feat(check): 需人工核对（P2）和分项 / 单价上限（C7）

- 需人工核对：份数、密封、递交、CA 签章加密、解密、保证金这类文件里看不出来的要求，
  从正式清单按主题合并、带招标原文出处；manual-checks 接口按项目存勾选、写勾选人和时间，
  复查沿用；查标页加「需人工核对」一块，覆盖范围写明列了几条
- 「报价」类清单项：读标从采购需求的清单表规则读出分项 / 单价上限（checklist/price.py）；
  查标交给规则 PRICE.OVER_CEILING_ITEM 比对分项报价表，不交给模型；规则集 2026.11-r2
- 版面工具（分页、分行、归栏）挪到 document/pipeline/layout.py，读标和查标共用
- 评测：price_defects.py 从规矩投标文件派生只带报价问题的 4 份（P01、P02、P03）

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

server/src/bid_doctor/document/pipeline/layout.py

@@ -0,0 +1,80 @@
+"""版面工具：按坐标把文字层的句子分成页、行、栏（读标认表格、查标认投标文件的表格都用）。
+
+有框线的表格按单元格的上沿分行（文字层给句子带了单元格框）：同一行的格子上沿相同，垂直居中的短格、
+折成几句的长格、纵向合并的格（只算进它的第一行）都不会串行；其余按句子上下范围重叠分行。
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+from bid_doctor.document.pipeline.text_layer import Sentence
+
+
+def box(s: Sentence) -> tuple[float, float, float, float]:
+    return (
+        min(q[0] for q in s.quads),
+        min(q[1] for q in s.quads),
+        max(q[2] for q in s.quads),
+        max(q[3] for q in s.quads),
+    )
+
+
+def span(s: Sentence) -> tuple[float, float]:
+    """左右范围：在表格里取单元格的，否则取文字的。"""
+    return (s.cell[0], s.cell[2]) if s.cell else (box(s)[0], box(s)[2])
+
+
+def pages(sentences: list[Sentence]) -> dict[int, list[Sentence]]:
+    out: dict[int, list[Sentence]] = {}
+    for s in sentences:
+        out.setdefault(s.page_no, []).append(s)
+    return out
+
+
+@dataclass(slots=True)
+class _Row:
+    top: float
+    bottom: float
+    in_cell: bool
+    items: list[Sentence]
+
+
+def rows(page: list[Sentence]) -> list[list[Sentence]]:
+    """一页的句子分成行，行内从左到右。"""
+    out: list[_Row] = []
+
+    def top_of(s: Sentence) -> float:
+        return s.cell[1] if s.cell else box(s)[1]
+
+    for s in sorted((x for x in page if x.quads), key=top_of):
+        if s.cell:
+            r = next((r for r in out if r.in_cell and abs(r.top - s.cell[1]) < 2), None)
+            if r:
+                r.items.append(s)
+            else:
+                out.append(_Row(s.cell[1], s.cell[1], True, [s]))
+            continue
+        _, top, _, bottom = box(s)
+        last = out[-1] if out else None
+        if last and not last.in_cell and top < last.bottom - 1:
+            last.bottom = max(last.bottom, bottom)
+            last.items.append(s)
+        else:
+            out.append(_Row(top, bottom, False, [s]))
+    return [sorted(r.items, key=lambda x: box(x)[0]) for r in sorted(out, key=lambda r: r.top)]
+
+
+def column(cell: Sentence, cols: dict[str, tuple[float, float]]) -> str | None:
+    """单元格落在哪一栏：和表头单元格左右范围重叠最多的那一栏；都不重叠（窄数字居中、表头字靠左）取中心最近的。
+    cols 要包含表头的每一格，否则别的栏的数字会被归到最近的已知栏。"""
+    x0, x1 = span(cell)
+    best, score = None, 0.0
+    for name, (c0, c1) in cols.items():
+        overlap = min(x1, c1) - max(x0, c0)
+        if overlap > score:
+            best, score = name, overlap
+    if best is None and cols:
+        mid = (x0 + x1) / 2
+        best = min(cols, key=lambda n: abs((cols[n][0] + cols[n][1]) / 2 - mid))
+    return best