        "contextSnapshot": {
+            "manifestId": "sha256:" + "1" * 64,
+            "contextSha256": "sha256:" + "2" * 64,
+            "generatedAt": "2026-07-20T00:00:00Z",
+        },
+        "sourceVersion": "raw-file-v1:sha256:" + "3" * 64,
+        "authorizationSnapshot": {
+            "snapshotId": "auth-001",
+            "allowedPurpose": "generation",
+            "verifiedAt": "2026-07-20T00:00:00Z",
+        },
+        "sourceStatus": "active",
+        "retrievalPlan": {
+            "planVersion": "writer-retrieval-plan-v1",
+            "planId": "sha256:" + "4" * 64,
+            "runId": "writer-run-001",
+            "asOf": 488,
+            "queries": [],
+            "cardIndexVersion": "card-v1",
+            "proseIndexVersion": "prose-v1",
+            "filters": {
+                "workId": 8,
+                "asOfChapter": 488,
+                "sourceStatus": "active",
+                "authorizationRequired": True,
+            },
+            "tieBreak": "score DESC, sourceVersion ASC, sourceId ASC, sourceOffset ASC",
+            "tokenBudget": {"maxContextChars": 20000},
+        },
+        "retrievalManifest": {
+            "manifestVersion": "writer-retrieval-manifest-v1",
+            "manifestId": "sha256:" + "1" * 64,
+            "planId": "sha256:" + "4" * 64,
+            "sources": [],
+            "omittedSources": [],
+        },
+        "fineOutline": {
+            "sourceRef": {
+                "sourceId": "fine-outline:489",
+                "sourceVersion": "fine-outline-v3",
+                "chapter": 489,
+            },
+            "hardConstraints": ["必须完成围攻突围"],
+            "adjustableBeats": [],
+            "declaredNewFacts": [],
+        },
+        "narrativeState": {
+            "time": "围攻当日",
+            "location": "圣蒂曼",
+            "characterPositions": {},
+            "immediateSituation": "战斗持续",
+        },
+        "factEvidence": [],
+        "proseEvidence": [],
+        "patternReferences": [],
+        "evidenceCoverage": [],
+        "outputContract": {
+            "targetChars": 4000,
+            "minChars": 3600,
+            "maxChars": 4400,
+            "frontmatterRequired": False,
+            "newSettingDeclarationRequired": True,
+        },
+        "tokenBudget": {"maxContextChars": 20000, "usedContextChars": 0},
+        "omittedSources": [],
+        "acceptanceEligible": True,
+    }
+    plan_payload = {key: value for key, value in context["retrievalPlan"].items() if key != "planId"}
+    context["retrievalPlan"]["planId"] = retrieval_identity(plan_payload)
+    context["retrievalManifest"]["planId"] = context["retrievalPlan"]["planId"]
+    manifest_payload = {key: value for key, value in context["retrievalManifest"].items() if key != "manifestId"}
+    context["retrievalManifest"]["manifestId"] = retrieval_identity(manifest_payload)
+    context["contextSnapshot"]["manifestId"] = context["retrievalManifest"]["manifestId"]
+    context["contextSnapshot"]["contextSha256"] = retrieval_identity(context)
+    return context
+
+
+def valid_output() -> dict:
+    """构造覆盖全部必填字段的最小合法写手输出。"""
+
+    body = normalize_text("第一段正文。")
+    digest = "sha256:" + hashlib.sha256(body.encode("utf-8")).hexdigest()
+    return {
+        "schemaVersion": "writer-output-v1",
+        "runId": "writer-run-001",
+        "attempt": 1,
+        "mode": "production",
+        "qualityPolicyVersion": "writer-production-v1",
+        "contextSnapshotId": "sha256:" + "1" * 64,
+        "contextSnapshotSha256": "sha256:" + "2" * 64,
+        "candidateVersion": 1,
+        "candidateSha256": digest,
+        "acceptanceEligible": True,
+        "candidateBody": body,
+        "claimLedger": [],
+        "evidenceRequests": [],
+        "newSettingDeclarations": [],
+        "selfCheck": {"hardConstraintsCovered": True, "notes": []},
+    }
+
+
+class WriterContractTest(unittest.TestCase):
+    def test_run_id_does_not_change_retrieval_identity(self):
+        first = {"runId": "run-a", "query": "咖啡\u0301", "nested": {"value": 1}}
+        second = {"runId": "run-b", "query": "咖啡\u0301", "nested": {"value": 1}}
+        self.assertEqual(retrieval_identity(first), retrieval_identity(second))
+
+    def test_text_is_nfc_and_lf_before_offsets_and_hash(self):
+        decomposed = "Cafe\u0301\r\n第二行\r第三行"
+        normalized = "Caf\u00e9\n第二行\n第三行"
+        self.assertEqual(normalize_text(decomposed), normalized)
+        self.assertEqual(canonical_json({"text": decomposed}), canonical_json({"text": normalized}))
+
+    def test_han_count_does_not_count_markdown_or_non_han_text(self):
+        self.assertEqual(han_count("# **正文** 123 ABC，扩展𠀀"), 5)
+
+    def test_target_chars_use_half_up_and_hard_bounds(self):
+        self.assertEqual(
+            calculate_target_chars(
+                recent_chapter_han_counts=[2501, 2502, 2503, 2504],
+                hard_event_count=3,
+                foreshadowing_action_count=0,
+                required_scene_count=0,
+            ),
+            2100,
+        )
+        self.assertEqual(calculate_target_chars(explicit_target_chars=1500), 2000)
+        self.assertEqual(calculate_target_chars(explicit_target_chars=11000), 10000)
+        self.assertEqual(
+            calculate_target_chars(
+                recent_chapter_han_counts=[3000, 3000, 3000],
+                hard_event_count=0,
+                foreshadowing_action_count=0,
+                required_scene_count=0,
+            ),
+            2600,
+        )
+
+    def test_evaluation_and_diagnostic_contexts_are_never_acceptable(self):
+        for purpose in ("evaluation", "diagnostic"):
+            context = valid_context()
+            context["mode"] = "diagnostic_only"
+            context["purpose"] = purpose
+            context["qualityPolicyVersion"] = "writer-eval-v1"
+            context["acceptanceEligible"] = True
+            with self.assertRaises(ContractError):
+                validate_writer_context(context)
+
+            context["acceptanceEligible"] = False
+            context["contextSnapshot"]["contextSha256"] = retrieval_identity(context)
+            validate_writer_context(context)
+
+    def test_unknown_missing_and_wrong_version_fail_closed(self):
+        for mutation in ("unknown", "missing", "version"):
+            context = copy.deepcopy(valid_context())
+            if mutation == "unknown":
+                context["unexpected"] = True
+            elif mutation == "missing":
+                del context["fineOutline"]
+            else:
+                context["schemaVersion"] = "writer-context-v2"
+            with self.subTest(mutation=mutation), self.assertRaises(ContractError):
+                validate_writer_context(context)
+
+        output = valid_output()
+        output["selfCheck"]["unknown"] = True
+        with self.assertRaises(ContractError):
+            validate_writer_output(output)
+
+    def test_candidate_hash_and_diagnostic_acceptance_are_checked(self):
+        output = valid_output()
+        validate_writer_output(output)
+
+        output["candidateBody"] = "被修改的正文"
+        with self.assertRaises(ContractError):
+            validate_writer_output(output)
+
+        diagnostic = valid_output()
+        diagnostic["mode"] = "diagnostic_only"
+        diagnostic["qualityPolicyVersion"] = "writer-eval-v1"
+        diagnostic["acceptanceEligible"] = True
+        with self.assertRaises(ContractError):
+            validate_writer_output(diagnostic)
+
+    def test_plan_manifest_and_context_identity_tampering_fails_closed(self):
+        for path in ("plan", "manifest", "context"):
+            context = valid_context()
+            if path == "plan":
+                context["retrievalPlan"]["cardIndexVersion"] = "tampered"
+            elif path == "manifest":
+                context["retrievalManifest"]["sources"].append(
+                    {"sourceId": "setting:1", "sourceVersion": "setting-v1"}
+                )
+            else:
+                context["fineOutline"]["hardConstraints"].append("被篡改的约束")
+            with self.subTest(path=path), self.assertRaises(ContractError):
+                validate_writer_context(context)
+
+
+if __name__ == "__main__":
+    unittest.main()

.claude/skills/read-context/scripts/writer_contract.py

@@ -0,0 +1,549 @@
+#!/usr/bin/env python3
+"""正文写手上下文与输出的严格合同。
+
+本模块只处理纯数据，不读取文件、数据库或网络。所有进入写手的文本先做
+Unicode NFC 与换行归一化，所有身份哈希都来自同一份规范 JSON。
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import math
+import re
+import unicodedata
+from decimal import Decimal, ROUND_HALF_UP
+from typing import Any, Mapping, Sequence
+
+
+CONTEXT_VERSION = "writer-context-v1"
+OUTPUT_VERSION = "writer-output-v1"
+PLAN_VERSION = "writer-retrieval-plan-v1"
+MANIFEST_VERSION = "writer-retrieval-manifest-v1"
+TIE_BREAK = "score DESC, sourceVersion ASC, sourceId ASC, sourceOffset ASC"
+
+_HASH_RE = re.compile(r"^sha256:[0-9a-f]{64}$")
+_RUN_ID_RE = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._:-]{0,127}$")
+_VOLATILE_IDENTITY_FIELDS = frozenset(
+    {"runId", "generatedAt", "timestamp", "executionNode", "contextSha256"}
+)
+
+# Unicode Script=Han 覆盖的标准区段。兼容表意文字也按“汉字”计数，
+# 但标点、Markdown、拉丁字母和数字不会落入这些区段。
+_HAN_RANGES = (
+    (0x3400, 0x4DBF),
+    (0x4E00, 0x9FFF),
+    (0xF900, 0xFAFF),
+    (0x20000, 0x2EBEF),
+    (0x2F800, 0x2FA1F),
+    (0x30000, 0x323AF),
+)
+
+
+class ContractError(ValueError):
+    """输入不符合严格合同时抛出，调用方必须失败关闭。"""
+
+
+def normalize_text(value: str) -> str:
+    """把文本统一为 NFC 与 LF，供哈希和 Unicode 偏移共同使用。"""
+
+    if not isinstance(value, str):
+        raise ContractError("待归一化文本必须是字符串")
+    return unicodedata.normalize("NFC", value.replace("\r\n", "\n").replace("\r", "\n"))
+
+
+def _normalize_json(value: Any) -> Any:
+    """递归归一化 JSON 值，并拒绝 JSON 之外或不可复现的数值。"""
+
+    if isinstance(value, str):
+        return normalize_text(value)
+    if value is None or isinstance(value, (bool, int)):
+        return value
+    if isinstance(value, float):
+        if not math.isfinite(value):
+            raise ContractError("规范 JSON 不允许 NaN 或 Infinity")
+        return value
+    if isinstance(value, list):
+        return [_normalize_json(item) for item in value]
+    if isinstance(value, tuple):
+        return [_normalize_json(item) for item in value]
+    if isinstance(value, Mapping):
+        result: dict[str, Any] = {}
+        for raw_key, item in value.items():
+            if not isinstance(raw_key, str):
+                raise ContractError("规范 JSON 的对象键必须是字符串")
+            key = normalize_text(raw_key)
+            if key in result:
+                raise ContractError(f"对象键在 NFC 归一化后冲突: {key}")
+            result[key] = _normalize_json(item)
+        return result
+    raise ContractError(f"值不是 JSON 类型: {type(value).__name__}")
+
+
+def canonical_json(value: Any) -> str:
+    """输出 UTF-8 语义、键排序、无额外空白的规范 JSON 文本。"""
+
+    return json.dumps(
+        _normalize_json(value),
+        ensure_ascii=False,
+        sort_keys=True,
+        separators=(",", ":"),
+        allow_nan=False,
+    )
+
+
+def _without_volatile_fields(value: Any) -> Any:
+    """递归移除运行元数据，防止同一检索输入得到不同身份。"""
+
+    if isinstance(value, list):
+        return [_without_volatile_fields(item) for item in value]
+    if isinstance(value, Mapping):
+        return {
+            key: _without_volatile_fields(item)
+            for key, item in value.items()
+            if key not in _VOLATILE_IDENTITY_FIELDS
+        }
+    return value
+
+
+def retrieval_identity(value: Any) -> str:
+    """计算计划、manifest 或上下文的稳定 SHA-256 身份。"""
+
+    encoded = canonical_json(_without_volatile_fields(value)).encode("utf-8")
+    return "sha256:" + hashlib.sha256(encoded).hexdigest()
+
+
+def han_count(value: str) -> int:
+    """统计 Unicode Han code point，不把标点或 Markdown 算入正文长度。"""
+
+    text = normalize_text(value)
+    return sum(any(start <= ord(char) <= end for start, end in _HAN_RANGES) for char in text)
+
+
+def _round_half_up(value: Decimal) -> int:
+    """以十进制 ROUND_HALF_UP 规则取整，避免 Python 银行家舍入。"""
+
+    return int(value.quantize(Decimal("1"), rounding=ROUND_HALF_UP))
+
+
+def _round_to_100(value: Decimal) -> int:
+    """以百字为单位执行十进制半入取整。"""
+
+    return int((value / Decimal(100)).quantize(Decimal("1"), rounding=ROUND_HALF_UP) * 100)
+
+
+def calculate_target_chars(
+    *,
+    explicit_target_chars: int | None = None,
+    recent_chapter_han_counts: Sequence[int] = (),
+    default_target_chars: int = 4000,
+    hard_event_count: int = 3,
+    foreshadowing_action_count: int = 0,
+    required_scene_count: int = 0,
+    min_chars: int = 2000,
+    max_chars: int = 10000,
+) -> int:
+    """按冻结历史中位数与细纲密度计算确定性目标汉字数。"""
+
+    counts = {
+        "default_target_chars": default_target_chars,
+        "hard_event_count": hard_event_count,
+        "foreshadowing_action_count": foreshadowing_action_count,
+        "required_scene_count": required_scene_count,
+        "min_chars": min_chars,
+        "max_chars": max_chars,
+    }
+    if any(isinstance(value, bool) or not isinstance(value, int) for value in counts.values()):
+        raise ContractError("篇幅参数必须是整数")
+    if default_target_chars <= 0 or min_chars <= 0 or max_chars < min_chars or any(value < 0 for key, value in counts.items() if "count" in key):
+        raise ContractError("篇幅边界或细纲计数非法")
+    if explicit_target_chars is not None:
+        if isinstance(explicit_target_chars, bool) or not isinstance(explicit_target_chars, int):
+            raise ContractError("显式 targetChars 必须是整数")
+        return min(max(explicit_target_chars, min_chars), max_chars)
+
+    valid_counts = list(recent_chapter_han_counts)[-20:]
+    if any(isinstance(value, bool) or not isinstance(value, int) or value < 500 for value in valid_counts):
+        raise ContractError("历史章汉字数必须是大于等于 500 的整数")
+    if len(valid_counts) >= 3:
+        ordered_counts = sorted(valid_counts)
+        midpoint = len(ordered_counts) // 2
+        if len(ordered_counts) % 2:
+            baseline = Decimal(ordered_counts[midpoint])
+        else:
+            baseline = (Decimal(ordered_counts[midpoint - 1]) + Decimal(ordered_counts[midpoint])) / 2
+    else:
+        baseline = Decimal(default_target_chars)
+    density = (
+        Decimal(hard_event_count)
+        + Decimal("0.5") * foreshadowing_action_count
+        + Decimal("0.5") * required_scene_count
+    )
+    factor = min(Decimal("1.15"), max(Decimal("0.85"), Decimal("0.85") + Decimal("0.05") * (density - 3)))
+    target = _round_to_100(Decimal(_round_half_up(baseline)) * factor)
+    return min(max(target, min_chars), max_chars)
+
+
+def _object(
+    value: Any,
+    path: str,
+    required: set[str] | frozenset[str],
+    optional: set[str] | frozenset[str] = frozenset(),
+) -> Mapping[str, Any]:
+    """校验严格对象，任何缺字段或未知字段都立即失败。"""
+
+    if not isinstance(value, Mapping):
+        raise ContractError(f"{path} 必须是对象")
+    missing = sorted(required - set(value))
+    unknown = sorted(set(value) - required - optional)
+    if missing:
+        raise ContractError(f"{path} 缺少字段: {','.join(missing)}")
+    if unknown:
+        raise ContractError(f"{path} 包含未知字段: {','.join(unknown)}")
+    return value
+
+
+def _string(value: Any, path: str, *, nonempty: bool = True) -> str:
+    """校验字符串，并在需要时拒绝空值。"""
+
+    if not isinstance(value, str) or (nonempty and not value.strip()):
+        raise ContractError(f"{path} 必须是非空字符串")
+    if value != normalize_text(value):
+        raise ContractError(f"{path} 必须预先归一化为 Unicode NFC/LF")
+    return value
+
+
+def _integer(value: Any, path: str, *, minimum: int = 0) -> int:
+    """校验整数，显式排除 bool 这一 Python int 子类。"""
+
+    if isinstance(value, bool) or not isinstance(value, int) or value < minimum:
+        raise ContractError(f"{path} 必须是大于等于 {minimum} 的整数")
+    return value
+
+
+def _boolean(value: Any, path: str) -> bool:
+    """校验严格布尔值。"""
+
+    if not isinstance(value, bool):
+        raise ContractError(f"{path} 必须是布尔值")
+    return value
+
+
+def _array(value: Any, path: str) -> list[Any]:
+    """校验数组并返回原值，拒绝元组等隐式转换。"""
+
+    if not isinstance(value, list):
+        raise ContractError(f"{path} 必须是数组")
+    return value
+
+
+def _hash(value: Any, path: str) -> str:
+    """校验带算法前缀的 SHA-256。"""
+
+    text = _string(value, path)
+    if not _HASH_RE.fullmatch(text):
+        raise ContractError(f"{path} 必须是 sha256: 加 64 位小写十六进制")
+    return text
+
+
+def _source_ref(value: Any, path: str) -> None:
+    """校验不可变来源引用；历史原文可额外携带块和字符区间。"""
+
+    ref = _object(
+        value,
+        path,
+        frozenset({"sourceId", "sourceVersion"}),
+        frozenset({"chapter", "blockId", "startCodePoint", "endCodePoint", "contentSha256", "sourceType"}),
+    )
+    _string(ref["sourceId"], f"{path}.sourceId")
+    _string(ref["sourceVersion"], f"{path}.sourceVersion")
+    for field in ("chapter", "blockId", "startCodePoint", "endCodePoint"):
+        if field in ref:
+            _integer(ref[field], f"{path}.{field}", minimum=0 if "CodePoint" in field else 1)
+    if "startCodePoint" in ref and "endCodePoint" in ref and ref["endCodePoint"] <= ref["startCodePoint"]:
+        raise ContractError(f"{path} 字符区间必须是非空左闭右开区间")
+    if "contentSha256" in ref:
+        _hash(ref["contentSha256"], f"{path}.contentSha256")
+    if "sourceType" in ref:
+        _string(ref["sourceType"], f"{path}.sourceType")
+
+
+def _validate_plan(value: Any, path: str) -> None:
+    """校验固定检索计划，不允许写手临场扩张查询。"""
+
+    plan = _object(
+        value,
+        path,
+        frozenset(
+            {"planVersion", "planId", "runId", "asOf", "queries", "cardIndexVersion", "proseIndexVersion", "filters", "tieBreak", "tokenBudget"}
+        ),
+    )
+    if plan["planVersion"] != PLAN_VERSION:
+        raise ContractError(f"{path}.planVersion 版本不支持")
+    _hash(plan["planId"], f"{path}.planId")
+    if not _RUN_ID_RE.fullmatch(_string(plan["runId"], f"{path}.runId")):
+        raise ContractError(f"{path}.runId 格式非法")
+    _integer(plan["asOf"], f"{path}.asOf", minimum=1)
+    for index, query in enumerate(_array(plan["queries"], f"{path}.queries")):
+        item = _object(query, f"{path}.queries[{index}]", frozenset({"queryId", "text", "entityTypes", "purpose", "topK"}))
+        _string(item["queryId"], f"{path}.queries[{index}].queryId")
+        _string(item["text"], f"{path}.queries[{index}].text")
+        if any(not isinstance(kind, str) or not kind for kind in _array(item["entityTypes"], f"{path}.queries[{index}].entityTypes")):
+            raise ContractError(f"{path}.queries[{index}].entityTypes 必须是非空字符串数组")
+        _string(item["purpose"], f"{path}.queries[{index}].purpose")
+        _integer(item["topK"], f"{path}.queries[{index}].topK", minimum=1)
+    _string(plan["cardIndexVersion"], f"{path}.cardIndexVersion")
+    _string(plan["proseIndexVersion"], f"{path}.proseIndexVersion")
+    filters = _object(plan["filters"], f"{path}.filters", frozenset({"workId", "asOfChapter", "sourceStatus", "authorizationRequired"}))
+    _integer(filters["workId"], f"{path}.filters.workId", minimum=1)
+    _integer(filters["asOfChapter"], f"{path}.filters.asOfChapter", minimum=1)
+    _string(filters["sourceStatus"], f"{path}.filters.sourceStatus")
+    _boolean(filters["authorizationRequired"], f"{path}.filters.authorizationRequired")
+    if plan["tieBreak"] != TIE_BREAK:
+        raise ContractError(f"{path}.tieBreak 不符合稳定排序合同")
+    budget = _object(plan["tokenBudget"], f"{path}.tokenBudget", frozenset({"maxContextChars"}), frozenset({"cardChars", "recentProseChars", "historicalProseChars", "patternChars"}))
+    for key, item in budget.items():
+        _integer(item, f"{path}.tokenBudget.{key}", minimum=0)
+    identity_payload = {key: item for key, item in plan.items() if key != "planId"}
+    if plan["planId"] != retrieval_identity(identity_payload):
+        raise ContractError(f"{path}.planId 与计划内容不匹配")
+
+
+def _validate_manifest(value: Any, path: str) -> None:
+    """校验检索清单的来源与裁剪回显。"""
+
+    manifest = _object(value, path, frozenset({"manifestVersion", "manifestId", "planId", "sources", "omittedSources"}))
+    if manifest["manifestVersion"] != MANIFEST_VERSION:
+        raise ContractError(f"{path}.manifestVersion 版本不支持")
+    _hash(manifest["manifestId"], f"{path}.manifestId")
+    _hash(manifest["planId"], f"{path}.planId")
+    for index, source in enumerate(_array(manifest["sources"], f"{path}.sources")):
+        _source_ref(source, f"{path}.sources[{index}]")
+    for index, omitted in enumerate(_array(manifest["omittedSources"], f"{path}.omittedSources")):
+        item = _object(omitted, f"{path}.omittedSources[{index}]", frozenset({"sourceId", "reason"}))
+        _string(item["sourceId"], f"{path}.omittedSources[{index}].sourceId")
+        _string(item["reason"], f"{path}.omittedSources[{index}].reason")
+    identity_payload = {key: item for key, item in manifest.items() if key != "manifestId"}
+    if manifest["manifestId"] != retrieval_identity(identity_payload):
+        raise ContractError(f"{path}.manifestId 与来源清单不匹配")
+
+
+def validate_writer_context(value: Any) -> dict[str, Any]:
+    """校验 WriterContext v1；成功时返回可安全复制的规范 JSON 对象。"""
+
+    required = frozenset(
+        {
+            "schemaVersion", "runId", "attempt", "mode", "purpose", "qualityPolicyVersion",
+            "workId", "targetChapter", "asOf", "contextSnapshot", "sourceVersion",
+            "authorizationSnapshot", "sourceStatus", "retrievalPlan", "retrievalManifest",
+            "fineOutline", "narrativeState", "factEvidence", "proseEvidence",
+            "patternReferences", "evidenceCoverage", "outputContract", "tokenBudget",
+            "omittedSources", "acceptanceEligible",
+        }
+    )
+    context = _object(value, "$", required)
+    if context["schemaVersion"] != CONTEXT_VERSION:
+        raise ContractError("$.schemaVersion 版本不支持")
+    run_id = _string(context["runId"], "$.runId")
+    if not _RUN_ID_RE.fullmatch(run_id):
+        raise ContractError("$.runId 格式非法")
+    _integer(context["attempt"], "$.attempt", minimum=1)
+    if context["mode"] not in {"production", "diagnostic_only"}:
+        raise ContractError("$.mode 枚举非法")
+    if context["purpose"] not in {"production", "evaluation", "diagnostic"}:
+        raise ContractError("$.purpose 枚举非法")
+    _string(context["qualityPolicyVersion"], "$.qualityPolicyVersion")
+    _integer(context["workId"], "$.workId", minimum=1)
+    target = _integer(context["targetChapter"], "$.targetChapter", minimum=1)
+    as_of = _integer(context["asOf"], "$.asOf", minimum=1)
+    if as_of >= target:
+        raise ContractError("$.asOf 必须早于 targetChapter")
+
+    snapshot = _object(context["contextSnapshot"], "$.contextSnapshot", frozenset({"manifestId", "contextSha256", "generatedAt"}))
+    _hash(snapshot["manifestId"], "$.contextSnapshot.manifestId")
+    _hash(snapshot["contextSha256"], "$.contextSnapshot.contextSha256")
+    _string(snapshot["generatedAt"], "$.contextSnapshot.generatedAt")
+    _string(context["sourceVersion"], "$.sourceVersion")
+    authorization = _object(
+        context["authorizationSnapshot"],
+        "$.authorizationSnapshot",
+        frozenset({"snapshotId", "allowedPurpose", "verifiedAt"}),
+        frozenset({"expiresAt", "sourceVersion"}),
+    )
+    for key, item in authorization.items():
+        _string(item, f"$.authorizationSnapshot.{key}")
+    if context["sourceStatus"] not in {"active", "authorized", "frozen_authorized"}:
+        raise ContractError("$.sourceStatus 不允许生成")
+
+    _validate_plan(context["retrievalPlan"], "$.retrievalPlan")
+    _validate_manifest(context["retrievalManifest"], "$.retrievalManifest")
+    if context["retrievalPlan"]["runId"] != run_id or context["retrievalPlan"]["asOf"] != as_of:
+        raise ContractError("检索计划与上下文的 runId/asOf 不一致")
+    if context["retrievalManifest"]["planId"] != context["retrievalPlan"]["planId"]:
+        raise ContractError("检索 manifest 未绑定当前计划")
+    if context["contextSnapshot"]["manifestId"] != context["retrievalManifest"]["manifestId"]:
+        raise ContractError("上下文快照未绑定当前 manifest")
+
+    outline = _object(context["fineOutline"], "$.fineOutline", frozenset({"sourceRef", "hardConstraints", "adjustableBeats", "declaredNewFacts"}))
+    _source_ref(outline["sourceRef"], "$.fineOutline.sourceRef")
+    for field in ("hardConstraints", "adjustableBeats"):
+        if any(not isinstance(item, str) or not item for item in _array(outline[field], f"$.fineOutline.{field}")):
+            raise ContractError(f"$.fineOutline.{field} 必须是非空字符串数组")
+    for index, fact in enumerate(_array(outline["declaredNewFacts"], "$.fineOutline.declaredNewFacts")):
+        item = _object(fact, f"$.fineOutline.declaredNewFacts[{index}]", frozenset({"factId", "text", "sourceRef"}))
+        _string(item["factId"], f"$.fineOutline.declaredNewFacts[{index}].factId")
+        _string(item["text"], f"$.fineOutline.declaredNewFacts[{index}].text")
+        _source_ref(item["sourceRef"], f"$.fineOutline.declaredNewFacts[{index}].sourceRef")
+
+    state = _object(context["narrativeState"], "$.narrativeState", frozenset({"time", "location", "characterPositions", "immediateSituation"}))
+    for field in ("time", "location", "immediateSituation"):
+        _string(state[field], f"$.narrativeState.{field}", nonempty=False)
+    positions = _object(state["characterPositions"], "$.narrativeState.characterPositions", frozenset(state["characterPositions"].keys()) if isinstance(state["characterPositions"], Mapping) else frozenset())
+    for key, item in positions.items():
+        _string(key, "$.narrativeState.characterPositions.<key>")
+        _string(item, f"$.narrativeState.characterPositions.{key}")
+
+    for index, evidence in enumerate(_array(context["factEvidence"], "$.factEvidence")):
+        item = _object(evidence, f"$.factEvidence[{index}]", frozenset({"evidenceId", "fact", "sourceType", "sourceRef", "contentSha256", "riskLevel"}))
+        for field in ("evidenceId", "fact", "sourceType"):
+            _string(item[field], f"$.factEvidence[{index}].{field}")
+        if item["sourceType"] not in {"historical_prose", "formal_setting", "canonical_state", "fine_outline_declared_new"}:
+            raise ContractError(f"$.factEvidence[{index}].sourceType 枚举非法")
+        _source_ref(item["sourceRef"], f"$.factEvidence[{index}].sourceRef")
+        if item["sourceType"] == "historical_prose":
+            required_location = {"chapter", "blockId", "startCodePoint", "endCodePoint"}
+            if not required_location.issubset(item["sourceRef"]):
+                raise ContractError(f"$.factEvidence[{index}] 历史事实必须回到章、块和字符区间")
+        _hash(item["contentSha256"], f"$.factEvidence[{index}].contentSha256")
+        if item["riskLevel"] not in {"low", "medium", "high"}:
+            raise ContractError(f"$.factEvidence[{index}].riskLevel 枚举非法")
+
+    for index, evidence in enumerate(_array(context["proseEvidence"], "$.proseEvidence")):
+        item = _object(evidence, f"$.proseEvidence[{index}]", frozenset({"evidenceId", "chapter", "sourceRef", "contentSha256", "purpose", "text", "isRecentBaseline"}))
+        _string(item["evidenceId"], f"$.proseEvidence[{index}].evidenceId")
+        chapter = _integer(item["chapter"], f"$.proseEvidence[{index}].chapter", minimum=1)
+        if chapter > as_of:
+            raise ContractError(f"$.proseEvidence[{index}] 超出冻结线")
+        _source_ref(item["sourceRef"], f"$.proseEvidence[{index}].sourceRef")
+        if not {"chapter", "blockId", "startCodePoint", "endCodePoint"}.issubset(item["sourceRef"]):
+            raise ContractError(f"$.proseEvidence[{index}] 必须带章、块和字符区间")
+        _hash(item["contentSha256"], f"$.proseEvidence[{index}].contentSha256")
+        _string(item["purpose"], f"$.proseEvidence[{index}].purpose")
+        text = _string(item["text"], f"$.proseEvidence[{index}].text")
+        _boolean(item["isRecentBaseline"], f"$.proseEvidence[{index}].isRecentBaseline")
+        expected = "sha256:" + hashlib.sha256(text.encode("utf-8")).hexdigest()
+        if item["contentSha256"] != expected:
+            raise ContractError(f"$.proseEvidence[{index}] 文本哈希不匹配")
+
+    for index, reference in enumerate(_array(context["patternReferences"], "$.patternReferences")):
+        _source_ref(reference, f"$.patternReferences[{index}]")
+    for index, coverage in enumerate(_array(context["evidenceCoverage"], "$.evidenceCoverage")):
+        item = _object(coverage, f"$.evidenceCoverage[{index}]", frozenset({"elementId", "elementType", "name", "status", "factEvidenceIds", "proseEvidenceIds", "gapReason"}))
+        for field in ("elementId", "elementType", "name", "gapReason"):
+            _string(item[field], f"$.evidenceCoverage[{index}].{field}", nonempty=field != "gapReason")
+        if item["status"] not in {"supported", "declared_new", "card_gap", "style_gap", "unsupported", "conflict"}:
+            raise ContractError(f"$.evidenceCoverage[{index}].status 枚举非法")
+        for field in ("factEvidenceIds", "proseEvidenceIds"):
+            if any(not isinstance(item_id, str) or not item_id for item_id in _array(item[field], f"$.evidenceCoverage[{index}].{field}")):
+                raise ContractError(f"$.evidenceCoverage[{index}].{field} 必须是字符串数组")
+
+    output = _object(context["outputContract"], "$.outputContract", frozenset({"targetChars", "minChars", "maxChars", "frontmatterRequired", "newSettingDeclarationRequired"}))
+    for field in ("targetChars", "minChars", "maxChars"):
+        _integer(output[field], f"$.outputContract.{field}", minimum=1)
+    if not output["minChars"] <= output["targetChars"] <= output["maxChars"]:
+        raise ContractError("$.outputContract 篇幅范围不包含目标值")
+    _boolean(output["frontmatterRequired"], "$.outputContract.frontmatterRequired")
+    _boolean(output["newSettingDeclarationRequired"], "$.outputContract.newSettingDeclarationRequired")
+    budget = _object(context["tokenBudget"], "$.tokenBudget", frozenset({"maxContextChars", "usedContextChars"}))
+    for field in budget:
+        _integer(budget[field], f"$.tokenBudget.{field}", minimum=0)
+    if budget["usedContextChars"] > budget["maxContextChars"]:
+        raise ContractError("$.tokenBudget 已超预算")
+    for index, omitted in enumerate(_array(context["omittedSources"], "$.omittedSources")):
+        item = _object(omitted, f"$.omittedSources[{index}]", frozenset({"sourceId", "reason"}))
+        _string(item["sourceId"], f"$.omittedSources[{index}].sourceId")
+        _string(item["reason"], f"$.omittedSources[{index}].reason")
+
+    eligible = _boolean(context["acceptanceEligible"], "$.acceptanceEligible")
+    if context["purpose"] in {"evaluation", "diagnostic"} or context["mode"] == "diagnostic_only":
+        if eligible:
+            raise ContractError("评测或诊断上下文必须 acceptanceEligible=false")
+    elif context["qualityPolicyVersion"] != "writer-production-v1":
+        raise ContractError("生产上下文必须绑定 writer-production-v1")
+    if context["contextSnapshot"]["contextSha256"] != retrieval_identity(context):
+        raise ContractError("$.contextSnapshot.contextSha256 与上下文内容不匹配")
+    return json.loads(canonical_json(context))
+
+
+def validate_writer_output(value: Any) -> dict[str, Any]:
+    """校验 WriterOutput v1、正文哈希和全部来源引用。"""
+
+    output = _object(
+        value,
+        "$",
+        frozenset(
+            {"schemaVersion", "runId", "attempt", "mode", "qualityPolicyVersion", "contextSnapshotId", "contextSnapshotSha256", "candidateVersion", "candidateSha256", "acceptanceEligible", "candidateBody", "claimLedger", "evidenceRequests", "newSettingDeclarations", "selfCheck"}
+        ),
+    )
+    if output["schemaVersion"] != OUTPUT_VERSION:
+        raise ContractError("$.schemaVersion 版本不支持")
+    if not _RUN_ID_RE.fullmatch(_string(output["runId"], "$.runId")):
+        raise ContractError("$.runId 格式非法")
+    _integer(output["attempt"], "$.attempt", minimum=1)
+    if output["mode"] not in {"production", "diagnostic_only"}:
+        raise ContractError("$.mode 枚举非法")
+    _string(output["qualityPolicyVersion"], "$.qualityPolicyVersion")
+    _hash(output["contextSnapshotId"], "$.contextSnapshotId")
+    _hash(output["contextSnapshotSha256"], "$.contextSnapshotSha256")
+    _integer(output["candidateVersion"], "$.candidateVersion", minimum=1)
+    body = _string(output["candidateBody"], "$.candidateBody")
+    candidate_hash = _hash(output["candidateSha256"], "$.candidateSha256")
+    expected_hash = "sha256:" + hashlib.sha256(body.encode("utf-8")).hexdigest()
+    if candidate_hash != expected_hash:
+        raise ContractError("$.candidateSha256 与规范正文不匹配")
+    eligible = _boolean(output["acceptanceEligible"], "$.acceptanceEligible")
+    if output["mode"] == "diagnostic_only" and eligible:
+        raise ContractError("诊断输出必须 acceptanceEligible=false")
+
+    for index, claim in enumerate(_array(output["claimLedger"], "$.claimLedger")):
+        item = _object(
+            claim,
+            f"$.claimLedger[{index}]",
+            frozenset({"claimId", "candidateSha256", "startCodePoint", "endCodePoint", "factType", "factEvidenceId", "coverageState"}),
+            frozenset({"proseEvidenceId"}),
+        )
+        for field in ("claimId", "factType", "factEvidenceId", "coverageState"):
+            _string(item[field], f"$.claimLedger[{index}].{field}")
+        _hash(item["candidateSha256"], f"$.claimLedger[{index}].candidateSha256")
+        if item["candidateSha256"] != candidate_hash:
+            raise ContractError(f"$.claimLedger[{index}] 未绑定当前候选")
+        start = _integer(item["startCodePoint"], f"$.claimLedger[{index}].startCodePoint")
+        end = _integer(item["endCodePoint"], f"$.claimLedger[{index}].endCodePoint", minimum=1)
+        if end <= start or end > len(body):
+            raise ContractError(f"$.claimLedger[{index}] Unicode 偏移越界")
+        if "proseEvidenceId" in item and item["proseEvidenceId"] is not None:
+            _string(item["proseEvidenceId"], f"$.claimLedger[{index}].proseEvidenceId")
+    for index, request in enumerate(_array(output["evidenceRequests"], "$.evidenceRequests")):
+        item = _object(request, f"$.evidenceRequests[{index}]", frozenset({"requestId", "query", "reason", "priority"}))
+        for field in ("requestId", "query", "reason", "priority"):
+            _string(item[field], f"$.evidenceRequests[{index}].{field}")
+    for index, declaration in enumerate(_array(output["newSettingDeclarations"], "$.newSettingDeclarations")):
+        item = _object(declaration, f"$.newSettingDeclarations[{index}]", frozenset({"declarationId", "factType", "text", "startCodePoint", "endCodePoint"}))
+        for field in ("declarationId", "factType", "text"):
+            _string(item[field], f"$.newSettingDeclarations[{index}].{field}")
+        start = _integer(item["startCodePoint"], f"$.newSettingDeclarations[{index}].startCodePoint")
+        end = _integer(item["endCodePoint"], f"$.newSettingDeclarations[{index}].endCodePoint", minimum=1)
+        if end <= start or end > len(body):
+            raise ContractError(f"$.newSettingDeclarations[{index}] Unicode 偏移越界")
+    self_check = _object(output["selfCheck"], "$.selfCheck", frozenset({"hardConstraintsCovered", "notes"}))
+    _boolean(self_check["hardConstraintsCovered"], "$.selfCheck.hardConstraintsCovered")
+    if any(not isinstance(note, str) for note in _array(self_check["notes"], "$.selfCheck.notes")):
+        raise ContractError("$.selfCheck.notes 必须是字符串数组")
+    return json.loads(canonical_json(output))
+
+
+__all__ = [
+    "CONTEXT_VERSION", "OUTPUT_VERSION", "PLAN_VERSION", "MANIFEST_VERSION", "TIE_BREAK",
+    "ContractError", "normalize_text", "canonical_json", "retrieval_identity", "han_count",
+    "calculate_target_chars", "validate_writer_context", "validate_writer_output",
+]

.claude/skills/replay-eval/scripts/load_reference_work.py

@@ -1,14 +1,16 @@
 #!/usr/bin/env python3
 """从实验库只读组装回放评测配置。
 
-本适配器只做 SELECT 和临时文件输出，不写数据库、不读取正文 block 的内容。
+本适配器只做 SELECT 和临时文件输出，不写数据库。正文读取仅允许通过
-参考作品的目标章只进入审计侧 proxy，历史规划上下文和三臂卡注入严格分开。
+`load_frozen_prose_rows()` 在只读快照内读取冻结线以前的 Canonical block；
+参考作品目标章仍只进入审计侧 proxy，历史上下文和三臂卡注入严格分开。
 """
 
 from __future__ import annotations
 
 import argparse
 import copy
+import hashlib
 import json
 import re
 from pathlib import Path
@@ -98,6 +100,31 @@ def _card_history(payload: Mapping[str, Any]) -> list[Mapping[str, Any]]:
     raise AdapterError("卡缺少可按绝对章号冻结的历史字段")
 
 
+def _structured_source_refs(payload: Mapping[str, Any], history: Sequence[Mapping[str, Any]], as_of: int) -> list[dict[str, Any]]:
+    """提取卡内结构化原文指针，文本“出处”不能替代块级引用。"""
+
+    candidates: list[Any] = [payload.get("sourceRefs")]
+    fields = payload.get("字段")
+    if isinstance(fields, Mapping):
+        candidates.append(fields.get("sourceRefs"))
+    candidates.extend(item.get("sourceRefs") for item in history if isinstance(item, Mapping))
+    refs: list[dict[str, Any]] = []
+    for candidate in candidates:
+        if not isinstance(candidate, list):
+            continue
+        for raw_ref in candidate:
+            if not isinstance(raw_ref, Mapping):
+                continue
+            chapter = normalize_chapter(raw_ref.get("chapter") or raw_ref.get("章"))
+            if chapter is None or chapter > as_of:
+                continue
+            normalized = copy.deepcopy(dict(raw_ref))
+            normalized["chapter"] = chapter
+            normalized.pop("章", None)
+            refs.append(normalized)
+    return refs
+
+
 def project_card(row: Mapping[str, Any], *, as_of: int, source_version: str) -> dict[str, Any]:
     """将候选卡投影为截至 as_of 的 eval-only 索引视图。"""
 
@@ -129,7 +156,13 @@ def project_card(row: Mapping[str, Any], *, as_of: int, source_version: str) ->
         "cardId": card_id,
         "type": card_type,
         "name": name,
+        "score": float(row.get("score") or 0),
+        "sourceId": f"eval-draft:{card_id}",
+        "sourceVersion": source_version,
+        "sourceOffset": 0,
         "milestones": copy.deepcopy(history),
+        "stateAsOf": copy.deepcopy(history),
+        "sourceRefs": _structured_source_refs(payload, history, normalized_as_of),
         "appearanceChapters": appearances,
         "derivedState": {
             "asOfChapter": normalized_as_of,
@@ -142,12 +175,116 @@ def project_card(row: Mapping[str, Any], *, as_of: int, source_version: str) ->
             "chapterRange": f"1-{normalized_as_of}",
         },
         "evaluationStatus": "eval_draft",
+        "sourceType": "upgrade_book",
+        "sourceKind": "eval_draft",
         "upstreamStatus": str(row.get("status") or "unknown"),
         "productionRetrievalEligible": False,
         "omittedHistoryCount": len(omitted),
     }
 
 
+def begin_read_snapshot(conn: Any) -> None:
+    """在第一条业务查询前固定可重复读、只读事务。"""
+
+    conn.execute("SET TRANSACTION ISOLATION LEVEL REPEATABLE READ READ ONLY")
+
+
+def load_frozen_prose_rows(
+    *,
+    dsn: str,
+    tenant_id: int,
+    work_id: int,
+    as_of: int,
+    source_refs: Sequence[Mapping[str, Any]] = (),
+    chapter_numbers: Sequence[int] = (),
+) -> list[dict[str, Any]]:
+    """从同一只读快照读取冻结线内的 Canonical 历史原文。
+
+    SQL 只存在于 replay-eval 读取适配器；read-context 的作品与回放仓储都
+    调用本入口，避免再造正文读取旁路。
+    """
+
+    normalized_as_of = _required_chapter(as_of, "as_of")
+    requested_chapters = {_required_chapter(item, "chapter_numbers[]") for item in chapter_numbers}
+    block_ids: set[int] = set()
+    refs_by_block: dict[int, list[Mapping[str, Any]]] = {}
+    for index, ref in enumerate(source_refs):
+        if not isinstance(ref, Mapping):
+            raise AdapterError(f"source_refs[{index}] 必须是对象")
+        chapter = _required_chapter(ref.get("chapter"), f"source_refs[{index}].chapter")
+        if chapter > normalized_as_of:
+            raise AdapterError(f"source_refs[{index}] 包含目标章或未来章")
+        block_id = ref.get("blockId")
+        if isinstance(block_id, bool) or not isinstance(block_id, int) or block_id <= 0:
+            raise AdapterError(f"source_refs[{index}].blockId 必须是正整数")
+        block_ids.add(block_id)
+        refs_by_block.setdefault(block_id, []).append(ref)
+    if any(chapter > normalized_as_of for chapter in requested_chapters):
+        raise AdapterError("chapter_numbers 包含目标章或未来章")
+    if not requested_chapters and not block_ids:
+        return []
+
+    with psycopg.connect(dsn, row_factory=dict_row) as conn:
+        begin_read_snapshot(conn)
+        rows = conn.execute(
+            """
+            SELECT ch.order_no AS chapter,ch.id AS chapter_id,ch.status AS chapter_status,
+                   b.id AS block_id,b.order_no AS block_order,b.revision,b.content_text
+            FROM muse_content_chapter ch
+            JOIN muse_content_block b ON b.chapter_id=ch.id
+            WHERE ch.tenant_id=%s AND ch.work_id=%s AND ch.deleted=FALSE
+              AND b.tenant_id=%s AND b.work_id=%s AND b.deleted=FALSE
+              AND ch.order_no<=%s AND ch.status IN ('published','confirmed','canonical')
+              AND (ch.order_no=ANY(%s) OR b.id=ANY(%s))
+            ORDER BY ch.order_no,b.order_no,b.id
+            """,
+            (tenant_id, work_id, tenant_id, work_id, normalized_as_of, sorted(requested_chapters), sorted(block_ids)),
+        ).fetchall()
+
+    result: list[dict[str, Any]] = []
+    for row in rows:
+        text = str(row.get("content_text") or "")
+        block_id = int(row["block_id"])
+        matching_refs = refs_by_block.get(block_id)
+        if matching_refs:
+            for ref in matching_refs:
+                start = int(ref.get("startCodePoint") or 0)
+                end = int(ref.get("endCodePoint") or len(text))
+                if start < 0 or end <= start or end > len(text):
+                    raise AdapterError(f"block {block_id} 的字符区间越界")
+                fragment = text[start:end]
+                result.append(
+                    {
+                        "chapter": int(row["chapter"]),
+                        "blockId": block_id,
+                        "blockOrder": int(row["block_order"]),
+                        "sourceRef": copy.deepcopy(dict(ref)),
+                        "text": fragment,
+                        "contentSha256": "sha256:" + hashlib.sha256(fragment.encode("utf-8")).hexdigest(),
+                    }
+                )
+        elif int(row["chapter"]) in requested_chapters:
+            source_version = f"chapter:{row['chapter']}:block:{block_id}:revision:{row.get('revision') or 0}"
+            result.append(
+                {
+                    "chapter": int(row["chapter"]),
+                    "blockId": block_id,
+                    "blockOrder": int(row["block_order"]),
+                    "sourceRef": {
+                        "sourceId": f"chapter:{row['chapter']}:block:{block_id}",
+                        "sourceVersion": source_version,
+                        "chapter": int(row["chapter"]),
+                        "blockId": block_id,
+                        "startCodePoint": 0,
+                        "endCodePoint": len(text),
+                    },
+                    "text": text,
+                    "contentSha256": "sha256:" + hashlib.sha256(text.encode("utf-8")).hexdigest(),
+                }
+            )
+    return result
+
+
 def _card_source_version(row: Mapping[str, Any], base_version: str) -> str:
     """把卡行 revision 纳入来源版本，防止卡内容变更复用旧版本。"""
 
@@ -435,7 +572,7 @@ def load_reference_rows(
     correct_ids, placebo_ids = _validate_selection(card_selection)
     selected_ids = [int(item) if str(item).isdigit() else item for item in correct_ids + placebo_ids]
     with psycopg.connect(dsn, row_factory=dict_row) as conn:
-        conn.execute("SET TRANSACTION READ ONLY")
+        begin_read_snapshot(conn)
         work = conn.execute(
             """
             SELECT id,title,revision,chapter_count,parse_status,import_status

.claude/skills/search/scripts/search.py

@@ -6,6 +6,7 @@
 import json
 import pathlib
 import sys
+from typing import Any, Callable
 
 import click
 import psycopg
@@ -38,6 +39,140 @@ def visible(ai_rule, purpose):
     return purpose in ai_rule
 
 
+def _default_embedder(intent: str) -> list[float]:
+    """复用 embed skill 生成单条查询向量，并统一失败语义。"""
+
+    vectors, bad = embed_texts(_session(), [intent])
+    if bad or not vectors:
+        raise ValueError("查询嵌入失败")
+    return vectors[0]
+
+
+def _structured_source_refs(payload: dict[str, Any], lineage: Any) -> list[dict[str, Any]]:
+    """只接收结构化来源指针；人类可读的“出处”文本不能冒充可回读引用。"""
+
+    candidates = []
+    if isinstance(lineage, dict):
+        candidates.append(lineage.get("sourceRefs"))
+    fields = payload.get("字段")
+    if isinstance(fields, dict):
+        candidates.append(fields.get("sourceRefs"))
+    candidates.append(payload.get("sourceRefs"))
+    for value in candidates:
+        if isinstance(value, list) and all(isinstance(item, dict) for item in value):
+            return value
+    return []
+
+
+def _milestones(payload: dict[str, Any]) -> list[dict[str, Any]]:
+    """从卡字段中提取历史里程碑，终态摘要不会进入冻结投影。"""
+
+    fields = payload.get("字段")
+    if not isinstance(fields, dict):
+        return []
+    for key in ("演变历程", "演变轨迹", "milestones"):
+        value = fields.get(key)
+        if isinstance(value, list) and all(isinstance(item, dict) for item in value):
+            return value
+    return []
+
+
+def search_cards(
+    intent: str,
+    *,
+    scope: str = "admin",
+    work_id: int | None = None,
+    ttype: str | None = None,
+    purpose: str = "generation",
+    top: int = 5,
+    dsn: str = DSN,
+    tenant_id: int = TENANT,
+    connection_factory: Callable[..., Any] = psycopg.connect,
+    embedder: Callable[[str], list[float]] = _default_embedder,
+) -> list[dict[str, Any]]:
+    """执行唯一的卡检索语义，CLI 与正文读取器共同调用本函数。
+
+    生产作品面只读 active Canonical entity、允许状态和有效绑定；治理面
+    保留原有 draft 能力，但正文生产适配器不会调用治理面。
+    """
+
+    if scope not in {"admin", "work"}:
+        raise ValueError("scope 只能是 admin 或 work")
+    if scope == "work" and not work_id:
+        raise ValueError("scope=work 必须提供 work_id")
+    if purpose not in {"generation", "planning", "detection", "extraction"}:
+        raise ValueError("purpose 非法")
+    if isinstance(top, bool) or not isinstance(top, int) or top <= 0:
+        raise ValueError("top 必须是正整数")
+
+    qvec = json.dumps(embedder(intent))
+    with connection_factory(dsn) as conn:
+        ai_rules = load_ai_context(conn)
+        if scope == "admin":
+            sql = """SELECT 'draft' AS src, d.id, d.draft_payload AS payload, d.status,
+                            1 - (e.embedding <=> %s::vector) AS score,
+                            d.revision, d.current_canonical_snapshot AS lineage,
+                            NULL::varchar AS binding_status, d.source_status
+                     FROM example_knowledge_embedding e
+                     JOIN muse_knowledge_draft d ON d.id = e.draft_id
+                     WHERE e.tenant_id=%s AND e.deleted=FALSE AND d.deleted=FALSE"""
+            args = [qvec, tenant_id]
+        else:
+            sql = """SELECT 'entity' AS src, en.id,
+                            jsonb_build_object('型', en.entity_type, '名称', en.normalized_name,
+                                               '一句话摘要', en.description, '字段', en.attributes) AS payload,
+                            en.status, 1 - (e.embedding <=> %s::vector) AS score,
+                            en.revision, en.lineage_payload AS lineage,
+                            b.binding_status, en.source_status
+                     FROM example_knowledge_embedding e
+                     JOIN muse_knowledge_entity en ON en.id = e.entity_id
+                     JOIN muse_knowledge_binding b ON b.kb_id = en.kb_id AND b.work_id = %s
+                          AND b.binding_status='active' AND b.deleted=FALSE AND b.tenant_id=%s
+                     WHERE e.tenant_id=%s AND e.deleted=FALSE AND en.deleted=FALSE
+                       AND en.status='active' AND en.source_status IN ('active','authorized')
+                       AND en.source_action_policy='allowed'"""
+            args = [qvec, work_id, tenant_id, tenant_id]
+        if ttype:
+            sql += (" AND d.draft_payload->>'型' = %s" if scope == "admin" else " AND en.entity_type = %s")
+            args.append(ttype)
+        id_column = "d.id" if scope == "admin" else "en.id"
+        revision_column = "d.revision" if scope == "admin" else "en.revision"
+        sql += f" ORDER BY score DESC, {revision_column}::text ASC, {id_column}::text ASC LIMIT %s"
+        args.append(top)
+        rows = conn.execute(sql, args).fetchall()
+
+    results = []
+    for src, row_id, raw_payload, status, score, revision, lineage, binding_status, source_status in rows:
+        payload = raw_payload or {}
+        card_type = payload.get("型") or payload.get("type") or "?"
+        rules = ai_rules.get(card_type, {})
+        fields = payload.get("字段") or {}
+        visible_fields = {key: item for key, item in fields.items() if visible(rules.get(key), purpose)}
+        source_version = f"{src}-revision:{revision or 0}"
+        source_id = f"canonical-entity:{row_id}" if src == "entity" else f"draft:{row_id}"
+        results.append(
+            {
+                "cardId": str(row_id),
+                "type": card_type,
+                "name": payload.get("名称"),
+                "score": float(score),
+                "summary": payload.get("一句话摘要"),
+                "visibleFields": visible_fields,
+                "omittedFields": sorted(set(fields) - set(visible_fields)),
+                "sourceId": source_id,
+                "sourceVersion": source_version,
+                "sourceOffset": 0,
+                "sourceRefs": _structured_source_refs(payload, lineage),
+                "milestones": _milestones(payload),
+                "sourceKind": "canonical_entity" if src == "entity" else "draft",
+                "sourceStatus": source_status or status,
+                "bindingStatus": binding_status,
+                "productionRetrievalEligible": src == "entity" and status == "active" and binding_status == "active",
+            }
+        )
+    return results
+
+
 @click.command()
 @click.argument("intent")
 @click.option("--scope", type=click.Choice(["admin", "work"]), default="admin", show_default=True,
@@ -49,55 +184,32 @@ def visible(ai_rule, purpose):
 @click.option("--top", default=5, show_default=True)
 @click.option("--json", "as_json", is_flag=True)
 def main(intent, scope, work_id, ttype, purpose, top, as_json):
-    if scope == "work" and not work_id:
+    try:
-        raise click.ClickException("--scope work 必须带 --work-id（授权过滤依赖绑定关系）")
+        cards = search_cards(
+            intent,
+            scope=scope,
+            work_id=work_id,
+            ttype=ttype,
+            purpose=purpose,
+            top=top,
+        )
+    except ValueError as error:
+        raise click.ClickException(str(error)) from error
 
-    vecs, bad = embed_texts(_session(), [intent])
+    results = [
-    if bad:
+        {
-        raise click.ClickException("查询嵌入失败")
+            "来源": item["sourceId"],
-    qvec = json.dumps(vecs[0])
+            "型": item["type"],
-
+            "名称": item["name"],
-    with psycopg.connect(DSN) as conn:
+            "状态": item["sourceStatus"],
-        ai_rules = load_ai_context(conn)
+            "相似度": round(item["score"], 4),
-        if scope == "admin":
+            "一句话摘要": item["summary"],
-            # 治理面：draft(pending/confirmed)+entity 全量
+            "可见字段": item["visibleFields"],
-            sql = """SELECT 'draft' AS src, d.id, d.draft_payload AS payload, d.status,
+            "出处": item["sourceRefs"],
-                            1 - (e.embedding <=> %s::vector) AS score
+            "裁剪回显": item["omittedFields"],
-                     FROM example_knowledge_embedding e
+        }
-                     JOIN muse_knowledge_draft d ON d.id = e.draft_id
+        for item in cards
-                     WHERE e.tenant_id=%s AND e.deleted=FALSE AND d.deleted=FALSE"""
+    ]
-            args = [qvec, TENANT]
-        else:
-            # 作品面：仅 active entity 且其 kb 已绑定到该作品（授权在查询层强制）
-            sql = """SELECT 'entity' AS src, en.id,
-                            jsonb_build_object('型', en.entity_type, '名称', en.normalized_name,
-                                               '一句话摘要', en.description, '字段', en.attributes) AS payload,
-                            en.status, 1 - (e.embedding <=> %s::vector) AS score
-                     FROM example_knowledge_embedding e
-                     JOIN muse_knowledge_entity en ON en.id = e.entity_id
-                     JOIN muse_knowledge_binding b ON b.kb_id = en.kb_id AND b.work_id = %s
-                          AND b.binding_status='active' AND b.deleted=FALSE AND b.tenant_id=%s
-                     WHERE e.tenant_id=%s AND e.deleted=FALSE AND en.deleted=FALSE AND en.status='active'"""
-            args = [qvec, work_id, TENANT, TENANT]
-        if ttype:
-            sql += (" AND d.draft_payload->>'型' = %s" if scope == "admin"
-                    else " AND en.entity_type = %s")
-            args.append(ttype)
-        sql += " ORDER BY score DESC LIMIT %s"
-        args.append(top)
-        rows = conn.execute(sql, args).fetchall()
-
-    results = []
-    for src, rid, payload, status, score in rows:
-        p = payload or {}
-        t = p.get("型", "?")
-        rules = ai_rules.get(t, {})
-        fields = p.get("字段") or {}
-        vis = {k: v for k, v in fields.items() if visible(rules.get(k), purpose)}
-        cut = sorted(set(fields) - set(vis))
-        results.append({"来源": f"{src}#{rid}", "型": t, "名称": p.get("名称"), "状态": status,
-                        "相似度": round(float(score), 4), "一句话摘要": p.get("一句话摘要"),
-                        "可见字段": vis, "出处": p.get("出处"), "裁剪回显": cut})
 
     if as_json:
         click.echo(json.dumps(results, ensure_ascii=False, indent=1))

docs/superpowers/plans/2026-07-20-writer-agent-v1.md

@@ -1,6 +1,6 @@
 # 正文智能体实验台 v1 Implementation Plan
 
-> **For agentic workers:** REQUIRED SUB-SKILL: Use `superpowers:subagent-driven-development` (recommended) or `superpowers:executing-plans` to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+> **执行约束（创始人 2026-07-20 确认）：** 不使用 worktree，不使用 superpower。执行类任务由子代理直接在当前 `main` 工作树实现，主代理负责文件边界、代码审查与机械验证；不得暂存或覆盖用户已有改动。
 
 **Goal:** 在参考作品回放环境中实现一套可机械验证的正文智能体：知识卡只承担索引职责，智能体必须顺着卡片证据回读冻结线内的历史原文，再依据大纲、细纲和事实证据生成正文，并经过审查、盲评及 Gate A/B 验收。
 
@@ -19,32 +19,30 @@
 - `抽取卡 -> 原文` 是强制链路：抽取卡若没有可追踪的历史原文来源，不得单独作为正文硬事实。作者已确认的正式设定、Canonical 状态与细纲声明的新事实可以直接成为事实证据，并标为 `declared_new` 或相应来源类型。
 - “完成”至少包含：测试通过、dry-run 通过、真实回放所需配置齐全；没有真实模型回放结果时只能称“实现完成”，不能称 Gate A/B 通过。
 
-## 环境准备：固定 worktree 本地依赖
+## 环境准备：当前主工作树本地依赖
 
-- [ ] 在内层 worktree 创建独立虚拟环境，不复用用户主工作树的 `.venv`：
+- [ ] 在 `agent-example` 当前 `main` 工作树复用已安装依赖的 `.venv`，先确认解释器版本与依赖可用：
 
 ```bash
-cd /private/tmp/agent-example-writer-v1
+cd /Users/qingse/Sync/local-git/oh-my-muse/agent-example
-uv venv --python 3.12 .venv
-uv pip install --python .venv/bin/python -r requirements.txt
 .venv/bin/python --version
 ```
 
-- [ ] 后续所有 Python 命令都在 `/private/tmp/agent-example-writer-v1` 执行并使用 `.venv/bin/python`；若 `requirements.txt` 变化，先重装依赖再验证。
+- [ ] 后续所有 Python 命令都在当前 `agent-example` 主工作树执行并使用 `.venv/bin/python`；若 `requirements.txt` 变化，先重装依赖再验证。
 
 ## 任务 0：回写稳定设计 SoT，标明实验边界
 
-> 本任务在已建立的外层独立 worktree `/private/tmp/oh-my-muse-writer-sot`（分支 `feature/writer-agent-sot`）执行并独立提交；后续任务均在当前 `agent-example` worktree `/private/tmp/agent-example-writer-v1` 执行。禁止在用户有未提交改动的外层主工作树直接编辑或提交。
+> 本任务直接在外层仓库 `/Users/qingse/Sync/local-git/oh-my-muse` 的当前 `main` 工作树执行。只提交下列精确路径，保留外层和内层工作树中的用户已有改动。
 
 **Files:**
-- Modify: `/private/tmp/oh-my-muse-writer-sot/design-docs/专题-01-正文建议接受(Accept Suggestion)实现规范.md`
+- Modify: `/Users/qingse/Sync/local-git/oh-my-muse/design-docs/专题-01-正文建议接受(Accept Suggestion)实现规范.md`
-- Modify: `/private/tmp/oh-my-muse-writer-sot/design-docs/专题-03-AI编排上下文与质量评测实现规范.md`
+- Modify: `/Users/qingse/Sync/local-git/oh-my-muse/design-docs/专题-03-AI编排上下文与质量评测实现规范.md`
-- Modify: `/private/tmp/oh-my-muse-writer-sot/design-docs/专题-04-生成质量门控与创作健康度设计方案.md`
+- Modify: `/Users/qingse/Sync/local-git/oh-my-muse/design-docs/专题-04-生成质量门控与创作健康度设计方案.md`
-- Modify: `/private/tmp/oh-my-muse-writer-sot/design-docs/专题-05-AI统一交互协议与外部AgentAdapter设计.md`
+- Modify: `/Users/qingse/Sync/local-git/oh-my-muse/design-docs/专题-05-AI统一交互协议与外部AgentAdapter设计.md`
-- Modify: `/private/tmp/oh-my-muse-writer-sot/design-docs/专题-06-元数据驱动的智能体架构.md`
+- Modify: `/Users/qingse/Sync/local-git/oh-my-muse/design-docs/专题-06-元数据驱动的智能体架构.md`
-- Modify: `/private/tmp/oh-my-muse-writer-sot/design-docs/专题-07-知识消费契约与质量闭环.md`
+- Modify: `/Users/qingse/Sync/local-git/oh-my-muse/design-docs/专题-07-知识消费契约与质量闭环.md`
-- Modify: `/private/tmp/oh-my-muse-writer-sot/design-docs/架构-04-状态机与约束清单.md`
+- Modify: `/Users/qingse/Sync/local-git/oh-my-muse/design-docs/架构-04-状态机与约束清单.md`
-- Modify: `/private/tmp/oh-my-muse-writer-sot/.agents/workflows/ai-development-protocol.md`
+- Modify: `/Users/qingse/Sync/local-git/oh-my-muse/.agents/workflows/ai-development-protocol.md`
 
 - [ ] 在专题-07 中把“卡是索引，根据卡回读原文”设为正文消费契约，明确卡不能替代原文。
 - [ ] 在专题-03 中补齐 `WriterContext v1`、`WriterOutput v1`、`RetrievalManifest` 和冻结语义。
@@ -59,7 +57,7 @@ uv pip install --python .venv/bin/python -r requirements.txt
 - [ ] 运行文档一致性检查：
 
 ```bash
-cd /private/tmp/oh-my-muse-writer-sot
+cd /Users/qingse/Sync/local-git/oh-my-muse
 rg -n "卡是索引|WriterContext v1|acceptanceEligible|Gate A|Gate B|Canonical" \
   'design-docs/专题-01-正文建议接受(Accept Suggestion)实现规范.md' \
   design-docs/专题-03-AI编排上下文与质量评测实现规范.md \
@@ -74,7 +72,7 @@ rg -n "卡是索引|WriterContext v1|acceptanceEligible|Gate A|Gate B|Canonical"
 - [ ] Commit：
 
 ```bash
-git -C /private/tmp/oh-my-muse-writer-sot add \
+git -C /Users/qingse/Sync/local-git/oh-my-muse add \
   'design-docs/专题-01-正文建议接受(Accept Suggestion)实现规范.md' \
   design-docs/专题-03-AI编排上下文与质量评测实现规范.md \
   design-docs/专题-04-生成质量门控与创作健康度设计方案.md \
@@ -83,7 +81,7 @@ git -C /private/tmp/oh-my-muse-writer-sot add \
   design-docs/专题-07-知识消费契约与质量闭环.md \
   design-docs/架构-04-状态机与约束清单.md \
   .agents/workflows/ai-development-protocol.md
-git -C /private/tmp/oh-my-muse-writer-sot commit \
+git -C /Users/qingse/Sync/local-git/oh-my-muse commit \
   -m "设计: 固化正文智能体卡索引原文回读契约"
 ```
 

meta/schemas/README.md

@@ -7,7 +7,7 @@
 - **两轴**：`domain` ∈ content / world / narrative / knowledge / ai_context；`scope` ∈ work / chapter / block / entity / relation / event / agent。
 - **aiContext 控制项**（阶段一只用这一个）：`true` 任何用途都可入 AI 上下文；`false` 一律不入；`[用途…]` 仅列出的用途可入。用途取值：`planning` / `generation` / `detection` / `extraction`。其余控制项（uiVisible/userEditable 等）阶段二随真后端启用。
 - **基础字段**（所有型共有，各 schema 不再重复）：`名称`、`别名`、`一句话摘要`、`标签`、`来源`（手工 / 抽取@第N章 / 拆书@书名）、`状态`（草稿 / 已确认）。
-- **状态**：`启用`（21 型）/ `待启用`（2 型：pacing、generation_context，首个用到的场景来临时再启用）。范式五型与参考书档案已于拆书场景（A8）启用并补全字段合同。
+- **状态**：23 型均已启用。`generation_context` 已于正文实验台启用；范式五型与参考书档案已于拆书场景（A8）启用并补全字段合同。
 - **演进**：增删型或字段先过专题-06 §4.4 的四判据与降级规则；变更靠 git 追溯。
 
 ## 实例落点表（哪个型的实例长在哪）
@@ -31,8 +31,8 @@
 | event | `知识/事件/*.md` |
 | reference_work | `knowledge/参考书/*/档案.md`（原文 txt 同目录） |
 | craft / combat / emotion / scene_pattern / trope | 公共面 `knowledge/范式/{技法,打斗,情感,通用桥段,套路}/`；作品面 `知识/` 对应子目录 |
-| generation_context（待启用） | 阶段一以上下文回显形式落 `works/*/评审/`，不建实例文件 |
+| generation_context | 阶段一以严格 JSON 上下文与 Markdown manifest 回显落 `works/*/评审/`，不建实例文件 |
-| pacing（待启用） | 卷中期节奏审计时启用 |
+| pacing | 卷中期节奏审计时使用 |
 
 知识卡「值得立卡」的门槛：有跨章戏份或跨章履约；一次性龙套与单场景道具不立卡，写在章内即可。
 

meta/schemas/generation_context.yaml

@@ -3,7 +3,63 @@ target_type: generation_context
 domain: ai_context
 scope: agent
 本体分组: 配套
-状态: 待启用
+状态: 启用
 判据: AI 上下文组装与输出合同的结构模具,运行时对象、用户不可见
-阶段一落法: 以 read-context 的上下文回显(works/*/评审/上下文-*.md)代替实例,不建卡
+实例落点: 以 read-context 的 JSON 上下文与 Markdown manifest 回显落 works/*/评审/,不建知识卡
-启用条件: 阶段二对齐统一读取器(专题-06 §7)的输出合同时正式启用
+严格合同:
+  版本:
+    WriterContext: writer-context-v1
+    WriterOutput: writer-output-v1
+    RetrievalPlan: writer-retrieval-plan-v1
+    RetrievalManifest: writer-retrieval-manifest-v1
+  规范化: UTF-8、Unicode NFC、LF 换行、对象键排序、无额外空白
+  身份规则: runId、generatedAt、timestamp、executionNode 不参与检索与上下文身份
+  未知字段: 拒绝
+  WriterContext必填字段:
+    - schemaVersion
+    - runId
+    - attempt
+    - mode
+    - purpose
+    - qualityPolicyVersion
+    - workId
+    - targetChapter
+    - asOf
+    - contextSnapshot
+    - sourceVersion
+    - authorizationSnapshot
+    - sourceStatus
+    - retrievalPlan
+    - retrievalManifest
+    - fineOutline
+    - narrativeState
+    - factEvidence
+    - proseEvidence
+    - patternReferences
+    - evidenceCoverage
+    - outputContract
+    - tokenBudget
+    - omittedSources
+    - acceptanceEligible
+  WriterOutput必填字段:
+    - schemaVersion
+    - runId
+    - attempt
+    - mode
+    - qualityPolicyVersion
+    - contextSnapshotId
+    - contextSnapshotSha256
+    - candidateVersion
+    - candidateSha256
+    - acceptanceEligible
+    - candidateBody
+    - claimLedger
+    - evidenceRequests
+    - newSettingDeclarations
+    - selfCheck
+  双证据:
+    factEvidence来源: [historical_prose, formal_setting, canonical_state, fine_outline_declared_new]
+    proseEvidence来源: 历史 Canonical 原文,必须带章号、块、字符区间与内容哈希
+  接受边界: evaluation、diagnostic、diagnostic_only 一律 acceptanceEligible=false
+  稳定排序: score DESC, sourceVersion ASC, sourceId ASC, sourceOffset ASC
+实现校验器: .claude/skills/read-context/scripts/writer_contract.py

meta/schemas/outline.yaml

@@ -13,8 +13,10 @@ scope: work
   - key: 近三章细纲
     说明: 每章:章目标/关键事件/出场角色/伏笔动作(埋·推·收)/章末钩子——续写的直接依据;章细纲字数≈章正文 3–5%(3000 字章→100–150 字);细纲是结构骨架不是缩写,超比例=退回重做
     aiContext: true
+  - { key: targetChars, 说明: 可选的本章目标汉字数;仍受 2000–10000 硬边界约束, aiContext: [generation] }
   - { key: 未来卷粗纲, 说明: 防续写提前收线,仅规划可见, aiContext: [planning] }
   - { key: 弃案记录, 说明: 改掉的旧方向,防被 AI 复活, aiContext: false }
 设计发现:
   - 2026-07-09 全书解析设计——存量作品需**规划逆向**(全文→章细纲→卷粗纲→主线,自底向上,与创作期自顶向下互为镜像);SoT 产品-03 §3.7 全书解析产出未含大纲/细纲,「细纲」粒度层级亦为本仓先行,均待回填
   - 同日——「细纲/正文字数比例」约束防解析退化成压缩复述,**创始人已拍板@2026-07-09**(章细纲 3–5%、卷粗纲 0.3–0.5%、主线 ≤50 字),已写入上方字段说明;D1 回填 design-docs 时随字段合同一并带走
+  - 2026-07-20 正文实验台——新增可选 targetChars;未提供时只依据目标章之前的有效 Canonical 章长中位数与细纲密度确定性计算