fix(check): 修国动真实投标文件暴露的 6 处（规则集 r5）

国动真实投标文件查标一次（check-eval §14）：44 个问题里有用的 9 个，致命、重要两级都是误报。按决策 E3 全部修（§15）： - 响应错位：「完全满足，无偏离。」这类复合肯定写法整体算肯定 - 附加条件：「不以……另行收费」是承诺 - 扫描件：文字层 v5 记每页图片占比，带一行说明的扫描页也认；提示写明扫描页范围 - 不需要这家投标人回答的：无特定资格要求、总括性资格要求、中标后才做的事、投标文件没声称适用的联合体和价格扣除 - 「方案应明确 A、B、C……」逐项再检索 - 同一批送出的片段都可引用（引文照样核对）；出处没过校验的不沿用国动重新检查 44 → 11（全是提示，有用的 9 个都在）；开发集召回 112 → 111 （模型重跑波动），致命级 25/32 不变，规矩的投标文件问题 239 → 213。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:16:12 +08:00
parent f5edc585f1
commit 808259da87
16 changed files with 373 additions and 21 deletions
--- a/server/tests/test_check_units.py
+++ b/server/tests/test_check_units.py
@@ -4,6 +4,8 @@ from __future__ import annotations

 from uuid import uuid4

+import pytest
+
 from bid_doctor.check import findings, match
 from bid_doctor.check.retrieval import Index, tokens
 from bid_doctor.document.pipeline.text_layer import Sentence
@@ -158,15 +160,16 @@ def test_scanned_fallback_turns_missing_evidence_into_a_hint() -> None:
        b: {"categories": ["tech_param"], "title": "内存"},
    }
    decisions = {a: applicability.Decision(True, evidence=True), b: applicability.Decision(True)}
-    docs = [match.BidDoc(uuid4(), 1, "资格证明册", [], [12, 13])]
+    docs = [match.BidDoc(uuid4(), 1, "资格证明册", [], [12, 13, 14, 20])]
    match.scanned_fallback(results, items, decisions, docs)
    spec = findings.spec(results[a], items[a])
    assert (
        results[a].verdict == "scanned"
        and spec is not None
        and spec.severity == "hint"
-        and "2 页扫描件" in spec.explanation
+        and "资格证明册第 12–14、20 页 是扫描件（共 4 页）" in spec.explanation
    )
+    assert match.page_ranges([5, 3, 4, 9]) == "3–5、9"
    assert results[b].verdict == "not_found"  # 技术参数不是证明材料，照常报未找到


@@ -240,6 +243,11 @@ def test_mismatch_read_only_when_response_cell_is_unrelated() -> None:
        ("支持端口汇聚、端口流量统计", "满足，支持端口汇聚和端口流量统计"),
        ("批量数据处理在非工作时段内完成", "满足，详见技术方案第12页"),
        ("系统支持国产操作系统", "满足"),
+        # 几个肯定写法连着写（国动真实投标文件，check-eval §14）：仍是肯定回答
+        ("支持按区域、距离和服务类型筛选并排序", "完全满足，无偏离。"),
+        ("后台支持批量导入并提示错误行", "符合，无负偏离；"),
+        ("支持离线缓存最近浏览的十条记录", "完全满足，无偏离；页面截图详见附件三"),
+        ("短信验证码有效期不超过五分钟", "完全满足，显示器升降行程：150mm"),
    ]
    sents, results, items = [], {}, {}
    for n, (req, resp) in enumerate(rows):
@@ -254,9 +262,18 @@ def test_mismatch_read_only_when_response_cell_is_unrelated() -> None:
        results[k] = (a, b)
    d = match.BidDoc(uuid4(), 1, "技术分册", sents)
    res = {k: match.Result("responded", "ai", "verified", evidence=[(d, a)]) for k, (a, _) in results.items()}
-    assert deviation.mismatch_read(res, items, [d]) == 1
+    assert deviation.mismatch_read(res, items, [d]) == 2
    verdicts = [r.verdict for r in res.values()]
-    assert verdicts == ["not_found", "responded", "responded", "responded"]
+    assert verdicts == [
+        "not_found",
+        "responded",
+        "responded",
+        "responded",
+        "responded",
+        "responded",
+        "responded",
+        "not_found",
+    ]
    first = next(iter(res.values()))
    assert first.detail["mismatch"] == "显示器升降行程：150mm" and first.reason_code == "response_mismatch"

@@ -346,3 +363,111 @@ def test_direct_read_uses_header_column_and_flags_unclear_values() -> None:
    assert results[b].verdict == "responded"
    spec = findings.spec(results[a], {"title": "支持双电源", "importance": "substantive", "categories": ["tech_param"]})
    assert spec is not None and spec.severity == "important" and spec.title.startswith("偏离栏需核对")
+
+
+def test_scanned_page_with_caption_counts_as_empty() -> None:
+    """扫描件常带一行文字说明（「合同扫描件第 2 页（共 15 页）」）：按图片占比和字数一起判断（check-eval §14）。"""
+    from bid_doctor.document.pipeline.text_layer import Page
+
+    def page(chars: int, cover: float) -> str:
+        return Page(1, 595, 842, 0, chars, [], cover).text_source
+
+    assert page(0, 0.0) == "empty"
+    assert page(26, 0.28) == "empty"  # 「2.1 营业执照」+ 执照扫描图
+    assert page(97, 0.53) == "empty"  # 一段说明 + 整页合同扫描
+    assert page(30, 0.0) == "text"  # 只有一行字的正文页（没有图）
+    assert page(495, 0.25) == "text"  # 带架构图的正文页
+    assert page(179, 0.45) == "text"  # 图多、字也不少
+    assert (
+        Page.from_json(
+            {"page_no": 1, "width_pt": 1, "height_pt": 1, "rotation": 0, "char_count": 30, "sentences": []}
+        ).text_source
+        == "text"
+    )  # v4 文字层没有图片占比
+
+
+def test_applicability_skips_conditions_the_bidder_does_not_claim() -> None:
+    """不需要这家投标人回答的条目（check-eval §14）：写明没有的、总括性的、中标后才做的、
+    投标文件没声称适用的条件性要求（联合体、价格扣除）。声称了（投标文件里有对应文件）就照常查。"""
+    from bid_doctor.check import applicability
+
+    def it(title: str, cats: list[str], req: str = "") -> dict[str, object]:
+        return {"title": title, "categories": cats, "requirement_text": req or title}
+
+    items = [
+        it("本项目第2包无特定资格要求", ["qualification"]),
+        it("提交特定资格要求的证明材料", ["qualification"], "特定资格要求的证明材料（如有，必须提供）"),
+        it("投标人应满足招标文件规定的资格条件", ["qualification"]),
+        it("中标后十日内提交详细实施计划", ["tech_param"]),
+        it("联合体投标的，须提交联合体协议书", ["qualification"]),
+        it("残疾人福利性单位价格扣除", ["scoring"], "提供《残疾人福利性单位声明函》，给予价格扣除。"),
+        it("提供营业执照副本复印件", ["qualification"]),
+    ]
+    ctx = applicability.context(items, ["营业执照副本复印件", "我公司不属于残疾人福利性单位。"])
+    got = [applicability.decide(i, {}, ctx).skip for i in items]
+    assert got == ["none_required", "none_required", "umbrella", "after_award", "not_claimed", None, None]
+    # 没有「无特定资格要求」时特定资格证明照查；不知道投标文件写了什么（ctx 为空）时条件性要求照查
+    assert applicability.decide(items[1], {}, applicability.Context(frozenset())).skip is None
+    assert applicability.decide(items[4], {}).skip is None
+    consortium = applicability.context(items, ["联合体协议书（附后）"])
+    assert applicability.decide(items[4], {}, consortium).skip is None
+
+
+async def test_citation_from_another_item_in_the_same_batch_is_verified(monkeypatch: pytest.MonkeyPatch) -> None:
+    """片段每批只送一次，模型引用了同一批别的条目的候选句：引文核对得上就算（check-eval §14）；
+    没送过的编号照样是 sid_unknown。"""
+    from bid_doctor.core.db import TenantContext
+
+    net = sent(1, "系统部署于互联网区与电子政务外网区，两区之间只经前置交换。", page=7)
+    mem = sent(2, "服务器内存配置为64GB，满足不低于32GB的要求。", page=9)
+    d = match.BidDoc(uuid4(), 1, "投标文件", [net, mem])
+    a, b = uuid4(), uuid4()
+    items = {
+        a: {"title": "系统部署在电子政务外网区", "categories": ["tech_param"]},
+        b: {"title": "服务器内存不低于32GB", "categories": ["tech_param"]},
+    }
+    # 前提：每条只取 1 个候选句时，b 的候选句里没有 net（net 只是 a 的候选句）
+    assert [h.sentence for h in Index([net, mem]).search(match.query_text(items[b]), 1)] == [mem]
+
+    async def fake(ctx: object, **kw: object) -> dict[str, object]:
+        ev = {"sid": "d1p7s1", "quote": "系统部署于互联网区与电子政务外网区"}
+        return {
+            "results": [
+                {"item": "I1", "verdict": "responded", "evidence": [ev]},
+                {"item": "I2", "verdict": "responded", "evidence": [ev]},  # 引的是 a 的候选句
+            ]
+        }
+
+    monkeypatch.setattr(match, "CANDIDATES", 1)
+    monkeypatch.setattr(match.ai, "run_json", fake)
+
+    async def progress(_: float) -> None:
+        return None
+
+    ctx = TenantContext(tenant_id=uuid4(), user_id=uuid4())
+    results, _ = await match.run(ctx, uuid4(), items, [d], progress)
+    assert results[a].verdict == "responded" and results[b].verdict == "responded"
+    assert results[b].evidence[0][1] is net
+
+    unknown = match.judge({"verdict": "responded", "evidence": [{"sid": "d1p8s1", "quote": "x" * 10}]}, {}, [])
+    assert unknown.reason_code == "sid_unknown"
+
+
+def test_outline_requirement_is_searched_item_by_item() -> None:
+    """「方案应明确 A、B、C……」：列出的每一项再各检索一次（check-eval §14）。"""
+    item = {"requirement_text": "运维方案应包括巡检频次、备份策略、故障上报流程以及应急演练安排，并确保措施可落地。"}
+    assert match.outline_terms(item) == ["巡检频次", "备份策略", "故障上报流程", "应急演练安排"]
+    assert match.outline_terms({"requirement_text": "方案应明确工期。"}) == []
+    assert match.outline_terms({"requirement_text": "支持巡检、备份和应急演练。"}) == []
+
+
+def test_reuse_skips_conclusions_whose_citation_failed() -> None:
+    """出处没过校验的结论不沿用（复查时重新问模型）；正常的结论按证据位置找回来。"""
+    from bid_doctor.check import reuse
+
+    d, allowed = bid()
+    (_, s) = allowed["d1p1s9"]
+    ok = match.Result("responded", "model", "verified", evidence=[(d, s)])
+    assert reuse.restore(reuse.snapshot(ok), [d], "r1") is not None
+    failed = match.judge({"verdict": "responded", "evidence": [{"sid": "d1p9s9", "quote": "x" * 10}]}, allowed, [])
+    assert reuse.restore(reuse.snapshot(failed), [d], "r1") is None