@topmindspace/tms-skills 0.1.0 → 0.1.2

package/top-ppt-html/scripts/negative_tests.py

@@ -1,307 +1,433 @@
-#!/usr/bin/env python3
-# -*- coding: utf-8 -*-
-"""TopPPT HTML · 门禁反向验证（故障注入）
-
-校验器最容易退化成「只会说 OK 的摆设」。本脚本往合格产物里**注入已知缺陷**，
-断言对应门禁必须报错——门禁抓不到即视为失败。
-
-覆盖：
-  N1  Exhibit 框漏编号（旧实现只校验已编号者的连续性 → 静默通过）
-  N2  参考资料两位数编号错配（旧正则 `ref-\\d` 两侧同时落空 → 静默通过）
-  N3  参考资料编号跳号
-  N4  PPTX 演讲者备注被剔离（dataTable=notes 声称数据入备注，不核对就只是一句承诺）
-  N5  模型图表数值篡改（cross_verify 数值核对）
-  N6  行动标题判断词鉴别力（正则本身）
-  N7  PPTX 字号越出比例尺
-  N8  data-chart 未登记类型（登记表白名单）
-  N9  图表多样性塌陷（全篇压成同型 → 不同类型数 < 模式下限）
-  N10 图片外链（零外链铁律）
-  N11 待核实有标色无图例（.tbd 必须配 .tbd-legend/.flagbar）
-  N12 模型 theme 与 data-theme 矛盾
-  N13 模型 mode 与 data-mode 矛盾
-  N14 内部锚点断裂（链接目标 id 不存在）
-
-用法: python scripts/negative_tests.py
-"""
-from __future__ import annotations
-
-import json
-import re
-import subprocess
-import sys
-import zipfile
-from pathlib import Path
-
-try:
-    sys.stdout.reconfigure(encoding='utf-8', errors='replace')
-except Exception:
-    pass
-
-ROOT = Path(__file__).resolve().parent.parent
-OUT = ROOT / 'dist' / 'negtest'
-PY = sys.executable
-
-
-def run_validate(path: Path) -> tuple[int, str]:
-    r = subprocess.run([PY, str(ROOT / 'scripts' / 'validate_report.py'), str(path), '--strict'],
-                       capture_output=True, text=True, encoding='utf-8', errors='replace')
-    return r.returncode, (r.stdout or '') + (r.stderr or '')
-
-
-def expect_fail(name: str, path: Path, keyword: str) -> bool:
-    code, out = run_validate(path)
-    # --strict 下 WARN 同样计失败（退出码 1），故 [WARN] 行也算「门禁抓到」
-    hit = [ln for ln in out.splitlines()
-           if keyword in ln and ('[FAIL]' in ln or '[WARN]' in ln)]
-    ok = code != 0 and bool(hit)
-    print(f'  [{"OK" if ok else "FAIL"}] {name}')
-    if hit:
-        print(f'        抓到: {hit[0].strip()[:120]}')
-    elif not ok:
-        print(f'        门禁未触发（期望 FAIL 含「{keyword}」）')
-    return ok
-
-
-def main() -> int:
-    base = ROOT / 'assets' / 'examples' / '2026-09-09-research-mckinsey.html'
-    if not base.exists():
-        print(f'SKIP：基准示例缺失 {base}')
-        return 0
-    txt = base.read_text(encoding='utf-8')
-    OUT.mkdir(parents=True, exist_ok=True)
-
-    code, out = run_validate(base)
-    if code != 0:
-        print('[FAIL] 基准示例本身未通过 strict，反向验证失去意义')
-        return 1
-    print('基准示例 strict 通过，开始故障注入：')
-
-    cases: list[tuple[str, str, str]] = []
-
-    # N1 抹掉一个 Exhibit 编号（框仍在）
-    cases.append(('N1 Exhibit 框漏编号',
-                  txt.replace('class="exhibit__no"', 'class="exhibit__no--x"', 1),
-                  'Exhibit 框均已编号'))
-
-    # N2 参考资料改成两位数编号，与正文引用错配
-    cases.append(('N2 引用编号两位数错配',
-                  re.sub(r'id="ref-1"', 'id="ref-10"', txt, count=1),
-                  '双向对齐'))
-
-    # N3 编号跳号（两侧同时改，集合仍对齐，但 1..N 不连续）
-    t3 = txt.replace('ref-2"', 'ref-9"')
-    cases.append(('N3 参考资料编号跳号', t3, '编号连续'))
-
-    # N8 data-chart 用了登记表外的类型
-    cases.append(('N8 图表类型未登记',
-                  txt.replace('data-chart="hbar"', 'data-chart="hbarx"', 1),
-                  '登记表内'))
-
-    # N9 图表多样性塌陷（全部压成 bar → 不同类型数 1 < research 下限 6）
-    t9 = txt
-    for ct in ('hbar', 'donut', 'waterfall', 'sankey', 'treemap', 'boxplot',
-               'network', 'marimekko', 'streamgraph', 'gantt', 'rose', 'candlestick'):
-        t9 = t9.replace(f'data-chart="{ct}"', 'data-chart="bar"')
-    cases.append(('N9 图表多样性塌陷', t9, '图表多样性'))
-
-    # N10 外链图片（零外链铁律）
-    cases.append(('N10 图片外链',
-                  txt.replace('</body>',
-                              '<img src="https://example.com/x.png" alt="外链注入"></body>', 1),
-                  '图片源无外链'))
-
-    # N11 待核实有标色无图例（改名不得包含原串——校验是子串判断）
-    t11 = txt.replace('tbd-legend', 'tbd-leg').replace('flagbar', 'flag-bar')
-    cases.append(('N11 待核实无图例', t11, '待核实标注'))
-
-    # N12 模型 theme 与页面 data-theme 矛盾
-    cases.append(('N12 模型主题不一致',
-                  txt.replace('"theme": "light"', '"theme": "dark"', 1),
-                  'data-theme 一致'))
-
-    # N13 模型 mode 与页面 data-mode 矛盾
-    cases.append(('N13 模型模式不一致',
-                  txt.replace('"mode": "research"', '"mode": "presentation"', 1),
-                  'data-mode 一致'))
-
-    # N14 内部锚点断裂（链接目标 id 不存在）
-    cases.append(('N14 锚点断裂',
-                  txt.replace('href="#s1"', 'href="#s404"'),
-                  '锚点闭环'))
-
-    # N15 HTML 标签泄漏进可见文本（截图级：转义标签当字面量显示）
-    cases.append(('N15 标签泄漏',
-                  txt.replace('</p>',
-                              ' &lt;a class="cite" href="#ref-1"&gt;[1]&lt;/a&gt;。</p>', 1),
-                  'HTML_TAG_IN_TEXT'))
-
-    # N16 极偏 donut（0.5% vs 99.5% 应改 KPI，禁环图）
-    t16 = txt
-    if '"type": "donut"' in t16 or '"type":"donut"' in t16:
-        t16 = re.sub(
-            r'("type"\s*:\s*"donut"[\s\S]{0,200}?"values"\s*:\s*\[)[^\]]*(\])',
-            r'\g<1>0.5, 99.5\g<2>', t16, count=1)
-        cases.append(('N16 极偏 donut', t16, 'CHART_SKEW'))
-
-    # N17 模型字段夹带 HTML 标签（extract 净化 + validate 扫描 REPORT_MODEL）
-    t17 = re.sub(
-        r'("soWhat"\s*:\s*")',
-        r'\g<1><a class=\\"cite\\" href=\\"#ref-1\\">[1]</a> ',
-        txt, count=1)
-    if t17 == txt:
-        t17 = txt.replace(
-            '"footnote":',
-            '"footnote": "<strong>x</strong> ', 1)
-    cases.append(('N17 模型夹带标签', t17, 'HTML_TAG_IN_TEXT'))
-
-    ok = True
-    for name, content, keyword in cases:
-        if content == txt:
-            print(f'  [SKIP] {name}（基准示例无对应结构）')
-            continue
-        p = OUT / (name.split()[0] + '.html')
-        p.write_text(content, encoding='utf-8')
-        ok = expect_fail(name, p, keyword) and ok
-
-    ok = _pptx_notes_case() and ok
-    ok = _chart_data_case() and ok
-    ok = _title_pattern_case() and ok
-    ok = _font_scale_case() and ok
-
-    print('反向验证通过：注入的缺陷都被门禁抓住。' if ok
-          else '反向验证失败：存在抓不到的缺陷，门禁有假阴性。')
-    return 0 if ok else 1
-
-
-def _font_scale_case() -> bool:
-    """N7：往 PPTX 里塞一个比例尺外的字号 → FONT_SIZE_OFF_SCALE 必须报。"""
-    name = '2026-09-09-research-mckinsey'
-    src = ROOT / 'dist' / 'regression' / f'{name}.pptx'
-    model = ROOT / 'assets' / 'examples' / f'{name}.model.json'
-    if not src.exists() or not model.exists():
-        print('  [SKIP] N7 字号越出比例尺（需先跑 regression）')
-        return True
-    out = OUT / 'N7-off-scale.pptx'
-    patched = False
-    with zipfile.ZipFile(src) as zin, zipfile.ZipFile(out, 'w', zipfile.ZIP_DEFLATED) as zout:
-        for item in zin.infolist():
-            data = zin.read(item.filename)
-            if not patched and item.filename == 'ppt/slides/slide4.xml':
-                text = data.decode('utf-8')
-                new = re.sub(r'sz="\d+"', 'sz="1234"', text, count=1)  # 12.34pt：任何档位都没有
-                if new != text:
-                    data = new.encode('utf-8')
-                    patched = True
-            zout.writestr(item, data)
-    if not patched:
-        print('  [SKIP] N7 字号越出比例尺（未找到可改写的 sz 属性）')
-        return True
-    report = OUT / 'N7.json'
-    subprocess.run([PY, str(ROOT / 'scripts' / 'validate_pptx.py'), str(out),
-                    '--strict', '--model=' + str(model), '--json-out', str(report)],
-                   capture_output=True, text=True, encoding='utf-8', errors='replace')
-    try:
-        data = json.loads(report.read_text(encoding='utf-8'))
-    except (OSError, json.JSONDecodeError):
-        print('  [FAIL] N7 字号越出比例尺：校验报告不可解析')
-        return False
-    codes = {i.get('code') for i in (data.get('errors') or []) + (data.get('warnings') or [])}
-    ok = 'FONT_SIZE_OFF_SCALE' in codes
-    print(f'  [{"OK" if ok else "FAIL"}] N7 字号越出比例尺')
-    if not ok:
-        print(f'        门禁未触发（期望 FONT_SIZE_OFF_SCALE，实际 {sorted(codes)[:6]}）')
-    return ok
-
-
-def _title_pattern_case() -> bool:
-    """N6：主题词式长标题必须被判为「无判断信号」，结论句标题必须放行。"""
-    lc = json.loads((ROOT / 'scripts' / 'layout-constants.json').read_text(encoding='utf-8'))
-    pat = ((lc.get('contentQuality') or {}).get('researchTitle') or {}).get('judgmentPattern')
-    if not pat:
-        print('  [SKIP] N6 行动标题判断词（未配置 judgmentPattern）')
-        return True
-    topic = ['智能体平台的应用架构与技术选型说明', '关于数据治理需求的整体情况介绍',
-             '企业知识库建设的相关背景与范围', '平台能力地图与模块组成一览表',
-             '项目实施过程中的应急预案汇总', '面向未来的技术路线图规划说明']
-    claim = ['投入产出比在第 3 季度首次转正', '三成团队已把智能体纳入日常流程',
-             '数据质量是当前最大的落地瓶颈', '推理成本一年内下降 62%',
-             '从试点到规模化：关键卡在治理', '自建方案的总成本高于采购方案']
-    leaked = [t for t in topic if re.search(pat, t)]
-    hurt = [t for t in claim if not re.search(pat, t)]
-    ok = not leaked and not hurt
-    print(f'  [{"OK" if ok else "FAIL"}] N6 行动标题判断词鉴别力'
-          f'（漏放主题词 {len(leaked)}/{len(topic)} · 误伤结论句 {len(hurt)}/{len(claim)}）')
-    for t in (leaked + hurt)[:3]:
-        print(f'        {t}')
-    return ok
-
-
-def _chart_data_case() -> bool:
-    """N5：篡改模型里的图表数值 → cross_verify 的数值核对必须报不符。"""
-    name = '2026-09-09-research-mckinsey'
-    pptx = ROOT / 'dist' / 'regression' / f'{name}.pptx'
-    mp = ROOT / 'assets' / 'examples' / f'{name}.model.json'
-    if not pptx.exists() or not mp.exists():
-        print('  [SKIP] N5 图表数值篡改（需先跑 regression）')
-        return True
-    sys.path.insert(0, str(ROOT / 'scripts'))
-    try:
-        from pptx import Presentation           # noqa: PLC0415
-        import cross_verify as CV               # noqa: PLC0415
-    except ImportError:
-        print('  [SKIP] N5 图表数值篡改（python-pptx 未安装）')
-        return True
-    model = json.loads(mp.read_text(encoding='utf-8'))
-    touched = False
-    for sec in (model.get('sections') or []):
-        c = CV._model_chart(sec) if isinstance(sec, dict) else None
-        if c and isinstance(c.get('values'), list) and c['values']:
-            c['values'] = [(v + 777) if isinstance(v, (int, float)) else v for v in c['values']]
-            touched = True
-            break
-    if not touched:
-        print('  [SKIP] N5 图表数值篡改（基准模型无可改图表）')
-        return True
-    issues = CV.chart_data_verify(Presentation(str(pptx)), model, name)
-    ok = bool(issues)
-    print(f'  [{"OK" if ok else "FAIL"}] N5 模型图表数值被篡改')
-    if ok:
-        print(f'        抓到: {issues[0][:120]}')
-    else:
-        print('        门禁未触发（数值核对形同虚设）')
-    return ok
-
-
-def _pptx_notes_case() -> bool:
-    """N4：剥离演讲者备注 → dataTable=notes 的图表数据不可追溯，必须被 strict 抓住。"""
-    name = '2026-09-09-research-mckinsey'
-    src = ROOT / 'dist' / 'regression' / f'{name}.pptx'
-    model = ROOT / 'assets' / 'examples' / f'{name}.model.json'
-    if not src.exists() or not model.exists():
-        print('  [SKIP] N4 备注剥离（需先跑 regression 生成 PPTX）')
-        return True
-    stripped = OUT / 'N4-no-notes.pptx'
-    with zipfile.ZipFile(src) as zin, zipfile.ZipFile(stripped, 'w', zipfile.ZIP_DEFLATED) as zout:
-        for item in zin.infolist():
-            if item.filename.startswith('ppt/notesSlides/'):
-                continue
-            zout.writestr(item, zin.read(item.filename))
-    report = OUT / 'N4.json'
-    subprocess.run([PY, str(ROOT / 'scripts' / 'validate_pptx.py'), str(stripped),
-                    '--strict', '--model=' + str(model), '--json-out', str(report)],
-                   capture_output=True, text=True, encoding='utf-8', errors='replace')
-    try:
-        data = json.loads(report.read_text(encoding='utf-8'))
-    except (OSError, json.JSONDecodeError):
-        print('  [FAIL] N4 备注剥离：校验报告不可解析')
-        return False
-    codes = {i.get('code') for i in (data.get('errors') or []) + (data.get('warnings') or [])}
-    ok = 'MODEL_CHART_NOTES_MISSING' in codes
-    print(f'  [{"OK" if ok else "FAIL"}] N4 演讲者备注被剥离')
-    if not ok:
-        print(f'        门禁未触发（期望 MODEL_CHART_NOTES_MISSING，实际 {sorted(codes)[:6]}）')
-    return ok
-
-
-if __name__ == '__main__':
-    sys.exit(main())
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""TopPPT HTML · 门禁反向验证（故障注入）
+
+校验器最容易退化成「只会说 OK 的摆设」。本脚本往合格产物里**注入已知缺陷**，
+断言对应门禁必须报错——门禁抓不到即视为失败。
+
+覆盖：
+  N1  Exhibit 框漏编号（旧实现只校验已编号者的连续性 → 静默通过）
+  N2  参考资料两位数编号错配（旧正则 `ref-\\d` 两侧同时落空 → 静默通过）
+  N3  参考资料编号跳号
+  N4  PPTX 演讲者备注被剔离（dataTable=notes 声称数据入备注，不核对就只是一句承诺）
+  N5  模型图表数值篡改（cross_verify 数值核对）
+  N6  行动标题判断词鉴别力（正则本身）
+  N7  PPTX 字号越出比例尺
+  N8  data-chart 未登记类型（登记表白名单）
+  N9  图表多样性塌陷（全篇压成同型 → 不同类型数 < 模式下限）
+  N10 图片外链（零外链铁律）
+  N11 待核实有标色无图例（.tbd 必须配 .tbd-legend/.flagbar）
+  N12 模型 theme 与 data-theme 矛盾
+  N13 模型 mode 与 data-mode 矛盾
+  N14 内部锚点断裂（链接目标 id 不存在）
+  L4  截断迹象（列表项省略号砍义）
+  L5  溢出未拆页
+  L6  混排缺对齐（ALIGN_RHYTHM）
+
+用法: python scripts/negative_tests.py
+"""
+from __future__ import annotations
+
+import json
+import re
+import subprocess
+import sys
+import zipfile
+from pathlib import Path
+
+try:
+    sys.stdout.reconfigure(encoding='utf-8', errors='replace')
+except Exception:
+    pass
+
+ROOT = Path(__file__).resolve().parent.parent
+OUT = ROOT / 'dist' / 'negtest'
+PY = sys.executable
+
+
+def run_validate(path: Path) -> tuple[int, str]:
+    r = subprocess.run([PY, str(ROOT / 'scripts' / 'validate_report.py'), str(path), '--strict'],
+                       capture_output=True, text=True, encoding='utf-8', errors='replace')
+    return r.returncode, (r.stdout or '') + (r.stderr or '')
+
+
+def expect_fail(name: str, path: Path, keyword: str) -> bool:
+    code, out = run_validate(path)
+    # --strict 下 WARN 同样计失败（退出码 1），故 [WARN] 行也算「门禁抓到」
+    hit = [ln for ln in out.splitlines()
+           if keyword in ln and ('[FAIL]' in ln or '[WARN]' in ln)]
+    ok = code != 0 and bool(hit)
+    print(f'  [{"OK" if ok else "FAIL"}] {name}')
+    if hit:
+        print(f'        抓到: {hit[0].strip()[:120]}')
+    elif not ok:
+        print(f'        门禁未触发（期望 FAIL 含「{keyword}」）')
+    return ok
+
+
+
+def expect_layout_qa_fail(name: str, path: Path, keyword: str) -> bool:
+    """Run validate_report --strict --layout-qa and require FAIL containing keyword."""
+    r = subprocess.run(
+        [PY, str(ROOT / 'scripts' / 'validate_report.py'), str(path),
+         '--strict', '--layout-qa'],
+        capture_output=True, text=True, encoding='utf-8', errors='replace')
+    out = (r.stdout or '') + (r.stderr or '')
+    hit = [ln for ln in out.splitlines()
+           if keyword in ln and ('[FAIL]' in ln or 'FAIL' in ln)]
+    ok = r.returncode != 0 and bool(hit)
+    print(f'  [{"OK" if ok else "FAIL"}] {name}')
+    if hit:
+        print(f'        抓到: {hit[0].strip()[:120]}')
+    elif not ok:
+        print(f'        layout-qa 未触发（期望含「{keyword}」）')
+    return ok
+
+
+def main() -> int:
+    base = ROOT / 'assets' / 'examples' / '2026-09-09-research-mckinsey.html'
+    if not base.exists():
+        print(f'SKIP：基准示例缺失 {base}')
+        return 0
+    txt = base.read_text(encoding='utf-8')
+    OUT.mkdir(parents=True, exist_ok=True)
+
+    code, out = run_validate(base)
+    if code != 0:
+        print('[FAIL] 基准示例本身未通过 strict，反向验证失去意义')
+        return 1
+    print('基准示例 strict 通过，开始故障注入：')
+
+    cases: list[tuple[str, str, str]] = []
+
+    # N1 抹掉一个 Exhibit 编号（框仍在）
+    cases.append(('N1 Exhibit 框漏编号',
+                  txt.replace('class="exhibit__no"', 'class="exhibit__no--x"', 1),
+                  'Exhibit 框均已编号'))
+
+    # N2 参考资料改成两位数编号，与正文引用错配
+    cases.append(('N2 引用编号两位数错配',
+                  re.sub(r'id="ref-1"', 'id="ref-10"', txt, count=1),
+                  '双向对齐'))
+
+    # N3 编号跳号（两侧同时改，集合仍对齐，但 1..N 不连续）
+    t3 = txt.replace('ref-2"', 'ref-9"')
+    cases.append(('N3 参考资料编号跳号', t3, '编号连续'))
+
+    # N8 data-chart 用了登记表外的类型
+    cases.append(('N8 图表类型未登记',
+                  txt.replace('data-chart="hbar"', 'data-chart="hbarx"', 1),
+                  '登记表内'))
+
+    # N9 图表多样性塌陷（全部压成 bar → 不同类型数 1 < research 下限 6）
+    t9 = txt
+    for ct in ('hbar', 'donut', 'waterfall', 'sankey', 'treemap', 'boxplot',
+               'network', 'marimekko', 'streamgraph', 'gantt', 'rose', 'candlestick'):
+        t9 = t9.replace(f'data-chart="{ct}"', 'data-chart="bar"')
+    cases.append(('N9 图表多样性塌陷', t9, '图表多样性'))
+
+    # N10 外链图片（零外链铁律）
+    cases.append(('N10 图片外链',
+                  txt.replace('</body>',
+                              '<img src="https://example.com/x.png" alt="外链注入"></body>', 1),
+                  '图片源无外链'))
+
+    # N11 待核实有标色无图例（改名不得包含原串——校验是子串判断）
+    t11 = txt.replace('tbd-legend', 'tbd-leg').replace('flagbar', 'flag-bar')
+    cases.append(('N11 待核实无图例', t11, '待核实标注'))
+
+    # N12 模型 theme 与页面 data-theme 矛盾
+    cases.append(('N12 模型主题不一致',
+                  txt.replace('"theme": "light"', '"theme": "dark"', 1),
+                  'data-theme 一致'))
+
+    # N13 模型 mode 与页面 data-mode 矛盾
+    cases.append(('N13 模型模式不一致',
+                  txt.replace('"mode": "research"', '"mode": "presentation"', 1),
+                  'data-mode 一致'))
+
+    # N14 内部锚点断裂（链接目标 id 不存在）
+    cases.append(('N14 锚点断裂',
+                  txt.replace('href="#s1"', 'href="#s404"'),
+                  '锚点闭环'))
+
+    # N15 HTML 标签泄漏进可见文本（截图级：转义标签当字面量显示）
+    cases.append(('N15 标签泄漏',
+                  txt.replace('</p>',
+                              ' &lt;a class="cite" href="#ref-1"&gt;[1]&lt;/a&gt;。</p>', 1),
+                  'HTML_TAG_IN_TEXT'))
+
+    # N16 极偏 donut（0.5% vs 99.5% 应改 KPI，禁环图）
+    t16 = txt
+    if '"type": "donut"' in t16 or '"type":"donut"' in t16:
+        t16 = re.sub(
+            r'("type"\s*:\s*"donut"[\s\S]{0,200}?"values"\s*:\s*\[)[^\]]*(\])',
+            r'\g<1>0.5, 99.5\g<2>', t16, count=1)
+        cases.append(('N16 极偏 donut', t16, 'CHART_SKEW'))
+
+    # N17 模型字段夹带 HTML 标签（extract 净化 + validate 扫描 REPORT_MODEL）
+    t17 = re.sub(
+        r'("soWhat"\s*:\s*")',
+        r'\g<1><a class=\\"cite\\" href=\\"#ref-1\\">[1]</a> ',
+        txt, count=1)
+    if t17 == txt:
+        t17 = txt.replace(
+            '"footnote":',
+            '"footnote": "<strong>x</strong> ', 1)
+    cases.append(('N17 模型夹带标签', t17, 'HTML_TAG_IN_TEXT'))
+
+    ok = True
+    for name, content, keyword in cases:
+        if content == txt:
+            print(f'  [SKIP] {name}（基准示例无对应结构）')
+            continue
+        p = OUT / (name.split()[0] + '.html')
+        p.write_text(content, encoding='utf-8')
+        ok = expect_fail(name, p, keyword) and ok
+
+
+    # ── Batch 2 · layout-qa 专用负例 ─────────────────────────────────────
+    # L1 管线产物缺 data-skel：先造带 skel 的页，再撕掉其中一页的标记
+    import tempfile
+    scaf = OUT / 'L1-scaffold.html'
+    subprocess.run(
+        [PY, str(ROOT / 'scripts' / 'scaffold_report.py'),
+         '--mode', 'presentation', '--style', 'business-blue', '--theme', 'light',
+         '--title', 'LayoutQA Neg', '--sections', '5', '--out', str(scaf)],
+        capture_output=True, text=True)
+    if scaf.exists() and 'data-skel="' in scaf.read_text(encoding='utf-8'):
+        t_l1 = scaf.read_text(encoding='utf-8')
+        # 撕掉第一个内容页的 data-skel
+        t_l1b = re.sub(r'(id="s1") data-skel="P\d+"', r'\1', t_l1, count=1)
+        p_l1 = OUT / 'L1-missing-skel.html'
+        p_l1.write_text(t_l1b, encoding='utf-8')
+        ok = expect_layout_qa_fail('L1 缺 data-skel', p_l1, 'LAYOUT_QA_MISSING_SKEL') and ok
+        # L2 连续 3 页同一 skel
+        t_l2 = t_l1
+        # force s1,s2,s3 all P4
+        t_l2 = re.sub(r'(id="s1"[^>]*data-skel=")P\d+"', r'\1P4"', t_l2, count=1)
+        t_l2 = re.sub(r'(id="s2"[^>]*data-skel=")P\d+"', r'\1P4"', t_l2, count=1)
+        t_l2 = re.sub(r'(id="s3"[^>]*data-skel=")P\d+"', r'\1P4"', t_l2, count=1)
+        # also rewrite if attribute order is data-skel before id
+        t_l2 = re.sub(r'(data-skel=")P\d+(" data-page-type="[^"]+" id="s1")', r'\1P4\2', t_l2)
+        t_l2 = re.sub(r'(data-skel=")P\d+("([^>]*) id="s1")', r'\1P4\2', t_l2)
+        # simpler: replace first three data-skel values
+        def _force_p4(src, n=3):
+            out, c = [], 0
+            for part in re.split(r'(data-skel="P\d+")', src):
+                if part.startswith('data-skel="') and c < n:
+                    out.append('data-skel="P4"'); c += 1
+                else:
+                    out.append(part)
+            return ''.join(out)
+        t_l2 = _force_p4(t_l1, 3)
+        p_l2 = OUT / 'L2-skel-streak.html'
+        p_l2.write_text(t_l2, encoding='utf-8')
+        ok = expect_layout_qa_fail('L2 连续同 skel', p_l2, 'LAYOUT_QA_SKEL_STREAK') and ok
+        # L3 极偏 donut in REPORT_MODEL
+        t_l3 = t_l1
+        if 'window.REPORT_MODEL' in t_l3:
+            # inject a skewed donut section into model JSON
+            inj = (
+                '{"type":"donut","title":"极偏环",'
+                '"chart":{"type":"donut","labels":["A","B"],"values":[0.5,99.5]}},'
+            )
+            t_l3 = re.sub(r'("sections"\s*:\s*\[)', r'\1' + inj, t_l3, count=1)
+            p_l3 = OUT / 'L3-skew-donut.html'
+            p_l3.write_text(t_l3, encoding='utf-8')
+            ok = expect_layout_qa_fail('L3 极偏 donut', p_l3, 'LAYOUT_QA_SKEW_DONUT') and ok
+
+        # L4 截断迹象：列表项以省略号砍义
+        t_l4 = t_l1
+        inj4 = '<div class="card"><ul><li class="card__li">这项证据其实很长但被故意截断了…</li></ul></div>'
+        t_l4 = re.sub(r'(id="s1"[^>]*>)', lambda m: m.group(1) + inj4, t_l4, count=1)
+        p_l4 = OUT / 'L4-truncation.html'
+        p_l4.write_text(t_l4, encoding='utf-8')
+        ok = expect_layout_qa_fail('L4 截断迹象', p_l4, 'LAYOUT_QA_TRUNCATION') and ok
+
+        # L5 溢出未拆页：塞入超长正文使页高估算爆掉且无续页信号
+        t_l5 = t_l1
+        wall = '论证要点' + ('详细证据与口径说明，必须完整保留不得删减。' * 40)
+        t_l5 = re.sub(
+            r'(id="s1"[^>]*>)',
+            lambda m: m.group(1) + '<div class="card"><p>' + wall + '</p></div>'
+                 + '<div class="card"><p>' + wall + '</p></div>'
+                 + '<div class="card"><p>' + wall + '</p></div>',
+            t_l5, count=1)
+        p_l5 = OUT / 'L5-overflow-nosplit.html'
+        p_l5.write_text(t_l5, encoding='utf-8')
+        ok = expect_layout_qa_fail('L5 溢出未拆页', p_l5, 'LAYOUT_QA_OVERFLOW_NO_SPLIT') and ok
+
+        # L6 混排缺对齐：图+卡同页但无对齐类
+        t_l6 = t_l1
+        inj6 = (
+            '<div class="grid g-2">'
+            '<div class="fig"><svg class="chart" data-chart="bar" style="height:200px"></svg></div>'
+            '<div class="card"><ul><li>证据甲</li><li>证据乙</li></ul></div>'
+            '</div>'
+        )
+        t_l6 = re.sub(r'(id="s1"[^>]*>)', lambda m: m.group(1) + inj6, t_l6, count=1)
+        # 撕掉 s1 段内对齐类，确保 ALIGN_RHYTHM 能抓住
+        def _strip_align_s1(src: str) -> str:
+            m = re.search(r'(<section[^>]*id="s1"[^>]*>)([\s\S]*?)(</section>)', src)
+            if not m:
+                # attribute order may be data-skel before id
+                m = re.search(r'(<section[^>]*id="s1"[^>]*>)([\s\S]*?)(</section>)', src)
+            if not m:
+                return src
+            body = m.group(2)
+            body = re.sub(r'\ba-start\b', '', body)
+            body = re.sub(r'\ba-c\b', '', body)
+            body = re.sub(r'\ba-end\b', '', body)
+            body = body.replace('align-items:start', '').replace('align-items: center', '')
+            return src[:m.start()] + m.group(1) + body + m.group(3) + src[m.end():]
+        t_l6 = _strip_align_s1(t_l6)
+        p_l6 = OUT / 'L6-align-rhythm.html'
+        p_l6.write_text(t_l6, encoding='utf-8')
+        ok = expect_layout_qa_fail('L6 混排缺对齐', p_l6, 'LAYOUT_QA_ALIGN_RHYTHM') and ok
+
+    else:
+        print('  [SKIP] layout-qa 负例（scaffold 未产出 data-skel）')
+
+        ok = _pptx_notes_case() and ok
+    ok = _chart_data_case() and ok
+    ok = _title_pattern_case() and ok
+    ok = _font_scale_case() and ok
+
+    print('反向验证通过：注入的缺陷都被门禁抓住。' if ok
+          else '反向验证失败：存在抓不到的缺陷，门禁有假阴性。')
+    return 0 if ok else 1
+
+
+def _font_scale_case() -> bool:
+    """N7：往 PPTX 里塞一个比例尺外的字号 → FONT_SIZE_OFF_SCALE 必须报。"""
+    name = '2026-09-09-research-mckinsey'
+    src = ROOT / 'dist' / 'regression' / f'{name}.pptx'
+    model = ROOT / 'assets' / 'examples' / f'{name}.model.json'
+    if not src.exists() or not model.exists():
+        print('  [SKIP] N7 字号越出比例尺（需先跑 regression）')
+        return True
+    out = OUT / 'N7-off-scale.pptx'
+    patched = False
+    with zipfile.ZipFile(src) as zin, zipfile.ZipFile(out, 'w', zipfile.ZIP_DEFLATED) as zout:
+        for item in zin.infolist():
+            data = zin.read(item.filename)
+            if not patched and item.filename == 'ppt/slides/slide4.xml':
+                text = data.decode('utf-8')
+                new = re.sub(r'sz="\d+"', 'sz="1234"', text, count=1)  # 12.34pt：任何档位都没有
+                if new != text:
+                    data = new.encode('utf-8')
+                    patched = True
+            zout.writestr(item, data)
+    if not patched:
+        print('  [SKIP] N7 字号越出比例尺（未找到可改写的 sz 属性）')
+        return True
+    report = OUT / 'N7.json'
+    subprocess.run([PY, str(ROOT / 'scripts' / 'validate_pptx.py'), str(out),
+                    '--strict', '--model=' + str(model), '--json-out', str(report)],
+                   capture_output=True, text=True, encoding='utf-8', errors='replace')
+    try:
+        data = json.loads(report.read_text(encoding='utf-8'))
+    except (OSError, json.JSONDecodeError):
+        print('  [FAIL] N7 字号越出比例尺：校验报告不可解析')
+        return False
+    codes = {i.get('code') for i in (data.get('errors') or []) + (data.get('warnings') or [])}
+    ok = 'FONT_SIZE_OFF_SCALE' in codes
+    print(f'  [{"OK" if ok else "FAIL"}] N7 字号越出比例尺')
+    if not ok:
+        print(f'        门禁未触发（期望 FONT_SIZE_OFF_SCALE，实际 {sorted(codes)[:6]}）')
+    return ok
+
+
+def _title_pattern_case() -> bool:
+    """N6：主题词式长标题必须被判为「无判断信号」，结论句标题必须放行。"""
+    lc = json.loads((ROOT / 'scripts' / 'layout-constants.json').read_text(encoding='utf-8'))
+    pat = ((lc.get('contentQuality') or {}).get('researchTitle') or {}).get('judgmentPattern')
+    if not pat:
+        print('  [SKIP] N6 行动标题判断词（未配置 judgmentPattern）')
+        return True
+    topic = ['智能体平台的应用架构与技术选型说明', '关于数据治理需求的整体情况介绍',
+             '企业知识库建设的相关背景与范围', '平台能力地图与模块组成一览表',
+             '项目实施过程中的应急预案汇总', '面向未来的技术路线图规划说明']
+    claim = ['投入产出比在第 3 季度首次转正', '三成团队已把智能体纳入日常流程',
+             '数据质量是当前最大的落地瓶颈', '推理成本一年内下降 62%',
+             '从试点到规模化：关键卡在治理', '自建方案的总成本高于采购方案']
+    leaked = [t for t in topic if re.search(pat, t)]
+    hurt = [t for t in claim if not re.search(pat, t)]
+    ok = not leaked and not hurt
+    print(f'  [{"OK" if ok else "FAIL"}] N6 行动标题判断词鉴别力'
+          f'（漏放主题词 {len(leaked)}/{len(topic)} · 误伤结论句 {len(hurt)}/{len(claim)}）')
+    for t in (leaked + hurt)[:3]:
+        print(f'        {t}')
+    return ok
+
+
+def _chart_data_case() -> bool:
+    """N5：篡改模型里的图表数值 → cross_verify 的数值核对必须报不符。"""
+    name = '2026-09-09-research-mckinsey'
+    pptx = ROOT / 'dist' / 'regression' / f'{name}.pptx'
+    mp = ROOT / 'assets' / 'examples' / f'{name}.model.json'
+    if not pptx.exists() or not mp.exists():
+        print('  [SKIP] N5 图表数值篡改（需先跑 regression）')
+        return True
+    sys.path.insert(0, str(ROOT / 'scripts'))
+    try:
+        from pptx import Presentation           # noqa: PLC0415
+        import cross_verify as CV               # noqa: PLC0415
+    except ImportError:
+        print('  [SKIP] N5 图表数值篡改（python-pptx 未安装）')
+        return True
+    model = json.loads(mp.read_text(encoding='utf-8'))
+    touched = False
+    for sec in (model.get('sections') or []):
+        c = CV._model_chart(sec) if isinstance(sec, dict) else None
+        if c and isinstance(c.get('values'), list) and c['values']:
+            c['values'] = [(v + 777) if isinstance(v, (int, float)) else v for v in c['values']]
+            touched = True
+            break
+    if not touched:
+        print('  [SKIP] N5 图表数值篡改（基准模型无可改图表）')
+        return True
+    issues = CV.chart_data_verify(Presentation(str(pptx)), model, name)
+    ok = bool(issues)
+    print(f'  [{"OK" if ok else "FAIL"}] N5 模型图表数值被篡改')
+    if ok:
+        print(f'        抓到: {issues[0][:120]}')
+    else:
+        print('        门禁未触发（数值核对形同虚设）')
+    return ok
+
+
+def _pptx_notes_case() -> bool:
+    """N4：剥离演讲者备注 → dataTable=notes 的图表数据不可追溯，必须被 strict 抓住。"""
+    name = '2026-09-09-research-mckinsey'
+    src = ROOT / 'dist' / 'regression' / f'{name}.pptx'
+    model = ROOT / 'assets' / 'examples' / f'{name}.model.json'
+    if not src.exists() or not model.exists():
+        print('  [SKIP] N4 备注剥离（需先跑 regression 生成 PPTX）')
+        return True
+    stripped = OUT / 'N4-no-notes.pptx'
+    with zipfile.ZipFile(src) as zin, zipfile.ZipFile(stripped, 'w', zipfile.ZIP_DEFLATED) as zout:
+        for item in zin.infolist():
+            if item.filename.startswith('ppt/notesSlides/'):
+                continue
+            zout.writestr(item, zin.read(item.filename))
+    report = OUT / 'N4.json'
+    subprocess.run([PY, str(ROOT / 'scripts' / 'validate_pptx.py'), str(stripped),
+                    '--strict', '--model=' + str(model), '--json-out', str(report)],
+                   capture_output=True, text=True, encoding='utf-8', errors='replace')
+    try:
+        data = json.loads(report.read_text(encoding='utf-8'))
+    except (OSError, json.JSONDecodeError):
+        print('  [FAIL] N4 备注剥离：校验报告不可解析')
+        return False
+    codes = {i.get('code') for i in (data.get('errors') or []) + (data.get('warnings') or [])}
+    ok = 'MODEL_CHART_NOTES_MISSING' in codes
+    print(f'  [{"OK" if ok else "FAIL"}] N4 演讲者备注被剥离')
+    if not ok:
+        print(f'        门禁未触发（期望 MODEL_CHART_NOTES_MISSING，实际 {sorted(codes)[:6]}）')
+    return ok
+
+
+if __name__ == '__main__':
+    sys.exit(main())