feat: 读标可以走完一份文件；审计待确认区，修规则缺陷、提示词出新版本

读标操作（FR-M2-07、08）
- 后端：清单项详情（ETag）、修改（If-Match 选填，记 edited_fields 和 AI 原值）、
  采纳（补出处或声明「人工确认、无出处」，可带修改）、删除 / 丢弃与恢复、人工新增、
  替换出处、人工建出处（选同一页的句子）、修改记录（来自审计）、确认清单（不可变版本，
  内容没变复用上一版）、版本列表；正式清单有变化时已确认的清单回到草稿；项目已递交只读
- 重新解析不推翻人工决定：丢弃的不再冒出来，采纳的留在正式清单，人工补过的出处不被覆盖
- 前端：待确认区采纳 / 补上出处后采纳 / 修改后采纳 / 丢弃，正式清单修改、补出处、删除
  （可撤销）、修改记录、新增条目、确认清单；pdf.js 原文查看器（点出处跳页、按坐标框高亮、
  选句补出处），整屏对照布局，J / K 上下条；跨页的一句话合成一处出处显示

待确认审计（架构 §13 SK11）
- 规则只修真正的缺陷：表格框线不算下划线（文字层 v3）；句号后的右括号并进同一句
  （切句 s3-v3）；中文编号小节和冒号编号条目的上级条款，依据沿上级链核对；取值认中文
  数字、同一行单元格和评分因素标题；补几种无效写法；比对引文前去掉我们加进输入的标记
- 不放宽：引文最短 8 个字、编号必须对、证据全部通过
- 提示词新版本：substantive v3，qualification / scoring / schedule / tech_param v2；
  引文复制整句、并列多项不拆、跨页只引上一行，除实质性条款外输入改成纯文本
- BD_AI_MODELS 按提示词声明的能力选模型；离线评测工具 evals/run_extract.py
- 待确认：国动 46 → 4，S02 17 → 13，S06 174 → 57

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

server/src/bid_doctor/document/service.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import json
 from datetime import UTC, datetime
 from typing import Any
 from uuid import UUID
@@ -12,14 +13,15 @@ from sqlalchemy.ext.asyncio import AsyncSession
 from bid_doctor.audit import service as audit
 from bid_doctor.billing import service as billing
 from bid_doctor.core import keys, queue
-from bid_doctor.core.crypto import seal_object, sha256
+from bid_doctor.core.crypto import open_object, seal_object, sha256
 from bid_doctor.core.db import TenantContext
 from bid_doctor.core.errors import AppError
 from bid_doctor.core.ids import uuid7
 from bid_doctor.core.storage import project_prefix, storage
-from bid_doctor.document import uploads
+from bid_doctor.document import citations, uploads
 from bid_doctor.document.citations import document_label
-from bid_doctor.document.models import ParseJob, SourceDocument
+from bid_doctor.document.models import Citation, ParseJob, SourceDocument
+from bid_doctor.document.pipeline.text_layer import Page
 from bid_doctor.project import service as projects
 
 PARSABLE = {"tender", "addendum", "clarification"}
@@ -227,3 +229,27 @@ async def summaries(s: AsyncSession, project_ids: list[UUID]) -> dict[UUID, dict
     ).scalars():
         out[job.project_id]["job"] = {"kind": "parse_job", "id": job.id, "progress": job.progress, "label": "解析中"}
     return out
+
+
+async def text_layer(s: AsyncSession, ctx: TenantContext, doc: SourceDocument) -> list[dict[str, Any]]:
+    """文字层（逐页的句子和坐标框），加密存在对象存储里（DM1）。"""
+    if not doc.text_layer_key:
+        raise AppError("document_not_ready")
+    dk = await keys.data_key(s, ctx, doc.project_id)
+    return json.loads(open_object(dk, storage().get(doc.text_layer_key), doc.text_layer_key))
+
+
+async def human_citation(
+    s: AsyncSession, ctx: TenantContext, project_id: UUID, document_id: UUID, sentence_keys: list[str]
+) -> Citation:
+    """人工在阅读器里选中句子建出处（api-design §4.4）：句子必须在这份文档的同一页上。"""
+    doc = await get(s, document_id)
+    if doc.project_id != project_id:
+        raise AppError("not_found")
+    pages = [Page.from_json(p) for p in await text_layer(s, ctx, doc)]
+    by_key = {x.key: x for p in pages for x in p.sentences}
+    chosen = [by_key[k] for k in dict.fromkeys(sentence_keys) if k in by_key]
+    if len(chosen) != len(set(sentence_keys)) or len({x.page_no for x in chosen}) != 1:
+        raise AppError("sentence_not_found")
+    (c,) = await citations.from_sentences(s, ctx, doc, chosen, "human", [])
+    return c