@topmindspace/tms-skills 0.1.0

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

@@ -0,0 +1,301 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""TopPPT HTML · 交付质量门禁（HTML/PPTX + Evals 确定性检查 + rubric 启发式）
+
+用法:
+    python scripts/quality_gate.py report.html
+    python scripts/quality_gate.py report.html --pptx report.pptx --model report.model.json
+    python scripts/quality_gate.py report.html --trace trace.json --json
+    python scripts/quality_gate.py report.html --require-rubric 60
+    python scripts/quality_gate.py report.html --deliver          # 末尾生成可粘贴的交付说明
+
+定位:
+    validate_* 回答「结构是否合格」；本脚本再合并：
+    ① evals/run_evals.py 的确定性/效率检查
+    ② 无 LLM 的 rubric 启发式（content/layout/chart/infographic/tone 五维）
+    ③ --deliver：按 SKILL.md「交付说明」七要素（路径/字节数/模式/风格/篇幅/格式/校验+引用）
+       生成结构化交付说明——七要素缺一即 gate FAIL，杜绝手拼遗漏
+    交付时一键跑完，避免「只过了结构、内容质量靠自觉」。
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import subprocess
+import sys
+import time
+from pathlib import Path
+
+try:
+    sys.stdout.reconfigure(encoding='utf-8', errors='replace')
+except Exception:
+    pass
+
+sys.path.insert(0, str(Path(__file__).resolve().parent))
+import checks_html  # noqa: E402  承载/多样性/组合版式单源
+
+ROOT = Path(__file__).resolve().parent.parent
+LC = json.loads((ROOT / 'scripts' / 'layout-constants.json').read_text(encoding='utf-8'))
+RUBRIC_PATH = ROOT / 'evals' / 'rubric.schema.json'
+
+
+def run(cmd: list[str]) -> tuple[int, str, float]:
+    """带计时的子进程执行（elapsed 秒用于交付耗时基线，v8.3 起）。"""
+    t0 = time.perf_counter()
+    r = subprocess.run(cmd, capture_output=True, text=True, encoding='utf-8')
+    out = (r.stdout or '') + (r.stderr or '')
+    return r.returncode, out, time.perf_counter() - t0
+
+
+def kb(n: int) -> str:
+    return f'{n / 1024:.1f}KB'
+
+
+def build_delivery(html: Path, pptx: Path | None, gates: list[dict], rubric: dict) -> dict:
+    """SKILL.md「交付说明」七要素：路径 / 字节数 / 模式 / 风格 / 篇幅 / 格式 / 校验+引用。"""
+    txt = html.read_text(encoding='utf-8')
+    head = txt[:3000]
+
+    def attr(name: str) -> str:
+        m = re.search(rf'{name}="([^"]+)"', head)
+        return m.group(1) if m else ''
+
+    all_bands = bands(txt)
+    content = checks_html.content_bands(txt)  # 与 composite_pages 同一口径（防两处维护）
+    refs = sorted(set(re.findall(r'id="(ref-\d+)"', txt)), key=lambda s: int(s.split('-')[1]))
+    note = {
+        'file': str(html),
+        'bytes': html.stat().st_size,
+        'mode': attr('data-mode'),
+        'style': attr('data-style'),
+        'theme': attr('data-theme'),
+        'pages': len(all_bands),
+        'content_pages': len(content),
+        'format': 'HTML' + (' + PPTX' if pptx else ''),
+        'refs': len(refs),
+        'validation': {g['name']: ('通过' if g['ok'] else '未通过') for g in gates},
+        'rubric': rubric['score'],
+    }
+    if pptx and pptx.exists():
+        note['pptx'] = {'file': str(pptx), 'bytes': pptx.stat().st_size}
+    return note
+
+
+def render_delivery(note: dict) -> str:
+    lines = [
+        f"- 文件：{note['file']}（{kb(note['bytes'])}）"
+        + (f"；{note['pptx']['file']}（{kb(note['pptx']['bytes'])}）" if note.get('pptx') else ''),
+        f"- 模式/风格/主题：{note['mode']} · {note['style']} · {note['theme']}",
+        f"- 篇幅：{note['content_pages']} 内容页（全篇 {note['pages']} 屏）",
+        f"- 交付格式：{note['format']}",
+        '- 校验：' + ' · '.join(f"{k} {v}" for k, v in note['validation'].items())
+        + f" · rubric {note['rubric']}",
+        f"- 引用：{note['refs']} 条（参考资料双向对齐）",
+    ]
+    return '\n'.join(lines)
+
+
+def delivery_complete(note: dict) -> tuple[bool, str]:
+    """七要素齐备性：字段缺失/为空串即不完整（数值 0 是合法值，如无外部引用的 refs=0）。"""
+    missing = [k for k in ('file', 'bytes', 'mode', 'style', 'theme', 'pages', 'format', 'refs')
+               if note.get(k) in (None, '')]
+    return (not missing), ('七要素齐备' if not missing else f'缺 {missing}')
+
+
+def bands(txt: str) -> list[str]:
+    return re.split(r'(?=<section class="band)', txt)[1:]
+
+
+def heuristic_rubric(html: Path) -> dict:
+    """五维启发式评分（0–100）。与 evals/rubric.schema.json 字段对齐，便于 LLM 复评时同构。"""
+    txt = html.read_text(encoding='utf-8')
+    mode_m = re.search(r'data-mode="([^"]+)"', txt)
+    mode = mode_m.group(1) if mode_m else 'presentation'
+    content_bands = checks_html.content_bands(txt)  # 与 composite_pages/delivery 同一口径
+
+    checks = []
+
+    def add(cid: str, score: int, ok: bool, notes: str) -> None:
+        checks.append({'id': cid, 'score': max(0, min(100, int(score))),
+                       'pass': bool(ok), 'notes': notes})
+
+    # ── content ──
+    n_sec = len(content_bands)
+    n_sowhat = txt.count('class="sowhat')
+    n_points = len(re.findall(r'<li\b', txt))
+    n_refs = len(set(re.findall(r'id="(ref-\d+)"', txt)))
+    n_tbd = len(re.findall(r'class="tbd(?:\s|")', txt))
+    c_score = 50
+    c_notes = [f'{n_sec} 内容页']
+    if n_sec >= 6:
+        c_score += 15
+    if n_sowhat >= max(1, n_sec // 4):
+        c_score += 15
+        c_notes.append(f'so-what {n_sowhat}')
+    else:
+        c_notes.append(f'so-what 偏少({n_sowhat})')
+    if n_points >= n_sec * 2:
+        c_score += 10
+    if n_refs:
+        c_score += 10
+        c_notes.append(f'引用 {n_refs}')
+    else:
+        c_notes.append('无外部引用')
+    if n_tbd and 'tbd-legend' not in txt and 'flagbar' not in txt:
+        c_score -= 10
+        c_notes.append('tbd 缺图例')
+    add('content', c_score, c_score >= 60, '；'.join(c_notes))
+
+    # ── layout ──
+    v = checks_html.chart_variety(LC)
+    multi, n_content = checks_html.composite_pages(txt)
+    ratio = (multi / n_content) if n_content else 0
+    l_score = 50 + int(ratio * 40)
+    if 'g-side' in txt or 'g-211' in txt or 'g-hero' in txt:
+        l_score += 10
+    add('layout', l_score, l_score >= 60,
+        f'组合版式 {multi}/{n_content}（{ratio:.0%}）')
+
+    # ── chart ──
+    used = checks_html.distinct_chart_types(txt)
+    n_chart_pages = len([1 for b in content_bands if re.search(r'data-chart="', b)])
+    ch_score = 40
+    ch_notes = [f'{len(used)} 型 / {n_chart_pages} 图表页']
+    if used:
+        floor = checks_html.variety_floor(mode, n_chart_pages, v)
+        if len(used) >= floor:
+            ch_score += 40
+        else:
+            ch_score += int(40 * len(used) / max(1, floor))
+        ch_notes.append(f'下限 {floor}')
+        if n_chart_pages >= 2:
+            ch_score += 10
+    else:
+        ch_notes.append('无图表')
+    add('chart', ch_score, ch_score >= 60, '；'.join(ch_notes))
+
+    # ── infographic ──
+    info = set(((LC.get('charts') or {}).get('scaffold') or {}).get('infoTypes') or {})
+    has_info = bool(set(used) & info)
+    has_struct = has_info or ('class="arch' in txt) or ('class="lane' in txt)
+    if mode == 'architecture':
+        i_score = 80 if has_struct else 30
+        i_notes = '架构模式结构主图：' + ('有' if has_struct else '缺')
+    else:
+        i_score = 70 if has_info else 55
+        i_notes = '信息图页型：' + ('有' if has_info else '未用（非必须）')
+    add('infographic', i_score, i_score >= 50, i_notes)
+
+    # ── tone ──
+    words = (LC.get('aiFlavor') or {}).get('words') or []
+    plain = re.sub(r'<[^>]+>', ' ', re.sub(r'<script\b[\s\S]*?</script>', ' ', txt))
+    hits = [w for w in words if re.search(w, plain)]
+    t_score = 100 if not hits else max(0, 100 - 15 * len(hits))
+    # 标题过短/空洞 so-what 轻微扣分
+    hollow = ((LC.get('contentQuality') or {}).get('sowhat') or {}).get('forbidden') or []
+    if any(h in txt for h in hollow):
+        t_score -= 15
+    add('tone', t_score, t_score >= 60,
+        'AI 腔命中 ' + (', '.join(hits[:3]) if hits else '无'))
+
+    total = round(sum(c['score'] for c in checks) / len(checks))
+    overall = all(c['pass'] for c in checks)
+    return {
+        'overall_pass': overall,
+        'score': total,
+        'checks': checks,
+        '$source': 'heuristic (quality_gate.py) · 可用 LLM 按 rubric.schema.json 复评',
+    }
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description='TopPPT HTML 交付质量门禁')
+    ap.add_argument('html', help='报告 HTML 路径')
+    ap.add_argument('--pptx', help='可选：PPTX 路径（跑 validate_pptx --strict）')
+    ap.add_argument('--model', help='可选：model.json（PPTX 往返保真）')
+    ap.add_argument('--trace', help='可选：过程 trace JSON（并入 evals）')
+    ap.add_argument('--require-rubric', type=int, default=60,
+                    help='启发式 rubric 总分下限（默认 60；传 0 表示只报告不卡死）')
+    ap.add_argument('--deliver', action='store_true',
+                    help='生成 SKILL.md「交付说明」七要素（缺一即 gate FAIL）')
+    ap.add_argument('--json', action='store_true')
+    args = ap.parse_args()
+
+    html = Path(args.html)
+    if not html.exists():
+        print(f'错误：HTML 不存在 {html}')
+        return 2
+
+    result: dict = {'html': str(html), 'gates': []}
+
+    def gate(name: str, ok: bool, note: str = '', elapsed: float = 0.0) -> None:
+        entry = {'name': name, 'ok': bool(ok), 'note': note}
+        if elapsed:
+            entry['elapsed_s'] = round(elapsed, 2)
+        result['gates'].append(entry)
+        mark = '✓' if ok else '✗'
+        tail = f' · {elapsed:.1f}s' if elapsed else ''
+        print(f'  {mark} {name}' + (f' — {note}' if note else '') + tail)
+
+    print(f'交付质量门禁 · {html.name}')
+    print('-' * 60)
+
+    # ① HTML strict
+    rc, out, el = run([sys.executable, str(ROOT / 'scripts' / 'validate_report.py'), str(html), '--strict'])
+    tail = [ln for ln in out.splitlines() if ln.startswith('PASS ') or ln.startswith('结论')]
+    gate('validate_report --strict', rc == 0, tail[-1] if tail else f'exit {rc}', el)
+
+    # ② PPTX strict（可选）
+    if args.pptx:
+        cmd = [sys.executable, str(ROOT / 'scripts' / 'validate_pptx.py'), args.pptx, '--strict']
+        if args.model:
+            cmd += ['--model=' + args.model]
+        rc, out, el = run(cmd)
+        tail = [ln for ln in out.splitlines() if ln.startswith('PASS ') or ln.startswith('结论')]
+        gate('validate_pptx --strict', rc == 0, tail[-1] if tail else f'exit {rc}', el)
+
+    # ③ evals 确定性
+    cmd = [sys.executable, str(ROOT / 'evals' / 'run_evals.py'), '--score', str(html)]
+    if args.trace:
+        cmd += ['--trace', args.trace]
+    rc, out, el = run(cmd)
+    gate('evals --score', rc == 0, '确定性 + 效率检查', el)
+
+    # ④ rubric 启发式
+    rubric = heuristic_rubric(html)
+    result['rubric'] = rubric
+    print(f'  · rubric 启发式总分 {rubric["score"]} · '
+          f'{"通过" if rubric["overall_pass"] else "未通过"}')
+    for c in rubric['checks']:
+        mark = '✓' if c['pass'] else '✗'
+        print(f'    {mark} [{c["id"]}] {c["score"]:>3}  {c["notes"][:70]}')
+    if args.require_rubric and args.require_rubric > 0:
+        gate(f'rubric ≥ {args.require_rubric}', rubric['score'] >= args.require_rubric,
+             f'{rubric["score"]}')
+    else:
+        print('  · rubric 仅报告（--require-rubric=0）')
+
+    # ⑤ 交付说明（SKILL.md 七要素 · 缺一即 FAIL）
+    if args.deliver:
+        note = build_delivery(html, Path(args.pptx) if args.pptx else None,
+                              result['gates'], rubric)
+        complete, why = delivery_complete(note)
+        gate('交付说明七要素', complete, why)
+        result['delivery'] = note
+        print('\n── 交付说明（可直接粘贴）──────────────────────')
+        print(render_delivery(note))
+        print('──────────────────────────────────────────────')
+
+    ok = all(g['ok'] for g in result['gates'])
+    result['pass'] = ok
+    print('-' * 60)
+    print('结论：' + ('可交付' if ok else '不可交付（修复后重跑）'))
+
+    if args.json:
+        print(json.dumps(result, ensure_ascii=False, indent=2))
+    return 0 if ok else 1
+
+
+if __name__ == '__main__':
+    sys.exit(main())

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

@@ -0,0 +1,308 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""TopPPT HTML · 全链路回归（维护工具，与机器环境无关）
+
+链路：
+  ① validate_report.py --strict          每份示例 HTML（三模式预算 + 模型一致性 + 版式断言）
+  ② extract_model.py → build_pptx.js → validate_pptx.py --strict --model   B 通道（交付通道）
+  ③ gen_channel_a.js（页面预览同引擎产物）→ validate_pptx.py --strict --model --allow-shape-charts
+  ④ probe_image_export.py（素材图片探针：示例矩阵零图片，真实素材路径/版式/门禁靠它覆盖）
+  ⑤ 双区组合页探针（split 左区 = 图表）：scaffold → extract → B 通道 + A 通道，两通道都过 strict
+  ⑥ cross_verify.py（python-pptx 第三方裁判，可选：未安装则 SKIP）——覆盖示例矩阵与 ⑤ 探针
+  ⑦ negative_tests.py（故障注入：往合格产物里埋已知缺陷，断言门禁必须报错）
+  ⑦-b 任务路由全量遍历（extract_snippet TASK_ROUTES 每条路由可实际抽取，防修复链路静默断裂）
+  ⑧ 深色主题探针（research / presentation × dark：示例矩阵只有 architecture 是 dark，
+     stylesDark / styleDataColorsDark 路径靠它覆盖）
+  ⑨ measure_height.py（浏览器真值：页高与容器裁切，可选：无 playwright 则 SKIP）
+
+用法：
+    python scripts/regression.py [--only <示例名关键字>]
+
+环境解析（全部自动探测，不绑定任何机器的固定路径）：
+    Node 可执行文件：TOP_PPT_NODE_EXE > PATH 上的 node > 常见托管目录最新版
+    pptxgenjs 所在 node_modules：TOP_PPT_NODE_PATH > 仓库内 node_modules > 逐级父目录 > 常见托管目录
+    python-pptx：未安装时跳过交叉裁判（打印 SKIP，不计失败）
+
+模板/常量/schema/运行时/示例任何改动后必跑。
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import subprocess
+import sys
+from pathlib import Path
+
+try:
+    sys.stdout.reconfigure(encoding='utf-8', errors='replace')
+    sys.stderr.reconfigure(encoding='utf-8', errors='replace')
+except Exception:
+    pass
+
+sys.path.insert(0, str(Path(__file__).resolve().parent))
+from env_probe import (NODE_HINT, NODE_MODULES_HINT, find_node,  # noqa: E402
+                       find_node_modules, has_python_pptx, node_env)
+
+ROOT = Path(__file__).resolve().parent.parent
+EX = ROOT / 'assets' / 'examples'
+B_OUT = ROOT / 'dist' / 'regression'
+A_OUT = ROOT / 'dist' / 'regression-a'
+PY = sys.executable
+
+
+# ── 执行 ─────────────────────────────────────────────────────────────────────
+
+def run(cmd: list[str], env: dict[str, str] | None = None) -> subprocess.CompletedProcess:
+    return subprocess.run(cmd, capture_output=True, text=True, encoding='utf-8', env=env)
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description='TopPPT HTML 全链路回归')
+    ap.add_argument('--only', help='只跑名字含该关键字的示例')
+    args = ap.parse_args()
+
+    node = find_node()
+    node_path = find_node_modules()
+    if not node:
+        print('错误: ' + NODE_HINT)
+        return 2
+    if not node_path:
+        print('错误: ' + NODE_MODULES_HINT)
+        return 2
+    env = node_env(node_path)
+    print(f'Node: {node}')
+    print(f'NODE_PATH: {node_path}')
+    print(f'python-pptx 第三方裁判: {"可用" if has_python_pptx() else "未安装（交叉裁判将跳过）"}')
+
+    B_OUT.mkdir(parents=True, exist_ok=True)
+    fails: list[str] = []
+    skipped: list[str] = []
+
+    examples = sorted(EX.glob('*.html'))
+    if args.only:
+        examples = [h for h in examples if args.only in h.stem]
+    if not examples:
+        print('错误: assets/examples/ 下没有可回归的示例。')
+        return 2
+
+    for html in examples:
+        stem = html.stem
+        model = html.with_suffix('.model.json')
+        print(f'=== {stem}')
+        r0 = run([PY, str(ROOT / 'scripts' / 'validate_report.py'), str(html), '--strict'])
+        print('  [validate_report --strict]', 'OK' if r0.returncode == 0 else 'FAIL')
+        if r0.returncode != 0:
+            fails.append(stem + ':report')
+            print((r0.stdout or '')[-800:])
+            continue
+        r1 = run([PY, str(ROOT / 'scripts' / 'extract_model.py'), str(html)])
+        if r1.returncode != 0:
+            fails.append(stem + ':extract')
+            print((r1.stdout or '')[-500:])
+            continue
+        bp = B_OUT / (stem + '.pptx')
+        r2 = run([node, str(ROOT / 'scripts' / 'build_pptx.js'), str(bp), '--model=' + str(model)], env=env)
+        if r2.returncode != 0 or not bp.exists():
+            fails.append(stem + ':build')
+            print((r2.stderr or r2.stdout or '')[-800:])
+            continue
+        r3 = run([PY, str(ROOT / 'scripts' / 'validate_pptx.py'), str(bp), '--strict', '--model=' + str(model)])
+        print('  [B 通道 strict]', 'OK (0/0)' if r3.returncode == 0 else f'FAIL exit={r3.returncode}')
+        if r3.returncode != 0:
+            fails.append(stem + ':pptx-b')
+            for line in (r3.stdout or '').splitlines()[-14:]:
+                print('   ', line)
+
+    print('\n=== A 通道（页面预览序列化引擎产物 · gen_channel_a 组装）')
+    rA = run([node, str(ROOT / 'scripts' / 'gen_channel_a.js')], env=env)
+    if rA.returncode != 0:
+        fails.append('A:generate')
+        print((rA.stderr or rA.stdout or '')[-800:])
+    else:
+        ppx_ok = has_python_pptx()
+        for ap in sorted(A_OUT.glob('*.pptx')):
+            model = EX / (ap.stem + '.model.json')
+            if not model.exists():
+                # 非示例产物（如 ⑤ 探针的 probe-split.pptx）由探针段单独校验，跳过避免误判
+                continue
+            # A 通道（预览引擎）图表为形状近似渲染，豁免 MODEL_CHART_COUNT（交付通道 B 不豁免）
+            r = run([PY, str(ROOT / 'scripts' / 'validate_pptx.py'), str(ap), '--strict',
+                     '--model=' + str(model), '--allow-shape-charts'])
+            ok = r.returncode == 0
+            ppx = 'SKIP'
+            if ppx_ok:
+                p = run([PY, '-c',
+                         'import sys;from pptx import Presentation;'
+                         'print(len(list(Presentation(sys.argv[1]).slides)))', str(ap)])
+                ppx = 'OK' if p.returncode == 0 else 'FAIL'
+            print(f'  {ap.name}: strict={"OK" if ok else "FAIL"} python-pptx={ppx}')
+            if not ok:
+                fails.append(ap.stem + ':pptx-a')
+                for line in (r.stdout or '').splitlines()[-10:]:
+                    print('   ', line)
+            if ppx == 'FAIL':
+                fails.append(ap.stem + ':ppx-a')
+            if ppx == 'SKIP':
+                skipped.append(ap.stem + ':ppx-a')
+
+    # ④ 素材图片导出探针：示例矩阵刻意零图片，真实素材路径/版式/门禁靠探针覆盖
+    r5 = run([PY, str(ROOT / 'scripts' / 'probe_image_export.py')], env=env)
+    print('  [probe_image_export]', 'OK' if r5.returncode == 0 else 'FAIL')
+    if r5.returncode != 0:
+        fails.append('probe-image')
+        for line in (r5.stdout or '').splitlines()[-8:]:
+            print('   ', line)
+
+    # ⑤ 双区组合页探针（split 左区 = 图表）：示例矩阵只覆盖"左文右图"缺省路径，
+    #    左区图表/表格/图片三个变体靠脚手架生成一次走完整交付链路（v8 新增能力）；
+    #    B 通道 + A 通道（同一模型）都跑，产物同名落两目录 → ⑥ 交叉裁判可逐页比对。
+    print('\n=== 双区组合页探针（split 左区 = 图表）')
+    probe_plan = B_OUT / 'probe-split.plan.json'
+    probe_html = B_OUT / 'probe-split.html'
+    probe_model = B_OUT / 'probe-split.model.json'
+    probe_pptx = B_OUT / 'probe-split.pptx'
+    probe_plan.write_text(json.dumps(
+        [{'type': 'points', 'eyebrow': '01 · 前置', 'title': '前置要点页（探针占位）'},
+         {'type': 'split', 'left_type': 'chart', 'eyebrow': '02 · 组合版式',
+          'title': '双区组合页探针：左图右文（左区图表路径）'}], ensure_ascii=False), encoding='utf-8')
+    ok6, why6 = False, ''
+    r6 = run([PY, str(ROOT / 'scripts' / 'scaffold_report.py'), '--mode', 'presentation',
+              '--plan', str(probe_plan), '--out', str(probe_html)])
+    if r6.returncode != 0:
+        why6 = 'scaffold'
+    else:
+        r7 = run([PY, str(ROOT / 'scripts' / 'extract_model.py'), str(probe_html), str(probe_model)])
+        if r7.returncode != 0:
+            why6 = 'extract'
+        else:
+            r8 = run([node, str(ROOT / 'scripts' / 'build_pptx.js'), str(probe_pptx),
+                      '--model=' + str(probe_model)], env=env)
+            if r8.returncode != 0 or not probe_pptx.exists():
+                why6 = 'build'
+            else:
+                r9 = run([PY, str(ROOT / 'scripts' / 'validate_pptx.py'), str(probe_pptx),
+                          '--strict', '--model=' + str(probe_model)])
+                ok6 = r9.returncode == 0
+                if not ok6:
+                    why6 = 'validate_pptx_b'
+                    for line in (r9.stdout or '').splitlines()[-10:]:
+                        print('   ', line)
+                else:
+                    # A 通道同模型（预览序列化引擎）：产物同名落 dist/regression-a/
+                    probe_pptx_a = A_OUT / 'probe-split.pptx'
+                    if probe_pptx_a.exists():
+                        probe_pptx_a.unlink()
+                    r10 = run([node, str(ROOT / 'scripts' / 'gen_channel_a.js'),
+                               '--model=' + str(probe_model), '--out=' + str(A_OUT)], env=env)
+                    if r10.returncode != 0 or not probe_pptx_a.exists():
+                        ok6, why6 = False, 'channel_a'
+                        print((r10.stderr or r10.stdout or '')[-400:])
+                    else:
+                        # 预览引擎图表为形状近似 → 豁免 MODEL_CHART_COUNT（同示例矩阵口径）
+                        r11 = run([PY, str(ROOT / 'scripts' / 'validate_pptx.py'), str(probe_pptx_a),
+                                   '--strict', '--model=' + str(probe_model),
+                                   '--allow-shape-charts'])
+                        if r11.returncode != 0:
+                            ok6, why6 = False, 'validate_pptx_a'
+                            for line in (r11.stdout or '').splitlines()[-10:]:
+                                print('   ', line)
+    print('  [split 左区图表 B/A 双通道]', 'OK' if ok6 else f'FAIL({why6})')
+    if not ok6:
+        fails.append('probe-split')
+
+    # ⑥b 深色主题导出探针：示例矩阵只有 architecture/graphite-dark 是 dark，
+    #     research / presentation 的 stylesDark + styleDataColorsDark 路径此前零覆盖。
+    print('\n=== 深色主题导出探针（research / presentation × dark）')
+    dark_fail = []
+    for stem in ('2026-09-09-research-mckinsey', '2026-09-09-presentation-business-blue'):
+        src_model = EX / f'{stem}.model.json'
+        if not src_model.exists():
+            continue
+        dm = B_OUT / f'{stem}-dark.model.json'
+        dp = B_OUT / f'{stem}-dark.pptx'
+        data = json.loads(src_model.read_text(encoding='utf-8'))
+        data['theme'] = 'dark'
+        dm.write_text(json.dumps(data, ensure_ascii=False), encoding='utf-8')
+        rb = run([node, str(ROOT / 'scripts' / 'build_pptx.js'), str(dp),
+                  '--model=' + str(dm)], env=env)
+        if rb.returncode != 0 or not dp.exists():
+            dark_fail.append(stem + ':build')
+            print((rb.stderr or rb.stdout or '')[-300:])
+            continue
+        rv = run([PY, str(ROOT / 'scripts' / 'validate_pptx.py'), str(dp),
+                  '--strict', '--model=' + str(dm)])
+        print(f'  {stem} dark:', 'OK' if rv.returncode == 0 else 'FAIL')
+        if rv.returncode != 0:
+            dark_fail.append(stem + ':validate')
+            for line in (rv.stdout or '').splitlines()[-8:]:
+                print('   ', line)
+    if dark_fail:
+        fails.append('dark-theme')
+
+    # ⑥ 双通道交叉裁判（python-pptx）：放在探针之后，A/B 产物齐全时才逐页比对
+    r4 = run([PY, str(ROOT / 'scripts' / 'cross_verify.py')])
+    if r4.returncode == 2:
+        print('  [cross_verify] SKIP（python-pptx 未安装，可选第三方裁判）')
+        skipped.append('cross-verify')
+    else:
+        print('  [cross_verify]', 'OK' if r4.returncode == 0 else 'FAIL')
+        if r4.returncode != 0:
+            fails.append('cross-verify')
+            print((r4.stdout or '')[-600:])
+
+    # ⑦ 门禁反向验证（故障注入）：正向全绿只能证明“没误报”，这里证明“真能报”
+    r12 = run([PY, str(ROOT / 'scripts' / 'negative_tests.py')])
+    print('  [negative_tests]', 'OK' if r12.returncode == 0 else 'FAIL')
+    if r12.returncode != 0:
+        fails.append('negative-tests')
+        print((r12.stdout or '')[-600:])
+
+    # ⑦-b 任务路由全量遍历：--task 与 validate_report FIX_GUIDE 的修复链路不得静默断裂
+    #     （v8.3 前教训：中文节号不被解析，7/11 修复命令执行即报错且无任何审计覆盖；
+    #      实现与 audit_docs ⑥ 共用 ES.verify_routes，防两处维护）
+    try:
+        import extract_snippet as ES  # noqa: PLC0415
+        route_bad = ES.verify_routes()
+        print('  [task_routes]', 'OK' if not route_bad else f'FAIL {route_bad}')
+        if route_bad:
+            fails.append('task-routes')
+    except ImportError:
+        print('  [task_routes] FAIL（extract_snippet 不可导入）')
+        fails.append('task-routes')
+
+    # ⑧ 浏览器真值测量（可选）：静态估算看不到「表格被半幅容器横向裁掉一列」这类真实布局缺陷。
+    #    当前解释器可能未装 playwright（托管运行时常见）；measure_height.py 会自动回落到
+    #    本机已装 playwright 的解释器，因此这里只要能跑 measure_height 就不再跳过。
+    mh_fail = []
+    for html in examples:
+        r13 = run([PY, str(ROOT / 'scripts' / 'measure_height.py'), str(html)])
+        out = (r13.stdout or '') + (r13.stderr or '')
+        if r13.returncode != 0:
+            mh_fail.append(html.stem)
+            print(out[-500:])
+        elif '跳过：' in out and 'playwright' in out.lower():
+            # 解释器与回落链都找不到 playwright → 整批跳过（可选依赖）
+            print('  [measure_height] SKIP（playwright 未安装，可选真值裁判）')
+            skipped.append('measure-height')
+            mh_fail = []
+            break
+    if 'measure-height' not in skipped:
+        print('  [measure_height 页高/容器裁切真值]', 'OK' if not mh_fail else f'FAIL {mh_fail}')
+        if mh_fail:
+            fails.append('measure-height')
+
+    print('=' * 56)
+    if skipped:
+        print('跳过（可选依赖缺失）:', sorted(set(skipped)))
+    if fails:
+        print('回归失败:', fails)
+        return 1
+    print('全链路回归通过：HTML strict + B 通道精导 strict 0/0 + A 通道（预览同引擎产物）strict 0/0'
+          + ' + 双区组合页双通道探针 + 深色主题探针 + 门禁反向验证 + 任务路由全量遍历'
+          + ('' if 'measure-height' in skipped else ' + 浏览器页高/裁切真值')
+          + (' + python-pptx 双裁判 + 双通道交叉一致' if 'cross-verify' not in skipped else ''))
+    return 0
+
+
+if __name__ == '__main__':
+    sys.exit(main())