feat(check): 候选句带同一行的单元格；提示词 check.match v2

检索常命中投标人照抄要求原文的那一格，响应和偏离写在同一行的别的格里，模型看不到。现在候选句在表格里
时，把同一行（同一页、上下范围重叠）的单元格一起交给模型，也可以引用；v2 说明怎么读一行。
开发集 17 份上查标召回 30% → 50%，带标志词的问题 2% → 54%，规矩投标文件上的误报略降，输入 token +44%；
漏响应和响应对象错位变差，下一轮看（docs/architecture/check-eval.md §3）。

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

server/src/bid_doctor/check/match.py

@@ -25,8 +25,9 @@ from bid_doctor.check.retrieval import Index
 from bid_doctor.core.db import TenantContext
 from bid_doctor.document.pipeline.text_layer import Sentence, norm
 
-PROMPT = ("check.match", 1)
+PROMPT = ("check.match", 2)
 CANDIDATES = 5  # 每条清单项送给模型的候选句数
+ROW_MATES = 6  # 候选句在表格里时，带上同一行的单元格（检索常命中投标人照抄要求的那一格，响应和偏离在同一行的别的格）
 BATCH = 8  # 每次调用判断的清单项数
 SPLIT_DEPTH = 2
 CONCURRENCY = 4  # TODO(第 3 步)：和解析共用「供应商 × 租户」统一限流（extract.py 的同一个 TODO）
@@ -99,15 +100,37 @@ def _item_block(key: str, item: dict[str, Any], sids: list[str]) -> str:
     return "\n".join(lines)
 
 
+def row_mates(s: Sentence, page: list[Sentence]) -> list[Sentence]:
+    """同一行的其他句子：同一页、上下范围重叠（和读标核对取值时找「同一行」是同一口径）。按从左到右排。"""
+    if not s.quads:
+        return []
+    top, bottom = min(q[1] for q in s.quads), max(q[3] for q in s.quads)
+    mates = [x for x in page if x is not s and x.quads and any(q[1] < bottom - 1 and q[3] > top + 1 for q in x.quads)]
+    return sorted(mates, key=lambda x: min(q[0] for q in x.quads))[:ROW_MATES]
+
+
 def build_user(
-    batch: list[tuple[str, dict[str, Any], list[str]]], fragments: dict[str, tuple[BidDoc, Sentence]]
+    batch: list[tuple[str, dict[str, Any], list[str]]],
+    fragments: dict[str, tuple[BidDoc, Sentence]],
+    rows: dict[str, list[str]] | None = None,
 ) -> str:
-    """数据区：待核对的要求 + 投标文件片段（每句只送一次）。系统提示词声明数据区里的指令一律不执行。"""
-    used = list(dict.fromkeys(sid for _, _, sids in batch for sid in sids))
+    """数据区：待核对的要求 + 投标文件片段（每句只送一次；表格里的句子后面跟着同一行的单元格）。
+    系统提示词声明数据区里的指令一律不执行。"""
+    rows = rows or {}
+    shown: set[str] = set()
+    frag_lines: list[str] = []
+    for _, _, sids in batch:
+        for sid in sids:
+            if sid in shown:
+                continue
+            shown.add(sid)
+            doc, sent = fragments[sid]
+            frag_lines.append(f"[{sid}]（{doc.label} 第{sent.page_no}页）{sent.text}")
+            mates = [m for m in rows.get(sid, []) if m not in shown]
+            if mates:
+                shown.update(mates)
+                frag_lines.append("　同一行：" + " ｜ ".join(f"[{m}] {fragments[m][1].text}" for m in mates))
     blocks = [_item_block(k, it, sids) for k, it, sids in batch]
-    frag_lines = [
-        f"[{sid}]（{fragments[sid][0].label} 第{fragments[sid][1].page_no}页）{fragments[sid][1].text}" for sid in used
-    ]
     return "## 待核对的要求\n\n" + "\n\n".join(blocks) + "\n\n## 投标文件片段\n\n" + "\n".join(frag_lines)
 
 
@@ -172,12 +195,30 @@ async def run(
     doc_of = {s.key: d for d in docs for s in d.sentences}
     results: dict[UUID, Result] = {}
     queue: list[tuple[UUID, dict[str, Any], dict[str, tuple[BidDoc, Sentence]]]] = []
+    pages: dict[tuple[UUID, int], list[Sentence]] = {}
+    for d in docs:
+        for s in d.sentences:
+            pages.setdefault((d.id, s.page_no), []).append(s)
+    rows: dict[str, list[str]] = {}  # 候选句 → 同一行的单元格（短编号）
+    sent_of = {
+        short_sid(d, x): (d, x) for d in docs for x in d.sentences
+    }  # 片段按编号查（同一行的单元格可能是别的条目的候选句）
+    hits_of: dict[UUID, list[str]] = {}  # 每条清单项自己的候选句（同一行的单元格跟在后面显示，不单列）
     for item_id, item in items.items():
         hits = index.search(query_text(item), CANDIDATES)
         if not hits:  # 检索不到候选句：不调模型（m3-plan §2.1）
             results[item_id] = Result("not_found", "retrieval", detail={"search_scope": scope})
             continue
-        allowed = {short_sid(doc_of[h.sentence.key], h.sentence): (doc_of[h.sentence.key], h.sentence) for h in hits}
+        allowed: dict[str, tuple[BidDoc, Sentence]] = {}
+        for h in hits:
+            doc = doc_of[h.sentence.key]
+            sid = short_sid(doc, h.sentence)
+            allowed[sid] = (doc, h.sentence)
+            hits_of.setdefault(item_id, []).append(sid)
+            mates = row_mates(h.sentence, pages[(doc.id, h.sentence.page_no)])
+            rows[sid] = [short_sid(doc, m) for m in mates]
+            for m in mates:  # 同一行的单元格也可以引用（响应、偏离常写在那里）
+                allowed.setdefault(short_sid(doc, m), (doc, m))
         queue.append((item_id, item, allowed))
 
     sem = asyncio.Semaphore(CONCURRENCY)
@@ -185,8 +226,12 @@ async def run(
 
     async def call(batch: list[tuple[UUID, dict[str, Any], dict[str, tuple[BidDoc, Sentence]]]], depth: int) -> None:
         keyed = [(f"I{i + 1}", item_id, item, allowed) for i, (item_id, item, allowed) in enumerate(batch)]
-        fragments = {sid: x for _, _, _, allowed in keyed for sid, x in allowed.items()}
-        user = build_user([(k, item, list(allowed)) for k, _, item, allowed in keyed], fragments)
+
+        user = build_user(
+            [(k, item, hits_of[item_id]) for k, item_id, item, _ in keyed],
+            sent_of,
+            {h: rows.get(h, []) for _, item_id, _, _ in keyed for h in hits_of[item_id]},
+        )
         try:
             async with sem:
                 stats.calls += 1