fix(server): 按 S01–S10 评测修复读标抽取（F1–F8、口径 Q1–Q4）

- 合并时待确认条目不再借正式条目进清单，没通过的类别记在 pending_categories - 结构识别：章（按章跳过不相干的部分）、表后「上述……否则无效」说明、★ 节标题 - INVALID 补写法；输出用到上限 92% 按截断处理，拆到底时采用已有输出 - 格式范本正文、整表概括句不收；项目信息多看评标方法所在分块 - WPS 旧版 DOC 直接修正字体表字符集；表格按单元格切句（s3-v4） - 口径：★ 节继承实质性、★ 定义为计分项时降为重要、「必须/应」声明记 keyword 规则、政府采购资格条件按 87 号令第六十三条带无效类别 - 提示词 substantive v5，qualification / schedule / tech_param v4 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 23:19:30 +08:00
parent f99bb8b358
commit f98fd51128
17 changed files with 912 additions and 62 deletions
--- a/server/tests/test_units.py
+++ b/server/tests/test_units.py
@@ -402,3 +402,155 @@ def test_project_info_status() -> None:
    assert info.status["budget_total"]["codes"] == ["ambiguous"]
    assert info.status["code"]["codes"] == ["not_found_with_evidence"]
    assert info.status["sme_only"]["codes"] == ["missing_answer"]
+
+
+# ---------------------------------------------------------------- S01–S10 评测后的修正（extract-eval §3、§5）
+
+
+def lines(*texts: str) -> list[Sentence]:
+    return [sent(t, seq=i + 1) for i, t in enumerate(texts)]
+
+
+def test_chapters_skip_toc_and_references() -> None:
+    from bid_doctor.document.pipeline.structure import chapters
+
+    ss = lines(
+        "第一章招标公告",  # 目录：章标题挨着
+        "第二章投标人须知",
+        "投标邀请",
+        "第一章招标公告",
+        "公告正文。",
+        "资格要求见第四章《采购需求》。",  # 引用，不是章标题
+        "第二章投标人须知",
+        "须知正文。",
+        "第八章投标文件格式要求另行说明",  # 正文里偶尔提到后面的章：章号要严格递增，取最长的序列
+        "须知正文。",
+        "第三章采购需求",
+        "参数。",
+        "第四章投标文件格式",
+    )
+    ch = chapters(ss)
+    assert [ch.get(s.key) for s in ss] == [
+        None, None, None, "notice", "notice", "notice", "instructions", "instructions", "instructions",
+        "instructions", "requirements", "requirements", "formats",
+    ]  # fmt: skip
+
+
+def test_trailing_note_covers_rows_above() -> None:
+    from bid_doctor.document.pipeline.structure import trailing_notes
+
+    ss = lines(
+        "一、资格证明文件",
+        "营业执照或相关部门登记证明",
+        "财务状况报告",
+        "说明：",
+        "（1）上述资格证明文件1～2项必须全部按要求提供，否则投标文件无效。",
+    )
+    notes = trailing_notes(ss)
+    assert set(notes) == {ss[1].key, ss[2].key}  # 到「一、」为止；「说明：」引导句不算
+    d = doc(sentences=ss)
+    d.trailing = notes
+    assert check_basis("explicit_invalid", [ss[1]], [ss[1]], d)
+
+
+def test_star_section_heading() -> None:
+    from bid_doctor.checklist.extract import importance_of
+    from bid_doctor.document.pipeline.structure import star_sections
+
+    ss = lines(
+        "三、★商务要求", "1.交货期", "自合同签定之日起60工作日内安装调试完成。", "四、其他说明", "本项目不接受分包。"
+    )
+    scope = star_sections(ss)
+    assert set(scope) == {ss[1].key, ss[2].key}  # 到下一个同级编号「四、」为止
+    d = doc(sentences=ss)
+    d.star_scope = scope
+    assert importance_of([ss[2]], d) == ("substantive", "marker")
+    assert check_basis("star", [ss[2]], [ss[2]], d)
+    assert importance_of([ss[4]], d) == ("normal", None)
+
+
+def test_star_defined_as_scored_is_important() -> None:
+    """文件把 ★ 定义成「核心技术参数」、不满足只扣分（S02）：★ 是「重要」，不是实质性（口径 Q2）。"""
+    from bid_doctor.checklist.extract import build_doc, importance_of
+
+    star, define = sent("★设备采用国产自主可控CPU芯片。", seq=1), sent("备注：加★为核心技术参数；", seq=2)
+    scored = sent("每有一项核心技术参数不满足的减4分。", seq=3)
+    d = build_doc([star, define, scored])
+    assert is_marked(star, d.declared) is None
+    assert importance_of([star], d) == ("important", "marker")
+    assert [(n.markers, n.meaning) for n in d.notes] == [({"star"}, "important")]
+    # 没有计分证据：维持实质性（降级的代价是把致命显示成重要）
+    assert is_marked(star, build_doc([star, define]).declared) == "star"
+    # 有一处说 ★ 是实质性要求、或者带 ★ 的句子写着后果，就按实质性（顾问审读第 1 条的反例）
+    for extra in ("本文件中打“★”内容为实质性要求。", "打★号的参数负偏离的，视为未响应。"):
+        assert is_marked(star, build_doc([star, define, scored, sent(extra, seq=4)]).declared) == "star"
+
+
+def test_heuristics_reject_lookalikes() -> None:
+    """顾问审读第 2、5、6 条：目录隔着页码、参数行带 ★ 编号、否定写法的表后说明。"""
+    from bid_doctor.document.pipeline.structure import chapters, star_sections, trailing_notes
+
+    toc = lines("第一章招标公告", "3", "第二章投标人须知", "10", "第三章采购需求", "20")
+    assert chapters(toc) == {}  # 只有目录：认不出可靠的章，就不跳过
+    ss = lines("★5、采用2*15W功放芯片，声音清晰；", "符合参数要求的产品彩页复印件。", "6、输出功率≥30W。")
+    assert star_sections(ss) == {}  # 参数行不是节标题
+    ss = lines("三、商务要求（★为实质性要求）", "1.交货期30天。")
+    assert star_sections(ss) == {}  # ★ 不紧挨编号：是图例，不是节标题
+    ss = lines("一、参数表", "内存不低于32GB", "以上参数为非实质性要求，不满足不导致投标无效。")
+    assert trailing_notes(ss) == {}
+
+
+def test_pending_item_does_not_upgrade_formal_item() -> None:
+    """待确认的条目不能借同一句的正式条目进正式清单（评测 §3.1：S07 普通参数被标成实质性）。"""
+    s = sent("支持SM2、SM3、SM4国密算法。")
+    tech = Extracted(task="tech_param", categories=["tech_param"], summary="国密算法", sentences=[s], zone="formal")
+    sub = Extracted(
+        task="substantive", categories=["invalidity"], summary="国密算法", sentences=[s], zone="pending",
+        importance="substantive", failures=["basis_unmatched"],
+    )  # fmt: skip
+    (merged,) = merge_items([tech, sub])
+    assert (merged.categories, merged.importance, merged.zone) == (["tech_param"], "normal", "formal")
+
+
+def test_near_output_limit_counts_as_incomplete() -> None:
+    from bid_doctor.ai.providers.base import near_limit
+
+    assert near_limit(7757) and near_limit(8192)  # S03：7757 / 8192 自己收了尾
+    assert not near_limit(5000) and not near_limit(None)
+
+
+def test_wps_doc_font_charset() -> None:
+    from bid_doctor.document.pipeline.convert import fix_wps_doc_charset
+
+    name = "宋体".encode("utf-16-le") + b"\x00\x00"
+    ffn = bytes([40 + len(name) - 1, 0, 0x90, 0x01, 0x02, 0]) + bytes(34) + name  # chs = 02（SYMBOL）
+    fixed = fix_wps_doc_charset(b"\xd0\xcf" + ffn + b"\x00" * 8)
+    assert fixed[2 + 4] == 0x86
+    assert fix_wps_doc_charset(b"\xd0\xcf" + ffn.replace(b"\x02", b"\x00", 1)) == b"\xd0\xcf" + ffn.replace(
+        b"\x02", b"\x00", 1
+    )  # 字符集本来没错：不动
+
+
+def test_pending_category_kept_on_formal_item() -> None:
+    """待确认的「无效」类别不并进正式条目，但也不丢：记在正式条目上（顾问审读第 3 条）。"""
+    s = sent("25.3.1投标人未按要求提供资格证明文件的。")
+    qual = Extracted(
+        task="qualification", categories=["qualification"], summary="资格证明", sentences=[s], zone="formal"
+    )
+    sub = Extracted(
+        task="substantive", categories=["invalidity"], summary="资格证明", sentences=[s], zone="pending",
+        importance="substantive", failures=["quote_not_original"],
+    )  # fmt: skip
+    (merged,) = merge_items([qual, sub])
+    assert merged.categories == ["qualification"]
+    assert merged.ai_values["pending_categories"] == {"invalidity": ["quote_not_original"]}
+
+
+def test_template_and_blanket_patterns() -> None:
+    from bid_doctor.checklist.extract import BLANKET, TEMPLATE
+
+    assert TEMPLATE.match("二、我单位具有良好的商业信誉和健全的财务会计制度。")
+    assert not TEMPLATE.match("投标人须具有良好的商业信誉和健全的财务会计制度。")
+    assert BLANKET.search("3投标人《资格证明文件》有任何一项不符合《资格审查要求》的，其投标无效。")
+    assert BLANKET.search("（1）上述资格证明文件1～6项必须全部提供，否则投标无效。")
+    assert not BLANKET.search("投标文件未按上述要求签章的，投标无效。")  # 这是一条无效情形，不是概括句