@topmindspace/tms-skills 2.0.1 → 2.0.2

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

@@ -1,1456 +1,1456 @@
-#!/usr/bin/env python3
-# -*- coding: utf-8 -*-
-"""
-TopPPT HTML· HTML 报告质量校验（交付闭环）
-用法:
-    python validate_report.py <报告.html> [--strict] [--json]
-
-输出每项 PASS/FAIL，结尾给汇总与结论；任一 FAIL 时退出码为 1（--strict 时 WARN 也计失败）。
---json 以 JSON 输出全部检查结果（供脚本/流水线读取）。
-生成流程：生成 → 跑本脚本 → 修复 FAIL → 再跑，直至全部 PASS 才交付。
-
-阈值全部来自单源 scripts/layout-constants.json（checkBudgets / charts / styleAccents / aiFlavor），
-本脚本不再内置任何风格色值或预算数字——改阈值只改 JSON。
-
-检查面：
-  结构闭合 / 主题风格 / 页面高度模型 / 宽屏翻页 / Agenda（architecture 极简形态可省略） / 粗体 /
-  零外链 / 引用与锚点闭环 / 单一强调色 / 内联写死色 / 内容密度 / 单页文字与组件预算 /
-  表格行数上限 / 页高溢出估算 / 去AI味 / 图表存在与动效 / 图表 data-chart 登记 / 图表最小尺寸 /
-  模式版式特征（含新页型 steps/heatmap/bullet/pyramid/image 与 research/architecture 专属） /
-  模型一致性（REPORT_MODEL ↔ 正文） / Exhibit 编号连续 / research so-what 与来源行 /
-  强调带约束（band--deep 反相页 ≤ 上限且不作末页） / 待核实标注（.tbd ↔ .tbd-legend/.flagbar） /
-  素材图片与配图占位（零外链 + alt + 内联体积 + 版式/比例锁定类对应 + 占位可见标签 + 模型对应） /
-  PPTX 预览配套（预览按钮 + 模型 + 运行时，页面无导出按钮） /
-  内容级质量（版式节奏连用上限 / so-what 实质与空洞套话 / research 标题含数字或判断词）
-"""
-import sys
-import re
-import json
-from html.parser import HTMLParser
-from pathlib import Path
-
-# Windows GBK 控制台兜底：强制 UTF-8 输出（含 ↔ 等符号）
-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))
-import checks_html  # noqa: E402  承载正则 / 图表通道 / 多样性 / 组合版式单源
-
-VOID = {'area', 'base', 'br', 'col', 'embed', 'hr', 'img', 'input', 'link', 'meta',
-        'param', 'source', 'track', 'wbr'}
-
-# ── 单源阈值（scripts/layout-constants.json）────────────────────────────────
-_DEFAULTS = {
-    'checkBudgets': {
-        'presentation': dict(char=1500, unit=8, empty=80, fit=1000, wrap=1400, trow=8),
-        'research': dict(char=3200, unit=12, empty=120, fit=2200, wrap=1240, trow=16),
-        'architecture': dict(char=600, unit=3, empty=60, fit=600, wrap=1600, trow=6),
-    },
-    'charts': {'types': [], 'minSize': {}},
-    'styleAccents': {},
-    'aiFlavor': {'words': []},
-    'imageSpec': {},
-    'contentQuality': {},
-}
-
-
-def load_constants():
-    path = Path(__file__).resolve().parent / 'layout-constants.json'
-    try:
-        lc = json.loads(path.read_text(encoding='utf-8'))
-    except (OSError, json.JSONDecodeError) as exc:
-        print(f'警告: 无法读取单源 {path}（{exc}），使用内置默认阈值。')
-        return _DEFAULTS
-    for key, fallback in _DEFAULTS.items():
-        lc.setdefault(key, fallback)
-    return lc
-
-
-LC = load_constants()
-MODE_BUDGETS = LC['checkBudgets']
-CHART_TYPES = set(LC['charts'].get('types') or [])
-CHART_MIN = LC['charts'].get('minSize') or {}
-# 图表多样性预算（治「登记 36 种、实际只用 bar/line/donut」）
-CHART_VARIETY = LC['charts'].get('variety') or {}
-AI_FLAVOR = LC['aiFlavor'].get('words') or []
-STYLE_ACCENTS = {k: {str(h).lower() for h in v}
-                 for k, v in (LC['styleAccents'] or {}).items()
-                 if not k.startswith('$') and isinstance(v, list)}
-ALL_ACCENTS = set().union(*STYLE_ACCENTS.values()) if STYLE_ACCENTS else set()
-# 图表登记四元组（charts.registry）与信息图页型（数据可追溯策略核对）
-CHART_REG = {k: v for k, v in (LC['charts'].get('registry') or {}).items()
-             if not str(k).startswith('$')}
-INFO_PAGE_TYPES = {'sankey', 'treemap', 'boxplot', 'network', 'marimekko', 'streamgraph'}
-# 素材图片规格与配图占位约定（单源 layout-constants.json imageSpec）
-IMAGE_SPEC = LC.get('imageSpec') or {}
-IMAGE_LAYOUTS = set(IMAGE_SPEC.get('layouts') or ['full', 'half', 'bleed', 'grid', 'compare', 'wall'])
-IMAGE_MULTI_LAYOUTS = set(IMAGE_SPEC.get('multiLayouts') or ['grid', 'compare', 'wall'])
-IMAGE_FIT = set(IMAGE_SPEC.get('fitEnum') or ['cover', 'contain'])
-IMAGE_MAX_PER_PAGE = int(IMAGE_SPEC.get('maxPerPage') or 6)
-IMAGE_MAX_INLINE = int(IMAGE_SPEC.get('maxInlineBytes') or 1572864)
-IMAGE_MAX_TOTAL = int(IMAGE_SPEC.get('maxTotalInlineBytes') or 8388608)
-
-AGENDA_SINGLE_MAX = 8        # 超过则建议 agenda--2col
-# 页高估算参数（单源 layout-constants.pageHeightEstimate）
-_HE = LC.get('pageHeightEstimate') or {}
-SCREEN_BUDGET_PX = int(_HE.get('screenBudgetPx') or 1000)
-WRAP_INSET_PX = int(_HE.get('wrapInsetPx') or 112)
-SHEAD_PX = int(_HE.get('sheadPx') or 190)
-UNIT_PX = int(_HE.get('unitPx') or 170)
-LINE_FACTOR = float(_HE.get('lineFactor') or 1.65)
-STRUCT_PAGE_IDS = ('refs', 'appendix')
-
-
-class Struct(HTMLParser):
-    def __init__(self):
-        super().__init__()
-        self.stack, self.err, self.mis = [], [], []
-
-    def handle_starttag(self, tag, attrs):
-        if tag not in VOID:
-            self.stack.append(tag)
-
-    def handle_endtag(self, tag):
-        if tag in VOID:
-            return
-        if self.stack and self.stack[-1] == tag:
-            self.stack.pop()
-        else:
-            try:
-                i = len(self.stack) - 1 - self.stack[::-1].index(tag)
-                self.mis.extend(self.stack[i:])
-                self.stack = self.stack[:i]
-            except ValueError:
-                self.err.append(tag)
-
-
-def _plain(html):
-    return re.sub(r'<[^>]+>', ' ', html)
-
-
-def _bands(txt):
-    parts = re.split(r'(?=<section class="band)', txt)
-    return [p for p in parts if p.startswith('<section class="band')]
-
-
-def _band_units(b):
-    """承载组件数：卡/指标/图/表 + 列表组/结构/步骤/媒体/矩阵（组合版式与 underfill 同口径）。"""
-    return (len(re.findall(r'class="(?:card|metric)[\s"]', b)) +
-            len(re.findall(r'class="fig[\s"]', b)) +
-            len(re.findall(r'class="tbl-wrap', b)) +
-            len(re.findall(r'class="(?:arch|lane|steps|heat|bul|pyr|media|matrix|exhibit)[\s"]', b)) +
-            len(re.findall(r'class="ul[\s"]', b)) +
-            len(re.findall(r'class="cols-[23]', b)))
-
-
-def _band_heavy_units(b):
-    """页高估算用「重承载行数」：卡/指标按栅格列数折行；图/表/结构各计 1 行级。"""
-    cards = len(re.findall(r'class="(?:card|metric)[\s"]', b))
-    figs = len(re.findall(r'class="fig[\s"]', b))
-    tables = len(re.findall(r'class="tbl-wrap', b))
-    structs = len(re.findall(r'class="(?:arch|lane|media|heat|exhibit)[\s"]', b))
-    if re.search(r'class="[^"]*\bg-4\b', b):
-        cols = 4
-    elif re.search(r'class="[^"]*\bg-3\b', b):
-        cols = 3
-    elif re.search(r'class="[^"]*\b(?:g-2|g-side|g-hero|g-half)\b', b):
-        cols = 2
-    else:
-        cols = 1
-    card_rows = (cards + cols - 1) // cols if cards else 0
-    return card_rows + figs + tables + structs
-
-
-def _check_charts(txt, chk):
-    """图表：存在性 / data-chart 登记 / 动效 / 最小尺寸。"""
-    svgs = re.findall(r'<svg\b[^>]*>', txt)
-    chart_tags = [t for t in svgs if 'class="chart' in t or "class='chart" in t or 'data-chart' in t]
-    chk("含内联 SVG 图表/图形", len(svgs) >= 1, f"{len(svgs)} 个", level="WARN")
-    if not chart_tags:
-        return
-
-    untyped = [t for t in chart_tags if 'data-chart' not in t]
-    chk("图表 svg 均有 data-chart 类型标记", not untyped,
-        f"{len(untyped)} 个未标记", level="WARN")
-
-    # data-chart 类型须在登记表内（layout-constants.json charts.types）
-    unknown = []
-    if CHART_TYPES:
-        for t in chart_tags:
-            m = re.search(r'data-chart="([^"]+)"', t)
-            if m and m.group(1) not in CHART_TYPES:
-                unknown.append(m.group(1))
-    chk("图表类型在登记表内（layout-constants.json charts.types）", not unknown,
-        f"未登记: {sorted(set(unknown))}", level="WARN")
-
-    has_io = 'IntersectionObserver' in txt
-    has_anim = 'data-anim' in txt or 'data-draw' in txt or 'data-sweep' in txt or 'data-count' in txt
-    chk("图表动效（IntersectionObserver + data-anim/draw/sweep/count）",
-        has_io and has_anim,
-        "" if (has_io and has_anim) else f"IO={'有' if has_io else '无'} anim={'有' if has_anim else '无'}",
-        level="WARN")
-
-    too_small, small_note = [], []
-    for t in chart_tags:
-        m = re.search(r'data-chart="([^"]+)"', t)
-        ctype = m.group(1) if m else None
-        rule = CHART_MIN.get(ctype, {}) if ctype else {}
-        wm = re.search(r'style="[^"]*?width:\s*(\d+(?:\.\d+)?)px', t)
-        vb = re.search(r'viewBox="[\d.\-]+ [\d.\-]+ ([\d.]+) ([\d.]+)"', t)
-        if wm and rule.get('pxWidthMin'):
-            w = float(wm.group(1))
-            if w < rule['pxWidthMin']:
-                too_small.append(f"{ctype} 宽{w:.0f}px<{rule['pxWidthMin']}px")
-        if vb and rule.get('vbHeightMin'):
-            h = float(vb.group(2))
-            if h < rule['vbHeightMin']:
-                small_note.append(f"{ctype} viewBox高{h:.0f}<{rule['vbHeightMin']}")
-    chk("图表最小尺寸·硬下限（环形/雷达/仪表盘显示宽）", not too_small,
-        "; ".join(too_small) if too_small else "")
-    chk("图表最小尺寸·建议（通宽图 viewBox 高）", not small_note,
-        "; ".join(small_note) if small_note else "", level="WARN")
-
-
-CARRIERS = checks_html.CARRIERS  # 单源 scripts/checks_html.py
-
-
-def _check_chart_variety(txt, chk, mode):
-    """图表与版式多样性（阈值单源 charts.variety；判定逻辑 checks_html）。"""
-    v = CHART_VARIETY
-    if not v:
-        return
-    per_page = checks_html.chart_types_per_page(txt)
-    used = [t for ts in per_page for t in ts]
-    if not used:
-        return
-    distinct = sorted(set(used))
-    n_chart_pages = len([1 for ts in per_page if ts])
-
-    floor = checks_html.variety_floor(mode, n_chart_pages, v)
-    chk("图表多样性（全篇不同 data-chart 类型数）", len(distinct) >= floor,
-        f"{len(distinct)} 种 / 下限 {floor}（{n_chart_pages} 个图表页）: {distinct}")
-
-    if v.get('noRepeatAdjacent'):
-        adj = checks_html.adjacent_same_type(per_page)
-        chk("图表不连续同型（相邻图表页不得同 data-chart）", not adj,
-            "; ".join(adj[:3]) if adj else "", level="WARN")
-
-    ratio = checks_html.composite_required(mode, v)
-    if ratio > 0:
-        multi, n_content = checks_html.composite_pages(txt)
-        need = max(1, int(n_content * ratio))
-        chk("组合版式（内容页含 ≥2 种承载类型的比例）", multi >= need,
-            f"{multi}/{n_content} 页（下限 {need}）——单件页是例外不是默认",
-            level="WARN")
-
-
-def _check_page_model(txt, chk, mode):
-    """页面高度模型（每页高度稳定）：所有 .band 至少一屏高，长结构页显式 --flow 退出。"""
-    has_model = ('--band-min' in txt) or ('calc(100svh - var(--bar-h))' in txt) or \
-                ('min-height:calc(100vh - var(--bar-h))' in txt)
-    chk("页面高度模型（--band-min / 一屏最小高度）", has_model,
-        "缺少每页一屏的 min-height 机制（见 design-system.md §1c）")
-    bands = _bands(txt)
-    if not bands:
-        return
-    # 固定页高页（非 --flow）应占多数；长结构页才允许 flow
-    flow_pages = []
-    for i, b in enumerate(bands, 1):
-        if 'band--flow' in b[:220]:
-            flow_pages.append(i)
-    bad_flow = []
-    for i in flow_pages:
-        b = bands[i - 1]
-        if not any(f'id="{pid}"' in b for pid in STRUCT_PAGE_IDS):
-            bad_flow.append(f"第{i}页")
-    chk("band--flow 仅用于长结构页（参考资料/附录）", not bad_flow,
-        f"{bad_flow} 使用 --flow 但非结构页（应改用默认一屏页）" if bad_flow else "", level="WARN")
-
-
-def _check_bands(txt, chk, mode="presentation"):
-    """单页预算与布局：文字量/组件数（防塞爆）· 过空页 · 页高溢出估算 · 混排对齐。"""
-    B = MODE_BUDGETS.get(mode, MODE_BUDGETS['presentation'])
-    bands = _bands(txt)
-    over_chars, over_units, too_empty, fit_overload = [], [], [], []
-    for i, b in enumerate(bands, 1):
-        if 'id="refs"' in b or 'id="appendix"' in b:
-            continue
-        b_clean = re.sub(r'<script\b[\s\S]*?</script>', ' ', b)
-        b_clean = re.sub(r'<style\b[\s\S]*?</style>', ' ', b_clean)
-        plain_len = len(_plain(b_clean).strip())
-        units = _band_units(b)
-        if plain_len > B["char"]:
-            over_chars.append(f"第{i}页 {plain_len}字")
-        if units > B["unit"]:
-            over_units.append(f"第{i}页 {units}个")
-        head_tag = b[:220]
-        is_fit = 'band--fit' in head_tag
-        is_deep = 'band--deep' in head_tag
-        if is_fit and plain_len > B["fit"]:
-            fit_overload.append(f"第{i}页 {plain_len}字")
-        if (not is_fit and not is_deep and units == 0 and plain_len < B["empty"]):
-            too_empty.append(f"第{i}页 {plain_len}字")
-    chk(f"单页文字量 ≤ {B['char']} 字（{mode} 模式 · 防溢出）", not over_chars,
-        "; ".join(over_chars) if over_chars else "", level="WARN")
-    chk(f"单页并列单元 ≤ {B['unit']} 个（{mode} 模式 · 防塞爆）", not over_units,
-        "; ".join(over_units) if over_units else "", level="WARN")
-    # v9：过空页升为 FAIL（截图级「只有标题」缺陷不得假过）
-    chk(f"无过空页（{mode} 模式 · 普通页 ≥{B['empty']} 字或有组件；防大面积留白）", not too_empty,
-        "; ".join(too_empty) if too_empty else "")
-    chk(f"满屏居中页 band--fit 内容 ≤ {B['fit']} 字（{mode} 模式 · 防垂直溢出）", not fit_overload,
-        "; ".join(fit_overload) if fit_overload else "", level="WARN")
-
-    # 页高溢出静态估算：字数×行高 + 组件固定高 vs 一屏预算（参数单源 pageHeightEstimate / checkBudgets）。
-    # 固定页高页（无 --flow）超预算即 FAIL——这是"每页高度稳定"的核心硬门禁。
-    fs_px = int(B.get('bodyPx') or 17)
-    wrap_px = int(B.get('wrap') or 1400) - WRAP_INSET_PX
-    cpl = max(10, int(wrap_px / fs_px))
-    est_over = []
-    for i, b in enumerate(bands, 1):
-        if 'id="refs"' in b or 'id="appendix"' in b:
-            continue
-        if 'band--flow' in b[:220]:
-            continue
-        b_clean = re.sub(r'<script\b[\s\S]*?</script>', ' ', b)
-        b_clean = re.sub(r'<style\b[\s\S]*?</style>', ' ', b_clean)
-        plen = len(_plain(b_clean).strip())
-        u = _band_heavy_units(b)
-        est_h = plen / cpl * fs_px * LINE_FACTOR + u * UNIT_PX + SHEAD_PX
-        # 静态估算容差 8%（真值以浏览器为准；轻微超出仍按 FAIL 会误伤密排页）
-        if est_h > SCREEN_BUDGET_PX * 1.08:
-            est_over.append(f"第{i}页≈{est_h:.0f}px")
-    chk(f"页高溢出估算（字×行高+组件 ≤ ~{SCREEN_BUDGET_PX}px/屏 · {mode}）", not est_over,
-        ("; ".join(est_over) + "（按「列表化/精炼→压缩→多列→拆页」处理，或给长结构页加 band--flow）")
-        if est_over else "")
-
-    bad_grids = 0
-    for m in re.finditer(r'<div class="grid[^"]*"([^>]*)>', txt):
-        attrs = m.group(1)
-        sec_end = txt.find('</section>', m.end())
-        ahead = txt[m.end(): sec_end if sec_end != -1 else m.end() + 4000]
-        if 'class="fig' in ahead and 'class="card' in ahead:
-            aligned = ('align-items' in attrs or
-                       re.search(r'\ba-(c|start|end)\b', m.group(0)))
-            if not aligned:
-                bad_grids += 1
-    chk("混排栅格（图+卡）显式对齐（a-start/a-c 或 align-items）", bad_grids == 0,
-        f"{bad_grids} 处未对齐" if bad_grids else "", level="WARN")
-
-
-def _check_icons(txt, chk):
-    """图标使用：长报告不应全文无图标（要点卡/关键列表至少一处）。"""
-    if len(_bands(txt)) < 6:
-        return
-    has_icon = ('ul--ico' in txt or 'metric__ico' in txt or
-                re.search(r'card__ico"[^>]*>\s*<svg', txt) or
-                re.search(r'card__ico">\s*<svg', txt))
-    chk("长报告（≥6页）要点卡/关键列表有图标", bool(has_icon),
-        "全文未用图标（见 icons.md 使用准则）", level="WARN")
-
-
-def _check_table_rows(txt, chk, mode="presentation"):
-    """表格行数上限（research 密表 ≤16 / presentation ≤8 / architecture ≤6；宁拆勿挤）。"""
-    B = MODE_BUDGETS.get(mode, MODE_BUDGETS['presentation'])
-    max_rows = B['trow']
-    over = []
-    for i, m in enumerate(re.finditer(r'<tbody>([\s\S]*?)</tbody>', txt), 1):
-        n = len(re.findall(r'<tr', m.group(1)))
-        if n > max_rows:
-            over.append(f"表{i} {n}行")
-    chk(f"表格行数 ≤ {max_rows} 行（{mode} 模式 · 宁拆勿挤）", not over,
-        "; ".join(over) if over else "", level="WARN")
-
-
-def _check_mode_layouts(txt, chk, mode):
-    """模式版式特征：三模式独立体系，各自应有标志性组件。"""
-    if mode == 'research':
-        has = ('class="cols-2' in txt or 'class="exhibit' in txt or
-               'class="cols-3' in txt or 'class="matrix' in txt)
-        chk("research 版式特征（cols-2/cols-3/exhibit/matrix 至少一处）", has,
-            "未见研究模式标志性组件", level="WARN")
-    elif mode == 'architecture':
-        has = ('class="arch' in txt or 'class="lane' in txt)
-        chk("architecture 版式特征（arch/lane 至少一处）", has,
-            "未见架构模式标志性组件", level="WARN")
-
-
-# 页型 → 正文标志性版式特征（正则；模型-正文同源抽查，见 `components.md` §46/§46b）
-TYPE_FEATURE = {
-    'steps': r'class="[^"]*\bsteps\b',
-    'heatmap': r'class="[^"]*\bheat\b',
-    'bullet': r'class="[^"]*\bbul\b',
-    'pyramid': r'class="[^"]*\bpyr\b',
-    'image': r'class="[^"]*\bmedia\b|<img\b',
-    'matrix': r'class="[^"]*\bmatrix\b',
-    'exhibit': r'class="[^"]*\bexhibit\b',
-    'twocol': r'class="[^"]*\bcols-2\b',
-    'threecol': r'class="[^"]*\bcols-3\b',
-    'halftable': r'class="[^"]*\bg-half\b',
-    'split': r'class="[^"]*\bg-side\b',
-    'diagram': r'class="[^"]*\barch\b',
-    'lane': r'class="[^"]*\blane\b',
-    'timeline': r'class="[^"]*\btl\b',
-    'comparison': r'class="[^"]*\bg-half\b',
-    'quote': r'band--accent|band--deep',
-    'donut': r'data-chart="donut"',
-    # 复杂信息图页型：正文以内联 SVG 承载（data-chart 登记与 charts.registry 同源）
-    'sankey': r'data-chart="sankey"',
-    'treemap': r'data-chart="treemap"',
-    'boxplot': r'data-chart="boxplot"',
-    'network': r'data-chart="network"',
-    'marimekko': r'data-chart="marimekko"',
-    'streamgraph': r'data-chart="streamgraph"',
-}
-
-
-def _check_type_features(txt, chk, model):
-    """模型声明的页型应在正文找到对应版式组件（防"模型填了页型但正文用旧版式"）。"""
-    if not model:
-        return
-    missing = []
-    for sec in (model.get('sections') or []):
-        t = (sec.get('type') or 'points')
-        feat = TYPE_FEATURE.get(t)
-        if feat and not re.search(feat, txt):
-            missing.append(t)
-    chk("模型页型 ↔ 正文版式组件对应", not missing,
-        f"{len(missing)} 处: {missing[:4]}" if missing else "", level="WARN")
-
-
-def _check_pptx_export(txt, chk):
-    """PPTX 预览配套：有预览按钮就必须有内容模型 + 预览运行时；页面不得残留导出按钮。"""
-    has_btn = 'id="pptPreviewBtn"' in txt
-    has_model = 'REPORT_MODEL' in txt
-    has_runtime = ('g.TopPptHtml = api' in txt) and ('function slidesXml' in txt)
-    if has_btn:
-        chk("PPTX 预览按钮配套（REPORT_MODEL + 预览运行时）",
-            has_model and has_runtime,
-            f"model={'有' if has_model else '无'} runtime={'有' if has_runtime else '无'}")
-    else:
-        chk("含 PPTX 预览按钮与内容模型（建议保留）",
-            has_model and has_runtime, "未集成预览按钮/模型", level="WARN")
-    chk("页面无 PPTX 导出按钮（仅预览 + 提示词）",
-        'id="pptxBtn"' not in txt and 'id="pptDownload"' not in txt,
-        "残留导出按钮 pptxBtn/pptDownload", level="WARN")
-
-
-def _extract_model(txt):
-    m = re.search(r'window\.REPORT_MODEL\s*=\s*', txt)
-    if not m:
-        return None, "未找到 window.REPORT_MODEL（PPTX 双通道导出将不可用）"
-    try:
-        model, _end = json.JSONDecoder().raw_decode(txt[m.end():])
-        return model, None
-    except json.JSONDecodeError as e:
-        return None, f"REPORT_MODEL 不是合法 JSON: {e}（必须双引号/无尾逗号/无注释）"
-
-
-def _chart_channel(chart_type) -> str:
-    """图表类型 → 交付通道（native / shape），与 charts.registry 同源（checks_html）。"""
-    return checks_html.chart_channel(chart_type, CHART_REG)
-
-
-def _chart_datatable_mode(chart: dict) -> str:
-    """数据表策略：图表级 > 登记表默认（appendix 收敛为 inline）。"""
-    return checks_html.chart_datatable_mode(chart, CHART_REG)
-
-
-def _check_chart_datatable(chk, model) -> None:
-    """数据可追溯：非原生（形状通道）图表与信息图页型的 dataTable 不得为 off。
-
-    与 validate_pptx.py 的 MODEL_CHART_DATATABLE 同源，但提前到 HTML 阶段暴露——
-    避免生成完 PPTX 才被拦（数据表策略是模型层决定，HTML 阶段即可判定）。
-    """
-    bad: list[str] = []
-    for sec in (model.get('sections') or []):
-        if not isinstance(sec, dict):
-            continue
-        st = sec.get('type') or ''
-        if st in INFO_PAGE_TYPES:
-            mode = str(((sec.get('chart') or {}).get('dataTable')) or 'notes').lower()
-            if mode == 'off':
-                bad.append(f'{st} 页信息图 dataTable=off')
-            continue
-        charts = []
-        if st in ('bar', 'donut', 'exhibit', 'halftable'):
-            c = sec.get('chart') or {}
-            if isinstance(c, dict) and c.get('labels') and c.get('values'):
-                charts.append(c)
-        elif st == 'split':
-            # 双区组合页：左/右两侧各可能是图表（left 缺省 points、right 缺省 bar）
-            for side, dflt in (('left', 'points'), ('right', 'bar')):
-                pane = sec.get(side) or {}
-                if not isinstance(pane, dict):
-                    continue
-                if (pane.get('type') or dflt) in ('table', 'image', 'points'):
-                    continue
-                if pane.get('labels') and pane.get('values'):
-                    charts.append(pane)
-        for chart in charts:
-            if _chart_channel(chart.get('type')) == 'shape' and _chart_datatable_mode(chart) == 'off':
-                bad.append(f"{st} 页 {chart.get('type')} 图表 dataTable=off")
-    chk("图表数据表策略（非原生图表不得 off，保证数据可追溯）", not bad,
-        '；'.join(bad[:3]) if bad else "", level="WARN")
-
-
-def _check_model_consistency(txt, chk, mode, style):
-    """模型-正文一致性：JSON 合法 / mode·style·theme 一致 / 章节标题抽查 / agenda / 页数。"""
-    model, err = _extract_model(txt)
-    if err:
-        chk("REPORT_MODEL 存在且为合法 JSON", False, err)
-        return None
-    chk("REPORT_MODEL 存在且为合法 JSON", True)
-    mmode = model.get('mode')
-    chk("REPORT_MODEL.mode 与 data-mode 一致", mmode == mode,
-        f"model={mmode!r} 页面={mode!r}")
-    mstyle = model.get('style')
-    chk("REPORT_MODEL.style 与 data-style 一致", mstyle == style,
-        f"model={mstyle!r} 页面={style!r}", level="WARN")
-    mtheme = model.get('theme') or 'light'
-    html_theme_m = re.search(r'<html[^>]*data-theme="([^"]+)"', txt)
-    html_theme = html_theme_m.group(1) if html_theme_m else 'light'
-    chk("REPORT_MODEL.theme 与 data-theme 一致（未声明默认 light）",
-        mtheme in ('light', 'dark') and mtheme == html_theme,
-        f"model={mtheme!r} 页面={html_theme!r}", level="WARN")
-    secs = model.get('sections') or []
-    chk("REPORT_MODEL.sections 非空（每章一页模型）", len(secs) >= 1,
-        f"{len(secs)} 页", level="WARN")
-    _check_chart_datatable(chk, model)
-    plain = _plain(txt)
-    missing = [str(s.get('title'))[:14] for s in secs
-               if not s.get('title') or str(s.get('title')) not in plain]
-    chk("模型章节标题均出现于正文（模型-正文同源）", not missing,
-        f"{len(missing)} 个缺失: {missing[:3]}" if missing else "", level="WARN")
-    n_html = txt.count('class="agenda__a"')
-    n_model = len(model.get('agenda') or [])
-    if n_html and n_model:
-        chk("模型 agenda 条数与正文一致", n_html == n_model,
-            f"model={n_model} 正文={n_html}", level="WARN")
-    n_bands = len(_bands(txt))
-    lo, hi = len(secs) + 3, len(secs) + 7
-    chk("页数粗匹配（正文 band ≈ sections + 结构页）", lo <= n_bands <= hi,
-        f"正文 {n_bands} 页 vs 模型 {len(secs)}+{lo - len(secs)}~{hi - len(secs)}", level="WARN")
-    return model
-
-
-def _check_anchors(txt, chk):
-    """锚点闭环：Agenda / 导航 / 正文内部链接指向的 id 必须存在，且 id 全篇唯一。
-    （重复 id 会让锚点跳到错误的页，且 HTML 规范不允许——硬拦。）"""
-    # 只看正文：内联运行时/UI 脚本里有 `id="' + nid() + '"` 这类拼接，不能算重复 id
-    body = re.sub(r'<script\b[\s\S]*?</script>', '', txt, flags=re.I)
-    all_ids = re.findall(r'\sid="([^"]+)"', body)
-    ids = set(all_ids)
-    dup = sorted({i for i in all_ids if all_ids.count(i) > 1})
-    chk("id 全篇唯一（无重复 id）", not dup,
-        f"重复 id {dup[:5]}（锚点会跳到错误的页）" if dup else "")
-    bad = []
-    for href in set(re.findall(r'href="#([^"]+)"', body)):
-        if href and href not in ids:
-            bad.append('#' + href)
-    chk("内部锚点闭环（agenda/nav/正文链接均有对应 id）", not bad,
-        f"悬空锚点 {sorted(bad)[:5]}" if bad else "")
-
-
-def _check_exhibits(txt, chk, mode):
-    """Exhibit 编号体系：每个 .exhibit 框都有编号，且全篇连续 1..N 无跳号无重复。"""
-    nos = [int(n) for n in re.findall(
-        r'class="exhibit__no"[^>]*>\s*Exhibit\s*(\d+)', txt, re.I)]
-    n_frame = len(re.findall(r'class="exhibit[\s"]', txt))
-    if not nos:
-        if mode == 'research':
-            chk("Exhibit 编号体系（research 深报告建议建立）", False,
-                "全文无 Exhibit 编号图表框", level="WARN")
-        return
-    # 只校验已编号者的连续性会漏掉「漏标编号的 exhibit 框」——先核框数再核编号。
-    chk(f"Exhibit 框均已编号（{n_frame} 框）", n_frame == len(nos),
-        f"{n_frame} 个 .exhibit 框但只有 {len(nos)} 个 .exhibit__no 编号" if n_frame != len(nos) else "")
-    expected = list(range(1, len(nos) + 1))
-    chk(f"Exhibit 编号连续（1..{len(nos)} 无跳号无重复）", sorted(nos) == expected,
-        f"读到 {nos}")
-
-
-def _check_research_extras(txt, chk, mode):
-    """research 模式：so-what 连续性（>3 页无结论条）与 Exhibit 来源行。"""
-    if mode != 'research':
-        return
-    bands = _bands(txt)
-    streak = max_streak = 0
-    for i, b in enumerate(bands, 1):
-        if i <= 2 or 'id="agenda"' in b or 'id="refs"' in b:
-            continue
-        if 'class="sowhat' in b or 'class="note' in b:
-            streak = 0
-        else:
-            streak += 1
-            max_streak = max(max_streak, streak)
-    chk("so-what 连续性（连续 ≤3 页无结论条）", max_streak <= 3,
-        f"连续 {max_streak} 页无 .sowhat/.note" if max_streak > 3 else "", level="WARN")
-    n_ex = len(re.findall(r'class="exhibit[\s"]', txt))
-    n_src = txt.count('class="exhibit__src"')
-    chk(f"Exhibit 图表框均有来源行（{n_ex} 框）", n_ex == n_src,
-        f"{n_ex - n_src} 个缺 .exhibit__src" if n_ex != n_src else "", level="WARN")
-
-
-def _check_emphasis(txt, chk):
-    """强调带约束：band--deep 是反相强调页（浅色主题下渲染为深色）——
-    全文受限使用，且不得作末页，否则浅色模式末尾会出现深色页（用户核心诉求）。"""
-    em = LC.get('emphasis') or {}
-    max_deep = em.get('deepMaxPages', 1)
-    forbid_last = em.get('deepForbiddenLast', True)
-    bands = _bands(txt)
-    deep_idx = [i for i, b in enumerate(bands, 1) if 'band--deep' in b[:220]]
-    chk(f"反相强调页 band--deep ≤ {max_deep} 页（浅色模式下渲染为深色）",
-        len(deep_idx) <= max_deep,
-        f"{len(deep_idx)} 页: {deep_idx}（收尾/金句页请用 band--accent）" if len(deep_idx) > max_deep else "",
-        level="WARN")
-    if forbid_last and bands:
-        chk("末页非反相深色页（band--deep）",
-            'band--deep' not in bands[-1][:220],
-            "末页为 band--deep：浅色模式下会以深色收尾（改用 band--accent）", level="WARN")
-
-
-def _layout_sig(band: str) -> str | None:
-    """从 band HTML 提取版式签名（优先级从具体到宽泛）。
-    图表类型并入签名：exhibit+sankey 与 exhibit+waterfall 视为不同版式。"""
-    pats = (
-        ('g-hero-full', r'g-hero-full'),
-        ('g-mosaic', r'g-mosaic'),
-        ('g-bento', r'g-bento'),
-        ('g-aside', r'g-aside'),
-        ('g-quad', r'g-quad'),
-        ('g-211', r'g-211'),
-        ('g-121', r'g-121'),
-        ('g-side', r'g-side'),
-        ('rows-2', r'rows-2'),
-        ('rows-3', r'rows-3'),
-        ('g-2', r'\bg-2\b'),
-        ('g-3', r'\bg-3\b'),
-        ('g-4', r'\bg-4\b'),
-        ('g-5', r'\bg-5\b'),
-        ('g-6', r'\bg-6\b'),
-        ('cols-3', r'cols-3'),
-        ('cols-2', r'cols-2'),
-        ('exhibit', r'class="exhibit'),
-        ('table', r'class="tbl-wrap'),
-        ('metrics', r'class="metric'),
-        ('arch', r'class="arch'),
-        ('lane', r'class="lane'),
-        ('matrix', r'class="matrix'),
-        ('heat', r'class="heat'),
-        ('bul', r'class="bul'),
-        ('pyr', r'class="pyr'),
-        ('steps', r'class="steps'),
-        ('timeline', r'class="tl[\s"]'),
-        ('media', r'class="media'),
-        ('stagger', r'stagger'),
-        ('quote', r'band--accent'),
-        ('card', r'class="card[\s"]'),
-    )
-    base = None
-    for name, pat in pats:
-        if re.search(pat, band):
-            base = name
-            break
-    if base is None:
-        return None
-    charts = re.findall(r'data-chart="([^"]+)"', band)
-    if charts:
-        # 去重保序，最多 2 类，避免顺序噪声
-        seen, uniq = set(), []
-        for c in charts:
-            if c not in seen:
-                seen.add(c)
-                uniq.append(c)
-        return f"{base}:{'+'.join(uniq[:2])}"
-    return base
-
-
-def _check_content_quality(txt, chk, mode, model):
-    """内容级质量：版式节奏 / so-what 非空洞 / research 行动标题含判断。"""
-    cq = LC.get('contentQuality') or {}
-    bands = _bands(txt)
-    skip = set((cq.get('rhythm') or {}).get('skipIds') or
-               ['cover', 'agenda', 'refs', 'appendix', 'next'])
-
-    # ① 版式节奏：同一签名不得连用超过上限
-    max_streak = int((cq.get('rhythm') or {}).get('maxSameLayoutStreak') or 2)
-    sigs: list[tuple[int, str]] = []
-    for i, b in enumerate(bands, 1):
-        head = b[:220]
-        if any(f'id="{sid}"' in head for sid in skip):
-            continue
-        sig = _layout_sig(b)
-        if sig:
-            sigs.append((i, sig))
-    streak_hits: list[str] = []
-    run_sig, run_start, run_len = None, 0, 0
-    for i, sig in sigs:
-        if sig == run_sig:
-            run_len += 1
-        else:
-            if run_sig and run_len > max_streak:
-                streak_hits.append(f'{run_sig}×{run_len}（起第{run_start}页）')
-            run_sig, run_start, run_len = sig, i, 1
-    if run_sig and run_len > max_streak:
-        streak_hits.append(f'{run_sig}×{run_len}（起第{run_start}页）')
-    chk(f"版式节奏（同一版式签名连续 ≤{max_streak} 页）", not streak_hits,
-        "; ".join(streak_hits[:3]) if streak_hits else "", level="WARN")
-
-    # ② so-what 禁空洞套话 + 最短长度（HTML 与模型双侧）
-    sw = cq.get('sowhat') or {}
-    min_chars = int(sw.get('minChars') or 12)
-    forbidden = list(sw.get('forbidden') or [])
-    texts: list[tuple[str, str]] = []
-    for m in re.finditer(r'class="sowhat__v"[^>]*>(.*?)</', txt, re.S):
-        body = re.sub(r'<[^>]+>', '', m.group(1) or '').strip()
-        if body:
-            texts.append(('html', body))
-    if model:
-        for idx, sec in enumerate(model.get('sections') or [], 1):
-            v = sec.get('soWhat')
-            if isinstance(v, str) and v.strip():
-                texts.append((f'model#{idx}', v.strip()))
-    short, hollow = [], []
-    for src, body in texts:
-        if len(body) < min_chars:
-            short.append(f'{src}:{body[:24]}')
-        if any(f in body for f in forbidden):
-            hollow.append(f'{src}:{body[:24]}')
-    if texts:
-        chk(f"so-what 实质（长度 ≥{min_chars} 字且无空洞套话）", not short and not hollow,
-            "; ".join((short + hollow)[:4]) if (short or hollow) else "", level="WARN")
-
-    # ③ research 行动标题须含数字或判断词
-    if mode == 'research':
-        rt = cq.get('researchTitle') or {}
-        if rt.get('requireDigitOrJudgment', True):
-            jpat = rt.get('judgmentPattern') or r'[\d]|是|应|需|将'
-            struct = ('报告大纲', '大纲', '参考资料', '数据来源', '下一步', '结论',
-                      '全文核心', 'agenda', '附录', '收尾')
-            h1s = re.findall(r'<h2 class="t-h1 shead__title"[^>]*>(.*?)</h2>', txt)
-            titles = [re.sub(r'<[^>]+>', '', h).strip() for h in h1s]
-            titles = [t for t in titles if len(t) >= 12 and
-                      not t.lower().startswith(struct) and not any(s in t for s in struct)]
-            no_judge = [t[:28] for t in titles if not re.search(jpat, t)]
-            chk("research 行动标题含数字或判断词（标题即结论）", not no_judge,
-                f"缺少判断信号: {no_judge[:3]}" if no_judge else "", level="WARN")
-
-
-def _check_v9_hard_gates(txt, chk, model):
-    """v9 硬门禁：标签泄漏 / 标题空页 / 极偏占比 / 简单图过大（治截图级缺陷）。"""
-    import html as _html
-    QG = LC.get('qualityGates') or {}
-    patterns = QG.get('tagLeakPatterns') or [
-        '<a ', '</a>', 'class="cite"', 'href=', '<strong', '</strong>'
-    ]
-    # ① HTML 标签泄漏进可见文本
-    # 先抓「转义后当字面量显示」的（&lt;a class=…&gt;）：源码里就有实体
-    leak_hits = []
-    for i, b in enumerate(_bands(txt), 1):
-        b2 = re.sub(r'<script\b[\s\S]*?</script>', ' ', b)
-        b2 = re.sub(r'<style\b[\s\S]*?</style>', ' ', b2)
-        # A. 转义标签字面量（用户截图形态：正文里直接可见 <a class="cite"…>）
-        if re.search(r'&lt;/?[a-zA-Z][^&]{0,60}&gt;', b2):
-            m = re.search(r'&lt;/?[a-zA-Z][^&]{0,60}&gt;', b2)
-            leak_hits.append(f"第{i}页含转义标签字面量 `{m.group(0)[:40]}`")
-            continue
-        # B. 反转义后剥真标签，残留源码片段
-        unesc = _html.unescape(b2)
-        plain = _plain(unesc)
-        for p in patterns:
-            if p in plain:
-                leak_hits.append(f"第{i}页可见文本含 `{p}`")
-                break
-    # C. REPORT_MODEL 字段内夹带 HTML 标签（在 <script> 里，正文剥离会漏掉）
-    model_tag_hits = []
-    mm = re.search(r'window\.REPORT_MODEL\s*=\s*(\{[\s\S]*?\})\s*;', txt)
-    if mm:
-        try:
-            import json as _json
-            mobj = _json.loads(mm.group(1))
-
-            def _scan(o, path=''):
-                if isinstance(o, dict):
-                    for k, v in o.items():
-                        _scan(v, f'{path}.{k}' if path else k)
-                elif isinstance(o, list):
-                    for i, v in enumerate(o):
-                        _scan(v, f'{path}[{i}]')
-                elif isinstance(o, str) and re.search(r'</?[a-zA-Z][^>]*>', o):
-                    model_tag_hits.append(path or '(root)')
-            _scan(mobj)
-        except Exception:
-            pass
-    if model_tag_hits:
-        leak_hits.append(f"REPORT_MODEL 字段夹带标签: {model_tag_hits[:3]}")
-    chk("HTML_TAG_IN_TEXT 可见文本无 HTML 标签源码泄漏", not leak_hits,
-        "; ".join(leak_hits[:5]) if leak_hits else "")
-
-    # ② 标题空页 / 承载不足
-    title_only, underfill = [], []
-    to_min = int((QG.get('titleOnly') or {}).get('minBodyChars') or 40)
-    uf = QG.get('underfill') or {}
-    uf_min = int(uf.get('minCarriers') or 2)
-    exempt = set(uf.get('exemptIds') or ['quote', 'cover', 'refs', 'appendix', 'next'])
-    for i, b in enumerate(_bands(txt), 1):
-        head = b[:240]
-        bid_m = re.search(r'id="([^"]+)"', head)
-        bid = bid_m.group(1) if bid_m else ''
-        if bid in exempt or any(f'id="{e}"' in head for e in exempt):
-            continue
-        if 'band--deep' in head or 'band--accent' in head:
-            continue
-        b2 = re.sub(r'<script\b[\s\S]*?</script>', ' ', b)
-        b2 = re.sub(r'<style\b[\s\S]*?</style>', ' ', b2)
-        # 去掉 shead（eyebrow+标题+导语）后的正文
-        body = re.sub(r'<div class="shead[\s\S]*?</div>\s*</div>', ' ', b2, count=1)
-        body = re.sub(r'<div class="shead[\s\S]*?</h2>\s*</div>', ' ', b2, count=1)
-        plain_body = _plain(body).strip()
-        # 再去掉与标题重复的大标题串
-        title_m = re.search(r'shead__title[^>]*>(.*?)</', b2)
-        if title_m:
-            plain_body = plain_body.replace(_plain(title_m.group(1)).strip(), '', 1).strip()
-        carriers = _band_units(b)
-        if len(plain_body) < to_min and carriers == 0:
-            title_only.append(f"第{i}页正文{len(plain_body)}字")
-        elif carriers < uf_min and len(plain_body) < to_min * 2:
-            underfill.append(f"第{i}页承载{carriers}")
-    chk(f"TITLE_ONLY_PAGE 非空页（去页头后正文 ≥{to_min} 字）", not title_only,
-        "; ".join(title_only[:5]) if title_only else "")
-    chk(f"UNDERFILL_PAGE 内容页承载 ≥{uf_min}（金句/章节幕豁免）", not underfill,
-        "; ".join(underfill[:5]) if underfill else "")
-
-    # ③ 极偏占比禁 donut/pie（治 0.5% vs 99.5% 环图叠字不可读）
-    skew = QG.get('chartSkew') or {}
-    skew_types = set(skew.get('types') or ['donut', 'pie', 'multidonut'])
-    min_pct = float(skew.get('minSectorPct') or 5)
-    max_ratio = float(skew.get('maxMinRatio') or 20)
-    skew_hits = []
-    if model and isinstance(model.get('sections'), list):
-        for si, sec in enumerate(model['sections'], 1):
-            if not isinstance(sec, dict):
-                continue
-            ch = sec.get('chart') or {}
-            ctype = str(ch.get('type') or sec.get('type') or '').lower()
-            if sec.get('type') == 'donut':
-                ctype = 'donut'
-            if ctype not in skew_types:
-                continue
-            vals = [float(v) for v in (ch.get('values') or []) if v is not None]
-            vals = [v for v in vals if v >= 0]
-            if len(vals) < 2:
-                continue
-            total = sum(vals) or 1
-            pcts = [v / total * 100 for v in vals]
-            mn, mx = min(pcts), max(pcts)
-            ratio = (mx / mn) if mn > 0 else 999
-            if mn < min_pct or ratio > max_ratio:
-                skew_hits.append(
-                    f"第{si}章 {ctype} 最小扇区 {mn:.1f}%（比 {ratio:.0f}:1）应改 KPI/进度/对比条")
-    chk(f"CHART_SKEW_INVALID 占比图最小扇区 ≥{min_pct:.0f}% 且 max/min ≤{max_ratio:.0f}",
-        not skew_hits, "; ".join(skew_hits[:4]) if skew_hits else "")
-
-    # ④ 简单图过大（类别≤2 或 数据点≤3 却近乎独占内容区）
-    ov = QG.get('chartOversize') or {}
-    simple_cats = int(ov.get('simpleMaxCats') or 2)
-    simple_pts = int(ov.get('simpleMaxPts') or 3)
-    max_pct_area = float(ov.get('maxContentAreaPct') or 55)
-    oversize = []
-    for i, b in enumerate(_bands(txt), 1):
-        for m in re.finditer(
-                r'<svg\b[^>]*data-chart="([^"]+)"[^>]*style="([^"]*)"[^>]*>',
-                b):
-            ctype, style = m.group(1), m.group(2)
-            # 解析 width / height（px 或 %）
-            wm = re.search(r'width:\s*(\d+(?:\.\d+)?)px', style)
-            hm = re.search(r'height:\s*(\d+(?:\.\d+)?)px', style)
-            vw = re.search(r'viewBox="0 0 (\d+(?:\.\d+)?) (\d+(?:\.\d+)?)"', m.group(0))
-            # 从 svg 块粗数数据点/类别
-            svg_end = b.find('</svg>', m.end())
-            block = b[m.start():svg_end if svg_end > 0 else m.end() + 800]
-            n_labels = len(re.findall(r'<text\b', block))
-            n_rects = len(re.findall(r'<rect\b', block))
-            n_circles = len(re.findall(r'<circle\b', block))
-            n_pts = max(n_labels, n_rects, n_circles)
-            is_simple = n_labels <= simple_cats or n_pts <= simple_pts
-            if not is_simple:
-                continue
-            # 简单图接近通宽且高度大 → 视为「简单大图」
-            h_px = float(hm.group(1)) if hm else (float(vw.group(2)) if vw else 0)
-            if h_px >= 320 or (not hm and not vw and 'width:100%' in style and 'height' not in style):
-                # 无显式高度的通宽简单图
-                if h_px >= 320 or ('width:100%' in style and n_pts <= simple_pts):
-                    oversize.append(f"第{i}页 {ctype} 简单图过大（{n_pts} 点/标记）")
-    chk(f"CHART_OVERSIZE 简单图（≤{simple_cats} 类或 ≤{simple_pts} 点）勿独占版面（上限 {max_pct_area:.0f}%）",
-        not oversize, "; ".join(oversize[:4]) if oversize else "")
-
-
-def _check_layout_grammar(txt, chk, mode):
-    """布局语法门禁：骨架类 / 单一重心 / 混排对齐 / 间距 token / 图标尺寸 / 标签防换行。"""
-    LS = LC.get('layoutSystem') or {}
-    presets = LS.get('presets') or {}
-    elem = LS.get('elements') or {}
-    mf = LS.get('multiFocus') or {}
-    align_cfg = LS.get('align') or {}
-    fill_t = (LS.get('fillTarget') or {}).get(mode) or [62, 78]
-    icon_sizes = set(elem.get('icon', {}).get('sizesPx') or [16, 18, 20, 24])
-    chart_w_pct = float(mf.get('chartWidthPct') or 55)
-    metric_fz = float(mf.get('metricFontPx') or 40)
-
-    # 收集预设骨架对应 class
-    skeleton_classes = set()
-    for p in (presets or {}).values():
-        for c in (p.get('html') or []):
-            skeleton_classes.add(c)
-    # 附加常见栅格类（grammar 白名单）
-    skeleton_classes.update({
-        'g-2', 'g-3', 'g-4', 'g-5', 'g-6', 'g-hero', 'g-hero--rev', 'g-side', 'g-side--rev',
-        'g-31', 'g-13', 'g-41', 'g-14', 'g-211', 'g-121', 'g-quad', 'g-aside', 'g-bento',
-        'g-hero-full', 'g-mosaic', 'rows-2', 'rows-3', 'stagger', 'media-grid', 'media-compare',
-    })
-
-    no_skel, multi_focus, align_miss, off_token = [], [], [], []
-    icon_bad, label_collapse = [], []
-    fill_low, fill_high = [], []
-    mixed_need = bool(align_cfg.get('requireOnMixed', True))
-    align_tokens = set(align_cfg.get('mixedGridClasses') or ['a-start', 'a-c', 'align-items'])
-
-    for i, b in enumerate(_bands(txt), 1):
-        head = b[:240]
-        if any(f'id="{x}' in head for x in ('refs', 'appendix', 'cover', 'agenda', 'quote')):
-            continue
-        if 'band--flow' in head:
-            continue
-        # 轻量合法页：金句/强调带/大数压场（fill 目标放宽）
-        is_light = ('band--fit' in head or 'band--accent' in head or
-                    'band--deep' in head or 'quote' in head)
-
-        # ① 骨架类（或 data-skel 标记）
-        has_skel = ('data-skel="' in b) or any(
-            re.search(rf'class="[^"]*\b{re.escape(c)}\b', b) for c in skeleton_classes)
-        # 手写 inline grid 且无预设类
-        inline_grid = re.findall(r'style="[^"]*display:\s*grid[^"]*"', b)
-        if not has_skel and inline_grid:
-            no_skel.append(f"第{i}页 inline grid 无 P1–P12 骨架类")
-
-        # ② 单一视觉重心（粗启发式）
-        n_big = 0
-        for m in re.finditer(r'<svg\b[^>]*>', b):
-            tag = m.group(0)
-            wm = re.search(r'width:\s*(\d+(?:\.\d+)?)px', tag)
-            wrap = int((re.search(r'--wrap:\s*(\d+)', txt) or [0, 1400])[1] or 1400) if False else 1400
-            if wm and float(wm.group(1)) / wrap * 100 >= chart_w_pct:
-                n_big += 1
-            if re.search(r'width:\s*100%', tag) and 'data-chart' in tag:
-                # 通宽图记 1 大件
-                n_big += 1
-        for m in re.finditer(r'font-size:\s*clamp\(\s*(\d+)px', b):
-            if float(m.group(1)) >= metric_fz:
-                n_big += 1
-                break
-        if re.search(r'class="[^"]*media--(?:full|bleed)', b):
-            n_big += 1
-        if n_big >= 2:
-            multi_focus.append(f"第{i}页大件×{n_big}")
-
-        # ③ 混排对齐（同时有图/媒体与卡/列表）
-        mixed = (re.search(r'class="[^"]*(?:fig|media|chart)', b) and
-                 re.search(r'class="[^"]*(?:card|ul|metric)', b))
-        if mixed and mixed_need:
-            if not any(t in b for t in align_tokens):
-                align_miss.append(f"第{i}页图卡混排缺 a-start/a-c")
-
-        # ④ 间距写死（margin/padding/gap 非 token / 非 clamp；≤8px 微调白名单）
-        for m in re.finditer(r'(?:margin|padding|gap)(?:-[a-z]+)?\s*:\s*([^;"]+)', b):
-            val = m.group(1).strip()
-            if not val or val.startswith('var(') or val.startswith('clamp(') or val.startswith('0'):
-                continue
-            if re.fullmatch(r'auto|inherit|initial|unset|normal', val):
-                continue
-            px_vals = [float(x) for x in re.findall(r'(\d+(?:\.\d+)?)px', val)]
-            if px_vals and all(v <= 8 for v in px_vals) and 'clamp' not in val and 'var(' not in val:
-                continue  # 4–8px 光学微调允许
-            if px_vals and 'clamp' not in val and 'var(' not in val:
-                off_token.append(f"第{i}页 `{val[:24]}`")
-
-        # ⑤ 图标尺寸
-        for m in re.finditer(
-                r'<svg\b[^>]*(?:class="[^"]*(?:ico|icon)[^"]*"|metric__ico|card__ico)[^>]*>',
-                b, re.I):
-            tag = m.group(0)
-            sm = re.search(r'width:\s*["\']?(\d+)', tag) or re.search(r'width="(\d+)"', tag)
-            if sm and int(sm.group(1)) not in icon_sizes:
-                icon_bad.append(f"第{i}页 {sm.group(1)}px")
-
-        # ⑥ 图表标签防换行（短盒 + 长 text）
-        for m in re.finditer(r'<text[^>]*width="(\d+)"[^>]*>([^<]{8,})</text>', b):
-            if int(m.group(1)) < 48:
-                label_collapse.append(f"第{i}页标签盒宽{m.group(1)}")
-
-        # ⑦ 结构图禁裸文字箭头（F13）
-        if re.search(r'class="[^"]*lane__arr[^"]*"[^>]*>\s*→', b) or re.search(
-                r'lane__arr">→', b):
-            label_collapse.append(f"第{i}页泳道裸文字→（改 .lane__arr）")
-
-        # ⑧ 填充率粗估（去空白字符 + 组件；轻量页不判过空）
-        b2 = re.sub(r'<script\b[\s\S]*?</script>', ' ', b)
-        b2 = re.sub(r'<style\b[\s\S]*?</style>', ' ', b2)
-        plen = len(re.sub(r'\s+', '', _plain(b2)))
-        units = _band_units(b)
-        budget = MODE_BUDGETS.get(mode, MODE_BUDGETS['presentation'])
-        est = (plen / max(1, budget['fit']) * 55) + (units / max(1, budget['unit']) * 45)
-        est = max(0, min(100, est))
-        lo_t = fill_t[0] * 0.55
-        hi_t = min(95, fill_t[1] * 1.15)
-        if (not is_light) and est < lo_t and units == 0 and plen < budget['empty']:
-            fill_low.append(f"第{i}页≈{est:.0f}%")
-        elif est > hi_t and not is_light:
-            fill_high.append(f"第{i}页≈{est:.0f}%")
-
-    chk(f"LAYOUT_NO_SKELETON 内容页使用 P1–P12 骨架类（layout-grammar）", not no_skel,
-        "; ".join(no_skel[:4]) if no_skel else "")
-    chk(f"LAYOUT_MULTI_FOCUS 一屏一视觉重心（大件 ≤{mf.get('maxPrimary', 1)}）", not multi_focus,
-        "; ".join(multi_focus[:4]) if multi_focus else "")
-    chk("LAYOUT_ALIGN_DRIFT 图卡混排显式对齐（a-start / a-c）", not align_miss,
-        "; ".join(align_miss[:4]) if align_miss else "")
-    chk("LAYOUT_SPACING_OFF_TOKEN 间距走 --sp-*/gap/clamp（禁游离 px）", not off_token,
-        "; ".join(off_token[:4]) if off_token else "", level="WARN")
-    chk("LAYOUT_ICON_SIZE 图标 ∈ {16,18,20,24}px", not icon_bad,
-        "; ".join(icon_bad[:4]) if icon_bad else "", level="WARN")
-    chk("LAYOUT_LABEL_COLLIDE 图表标签盒足够宽 / 结构图无裸文字箭头", not label_collapse,
-        "; ".join(label_collapse[:4]) if label_collapse else "")
-    chk(f"LAYOUT_FILL 填充率目标 {fill_t[0]}–{fill_t[1]}%（{mode}）",
-        not fill_low and not fill_high,
-        ("过空: " + "; ".join(fill_low[:3]) if fill_low else "") +
-        ((" 过满: " + "; ".join(fill_high[:3])) if fill_high else ""))
-
-
-def _check_annotations(txt, chk, model):
-    """待核实标注：.tbd 内联标色须配 .tbd-legend / .flagbar 说明；单页数量上限。"""
-    an = LC.get('annotations') or {}
-    max_per_page = an.get('flagMaxPerPage', 12)
-    require_legend = an.get('requireLegend', True)
-    tbd_re = re.compile(r'class="tbd(?:\s|")')
-    no_legend, over = [], []
-    for i, b in enumerate(_bands(txt), 1):
-        n = len(tbd_re.findall(b))
-        if n and require_legend and ('tbd-legend' not in b and 'flagbar' not in b):
-            no_legend.append(f"第{i}页")
-        if n > max_per_page:
-            over.append(f"第{i}页 {n}处")
-    chk("待核实标注 .tbd 均配 .tbd-legend/.flagbar 说明", not no_legend,
-        f"{no_legend} 有标色项但无说明（用户不知为何标色）" if no_legend else "", level="WARN")
-    chk(f"单页待核实标注 ≤ {max_per_page} 处（过密则转 .flagbar 清单）", not over,
-        "; ".join(over) if over else "", level="WARN")
-    if model:
-        mflags = [s for s in (model.get('sections') or []) if s.get('flags')]
-        if mflags and 'flagbar' not in txt and 'tbd-legend' not in txt:
-            chk("模型 flags ↔ 正文待核实条对应", False,
-                f"{len(mflags)} 页模型含 flags，正文无 .flagbar/.tbd-legend", level="WARN")
-
-
-def _img_holders(sec):
-    """页内可能承载素材图片的容器（image 页型 / split 右栏）。"""
-    out = []
-    if isinstance(sec.get('image'), dict):
-        out.append(sec['image'])
-    right = sec.get('right')
-    if isinstance(right, dict) and isinstance(right.get('image'), dict):
-        out.append(right['image'])
-    return out
-
-
-def _img_srcs(img):
-    """图片对象的全部真实 src（image.src + image.items[].src），排除占位符。"""
-    if img.get('placeholder'):
-        return []
-    srcs = []
-    if isinstance(img.get('src'), str) and img['src'].strip():
-        srcs.append(img['src'])
-    for it in (img.get('items') or []):
-        if isinstance(it, dict) and isinstance(it.get('src'), str) and it['src'].strip():
-            srcs.append(it['src'])
-        elif isinstance(it, str) and it.strip():
-            srcs.append(it)
-    return srcs
-
-
-def _check_media(txt, chk, model):
-    """素材图片与配图占位：零外链铁律（<img> 只允许 data: 内联或相对路径）；
-    alt 可访问性；配图占位必须有可见标签；模型 image 三选一（src/items/placeholder）
-    与正文版式、数量、裁切策略一一对应。"""
-    imgs = re.findall(r'<img\b[^>]*>', txt)
-    if imgs:
-        bad = [t for t in imgs if re.search(r'src\s*=\s*["\']\s*(?:https?:)?//', t)]
-        chk("图片源无外链（data: 内联或相对路径，零外链铁律）", not bad,
-            f"{len(bad)} 处外链图片" if bad else "")
-        noalt = [t for t in imgs if 'alt=' not in t]
-        chk("图片均带 alt（可访问性）", not noalt,
-            f"{len(noalt)} 处缺 alt" if noalt else "", level="WARN")
-        data_imgs = [t for t in imgs if re.search(r'src\s*=\s*["\']\s*data:', t)]
-        inline_bytes = sum(len(t) for t in data_imgs)
-        chk("单图 data: 内联体积在上限内", all(len(t) <= IMAGE_MAX_INLINE for t in data_imgs),
-            f"最大 {max((len(t) for t in data_imgs), default=0) // 1024}KB > 上限 {IMAGE_MAX_INLINE // 1024}KB",
-            level="WARN")
-        chk("报告内联图片总量在上限内", inline_bytes <= IMAGE_MAX_TOTAL,
-            f"{inline_bytes // 1024}KB > 上限 {IMAGE_MAX_TOTAL // 1024}KB（改用相对路径）",
-            level="WARN")
-    ph_blocks = re.findall(r'class="[^"]*\bmedia--ph\b[^"]*"', txt)
-    if ph_blocks:
-        chk("配图占位含可见标签（.media__ph）", 'media__ph' in txt,
-            f"{len(ph_blocks)} 处 .media--ph 缺 .media__ph 标签（空占位 = 不合格）")
-    if not model:
-        return
-    holders = [(i, img) for i, s in enumerate(model.get('sections') or [])
-               if isinstance(s, dict) for img in _img_holders(s)]
-    if not holders:
-        return
-    bad_src, bad_layout, bad_fit, too_many = [], [], [], []
-    n_declared = 0
-    n_ph = 0
-    for idx, img in holders:
-        srcs = _img_srcs(img)
-        n_declared += len(srcs)
-        if img.get('placeholder'):
-            n_ph += 1
-        for s in srcs:
-            if re.match(r'\s*(?:https?:)?//', str(s)):
-                bad_src.append(f"sections[{idx}]")
-        layout = str(img.get('layout') or ('grid' if len(img.get('items') or []) > 1 else 'full')).lower()
-        if layout not in IMAGE_LAYOUTS:
-            bad_layout.append(f"sections[{idx}]:{layout}")
-        if img.get('fit') and str(img['fit']).lower() not in IMAGE_FIT:
-            bad_fit.append(f"sections[{idx}]:{img['fit']}")
-        items = img.get('items') or []
-        if len(items) > IMAGE_MAX_PER_PAGE:
-            too_many.append(f"sections[{idx}]={len(items)}")
-        if layout in IMAGE_MULTI_LAYOUTS and not img.get('placeholder') and len(items) < 2:
-            too_many.append(f"sections[{idx}] {layout} 需 ≥2 张图（当前 {len(items)}）")
-    chk("模型 image.src 非外链（data:/相对路径）", not bad_src,
-        f"{len(bad_src)} 处外链 src" if bad_src else "")
-    chk("模型 image.layout 合法（imageSpec.layouts）", not bad_layout,
-        f"非法版式 {sorted(set(bad_layout))}" if bad_layout else "")
-    chk("模型 image.fit 合法（cover/contain）", not bad_fit,
-        f"非法裁切 {sorted(set(bad_fit))}" if bad_fit else "")
-    chk("单页图片数/多图版式图数合规（imageSpec.maxPerPage · 多图版式 ≥2 张）", not too_many,
-        f"{too_many}" if too_many else "", level="WARN")
-    layout_class = {'grid': 'media-grid', 'compare': 'media-compare', 'wall': 'media-wall',
-                    'bleed': 'media--bleed', 'ph': 'media--ph'}
-    ratio_css = IMAGE_SPEC.get('ratioCssClass') or {}
-    miss_cls = []
-    for idx, img in holders:
-        lay = str(img.get('layout') or ('grid' if len(img.get('items') or []) > 1 else 'full')).lower()
-        cls = layout_class.get(lay)
-        if cls and cls not in txt:
-            miss_cls.append(f"sections[{idx}]:{lay}→.{cls}")
-        # 比例锁定类：HTML 与 PPTX 用同一份 imageSpec.ratioCssClass，缺了就会出现"比例走样"
-        rcls = ratio_css.get(lay)
-        if rcls and ('.' + rcls) not in txt:
-            miss_cls.append(f"sections[{idx}]:{lay}→.{rcls}（比例未锁定）")
-        if img.get('placeholder') and 'media--ph' not in txt:
-            miss_cls.append(f"sections[{idx}]:placeholder→.media--ph")
-    chk("模型图片版式 ↔ 正文版式类/比例锁定类对应", not miss_cls,
-        f"缺 {sorted(set(miss_cls))}" if miss_cls else "", level="WARN")
-    if n_declared and len(imgs) < n_declared:
-        chk("模型 image ↔ 正文图片数量对应", False,
-            f"模型声明 {n_declared} 张但正文仅 {len(imgs)} 个 <img>", level="WARN")
-    if n_ph and not ph_blocks:
-        chk("模型配图占位 ↔ 正文 .media--ph 对应", False,
-            f"{n_ph} 页模型声明 image.placeholder 但正文无 .media--ph", level="WARN")
-
-
-# 失败检查项 → 失败模式 / 处置动作 / 精确取码命令。
-# 目的：校验失败时直接给出「改什么、按什么顺序改、去哪取代码」，
-# 使智能体不必整读 references/failure-modes.md（17KB）就能收敛。
-# 关键词按检查项名称匹配；新增检查项时同步在此登记，否则只回落通用处置顺序。
-FIX_GUIDE = [
-    (("页高", "溢出", "满屏", "文字预算"), "F4 容器溢出",
-     "① 列表化/精炼 ② 升级承载形态 ③ 换/扩组合版式 ④ 分区 ⑤ 拆页 ⑥ 最后才有限缩字号",
-     "--task content-rules"),
-    (("过空", "内容不足", "密度"), "F1 内容不足 / F3 密度塌陷",
-     "补证据与含义（数字+口径+so-what），不要用装饰或放大字号填空",
-     "--task content-rules"),
-    (("图表", "多样性", "登记"), "F6 图表降级 / F12 图表单一",
-     "换图表类型拉开多样性；连续两页不得同型；类型须在 charts.registry 登记",
-     "--task chart-pick"),
-    (("最小尺寸",), "图表尺寸不足",
-     "把 svg 的 viewBox 高/显示宽度提到 charts.minSize 之上（尺寸问题改尺寸，不要靠换图型回避）",
-     "--task chart-pick"),
-    (("Exhibit",), "研究模式证据编号",
-     "每个 .exhibit 框都要有 .exhibit__no（Exhibit N，全篇连续）与 .exhibit__src 来源行",
-     "--task research-evidence"),
-    (("引用", "参考资料", "锚点"), "引用闭环",
-     "正文 [n] 上标与文末条目双向对齐且编号从 1 连续；ref-link 带 target/rel",
-     "--file components-atoms.md --section 7"),
-    (("待核实", "tbd"), "待核实标注",
-     "每处 .tbd 必须配 .tbd-legend 或 .flagbar 说明口径，只标色不解释即不合格",
-     "--file components-atoms.md --section 11b"),
-    (("强调", "风格", "主题", "配色"), "F8 主题/风格漂移",
-     "风格是皮肤不是解药：末页禁 band--deep，收尾用 band--accent；单一强调色",
-     "--file components-atoms.md --section 11"),
-    (("so-what", "结论", "标题"), "F15 空洞结论",
-     "research 主标题须是结论句（≥12 字含数字或判断词）；so-what 禁套话",
-     "--task content-rules"),
-    (("版式", "节奏", "组合"), "F11 单件页默认",
-     "默认一页=主件+从件+注释；同一版式不连用超 2 页，密度 L/M/H 交替",
-     "--task presentation-combo"),
-    (("图片", "占位", "media"), "F9 配图走样 / F10 空占位",
-     "无素材用 image.placeholder + .media--ph 锁版式；路径以模型目录为锚；禁外链",
-     "--task image-layout"),
-    (("AI", "去AI味"), "文风",
-     "改写命中的高危词；大段文字转列表",
-     "--task content-rules"),
-]
-
-
-def _print_fix_guide(results, strict, width):
-    """校验未通过时输出定向修复指引（只列命中的失败模式，不做全量倾倒）。"""
-    failed = [(lv, nm, nt) for lv, nm, ok, nt in results
-              if not ok and (lv == "FAIL" or strict)]
-    if not failed:
-        return
-    seen, guides = set(), []
-    for _, name, _ in failed:
-        nl = name.lower()
-        # 最长关键词优先（而非首个命中即 break）：避免「图表最小尺寸」被泛化的
-        #「图表」关键词抢走、误路由到「换图表类型」
-        best = None
-        for keys, mode, action, cmd in FIX_GUIDE:
-            hit = max((len(k) for k in keys if k.lower() in nl), default=0)
-            if hit and (best is None or hit > best[0]):
-                best = (hit, mode, action, cmd)
-        if best and best[1] not in seen:
-            seen.add(best[1])
-            guides.append((best[1], best[2], best[3]))
-    print("-" * width)
-    print("修复指引（按下列顺序改；跳步直接缩字号/砍内容会把问题推给下一环）：")
-    if guides:
-        for i, (mode, action, cmd) in enumerate(guides, 1):
-            print(f"  {i}. [{mode}] {action}")
-            print(f"     取码: python scripts/extract_snippet.py {cmd}")
-    else:
-        print("  未匹配到已登记的失败模式，按通用顺序处置：")
-        print("  ① 补内容 ② 优化形态（列表化/精炼）③ 换承载形态 "
-              "④ 调容器/网格 ⑤ 有限缩字号 ⑥ 最后拆页")
-    print("  完整失败模式库与错误解释纠正表: references/failure-modes.md")
-
-
-def main():
-    if len(sys.argv) < 2:
-        print(__doc__)
-        return 2
-    path = Path(sys.argv[1])
-    strict = '--strict' in sys.argv
-    as_json = '--json' in sys.argv
-    if not path.exists():
-        print(f"文件不存在: {path}")
-        return 2
-    txt = path.read_text(encoding='utf-8')
-
-    v = Struct()
-    v.feed(txt)
-
-    results = []
-
-    def chk(name, ok, note="", level="FAIL"):
-        results.append((level, name, bool(ok), note))
-
-    # ── 结构 ──
-    chk("HTMLParser 结构 0 未闭合", not v.stack, f"剩余 {v.stack}" if v.stack else "")
-    chk("HTMLParser 0 失配/游离", not v.mis and not v.err,
-        f"mis={v.mis} err={v.err}" if (v.mis or v.err) else "")
-
-    # ── 主题与风格 ──
-    chk("data-style 已指定风格", 'data-style="' in txt)
-    chk('data-theme="light" + dark 变量块',
-        'data-theme="light"' in txt and '[data-theme="dark"]' in txt)
-    chk("主题切换按钮 + localStorage + prefers-color-scheme",
-        'id="themeBtn"' in txt and 'localStorage' in txt and 'prefers-color-scheme' in txt)
-
-    # ── 模式 ──
-    mode_m = re.search(r'<html[^>]*data-mode="([^"]+)"', txt)
-    mode = mode_m.group(1) if mode_m else None
-    chk("data-mode 已声明模式（presentation/research/architecture）",
-        mode in MODE_BUDGETS,
-        f"读到 {mode!r}（未声明按 presentation 处理）" if mode not in MODE_BUDGETS else "",
-        level="WARN" if mode is None else "FAIL")
-    mode = mode if mode in MODE_BUDGETS else 'presentation'
-
-    # ── 宽屏与页面高度模型 ──
-    B0 = MODE_BUDGETS[mode]
-    _w = B0['wrap']
-    chk(f"版心 --wrap: {_w}px（{mode} 模式）",
-        f'--wrap:{_w}px' in txt or f'--wrap: {_w}px' in txt)
-    _check_page_model(txt, chk, mode)
-    chk("scroll-snap 翻页停靠", 'scroll-snap-type' in txt)
-
-    # ── 翻页与大纲 ──
-    chk("翻页 JS（方向键 + scrollIntoView）",
-        'scrollIntoView' in txt and 'ArrowRight' in txt)
-    chk("页码指示 pager", 'pager__dot' in txt and 'id="pagerDots"' in txt)
-    has_agenda = 'id="agenda"' in txt
-    if mode == 'architecture':
-        # architecture 极简形态：内容页 ≤4 时省略 Agenda 合法（封面/收尾/参考资料不计内容页）
-        struct_ids = ('agenda', 'next', 'refs', 'appendix')
-        bands_a = _bands(txt)
-        content_n = 0
-        for i, b in enumerate(bands_a):
-            if i == 0:
-                continue
-            if any(f'id="{sid}"' in b[:220] for sid in struct_ids):
-                continue
-            content_n += 1
-        if content_n > 4:
-            chk("存在 Agenda 大纲页（architecture 内容页 >4 时建议补）", has_agenda,
-                f"内容页 {content_n} 页无 Agenda", level="WARN")
-    else:
-        chk("存在 Agenda 大纲页", has_agenda)
-    if has_agenda:
-        band_secs = [m.start() for m in re.finditer(r'<section class="band', txt)]
-        agenda_pos = txt.find('id="agenda"')
-        second_band_pos = band_secs[1] if len(band_secs) > 1 else -1
-        ok_second = (second_band_pos != -1 and
-                     band_secs[0] < agenda_pos < (band_secs[2] if len(band_secs) > 2 else len(txt)))
-        chk("Agenda 位于第二页", ok_second,
-            f"共{len(band_secs)}页 agenda@{agenda_pos} 第2页@{second_band_pos}")
-        n_items = txt.count('class="agenda__a"')
-        chk("Agenda 条目可点击跳转 (≥4)", n_items >= 4, f"{n_items} 条")
-        if n_items > AGENDA_SINGLE_MAX:
-            chk(f"Agenda >{AGENDA_SINGLE_MAX} 条时已用 agenda--2col 双列",
-                'agenda--2col' in txt, f"{n_items} 条未双列", level="WARN")
-
-    # ── 主标题粗体 ──
-    chk("主标题粗体 --fw-title/--fw-display",
-        ('--fw-title' in txt and 'font-weight:var(--fw-title)' in txt) or
-        re.search(r'\.t-h1\{[^}]*font-weight:\s*(600|700)', txt))
-
-    # ── 零外链 ──
-    ext = re.findall(r'<(?:link|script)[^>]+(?:href|src)=["\']https?://[^"\']+', txt)
-    chk("无 <link>/<script src> 外链", not ext, f"{len(ext)} 处" if ext else "")
-    img_ext = re.findall(r'<img[^>]+src=["\']https?://', txt)
-    chk("无外链图片", not img_ext, f"{len(img_ext)} 处" if img_ext else "")
-
-    # ── 引用与锚点闭环 ──
-    # 编号须匹配多位数：`ref-\d` 会让 ref-10 及以后的条目在两侧同时落空，
-    # 使双向对齐检查在 ≥10 条引用时静默通过（假阴性）。
-    cites = re.findall(r'class="cite" href="#(ref-\d+)"', txt)
-    refs = re.findall(r'id="(ref-\d+)"', txt)
-    if cites or refs:
-        chk("引用标记 ↔ 参考资料条目 双向对齐", set(cites) == set(refs),
-            f"引用{sorted(set(cites))} 条目{sorted(set(refs))}")
-        nums = sorted(int(r.split('-')[1]) for r in set(refs))
-        chk(f"参考资料编号连续（1..{len(nums)} 无跳号）", nums == list(range(1, len(nums) + 1)),
-            f"读到 {nums}" if nums != list(range(1, len(nums) + 1)) else "")
-    else:
-        chk("引用标记 ↔ 参考资料条目 双向对齐", True, "（无外部引用）", level="WARN")
-    bad_links = re.findall(r'<a class="ref-link"(?![^>]*target="_blank")[^>]*>', txt) + \
-                re.findall(r'<a class="ref-link"(?![^>]*rel="noopener")[^>]*>', txt)
-    chk('.ref-link 均带 target="_blank" rel="noopener"', not bad_links,
-        f"{len(bad_links)} 处" if bad_links else "")
-    _check_anchors(txt, chk)
-
-    # ── 单一强调色（色值表来自 layout-constants.json styleAccents 单源） ──
-    style_m = re.search(r'<html[^>]*data-style="([^"]+)"', txt)
-    style = style_m.group(1) if style_m else 'business-blue'
-    body_txt = txt[txt.find('</style>'):] if '</style>' in txt else txt
-    body_txt = re.sub(r'<script\b[\s\S]*?</script>', ' ', body_txt)
-    if STYLE_ACCENTS:
-        own = STYLE_ACCENTS.get(style, set())
-        others = ALL_ACCENTS - own
-        bad_hues = [h for h in others if h in body_txt.lower()]
-        if style == 'spectrum':
-            chk("彩色模块边界（spectrum：数据色限 c1–c5，无其它风格强调色）", not bad_hues,
-                f"残留 {bad_hues}" if bad_hues else "")
-        else:
-            chk(f"单一强调色（{style}，正文无第二色相）", not bad_hues,
-                f"残留 {bad_hues}" if bad_hues else "")
-
-    hard_hex = re.findall(r'style="[^"]*(?:color|background|border-color|fill|stroke)\s*:\s*#',
-                          body_txt)
-    chk("正文无内联写死 hex 色（用 var()/语义类）", not hard_hex,
-        f"{len(hard_hex)} 处" if hard_hex else "", level="WARN")
-
-    # ── 内容与密度 ──
-    paras = re.findall(r'<p class="t-body"[^>]*>(.*?)</p>', txt, re.S)
-    long_paras = [x for x in paras if len(re.sub(r'<[^>]+>', '', x).strip()) > 110]
-    chk("卡片内无超 110 字纯段落（应转列表）", not long_paras,
-        f"{len(long_paras)} 段" if long_paras else "")
-    _check_bands(txt, chk, mode)
-    _check_table_rows(txt, chk, mode)
-    _check_mode_layouts(txt, chk, mode)
-    if mode != 'architecture':
-        _check_icons(txt, chk)
-    _check_pptx_export(txt, chk)
-    model = _check_model_consistency(txt, chk, mode, style)
-    _check_type_features(txt, chk, model)
-    _check_exhibits(txt, chk, mode)
-    _check_research_extras(txt, chk, mode)
-    _check_emphasis(txt, chk)
-    _check_annotations(txt, chk, model)
-    _check_media(txt, chk, model)
-
-    if mode == 'research':
-        STRUCT = ('报告大纲', '大纲', '参考资料', '数据来源', '下一步', '结论', '全文核心', 'agenda', '附录')
-        h1s = re.findall(r'<h2 class="t-h1 shead__title"[^>]*>(.*?)</h2>', txt)
-        short = [re.sub(r'<[^>]+>', '', h).strip() for h in h1s]
-        short = [h for h in short if 0 < len(h) < 12 and not h.lower().startswith(STRUCT)]
-        chk("research 行动标题（章节主标题 ≥12 字，标题即结论）", not short,
-            f"过短: {short[:3]}" if short else "", level="WARN")
-
-    _check_content_quality(txt, chk, mode, model)
-    _check_v9_hard_gates(txt, chk, model)
-    _check_layout_grammar(txt, chk, mode)
-
-    # ── 去AI味（词表来自单源） ──
-    body_plain = _plain(body_txt)
-    hits = [w for w in AI_FLAVOR if re.search(w, body_plain)]
-    chk("去AI味（无高危 AI 腔词汇）", not hits, f"命中 {hits}" if hits else "", level="WARN")
-
-    # ── 图表 ──
-    _check_charts(txt, chk)
-    _check_chart_variety(txt, chk, mode)
-
-    # ── 汇总 ──
-    n_fail = sum(1 for lv, _, ok, _ in results if not ok and lv == "FAIL")
-    n_warn = sum(1 for lv, _, ok, _ in results if not ok and lv == "WARN")
-    n_pass = sum(1 for _, _, ok, _ in results if ok)
-    verdict = "通过，可交付" if n_fail == 0 and (not strict or n_warn == 0) else \
-              ("有警告" if n_fail == 0 else "不通过，需修复后重跑")
-    exit_code = 1 if (n_fail > 0 or (strict and n_warn > 0)) else 0
-
-    if as_json:
-        print(json.dumps({
-            "file": str(path),
-            "strict": strict,
-            "pass": n_pass, "warn": n_warn, "fail": n_fail,
-            "verdict": verdict,
-            "checks": [{"level": lv, "name": nm, "ok": ok, "note": nt}
-                       for lv, nm, ok, nt in results],
-        }, ensure_ascii=False, indent=2))
-        return exit_code
-
-    W = 68
-    print("=" * W)
-    print(f"TopPPT HTML· 质量校验 {LC.get('version', '')}  {'[STRICT]' if strict else ''}")
-    print(f"文件: {path.name}  ({len(txt.encode('utf-8'))/1024:.1f} KB)")
-    print("=" * W)
-    for level, name, ok, note in results:
-        tag = "PASS" if ok else ("WARN" if level == "WARN" else "FAIL")
-        line = f"[{tag}] {name}"
-        if note and not ok:
-            line += f"  -> {note}"
-        print(line)
-    print("-" * W)
-    print(f"PASS {n_pass}  WARN {n_warn}  FAIL {n_fail}")
-    print(f"结论: {verdict}")
-    if exit_code != 0:
-        _print_fix_guide(results, strict, W)
-    print("=" * W)
-    return exit_code
-
-
-if __name__ == '__main__':
-    sys.exit(main())
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+TopPPT HTML· HTML 报告质量校验（交付闭环）
+用法:
+    python validate_report.py <报告.html> [--strict] [--json]
+
+输出每项 PASS/FAIL，结尾给汇总与结论；任一 FAIL 时退出码为 1（--strict 时 WARN 也计失败）。
+--json 以 JSON 输出全部检查结果（供脚本/流水线读取）。
+生成流程：生成 → 跑本脚本 → 修复 FAIL → 再跑，直至全部 PASS 才交付。
+
+阈值全部来自单源 scripts/layout-constants.json（checkBudgets / charts / styleAccents / aiFlavor），
+本脚本不再内置任何风格色值或预算数字——改阈值只改 JSON。
+
+检查面：
+  结构闭合 / 主题风格 / 页面高度模型 / 宽屏翻页 / Agenda（architecture 极简形态可省略） / 粗体 /
+  零外链 / 引用与锚点闭环 / 单一强调色 / 内联写死色 / 内容密度 / 单页文字与组件预算 /
+  表格行数上限 / 页高溢出估算 / 去AI味 / 图表存在与动效 / 图表 data-chart 登记 / 图表最小尺寸 /
+  模式版式特征（含新页型 steps/heatmap/bullet/pyramid/image 与 research/architecture 专属） /
+  模型一致性（REPORT_MODEL ↔ 正文） / Exhibit 编号连续 / research so-what 与来源行 /
+  强调带约束（band--deep 反相页 ≤ 上限且不作末页） / 待核实标注（.tbd ↔ .tbd-legend/.flagbar） /
+  素材图片与配图占位（零外链 + alt + 内联体积 + 版式/比例锁定类对应 + 占位可见标签 + 模型对应） /
+  PPTX 预览配套（预览按钮 + 模型 + 运行时，页面无导出按钮） /
+  内容级质量（版式节奏连用上限 / so-what 实质与空洞套话 / research 标题含数字或判断词）
+"""
+import sys
+import re
+import json
+from html.parser import HTMLParser
+from pathlib import Path
+
+# Windows GBK 控制台兜底：强制 UTF-8 输出（含 ↔ 等符号）
+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))
+import checks_html  # noqa: E402  承载正则 / 图表通道 / 多样性 / 组合版式单源
+
+VOID = {'area', 'base', 'br', 'col', 'embed', 'hr', 'img', 'input', 'link', 'meta',
+        'param', 'source', 'track', 'wbr'}
+
+# ── 单源阈值（scripts/layout-constants.json）────────────────────────────────
+_DEFAULTS = {
+    'checkBudgets': {
+        'presentation': dict(char=1500, unit=8, empty=80, fit=1000, wrap=1400, trow=8),
+        'research': dict(char=3200, unit=12, empty=120, fit=2200, wrap=1240, trow=16),
+        'architecture': dict(char=600, unit=3, empty=60, fit=600, wrap=1600, trow=6),
+    },
+    'charts': {'types': [], 'minSize': {}},
+    'styleAccents': {},
+    'aiFlavor': {'words': []},
+    'imageSpec': {},
+    'contentQuality': {},
+}
+
+
+def load_constants():
+    path = Path(__file__).resolve().parent / 'layout-constants.json'
+    try:
+        lc = json.loads(path.read_text(encoding='utf-8'))
+    except (OSError, json.JSONDecodeError) as exc:
+        print(f'警告: 无法读取单源 {path}（{exc}），使用内置默认阈值。')
+        return _DEFAULTS
+    for key, fallback in _DEFAULTS.items():
+        lc.setdefault(key, fallback)
+    return lc
+
+
+LC = load_constants()
+MODE_BUDGETS = LC['checkBudgets']
+CHART_TYPES = set(LC['charts'].get('types') or [])
+CHART_MIN = LC['charts'].get('minSize') or {}
+# 图表多样性预算（治「登记 36 种、实际只用 bar/line/donut」）
+CHART_VARIETY = LC['charts'].get('variety') or {}
+AI_FLAVOR = LC['aiFlavor'].get('words') or []
+STYLE_ACCENTS = {k: {str(h).lower() for h in v}
+                 for k, v in (LC['styleAccents'] or {}).items()
+                 if not k.startswith('$') and isinstance(v, list)}
+ALL_ACCENTS = set().union(*STYLE_ACCENTS.values()) if STYLE_ACCENTS else set()
+# 图表登记四元组（charts.registry）与信息图页型（数据可追溯策略核对）
+CHART_REG = {k: v for k, v in (LC['charts'].get('registry') or {}).items()
+             if not str(k).startswith('$')}
+INFO_PAGE_TYPES = {'sankey', 'treemap', 'boxplot', 'network', 'marimekko', 'streamgraph'}
+# 素材图片规格与配图占位约定（单源 layout-constants.json imageSpec）
+IMAGE_SPEC = LC.get('imageSpec') or {}
+IMAGE_LAYOUTS = set(IMAGE_SPEC.get('layouts') or ['full', 'half', 'bleed', 'grid', 'compare', 'wall'])
+IMAGE_MULTI_LAYOUTS = set(IMAGE_SPEC.get('multiLayouts') or ['grid', 'compare', 'wall'])
+IMAGE_FIT = set(IMAGE_SPEC.get('fitEnum') or ['cover', 'contain'])
+IMAGE_MAX_PER_PAGE = int(IMAGE_SPEC.get('maxPerPage') or 6)
+IMAGE_MAX_INLINE = int(IMAGE_SPEC.get('maxInlineBytes') or 1572864)
+IMAGE_MAX_TOTAL = int(IMAGE_SPEC.get('maxTotalInlineBytes') or 8388608)
+
+AGENDA_SINGLE_MAX = 8        # 超过则建议 agenda--2col
+# 页高估算参数（单源 layout-constants.pageHeightEstimate）
+_HE = LC.get('pageHeightEstimate') or {}
+SCREEN_BUDGET_PX = int(_HE.get('screenBudgetPx') or 1000)
+WRAP_INSET_PX = int(_HE.get('wrapInsetPx') or 112)
+SHEAD_PX = int(_HE.get('sheadPx') or 190)
+UNIT_PX = int(_HE.get('unitPx') or 170)
+LINE_FACTOR = float(_HE.get('lineFactor') or 1.65)
+STRUCT_PAGE_IDS = ('refs', 'appendix')
+
+
+class Struct(HTMLParser):
+    def __init__(self):
+        super().__init__()
+        self.stack, self.err, self.mis = [], [], []
+
+    def handle_starttag(self, tag, attrs):
+        if tag not in VOID:
+            self.stack.append(tag)
+
+    def handle_endtag(self, tag):
+        if tag in VOID:
+            return
+        if self.stack and self.stack[-1] == tag:
+            self.stack.pop()
+        else:
+            try:
+                i = len(self.stack) - 1 - self.stack[::-1].index(tag)
+                self.mis.extend(self.stack[i:])
+                self.stack = self.stack[:i]
+            except ValueError:
+                self.err.append(tag)
+
+
+def _plain(html):
+    return re.sub(r'<[^>]+>', ' ', html)
+
+
+def _bands(txt):
+    parts = re.split(r'(?=<section class="band)', txt)
+    return [p for p in parts if p.startswith('<section class="band')]
+
+
+def _band_units(b):
+    """承载组件数：卡/指标/图/表 + 列表组/结构/步骤/媒体/矩阵（组合版式与 underfill 同口径）。"""
+    return (len(re.findall(r'class="(?:card|metric)[\s"]', b)) +
+            len(re.findall(r'class="fig[\s"]', b)) +
+            len(re.findall(r'class="tbl-wrap', b)) +
+            len(re.findall(r'class="(?:arch|lane|steps|heat|bul|pyr|media|matrix|exhibit)[\s"]', b)) +
+            len(re.findall(r'class="ul[\s"]', b)) +
+            len(re.findall(r'class="cols-[23]', b)))
+
+
+def _band_heavy_units(b):
+    """页高估算用「重承载行数」：卡/指标按栅格列数折行；图/表/结构各计 1 行级。"""
+    cards = len(re.findall(r'class="(?:card|metric)[\s"]', b))
+    figs = len(re.findall(r'class="fig[\s"]', b))
+    tables = len(re.findall(r'class="tbl-wrap', b))
+    structs = len(re.findall(r'class="(?:arch|lane|media|heat|exhibit)[\s"]', b))
+    if re.search(r'class="[^"]*\bg-4\b', b):
+        cols = 4
+    elif re.search(r'class="[^"]*\bg-3\b', b):
+        cols = 3
+    elif re.search(r'class="[^"]*\b(?:g-2|g-side|g-hero|g-half)\b', b):
+        cols = 2
+    else:
+        cols = 1
+    card_rows = (cards + cols - 1) // cols if cards else 0
+    return card_rows + figs + tables + structs
+
+
+def _check_charts(txt, chk):
+    """图表：存在性 / data-chart 登记 / 动效 / 最小尺寸。"""
+    svgs = re.findall(r'<svg\b[^>]*>', txt)
+    chart_tags = [t for t in svgs if 'class="chart' in t or "class='chart" in t or 'data-chart' in t]
+    chk("含内联 SVG 图表/图形", len(svgs) >= 1, f"{len(svgs)} 个", level="WARN")
+    if not chart_tags:
+        return
+
+    untyped = [t for t in chart_tags if 'data-chart' not in t]
+    chk("图表 svg 均有 data-chart 类型标记", not untyped,
+        f"{len(untyped)} 个未标记", level="WARN")
+
+    # data-chart 类型须在登记表内（layout-constants.json charts.types）
+    unknown = []
+    if CHART_TYPES:
+        for t in chart_tags:
+            m = re.search(r'data-chart="([^"]+)"', t)
+            if m and m.group(1) not in CHART_TYPES:
+                unknown.append(m.group(1))
+    chk("图表类型在登记表内（layout-constants.json charts.types）", not unknown,
+        f"未登记: {sorted(set(unknown))}", level="WARN")
+
+    has_io = 'IntersectionObserver' in txt
+    has_anim = 'data-anim' in txt or 'data-draw' in txt or 'data-sweep' in txt or 'data-count' in txt
+    chk("图表动效（IntersectionObserver + data-anim/draw/sweep/count）",
+        has_io and has_anim,
+        "" if (has_io and has_anim) else f"IO={'有' if has_io else '无'} anim={'有' if has_anim else '无'}",
+        level="WARN")
+
+    too_small, small_note = [], []
+    for t in chart_tags:
+        m = re.search(r'data-chart="([^"]+)"', t)
+        ctype = m.group(1) if m else None
+        rule = CHART_MIN.get(ctype, {}) if ctype else {}
+        wm = re.search(r'style="[^"]*?width:\s*(\d+(?:\.\d+)?)px', t)
+        vb = re.search(r'viewBox="[\d.\-]+ [\d.\-]+ ([\d.]+) ([\d.]+)"', t)
+        if wm and rule.get('pxWidthMin'):
+            w = float(wm.group(1))
+            if w < rule['pxWidthMin']:
+                too_small.append(f"{ctype} 宽{w:.0f}px<{rule['pxWidthMin']}px")
+        if vb and rule.get('vbHeightMin'):
+            h = float(vb.group(2))
+            if h < rule['vbHeightMin']:
+                small_note.append(f"{ctype} viewBox高{h:.0f}<{rule['vbHeightMin']}")
+    chk("图表最小尺寸·硬下限（环形/雷达/仪表盘显示宽）", not too_small,
+        "; ".join(too_small) if too_small else "")
+    chk("图表最小尺寸·建议（通宽图 viewBox 高）", not small_note,
+        "; ".join(small_note) if small_note else "", level="WARN")
+
+
+CARRIERS = checks_html.CARRIERS  # 单源 scripts/checks_html.py
+
+
+def _check_chart_variety(txt, chk, mode):
+    """图表与版式多样性（阈值单源 charts.variety；判定逻辑 checks_html）。"""
+    v = CHART_VARIETY
+    if not v:
+        return
+    per_page = checks_html.chart_types_per_page(txt)
+    used = [t for ts in per_page for t in ts]
+    if not used:
+        return
+    distinct = sorted(set(used))
+    n_chart_pages = len([1 for ts in per_page if ts])
+
+    floor = checks_html.variety_floor(mode, n_chart_pages, v)
+    chk("图表多样性（全篇不同 data-chart 类型数）", len(distinct) >= floor,
+        f"{len(distinct)} 种 / 下限 {floor}（{n_chart_pages} 个图表页）: {distinct}")
+
+    if v.get('noRepeatAdjacent'):
+        adj = checks_html.adjacent_same_type(per_page)
+        chk("图表不连续同型（相邻图表页不得同 data-chart）", not adj,
+            "; ".join(adj[:3]) if adj else "", level="WARN")
+
+    ratio = checks_html.composite_required(mode, v)
+    if ratio > 0:
+        multi, n_content = checks_html.composite_pages(txt)
+        need = max(1, int(n_content * ratio))
+        chk("组合版式（内容页含 ≥2 种承载类型的比例）", multi >= need,
+            f"{multi}/{n_content} 页（下限 {need}）——单件页是例外不是默认",
+            level="WARN")
+
+
+def _check_page_model(txt, chk, mode):
+    """页面高度模型（每页高度稳定）：所有 .band 至少一屏高，长结构页显式 --flow 退出。"""
+    has_model = ('--band-min' in txt) or ('calc(100svh - var(--bar-h))' in txt) or \
+                ('min-height:calc(100vh - var(--bar-h))' in txt)
+    chk("页面高度模型（--band-min / 一屏最小高度）", has_model,
+        "缺少每页一屏的 min-height 机制（见 design-system.md §1c）")
+    bands = _bands(txt)
+    if not bands:
+        return
+    # 固定页高页（非 --flow）应占多数；长结构页才允许 flow
+    flow_pages = []
+    for i, b in enumerate(bands, 1):
+        if 'band--flow' in b[:220]:
+            flow_pages.append(i)
+    bad_flow = []
+    for i in flow_pages:
+        b = bands[i - 1]
+        if not any(f'id="{pid}"' in b for pid in STRUCT_PAGE_IDS):
+            bad_flow.append(f"第{i}页")
+    chk("band--flow 仅用于长结构页（参考资料/附录）", not bad_flow,
+        f"{bad_flow} 使用 --flow 但非结构页（应改用默认一屏页）" if bad_flow else "", level="WARN")
+
+
+def _check_bands(txt, chk, mode="presentation"):
+    """单页预算与布局：文字量/组件数（防塞爆）· 过空页 · 页高溢出估算 · 混排对齐。"""
+    B = MODE_BUDGETS.get(mode, MODE_BUDGETS['presentation'])
+    bands = _bands(txt)
+    over_chars, over_units, too_empty, fit_overload = [], [], [], []
+    for i, b in enumerate(bands, 1):
+        if 'id="refs"' in b or 'id="appendix"' in b:
+            continue
+        b_clean = re.sub(r'<script\b[\s\S]*?</script>', ' ', b)
+        b_clean = re.sub(r'<style\b[\s\S]*?</style>', ' ', b_clean)
+        plain_len = len(_plain(b_clean).strip())
+        units = _band_units(b)
+        if plain_len > B["char"]:
+            over_chars.append(f"第{i}页 {plain_len}字")
+        if units > B["unit"]:
+            over_units.append(f"第{i}页 {units}个")
+        head_tag = b[:220]
+        is_fit = 'band--fit' in head_tag
+        is_deep = 'band--deep' in head_tag
+        if is_fit and plain_len > B["fit"]:
+            fit_overload.append(f"第{i}页 {plain_len}字")
+        if (not is_fit and not is_deep and units == 0 and plain_len < B["empty"]):
+            too_empty.append(f"第{i}页 {plain_len}字")
+    chk(f"单页文字量 ≤ {B['char']} 字（{mode} 模式 · 防溢出）", not over_chars,
+        "; ".join(over_chars) if over_chars else "", level="WARN")
+    chk(f"单页并列单元 ≤ {B['unit']} 个（{mode} 模式 · 防塞爆）", not over_units,
+        "; ".join(over_units) if over_units else "", level="WARN")
+    # v9：过空页升为 FAIL（截图级「只有标题」缺陷不得假过）
+    chk(f"无过空页（{mode} 模式 · 普通页 ≥{B['empty']} 字或有组件；防大面积留白）", not too_empty,
+        "; ".join(too_empty) if too_empty else "")
+    chk(f"满屏居中页 band--fit 内容 ≤ {B['fit']} 字（{mode} 模式 · 防垂直溢出）", not fit_overload,
+        "; ".join(fit_overload) if fit_overload else "", level="WARN")
+
+    # 页高溢出静态估算：字数×行高 + 组件固定高 vs 一屏预算（参数单源 pageHeightEstimate / checkBudgets）。
+    # 固定页高页（无 --flow）超预算即 FAIL——这是"每页高度稳定"的核心硬门禁。
+    fs_px = int(B.get('bodyPx') or 17)
+    wrap_px = int(B.get('wrap') or 1400) - WRAP_INSET_PX
+    cpl = max(10, int(wrap_px / fs_px))
+    est_over = []
+    for i, b in enumerate(bands, 1):
+        if 'id="refs"' in b or 'id="appendix"' in b:
+            continue
+        if 'band--flow' in b[:220]:
+            continue
+        b_clean = re.sub(r'<script\b[\s\S]*?</script>', ' ', b)
+        b_clean = re.sub(r'<style\b[\s\S]*?</style>', ' ', b_clean)
+        plen = len(_plain(b_clean).strip())
+        u = _band_heavy_units(b)
+        est_h = plen / cpl * fs_px * LINE_FACTOR + u * UNIT_PX + SHEAD_PX
+        # 静态估算容差 8%（真值以浏览器为准；轻微超出仍按 FAIL 会误伤密排页）
+        if est_h > SCREEN_BUDGET_PX * 1.08:
+            est_over.append(f"第{i}页≈{est_h:.0f}px")
+    chk(f"页高溢出估算（字×行高+组件 ≤ ~{SCREEN_BUDGET_PX}px/屏 · {mode}）", not est_over,
+        ("; ".join(est_over) + "（按「列表化/精炼→压缩→多列→拆页」处理，或给长结构页加 band--flow）")
+        if est_over else "")
+
+    bad_grids = 0
+    for m in re.finditer(r'<div class="grid[^"]*"([^>]*)>', txt):
+        attrs = m.group(1)
+        sec_end = txt.find('</section>', m.end())
+        ahead = txt[m.end(): sec_end if sec_end != -1 else m.end() + 4000]
+        if 'class="fig' in ahead and 'class="card' in ahead:
+            aligned = ('align-items' in attrs or
+                       re.search(r'\ba-(c|start|end)\b', m.group(0)))
+            if not aligned:
+                bad_grids += 1
+    chk("混排栅格（图+卡）显式对齐（a-start/a-c 或 align-items）", bad_grids == 0,
+        f"{bad_grids} 处未对齐" if bad_grids else "", level="WARN")
+
+
+def _check_icons(txt, chk):
+    """图标使用：长报告不应全文无图标（要点卡/关键列表至少一处）。"""
+    if len(_bands(txt)) < 6:
+        return
+    has_icon = ('ul--ico' in txt or 'metric__ico' in txt or
+                re.search(r'card__ico"[^>]*>\s*<svg', txt) or
+                re.search(r'card__ico">\s*<svg', txt))
+    chk("长报告（≥6页）要点卡/关键列表有图标", bool(has_icon),
+        "全文未用图标（见 icons.md 使用准则）", level="WARN")
+
+
+def _check_table_rows(txt, chk, mode="presentation"):
+    """表格行数上限（research 密表 ≤16 / presentation ≤8 / architecture ≤6；宁拆勿挤）。"""
+    B = MODE_BUDGETS.get(mode, MODE_BUDGETS['presentation'])
+    max_rows = B['trow']
+    over = []
+    for i, m in enumerate(re.finditer(r'<tbody>([\s\S]*?)</tbody>', txt), 1):
+        n = len(re.findall(r'<tr', m.group(1)))
+        if n > max_rows:
+            over.append(f"表{i} {n}行")
+    chk(f"表格行数 ≤ {max_rows} 行（{mode} 模式 · 宁拆勿挤）", not over,
+        "; ".join(over) if over else "", level="WARN")
+
+
+def _check_mode_layouts(txt, chk, mode):
+    """模式版式特征：三模式独立体系，各自应有标志性组件。"""
+    if mode == 'research':
+        has = ('class="cols-2' in txt or 'class="exhibit' in txt or
+               'class="cols-3' in txt or 'class="matrix' in txt)
+        chk("research 版式特征（cols-2/cols-3/exhibit/matrix 至少一处）", has,
+            "未见研究模式标志性组件", level="WARN")
+    elif mode == 'architecture':
+        has = ('class="arch' in txt or 'class="lane' in txt)
+        chk("architecture 版式特征（arch/lane 至少一处）", has,
+            "未见架构模式标志性组件", level="WARN")
+
+
+# 页型 → 正文标志性版式特征（正则；模型-正文同源抽查，见 `components.md` §46/§46b）
+TYPE_FEATURE = {
+    'steps': r'class="[^"]*\bsteps\b',
+    'heatmap': r'class="[^"]*\bheat\b',
+    'bullet': r'class="[^"]*\bbul\b',
+    'pyramid': r'class="[^"]*\bpyr\b',
+    'image': r'class="[^"]*\bmedia\b|<img\b',
+    'matrix': r'class="[^"]*\bmatrix\b',
+    'exhibit': r'class="[^"]*\bexhibit\b',
+    'twocol': r'class="[^"]*\bcols-2\b',
+    'threecol': r'class="[^"]*\bcols-3\b',
+    'halftable': r'class="[^"]*\bg-half\b',
+    'split': r'class="[^"]*\bg-side\b',
+    'diagram': r'class="[^"]*\barch\b',
+    'lane': r'class="[^"]*\blane\b',
+    'timeline': r'class="[^"]*\btl\b',
+    'comparison': r'class="[^"]*\bg-half\b',
+    'quote': r'band--accent|band--deep',
+    'donut': r'data-chart="donut"',
+    # 复杂信息图页型：正文以内联 SVG 承载（data-chart 登记与 charts.registry 同源）
+    'sankey': r'data-chart="sankey"',
+    'treemap': r'data-chart="treemap"',
+    'boxplot': r'data-chart="boxplot"',
+    'network': r'data-chart="network"',
+    'marimekko': r'data-chart="marimekko"',
+    'streamgraph': r'data-chart="streamgraph"',
+}
+
+
+def _check_type_features(txt, chk, model):
+    """模型声明的页型应在正文找到对应版式组件（防"模型填了页型但正文用旧版式"）。"""
+    if not model:
+        return
+    missing = []
+    for sec in (model.get('sections') or []):
+        t = (sec.get('type') or 'points')
+        feat = TYPE_FEATURE.get(t)
+        if feat and not re.search(feat, txt):
+            missing.append(t)
+    chk("模型页型 ↔ 正文版式组件对应", not missing,
+        f"{len(missing)} 处: {missing[:4]}" if missing else "", level="WARN")
+
+
+def _check_pptx_export(txt, chk):
+    """PPTX 预览配套：有预览按钮就必须有内容模型 + 预览运行时；页面不得残留导出按钮。"""
+    has_btn = 'id="pptPreviewBtn"' in txt
+    has_model = 'REPORT_MODEL' in txt
+    has_runtime = ('g.TopPptHtml = api' in txt) and ('function slidesXml' in txt)
+    if has_btn:
+        chk("PPTX 预览按钮配套（REPORT_MODEL + 预览运行时）",
+            has_model and has_runtime,
+            f"model={'有' if has_model else '无'} runtime={'有' if has_runtime else '无'}")
+    else:
+        chk("含 PPTX 预览按钮与内容模型（建议保留）",
+            has_model and has_runtime, "未集成预览按钮/模型", level="WARN")
+    chk("页面无 PPTX 导出按钮（仅预览 + 提示词）",
+        'id="pptxBtn"' not in txt and 'id="pptDownload"' not in txt,
+        "残留导出按钮 pptxBtn/pptDownload", level="WARN")
+
+
+def _extract_model(txt):
+    m = re.search(r'window\.REPORT_MODEL\s*=\s*', txt)
+    if not m:
+        return None, "未找到 window.REPORT_MODEL（PPTX 双通道导出将不可用）"
+    try:
+        model, _end = json.JSONDecoder().raw_decode(txt[m.end():])
+        return model, None
+    except json.JSONDecodeError as e:
+        return None, f"REPORT_MODEL 不是合法 JSON: {e}（必须双引号/无尾逗号/无注释）"
+
+
+def _chart_channel(chart_type) -> str:
+    """图表类型 → 交付通道（native / shape），与 charts.registry 同源（checks_html）。"""
+    return checks_html.chart_channel(chart_type, CHART_REG)
+
+
+def _chart_datatable_mode(chart: dict) -> str:
+    """数据表策略：图表级 > 登记表默认（appendix 收敛为 inline）。"""
+    return checks_html.chart_datatable_mode(chart, CHART_REG)
+
+
+def _check_chart_datatable(chk, model) -> None:
+    """数据可追溯：非原生（形状通道）图表与信息图页型的 dataTable 不得为 off。
+
+    与 validate_pptx.py 的 MODEL_CHART_DATATABLE 同源，但提前到 HTML 阶段暴露——
+    避免生成完 PPTX 才被拦（数据表策略是模型层决定，HTML 阶段即可判定）。
+    """
+    bad: list[str] = []
+    for sec in (model.get('sections') or []):
+        if not isinstance(sec, dict):
+            continue
+        st = sec.get('type') or ''
+        if st in INFO_PAGE_TYPES:
+            mode = str(((sec.get('chart') or {}).get('dataTable')) or 'notes').lower()
+            if mode == 'off':
+                bad.append(f'{st} 页信息图 dataTable=off')
+            continue
+        charts = []
+        if st in ('bar', 'donut', 'exhibit', 'halftable'):
+            c = sec.get('chart') or {}
+            if isinstance(c, dict) and c.get('labels') and c.get('values'):
+                charts.append(c)
+        elif st == 'split':
+            # 双区组合页：左/右两侧各可能是图表（left 缺省 points、right 缺省 bar）
+            for side, dflt in (('left', 'points'), ('right', 'bar')):
+                pane = sec.get(side) or {}
+                if not isinstance(pane, dict):
+                    continue
+                if (pane.get('type') or dflt) in ('table', 'image', 'points'):
+                    continue
+                if pane.get('labels') and pane.get('values'):
+                    charts.append(pane)
+        for chart in charts:
+            if _chart_channel(chart.get('type')) == 'shape' and _chart_datatable_mode(chart) == 'off':
+                bad.append(f"{st} 页 {chart.get('type')} 图表 dataTable=off")
+    chk("图表数据表策略（非原生图表不得 off，保证数据可追溯）", not bad,
+        '；'.join(bad[:3]) if bad else "", level="WARN")
+
+
+def _check_model_consistency(txt, chk, mode, style):
+    """模型-正文一致性：JSON 合法 / mode·style·theme 一致 / 章节标题抽查 / agenda / 页数。"""
+    model, err = _extract_model(txt)
+    if err:
+        chk("REPORT_MODEL 存在且为合法 JSON", False, err)
+        return None
+    chk("REPORT_MODEL 存在且为合法 JSON", True)
+    mmode = model.get('mode')
+    chk("REPORT_MODEL.mode 与 data-mode 一致", mmode == mode,
+        f"model={mmode!r} 页面={mode!r}")
+    mstyle = model.get('style')
+    chk("REPORT_MODEL.style 与 data-style 一致", mstyle == style,
+        f"model={mstyle!r} 页面={style!r}", level="WARN")
+    mtheme = model.get('theme') or 'light'
+    html_theme_m = re.search(r'<html[^>]*data-theme="([^"]+)"', txt)
+    html_theme = html_theme_m.group(1) if html_theme_m else 'light'
+    chk("REPORT_MODEL.theme 与 data-theme 一致（未声明默认 light）",
+        mtheme in ('light', 'dark') and mtheme == html_theme,
+        f"model={mtheme!r} 页面={html_theme!r}", level="WARN")
+    secs = model.get('sections') or []
+    chk("REPORT_MODEL.sections 非空（每章一页模型）", len(secs) >= 1,
+        f"{len(secs)} 页", level="WARN")
+    _check_chart_datatable(chk, model)
+    plain = _plain(txt)
+    missing = [str(s.get('title'))[:14] for s in secs
+               if not s.get('title') or str(s.get('title')) not in plain]
+    chk("模型章节标题均出现于正文（模型-正文同源）", not missing,
+        f"{len(missing)} 个缺失: {missing[:3]}" if missing else "", level="WARN")
+    n_html = txt.count('class="agenda__a"')
+    n_model = len(model.get('agenda') or [])
+    if n_html and n_model:
+        chk("模型 agenda 条数与正文一致", n_html == n_model,
+            f"model={n_model} 正文={n_html}", level="WARN")
+    n_bands = len(_bands(txt))
+    lo, hi = len(secs) + 3, len(secs) + 7
+    chk("页数粗匹配（正文 band ≈ sections + 结构页）", lo <= n_bands <= hi,
+        f"正文 {n_bands} 页 vs 模型 {len(secs)}+{lo - len(secs)}~{hi - len(secs)}", level="WARN")
+    return model
+
+
+def _check_anchors(txt, chk):
+    """锚点闭环：Agenda / 导航 / 正文内部链接指向的 id 必须存在，且 id 全篇唯一。
+    （重复 id 会让锚点跳到错误的页，且 HTML 规范不允许——硬拦。）"""
+    # 只看正文：内联运行时/UI 脚本里有 `id="' + nid() + '"` 这类拼接，不能算重复 id
+    body = re.sub(r'<script\b[\s\S]*?</script>', '', txt, flags=re.I)
+    all_ids = re.findall(r'\sid="([^"]+)"', body)
+    ids = set(all_ids)
+    dup = sorted({i for i in all_ids if all_ids.count(i) > 1})
+    chk("id 全篇唯一（无重复 id）", not dup,
+        f"重复 id {dup[:5]}（锚点会跳到错误的页）" if dup else "")
+    bad = []
+    for href in set(re.findall(r'href="#([^"]+)"', body)):
+        if href and href not in ids:
+            bad.append('#' + href)
+    chk("内部锚点闭环（agenda/nav/正文链接均有对应 id）", not bad,
+        f"悬空锚点 {sorted(bad)[:5]}" if bad else "")
+
+
+def _check_exhibits(txt, chk, mode):
+    """Exhibit 编号体系：每个 .exhibit 框都有编号，且全篇连续 1..N 无跳号无重复。"""
+    nos = [int(n) for n in re.findall(
+        r'class="exhibit__no"[^>]*>\s*Exhibit\s*(\d+)', txt, re.I)]
+    n_frame = len(re.findall(r'class="exhibit[\s"]', txt))
+    if not nos:
+        if mode == 'research':
+            chk("Exhibit 编号体系（research 深报告建议建立）", False,
+                "全文无 Exhibit 编号图表框", level="WARN")
+        return
+    # 只校验已编号者的连续性会漏掉「漏标编号的 exhibit 框」——先核框数再核编号。
+    chk(f"Exhibit 框均已编号（{n_frame} 框）", n_frame == len(nos),
+        f"{n_frame} 个 .exhibit 框但只有 {len(nos)} 个 .exhibit__no 编号" if n_frame != len(nos) else "")
+    expected = list(range(1, len(nos) + 1))
+    chk(f"Exhibit 编号连续（1..{len(nos)} 无跳号无重复）", sorted(nos) == expected,
+        f"读到 {nos}")
+
+
+def _check_research_extras(txt, chk, mode):
+    """research 模式：so-what 连续性（>3 页无结论条）与 Exhibit 来源行。"""
+    if mode != 'research':
+        return
+    bands = _bands(txt)
+    streak = max_streak = 0
+    for i, b in enumerate(bands, 1):
+        if i <= 2 or 'id="agenda"' in b or 'id="refs"' in b:
+            continue
+        if 'class="sowhat' in b or 'class="note' in b:
+            streak = 0
+        else:
+            streak += 1
+            max_streak = max(max_streak, streak)
+    chk("so-what 连续性（连续 ≤3 页无结论条）", max_streak <= 3,
+        f"连续 {max_streak} 页无 .sowhat/.note" if max_streak > 3 else "", level="WARN")
+    n_ex = len(re.findall(r'class="exhibit[\s"]', txt))
+    n_src = txt.count('class="exhibit__src"')
+    chk(f"Exhibit 图表框均有来源行（{n_ex} 框）", n_ex == n_src,
+        f"{n_ex - n_src} 个缺 .exhibit__src" if n_ex != n_src else "", level="WARN")
+
+
+def _check_emphasis(txt, chk):
+    """强调带约束：band--deep 是反相强调页（浅色主题下渲染为深色）——
+    全文受限使用，且不得作末页，否则浅色模式末尾会出现深色页（用户核心诉求）。"""
+    em = LC.get('emphasis') or {}
+    max_deep = em.get('deepMaxPages', 1)
+    forbid_last = em.get('deepForbiddenLast', True)
+    bands = _bands(txt)
+    deep_idx = [i for i, b in enumerate(bands, 1) if 'band--deep' in b[:220]]
+    chk(f"反相强调页 band--deep ≤ {max_deep} 页（浅色模式下渲染为深色）",
+        len(deep_idx) <= max_deep,
+        f"{len(deep_idx)} 页: {deep_idx}（收尾/金句页请用 band--accent）" if len(deep_idx) > max_deep else "",
+        level="WARN")
+    if forbid_last and bands:
+        chk("末页非反相深色页（band--deep）",
+            'band--deep' not in bands[-1][:220],
+            "末页为 band--deep：浅色模式下会以深色收尾（改用 band--accent）", level="WARN")
+
+
+def _layout_sig(band: str) -> str | None:
+    """从 band HTML 提取版式签名（优先级从具体到宽泛）。
+    图表类型并入签名：exhibit+sankey 与 exhibit+waterfall 视为不同版式。"""
+    pats = (
+        ('g-hero-full', r'g-hero-full'),
+        ('g-mosaic', r'g-mosaic'),
+        ('g-bento', r'g-bento'),
+        ('g-aside', r'g-aside'),
+        ('g-quad', r'g-quad'),
+        ('g-211', r'g-211'),
+        ('g-121', r'g-121'),
+        ('g-side', r'g-side'),
+        ('rows-2', r'rows-2'),
+        ('rows-3', r'rows-3'),
+        ('g-2', r'\bg-2\b'),
+        ('g-3', r'\bg-3\b'),
+        ('g-4', r'\bg-4\b'),
+        ('g-5', r'\bg-5\b'),
+        ('g-6', r'\bg-6\b'),
+        ('cols-3', r'cols-3'),
+        ('cols-2', r'cols-2'),
+        ('exhibit', r'class="exhibit'),
+        ('table', r'class="tbl-wrap'),
+        ('metrics', r'class="metric'),
+        ('arch', r'class="arch'),
+        ('lane', r'class="lane'),
+        ('matrix', r'class="matrix'),
+        ('heat', r'class="heat'),
+        ('bul', r'class="bul'),
+        ('pyr', r'class="pyr'),
+        ('steps', r'class="steps'),
+        ('timeline', r'class="tl[\s"]'),
+        ('media', r'class="media'),
+        ('stagger', r'stagger'),
+        ('quote', r'band--accent'),
+        ('card', r'class="card[\s"]'),
+    )
+    base = None
+    for name, pat in pats:
+        if re.search(pat, band):
+            base = name
+            break
+    if base is None:
+        return None
+    charts = re.findall(r'data-chart="([^"]+)"', band)
+    if charts:
+        # 去重保序，最多 2 类，避免顺序噪声
+        seen, uniq = set(), []
+        for c in charts:
+            if c not in seen:
+                seen.add(c)
+                uniq.append(c)
+        return f"{base}:{'+'.join(uniq[:2])}"
+    return base
+
+
+def _check_content_quality(txt, chk, mode, model):
+    """内容级质量：版式节奏 / so-what 非空洞 / research 行动标题含判断。"""
+    cq = LC.get('contentQuality') or {}
+    bands = _bands(txt)
+    skip = set((cq.get('rhythm') or {}).get('skipIds') or
+               ['cover', 'agenda', 'refs', 'appendix', 'next'])
+
+    # ① 版式节奏：同一签名不得连用超过上限
+    max_streak = int((cq.get('rhythm') or {}).get('maxSameLayoutStreak') or 2)
+    sigs: list[tuple[int, str]] = []
+    for i, b in enumerate(bands, 1):
+        head = b[:220]
+        if any(f'id="{sid}"' in head for sid in skip):
+            continue
+        sig = _layout_sig(b)
+        if sig:
+            sigs.append((i, sig))
+    streak_hits: list[str] = []
+    run_sig, run_start, run_len = None, 0, 0
+    for i, sig in sigs:
+        if sig == run_sig:
+            run_len += 1
+        else:
+            if run_sig and run_len > max_streak:
+                streak_hits.append(f'{run_sig}×{run_len}（起第{run_start}页）')
+            run_sig, run_start, run_len = sig, i, 1
+    if run_sig and run_len > max_streak:
+        streak_hits.append(f'{run_sig}×{run_len}（起第{run_start}页）')
+    chk(f"版式节奏（同一版式签名连续 ≤{max_streak} 页）", not streak_hits,
+        "; ".join(streak_hits[:3]) if streak_hits else "", level="WARN")
+
+    # ② so-what 禁空洞套话 + 最短长度（HTML 与模型双侧）
+    sw = cq.get('sowhat') or {}
+    min_chars = int(sw.get('minChars') or 12)
+    forbidden = list(sw.get('forbidden') or [])
+    texts: list[tuple[str, str]] = []
+    for m in re.finditer(r'class="sowhat__v"[^>]*>(.*?)</', txt, re.S):
+        body = re.sub(r'<[^>]+>', '', m.group(1) or '').strip()
+        if body:
+            texts.append(('html', body))
+    if model:
+        for idx, sec in enumerate(model.get('sections') or [], 1):
+            v = sec.get('soWhat')
+            if isinstance(v, str) and v.strip():
+                texts.append((f'model#{idx}', v.strip()))
+    short, hollow = [], []
+    for src, body in texts:
+        if len(body) < min_chars:
+            short.append(f'{src}:{body[:24]}')
+        if any(f in body for f in forbidden):
+            hollow.append(f'{src}:{body[:24]}')
+    if texts:
+        chk(f"so-what 实质（长度 ≥{min_chars} 字且无空洞套话）", not short and not hollow,
+            "; ".join((short + hollow)[:4]) if (short or hollow) else "", level="WARN")
+
+    # ③ research 行动标题须含数字或判断词
+    if mode == 'research':
+        rt = cq.get('researchTitle') or {}
+        if rt.get('requireDigitOrJudgment', True):
+            jpat = rt.get('judgmentPattern') or r'[\d]|是|应|需|将'
+            struct = ('报告大纲', '大纲', '参考资料', '数据来源', '下一步', '结论',
+                      '全文核心', 'agenda', '附录', '收尾')
+            h1s = re.findall(r'<h2 class="t-h1 shead__title"[^>]*>(.*?)</h2>', txt)
+            titles = [re.sub(r'<[^>]+>', '', h).strip() for h in h1s]
+            titles = [t for t in titles if len(t) >= 12 and
+                      not t.lower().startswith(struct) and not any(s in t for s in struct)]
+            no_judge = [t[:28] for t in titles if not re.search(jpat, t)]
+            chk("research 行动标题含数字或判断词（标题即结论）", not no_judge,
+                f"缺少判断信号: {no_judge[:3]}" if no_judge else "", level="WARN")
+
+
+def _check_v9_hard_gates(txt, chk, model):
+    """v9 硬门禁：标签泄漏 / 标题空页 / 极偏占比 / 简单图过大（治截图级缺陷）。"""
+    import html as _html
+    QG = LC.get('qualityGates') or {}
+    patterns = QG.get('tagLeakPatterns') or [
+        '<a ', '</a>', 'class="cite"', 'href=', '<strong', '</strong>'
+    ]
+    # ① HTML 标签泄漏进可见文本
+    # 先抓「转义后当字面量显示」的（&lt;a class=…&gt;）：源码里就有实体
+    leak_hits = []
+    for i, b in enumerate(_bands(txt), 1):
+        b2 = re.sub(r'<script\b[\s\S]*?</script>', ' ', b)
+        b2 = re.sub(r'<style\b[\s\S]*?</style>', ' ', b2)
+        # A. 转义标签字面量（用户截图形态：正文里直接可见 <a class="cite"…>）
+        if re.search(r'&lt;/?[a-zA-Z][^&]{0,60}&gt;', b2):
+            m = re.search(r'&lt;/?[a-zA-Z][^&]{0,60}&gt;', b2)
+            leak_hits.append(f"第{i}页含转义标签字面量 `{m.group(0)[:40]}`")
+            continue
+        # B. 反转义后剥真标签，残留源码片段
+        unesc = _html.unescape(b2)
+        plain = _plain(unesc)
+        for p in patterns:
+            if p in plain:
+                leak_hits.append(f"第{i}页可见文本含 `{p}`")
+                break
+    # C. REPORT_MODEL 字段内夹带 HTML 标签（在 <script> 里，正文剥离会漏掉）
+    model_tag_hits = []
+    mm = re.search(r'window\.REPORT_MODEL\s*=\s*(\{[\s\S]*?\})\s*;', txt)
+    if mm:
+        try:
+            import json as _json
+            mobj = _json.loads(mm.group(1))
+
+            def _scan(o, path=''):
+                if isinstance(o, dict):
+                    for k, v in o.items():
+                        _scan(v, f'{path}.{k}' if path else k)
+                elif isinstance(o, list):
+                    for i, v in enumerate(o):
+                        _scan(v, f'{path}[{i}]')
+                elif isinstance(o, str) and re.search(r'</?[a-zA-Z][^>]*>', o):
+                    model_tag_hits.append(path or '(root)')
+            _scan(mobj)
+        except Exception:
+            pass
+    if model_tag_hits:
+        leak_hits.append(f"REPORT_MODEL 字段夹带标签: {model_tag_hits[:3]}")
+    chk("HTML_TAG_IN_TEXT 可见文本无 HTML 标签源码泄漏", not leak_hits,
+        "; ".join(leak_hits[:5]) if leak_hits else "")
+
+    # ② 标题空页 / 承载不足
+    title_only, underfill = [], []
+    to_min = int((QG.get('titleOnly') or {}).get('minBodyChars') or 40)
+    uf = QG.get('underfill') or {}
+    uf_min = int(uf.get('minCarriers') or 2)
+    exempt = set(uf.get('exemptIds') or ['quote', 'cover', 'refs', 'appendix', 'next'])
+    for i, b in enumerate(_bands(txt), 1):
+        head = b[:240]
+        bid_m = re.search(r'id="([^"]+)"', head)
+        bid = bid_m.group(1) if bid_m else ''
+        if bid in exempt or any(f'id="{e}"' in head for e in exempt):
+            continue
+        if 'band--deep' in head or 'band--accent' in head:
+            continue
+        b2 = re.sub(r'<script\b[\s\S]*?</script>', ' ', b)
+        b2 = re.sub(r'<style\b[\s\S]*?</style>', ' ', b2)
+        # 去掉 shead（eyebrow+标题+导语）后的正文
+        body = re.sub(r'<div class="shead[\s\S]*?</div>\s*</div>', ' ', b2, count=1)
+        body = re.sub(r'<div class="shead[\s\S]*?</h2>\s*</div>', ' ', b2, count=1)
+        plain_body = _plain(body).strip()
+        # 再去掉与标题重复的大标题串
+        title_m = re.search(r'shead__title[^>]*>(.*?)</', b2)
+        if title_m:
+            plain_body = plain_body.replace(_plain(title_m.group(1)).strip(), '', 1).strip()
+        carriers = _band_units(b)
+        if len(plain_body) < to_min and carriers == 0:
+            title_only.append(f"第{i}页正文{len(plain_body)}字")
+        elif carriers < uf_min and len(plain_body) < to_min * 2:
+            underfill.append(f"第{i}页承载{carriers}")
+    chk(f"TITLE_ONLY_PAGE 非空页（去页头后正文 ≥{to_min} 字）", not title_only,
+        "; ".join(title_only[:5]) if title_only else "")
+    chk(f"UNDERFILL_PAGE 内容页承载 ≥{uf_min}（金句/章节幕豁免）", not underfill,
+        "; ".join(underfill[:5]) if underfill else "")
+
+    # ③ 极偏占比禁 donut/pie（治 0.5% vs 99.5% 环图叠字不可读）
+    skew = QG.get('chartSkew') or {}
+    skew_types = set(skew.get('types') or ['donut', 'pie', 'multidonut'])
+    min_pct = float(skew.get('minSectorPct') or 5)
+    max_ratio = float(skew.get('maxMinRatio') or 20)
+    skew_hits = []
+    if model and isinstance(model.get('sections'), list):
+        for si, sec in enumerate(model['sections'], 1):
+            if not isinstance(sec, dict):
+                continue
+            ch = sec.get('chart') or {}
+            ctype = str(ch.get('type') or sec.get('type') or '').lower()
+            if sec.get('type') == 'donut':
+                ctype = 'donut'
+            if ctype not in skew_types:
+                continue
+            vals = [float(v) for v in (ch.get('values') or []) if v is not None]
+            vals = [v for v in vals if v >= 0]
+            if len(vals) < 2:
+                continue
+            total = sum(vals) or 1
+            pcts = [v / total * 100 for v in vals]
+            mn, mx = min(pcts), max(pcts)
+            ratio = (mx / mn) if mn > 0 else 999
+            if mn < min_pct or ratio > max_ratio:
+                skew_hits.append(
+                    f"第{si}章 {ctype} 最小扇区 {mn:.1f}%（比 {ratio:.0f}:1）应改 KPI/进度/对比条")
+    chk(f"CHART_SKEW_INVALID 占比图最小扇区 ≥{min_pct:.0f}% 且 max/min ≤{max_ratio:.0f}",
+        not skew_hits, "; ".join(skew_hits[:4]) if skew_hits else "")
+
+    # ④ 简单图过大（类别≤2 或 数据点≤3 却近乎独占内容区）
+    ov = QG.get('chartOversize') or {}
+    simple_cats = int(ov.get('simpleMaxCats') or 2)
+    simple_pts = int(ov.get('simpleMaxPts') or 3)
+    max_pct_area = float(ov.get('maxContentAreaPct') or 55)
+    oversize = []
+    for i, b in enumerate(_bands(txt), 1):
+        for m in re.finditer(
+                r'<svg\b[^>]*data-chart="([^"]+)"[^>]*style="([^"]*)"[^>]*>',
+                b):
+            ctype, style = m.group(1), m.group(2)
+            # 解析 width / height（px 或 %）
+            wm = re.search(r'width:\s*(\d+(?:\.\d+)?)px', style)
+            hm = re.search(r'height:\s*(\d+(?:\.\d+)?)px', style)
+            vw = re.search(r'viewBox="0 0 (\d+(?:\.\d+)?) (\d+(?:\.\d+)?)"', m.group(0))
+            # 从 svg 块粗数数据点/类别
+            svg_end = b.find('</svg>', m.end())
+            block = b[m.start():svg_end if svg_end > 0 else m.end() + 800]
+            n_labels = len(re.findall(r'<text\b', block))
+            n_rects = len(re.findall(r'<rect\b', block))
+            n_circles = len(re.findall(r'<circle\b', block))
+            n_pts = max(n_labels, n_rects, n_circles)
+            is_simple = n_labels <= simple_cats or n_pts <= simple_pts
+            if not is_simple:
+                continue
+            # 简单图接近通宽且高度大 → 视为「简单大图」
+            h_px = float(hm.group(1)) if hm else (float(vw.group(2)) if vw else 0)
+            if h_px >= 320 or (not hm and not vw and 'width:100%' in style and 'height' not in style):
+                # 无显式高度的通宽简单图
+                if h_px >= 320 or ('width:100%' in style and n_pts <= simple_pts):
+                    oversize.append(f"第{i}页 {ctype} 简单图过大（{n_pts} 点/标记）")
+    chk(f"CHART_OVERSIZE 简单图（≤{simple_cats} 类或 ≤{simple_pts} 点）勿独占版面（上限 {max_pct_area:.0f}%）",
+        not oversize, "; ".join(oversize[:4]) if oversize else "")
+
+
+def _check_layout_grammar(txt, chk, mode):
+    """布局语法门禁：骨架类 / 单一重心 / 混排对齐 / 间距 token / 图标尺寸 / 标签防换行。"""
+    LS = LC.get('layoutSystem') or {}
+    presets = LS.get('presets') or {}
+    elem = LS.get('elements') or {}
+    mf = LS.get('multiFocus') or {}
+    align_cfg = LS.get('align') or {}
+    fill_t = (LS.get('fillTarget') or {}).get(mode) or [62, 78]
+    icon_sizes = set(elem.get('icon', {}).get('sizesPx') or [16, 18, 20, 24])
+    chart_w_pct = float(mf.get('chartWidthPct') or 55)
+    metric_fz = float(mf.get('metricFontPx') or 40)
+
+    # 收集预设骨架对应 class
+    skeleton_classes = set()
+    for p in (presets or {}).values():
+        for c in (p.get('html') or []):
+            skeleton_classes.add(c)
+    # 附加常见栅格类（grammar 白名单）
+    skeleton_classes.update({
+        'g-2', 'g-3', 'g-4', 'g-5', 'g-6', 'g-hero', 'g-hero--rev', 'g-side', 'g-side--rev',
+        'g-31', 'g-13', 'g-41', 'g-14', 'g-211', 'g-121', 'g-quad', 'g-aside', 'g-bento',
+        'g-hero-full', 'g-mosaic', 'rows-2', 'rows-3', 'stagger', 'media-grid', 'media-compare',
+    })
+
+    no_skel, multi_focus, align_miss, off_token = [], [], [], []
+    icon_bad, label_collapse = [], []
+    fill_low, fill_high = [], []
+    mixed_need = bool(align_cfg.get('requireOnMixed', True))
+    align_tokens = set(align_cfg.get('mixedGridClasses') or ['a-start', 'a-c', 'align-items'])
+
+    for i, b in enumerate(_bands(txt), 1):
+        head = b[:240]
+        if any(f'id="{x}' in head for x in ('refs', 'appendix', 'cover', 'agenda', 'quote')):
+            continue
+        if 'band--flow' in head:
+            continue
+        # 轻量合法页：金句/强调带/大数压场（fill 目标放宽）
+        is_light = ('band--fit' in head or 'band--accent' in head or
+                    'band--deep' in head or 'quote' in head)
+
+        # ① 骨架类（或 data-skel 标记）
+        has_skel = ('data-skel="' in b) or any(
+            re.search(rf'class="[^"]*\b{re.escape(c)}\b', b) for c in skeleton_classes)
+        # 手写 inline grid 且无预设类
+        inline_grid = re.findall(r'style="[^"]*display:\s*grid[^"]*"', b)
+        if not has_skel and inline_grid:
+            no_skel.append(f"第{i}页 inline grid 无 P1–P12 骨架类")
+
+        # ② 单一视觉重心（粗启发式）
+        n_big = 0
+        for m in re.finditer(r'<svg\b[^>]*>', b):
+            tag = m.group(0)
+            wm = re.search(r'width:\s*(\d+(?:\.\d+)?)px', tag)
+            wrap = int((re.search(r'--wrap:\s*(\d+)', txt) or [0, 1400])[1] or 1400) if False else 1400
+            if wm and float(wm.group(1)) / wrap * 100 >= chart_w_pct:
+                n_big += 1
+            if re.search(r'width:\s*100%', tag) and 'data-chart' in tag:
+                # 通宽图记 1 大件
+                n_big += 1
+        for m in re.finditer(r'font-size:\s*clamp\(\s*(\d+)px', b):
+            if float(m.group(1)) >= metric_fz:
+                n_big += 1
+                break
+        if re.search(r'class="[^"]*media--(?:full|bleed)', b):
+            n_big += 1
+        if n_big >= 2:
+            multi_focus.append(f"第{i}页大件×{n_big}")
+
+        # ③ 混排对齐（同时有图/媒体与卡/列表）
+        mixed = (re.search(r'class="[^"]*(?:fig|media|chart)', b) and
+                 re.search(r'class="[^"]*(?:card|ul|metric)', b))
+        if mixed and mixed_need:
+            if not any(t in b for t in align_tokens):
+                align_miss.append(f"第{i}页图卡混排缺 a-start/a-c")
+
+        # ④ 间距写死（margin/padding/gap 非 token / 非 clamp；≤8px 微调白名单）
+        for m in re.finditer(r'(?:margin|padding|gap)(?:-[a-z]+)?\s*:\s*([^;"]+)', b):
+            val = m.group(1).strip()
+            if not val or val.startswith('var(') or val.startswith('clamp(') or val.startswith('0'):
+                continue
+            if re.fullmatch(r'auto|inherit|initial|unset|normal', val):
+                continue
+            px_vals = [float(x) for x in re.findall(r'(\d+(?:\.\d+)?)px', val)]
+            if px_vals and all(v <= 8 for v in px_vals) and 'clamp' not in val and 'var(' not in val:
+                continue  # 4–8px 光学微调允许
+            if px_vals and 'clamp' not in val and 'var(' not in val:
+                off_token.append(f"第{i}页 `{val[:24]}`")
+
+        # ⑤ 图标尺寸
+        for m in re.finditer(
+                r'<svg\b[^>]*(?:class="[^"]*(?:ico|icon)[^"]*"|metric__ico|card__ico)[^>]*>',
+                b, re.I):
+            tag = m.group(0)
+            sm = re.search(r'width:\s*["\']?(\d+)', tag) or re.search(r'width="(\d+)"', tag)
+            if sm and int(sm.group(1)) not in icon_sizes:
+                icon_bad.append(f"第{i}页 {sm.group(1)}px")
+
+        # ⑥ 图表标签防换行（短盒 + 长 text）
+        for m in re.finditer(r'<text[^>]*width="(\d+)"[^>]*>([^<]{8,})</text>', b):
+            if int(m.group(1)) < 48:
+                label_collapse.append(f"第{i}页标签盒宽{m.group(1)}")
+
+        # ⑦ 结构图禁裸文字箭头（F13）
+        if re.search(r'class="[^"]*lane__arr[^"]*"[^>]*>\s*→', b) or re.search(
+                r'lane__arr">→', b):
+            label_collapse.append(f"第{i}页泳道裸文字→（改 .lane__arr）")
+
+        # ⑧ 填充率粗估（去空白字符 + 组件；轻量页不判过空）
+        b2 = re.sub(r'<script\b[\s\S]*?</script>', ' ', b)
+        b2 = re.sub(r'<style\b[\s\S]*?</style>', ' ', b2)
+        plen = len(re.sub(r'\s+', '', _plain(b2)))
+        units = _band_units(b)
+        budget = MODE_BUDGETS.get(mode, MODE_BUDGETS['presentation'])
+        est = (plen / max(1, budget['fit']) * 55) + (units / max(1, budget['unit']) * 45)
+        est = max(0, min(100, est))
+        lo_t = fill_t[0] * 0.55
+        hi_t = min(95, fill_t[1] * 1.15)
+        if (not is_light) and est < lo_t and units == 0 and plen < budget['empty']:
+            fill_low.append(f"第{i}页≈{est:.0f}%")
+        elif est > hi_t and not is_light:
+            fill_high.append(f"第{i}页≈{est:.0f}%")
+
+    chk(f"LAYOUT_NO_SKELETON 内容页使用 P1–P12 骨架类（layout-grammar）", not no_skel,
+        "; ".join(no_skel[:4]) if no_skel else "")
+    chk(f"LAYOUT_MULTI_FOCUS 一屏一视觉重心（大件 ≤{mf.get('maxPrimary', 1)}）", not multi_focus,
+        "; ".join(multi_focus[:4]) if multi_focus else "")
+    chk("LAYOUT_ALIGN_DRIFT 图卡混排显式对齐（a-start / a-c）", not align_miss,
+        "; ".join(align_miss[:4]) if align_miss else "")
+    chk("LAYOUT_SPACING_OFF_TOKEN 间距走 --sp-*/gap/clamp（禁游离 px）", not off_token,
+        "; ".join(off_token[:4]) if off_token else "", level="WARN")
+    chk("LAYOUT_ICON_SIZE 图标 ∈ {16,18,20,24}px", not icon_bad,
+        "; ".join(icon_bad[:4]) if icon_bad else "", level="WARN")
+    chk("LAYOUT_LABEL_COLLIDE 图表标签盒足够宽 / 结构图无裸文字箭头", not label_collapse,
+        "; ".join(label_collapse[:4]) if label_collapse else "")
+    chk(f"LAYOUT_FILL 填充率目标 {fill_t[0]}–{fill_t[1]}%（{mode}）",
+        not fill_low and not fill_high,
+        ("过空: " + "; ".join(fill_low[:3]) if fill_low else "") +
+        ((" 过满: " + "; ".join(fill_high[:3])) if fill_high else ""))
+
+
+def _check_annotations(txt, chk, model):
+    """待核实标注：.tbd 内联标色须配 .tbd-legend / .flagbar 说明；单页数量上限。"""
+    an = LC.get('annotations') or {}
+    max_per_page = an.get('flagMaxPerPage', 12)
+    require_legend = an.get('requireLegend', True)
+    tbd_re = re.compile(r'class="tbd(?:\s|")')
+    no_legend, over = [], []
+    for i, b in enumerate(_bands(txt), 1):
+        n = len(tbd_re.findall(b))
+        if n and require_legend and ('tbd-legend' not in b and 'flagbar' not in b):
+            no_legend.append(f"第{i}页")
+        if n > max_per_page:
+            over.append(f"第{i}页 {n}处")
+    chk("待核实标注 .tbd 均配 .tbd-legend/.flagbar 说明", not no_legend,
+        f"{no_legend} 有标色项但无说明（用户不知为何标色）" if no_legend else "", level="WARN")
+    chk(f"单页待核实标注 ≤ {max_per_page} 处（过密则转 .flagbar 清单）", not over,
+        "; ".join(over) if over else "", level="WARN")
+    if model:
+        mflags = [s for s in (model.get('sections') or []) if s.get('flags')]
+        if mflags and 'flagbar' not in txt and 'tbd-legend' not in txt:
+            chk("模型 flags ↔ 正文待核实条对应", False,
+                f"{len(mflags)} 页模型含 flags，正文无 .flagbar/.tbd-legend", level="WARN")
+
+
+def _img_holders(sec):
+    """页内可能承载素材图片的容器（image 页型 / split 右栏）。"""
+    out = []
+    if isinstance(sec.get('image'), dict):
+        out.append(sec['image'])
+    right = sec.get('right')
+    if isinstance(right, dict) and isinstance(right.get('image'), dict):
+        out.append(right['image'])
+    return out
+
+
+def _img_srcs(img):
+    """图片对象的全部真实 src（image.src + image.items[].src），排除占位符。"""
+    if img.get('placeholder'):
+        return []
+    srcs = []
+    if isinstance(img.get('src'), str) and img['src'].strip():
+        srcs.append(img['src'])
+    for it in (img.get('items') or []):
+        if isinstance(it, dict) and isinstance(it.get('src'), str) and it['src'].strip():
+            srcs.append(it['src'])
+        elif isinstance(it, str) and it.strip():
+            srcs.append(it)
+    return srcs
+
+
+def _check_media(txt, chk, model):
+    """素材图片与配图占位：零外链铁律（<img> 只允许 data: 内联或相对路径）；
+    alt 可访问性；配图占位必须有可见标签；模型 image 三选一（src/items/placeholder）
+    与正文版式、数量、裁切策略一一对应。"""
+    imgs = re.findall(r'<img\b[^>]*>', txt)
+    if imgs:
+        bad = [t for t in imgs if re.search(r'src\s*=\s*["\']\s*(?:https?:)?//', t)]
+        chk("图片源无外链（data: 内联或相对路径，零外链铁律）", not bad,
+            f"{len(bad)} 处外链图片" if bad else "")
+        noalt = [t for t in imgs if 'alt=' not in t]
+        chk("图片均带 alt（可访问性）", not noalt,
+            f"{len(noalt)} 处缺 alt" if noalt else "", level="WARN")
+        data_imgs = [t for t in imgs if re.search(r'src\s*=\s*["\']\s*data:', t)]
+        inline_bytes = sum(len(t) for t in data_imgs)
+        chk("单图 data: 内联体积在上限内", all(len(t) <= IMAGE_MAX_INLINE for t in data_imgs),
+            f"最大 {max((len(t) for t in data_imgs), default=0) // 1024}KB > 上限 {IMAGE_MAX_INLINE // 1024}KB",
+            level="WARN")
+        chk("报告内联图片总量在上限内", inline_bytes <= IMAGE_MAX_TOTAL,
+            f"{inline_bytes // 1024}KB > 上限 {IMAGE_MAX_TOTAL // 1024}KB（改用相对路径）",
+            level="WARN")
+    ph_blocks = re.findall(r'class="[^"]*\bmedia--ph\b[^"]*"', txt)
+    if ph_blocks:
+        chk("配图占位含可见标签（.media__ph）", 'media__ph' in txt,
+            f"{len(ph_blocks)} 处 .media--ph 缺 .media__ph 标签（空占位 = 不合格）")
+    if not model:
+        return
+    holders = [(i, img) for i, s in enumerate(model.get('sections') or [])
+               if isinstance(s, dict) for img in _img_holders(s)]
+    if not holders:
+        return
+    bad_src, bad_layout, bad_fit, too_many = [], [], [], []
+    n_declared = 0
+    n_ph = 0
+    for idx, img in holders:
+        srcs = _img_srcs(img)
+        n_declared += len(srcs)
+        if img.get('placeholder'):
+            n_ph += 1
+        for s in srcs:
+            if re.match(r'\s*(?:https?:)?//', str(s)):
+                bad_src.append(f"sections[{idx}]")
+        layout = str(img.get('layout') or ('grid' if len(img.get('items') or []) > 1 else 'full')).lower()
+        if layout not in IMAGE_LAYOUTS:
+            bad_layout.append(f"sections[{idx}]:{layout}")
+        if img.get('fit') and str(img['fit']).lower() not in IMAGE_FIT:
+            bad_fit.append(f"sections[{idx}]:{img['fit']}")
+        items = img.get('items') or []
+        if len(items) > IMAGE_MAX_PER_PAGE:
+            too_many.append(f"sections[{idx}]={len(items)}")
+        if layout in IMAGE_MULTI_LAYOUTS and not img.get('placeholder') and len(items) < 2:
+            too_many.append(f"sections[{idx}] {layout} 需 ≥2 张图（当前 {len(items)}）")
+    chk("模型 image.src 非外链（data:/相对路径）", not bad_src,
+        f"{len(bad_src)} 处外链 src" if bad_src else "")
+    chk("模型 image.layout 合法（imageSpec.layouts）", not bad_layout,
+        f"非法版式 {sorted(set(bad_layout))}" if bad_layout else "")
+    chk("模型 image.fit 合法（cover/contain）", not bad_fit,
+        f"非法裁切 {sorted(set(bad_fit))}" if bad_fit else "")
+    chk("单页图片数/多图版式图数合规（imageSpec.maxPerPage · 多图版式 ≥2 张）", not too_many,
+        f"{too_many}" if too_many else "", level="WARN")
+    layout_class = {'grid': 'media-grid', 'compare': 'media-compare', 'wall': 'media-wall',
+                    'bleed': 'media--bleed', 'ph': 'media--ph'}
+    ratio_css = IMAGE_SPEC.get('ratioCssClass') or {}
+    miss_cls = []
+    for idx, img in holders:
+        lay = str(img.get('layout') or ('grid' if len(img.get('items') or []) > 1 else 'full')).lower()
+        cls = layout_class.get(lay)
+        if cls and cls not in txt:
+            miss_cls.append(f"sections[{idx}]:{lay}→.{cls}")
+        # 比例锁定类：HTML 与 PPTX 用同一份 imageSpec.ratioCssClass，缺了就会出现"比例走样"
+        rcls = ratio_css.get(lay)
+        if rcls and ('.' + rcls) not in txt:
+            miss_cls.append(f"sections[{idx}]:{lay}→.{rcls}（比例未锁定）")
+        if img.get('placeholder') and 'media--ph' not in txt:
+            miss_cls.append(f"sections[{idx}]:placeholder→.media--ph")
+    chk("模型图片版式 ↔ 正文版式类/比例锁定类对应", not miss_cls,
+        f"缺 {sorted(set(miss_cls))}" if miss_cls else "", level="WARN")
+    if n_declared and len(imgs) < n_declared:
+        chk("模型 image ↔ 正文图片数量对应", False,
+            f"模型声明 {n_declared} 张但正文仅 {len(imgs)} 个 <img>", level="WARN")
+    if n_ph and not ph_blocks:
+        chk("模型配图占位 ↔ 正文 .media--ph 对应", False,
+            f"{n_ph} 页模型声明 image.placeholder 但正文无 .media--ph", level="WARN")
+
+
+# 失败检查项 → 失败模式 / 处置动作 / 精确取码命令。
+# 目的：校验失败时直接给出「改什么、按什么顺序改、去哪取代码」，
+# 使智能体不必整读 references/failure-modes.md（17KB）就能收敛。
+# 关键词按检查项名称匹配；新增检查项时同步在此登记，否则只回落通用处置顺序。
+FIX_GUIDE = [
+    (("页高", "溢出", "满屏", "文字预算"), "F4 容器溢出",
+     "① 列表化/精炼 ② 升级承载形态 ③ 换/扩组合版式 ④ 分区 ⑤ 拆页 ⑥ 最后才有限缩字号",
+     "--task content-rules"),
+    (("过空", "内容不足", "密度"), "F1 内容不足 / F3 密度塌陷",
+     "补证据与含义（数字+口径+so-what），不要用装饰或放大字号填空",
+     "--task content-rules"),
+    (("图表", "多样性", "登记"), "F6 图表降级 / F12 图表单一",
+     "换图表类型拉开多样性；连续两页不得同型；类型须在 charts.registry 登记",
+     "--task chart-pick"),
+    (("最小尺寸",), "图表尺寸不足",
+     "把 svg 的 viewBox 高/显示宽度提到 charts.minSize 之上（尺寸问题改尺寸，不要靠换图型回避）",
+     "--task chart-pick"),
+    (("Exhibit",), "研究模式证据编号",
+     "每个 .exhibit 框都要有 .exhibit__no（Exhibit N，全篇连续）与 .exhibit__src 来源行",
+     "--task research-evidence"),
+    (("引用", "参考资料", "锚点"), "引用闭环",
+     "正文 [n] 上标与文末条目双向对齐且编号从 1 连续；ref-link 带 target/rel",
+     "--file components-atoms.md --section 7"),
+    (("待核实", "tbd"), "待核实标注",
+     "每处 .tbd 必须配 .tbd-legend 或 .flagbar 说明口径，只标色不解释即不合格",
+     "--file components-atoms.md --section 11b"),
+    (("强调", "风格", "主题", "配色"), "F8 主题/风格漂移",
+     "风格是皮肤不是解药：末页禁 band--deep，收尾用 band--accent；单一强调色",
+     "--file components-atoms.md --section 11"),
+    (("so-what", "结论", "标题"), "F15 空洞结论",
+     "research 主标题须是结论句（≥12 字含数字或判断词）；so-what 禁套话",
+     "--task content-rules"),
+    (("版式", "节奏", "组合"), "F11 单件页默认",
+     "默认一页=主件+从件+注释；同一版式不连用超 2 页，密度 L/M/H 交替",
+     "--task presentation-combo"),
+    (("图片", "占位", "media"), "F9 配图走样 / F10 空占位",
+     "无素材用 image.placeholder + .media--ph 锁版式；路径以模型目录为锚；禁外链",
+     "--task image-layout"),
+    (("AI", "去AI味"), "文风",
+     "改写命中的高危词；大段文字转列表",
+     "--task content-rules"),
+]
+
+
+def _print_fix_guide(results, strict, width):
+    """校验未通过时输出定向修复指引（只列命中的失败模式，不做全量倾倒）。"""
+    failed = [(lv, nm, nt) for lv, nm, ok, nt in results
+              if not ok and (lv == "FAIL" or strict)]
+    if not failed:
+        return
+    seen, guides = set(), []
+    for _, name, _ in failed:
+        nl = name.lower()
+        # 最长关键词优先（而非首个命中即 break）：避免「图表最小尺寸」被泛化的
+        #「图表」关键词抢走、误路由到「换图表类型」
+        best = None
+        for keys, mode, action, cmd in FIX_GUIDE:
+            hit = max((len(k) for k in keys if k.lower() in nl), default=0)
+            if hit and (best is None or hit > best[0]):
+                best = (hit, mode, action, cmd)
+        if best and best[1] not in seen:
+            seen.add(best[1])
+            guides.append((best[1], best[2], best[3]))
+    print("-" * width)
+    print("修复指引（按下列顺序改；跳步直接缩字号/砍内容会把问题推给下一环）：")
+    if guides:
+        for i, (mode, action, cmd) in enumerate(guides, 1):
+            print(f"  {i}. [{mode}] {action}")
+            print(f"     取码: python scripts/extract_snippet.py {cmd}")
+    else:
+        print("  未匹配到已登记的失败模式，按通用顺序处置：")
+        print("  ① 补内容 ② 优化形态（列表化/精炼）③ 换承载形态 "
+              "④ 调容器/网格 ⑤ 有限缩字号 ⑥ 最后拆页")
+    print("  完整失败模式库与错误解释纠正表: references/failure-modes.md")
+
+
+def main():
+    if len(sys.argv) < 2:
+        print(__doc__)
+        return 2
+    path = Path(sys.argv[1])
+    strict = '--strict' in sys.argv
+    as_json = '--json' in sys.argv
+    if not path.exists():
+        print(f"文件不存在: {path}")
+        return 2
+    txt = path.read_text(encoding='utf-8')
+
+    v = Struct()
+    v.feed(txt)
+
+    results = []
+
+    def chk(name, ok, note="", level="FAIL"):
+        results.append((level, name, bool(ok), note))
+
+    # ── 结构 ──
+    chk("HTMLParser 结构 0 未闭合", not v.stack, f"剩余 {v.stack}" if v.stack else "")
+    chk("HTMLParser 0 失配/游离", not v.mis and not v.err,
+        f"mis={v.mis} err={v.err}" if (v.mis or v.err) else "")
+
+    # ── 主题与风格 ──
+    chk("data-style 已指定风格", 'data-style="' in txt)
+    chk('data-theme="light" + dark 变量块',
+        'data-theme="light"' in txt and '[data-theme="dark"]' in txt)
+    chk("主题切换按钮 + localStorage + prefers-color-scheme",
+        'id="themeBtn"' in txt and 'localStorage' in txt and 'prefers-color-scheme' in txt)
+
+    # ── 模式 ──
+    mode_m = re.search(r'<html[^>]*data-mode="([^"]+)"', txt)
+    mode = mode_m.group(1) if mode_m else None
+    chk("data-mode 已声明模式（presentation/research/architecture）",
+        mode in MODE_BUDGETS,
+        f"读到 {mode!r}（未声明按 presentation 处理）" if mode not in MODE_BUDGETS else "",
+        level="WARN" if mode is None else "FAIL")
+    mode = mode if mode in MODE_BUDGETS else 'presentation'
+
+    # ── 宽屏与页面高度模型 ──
+    B0 = MODE_BUDGETS[mode]
+    _w = B0['wrap']
+    chk(f"版心 --wrap: {_w}px（{mode} 模式）",
+        f'--wrap:{_w}px' in txt or f'--wrap: {_w}px' in txt)
+    _check_page_model(txt, chk, mode)
+    chk("scroll-snap 翻页停靠", 'scroll-snap-type' in txt)
+
+    # ── 翻页与大纲 ──
+    chk("翻页 JS（方向键 + scrollIntoView）",
+        'scrollIntoView' in txt and 'ArrowRight' in txt)
+    chk("页码指示 pager", 'pager__dot' in txt and 'id="pagerDots"' in txt)
+    has_agenda = 'id="agenda"' in txt
+    if mode == 'architecture':
+        # architecture 极简形态：内容页 ≤4 时省略 Agenda 合法（封面/收尾/参考资料不计内容页）
+        struct_ids = ('agenda', 'next', 'refs', 'appendix')
+        bands_a = _bands(txt)
+        content_n = 0
+        for i, b in enumerate(bands_a):
+            if i == 0:
+                continue
+            if any(f'id="{sid}"' in b[:220] for sid in struct_ids):
+                continue
+            content_n += 1
+        if content_n > 4:
+            chk("存在 Agenda 大纲页（architecture 内容页 >4 时建议补）", has_agenda,
+                f"内容页 {content_n} 页无 Agenda", level="WARN")
+    else:
+        chk("存在 Agenda 大纲页", has_agenda)
+    if has_agenda:
+        band_secs = [m.start() for m in re.finditer(r'<section class="band', txt)]
+        agenda_pos = txt.find('id="agenda"')
+        second_band_pos = band_secs[1] if len(band_secs) > 1 else -1
+        ok_second = (second_band_pos != -1 and
+                     band_secs[0] < agenda_pos < (band_secs[2] if len(band_secs) > 2 else len(txt)))
+        chk("Agenda 位于第二页", ok_second,
+            f"共{len(band_secs)}页 agenda@{agenda_pos} 第2页@{second_band_pos}")
+        n_items = txt.count('class="agenda__a"')
+        chk("Agenda 条目可点击跳转 (≥4)", n_items >= 4, f"{n_items} 条")
+        if n_items > AGENDA_SINGLE_MAX:
+            chk(f"Agenda >{AGENDA_SINGLE_MAX} 条时已用 agenda--2col 双列",
+                'agenda--2col' in txt, f"{n_items} 条未双列", level="WARN")
+
+    # ── 主标题粗体 ──
+    chk("主标题粗体 --fw-title/--fw-display",
+        ('--fw-title' in txt and 'font-weight:var(--fw-title)' in txt) or
+        re.search(r'\.t-h1\{[^}]*font-weight:\s*(600|700)', txt))
+
+    # ── 零外链 ──
+    ext = re.findall(r'<(?:link|script)[^>]+(?:href|src)=["\']https?://[^"\']+', txt)
+    chk("无 <link>/<script src> 外链", not ext, f"{len(ext)} 处" if ext else "")
+    img_ext = re.findall(r'<img[^>]+src=["\']https?://', txt)
+    chk("无外链图片", not img_ext, f"{len(img_ext)} 处" if img_ext else "")
+
+    # ── 引用与锚点闭环 ──
+    # 编号须匹配多位数：`ref-\d` 会让 ref-10 及以后的条目在两侧同时落空，
+    # 使双向对齐检查在 ≥10 条引用时静默通过（假阴性）。
+    cites = re.findall(r'class="cite" href="#(ref-\d+)"', txt)
+    refs = re.findall(r'id="(ref-\d+)"', txt)
+    if cites or refs:
+        chk("引用标记 ↔ 参考资料条目 双向对齐", set(cites) == set(refs),
+            f"引用{sorted(set(cites))} 条目{sorted(set(refs))}")
+        nums = sorted(int(r.split('-')[1]) for r in set(refs))
+        chk(f"参考资料编号连续（1..{len(nums)} 无跳号）", nums == list(range(1, len(nums) + 1)),
+            f"读到 {nums}" if nums != list(range(1, len(nums) + 1)) else "")
+    else:
+        chk("引用标记 ↔ 参考资料条目 双向对齐", True, "（无外部引用）", level="WARN")
+    bad_links = re.findall(r'<a class="ref-link"(?![^>]*target="_blank")[^>]*>', txt) + \
+                re.findall(r'<a class="ref-link"(?![^>]*rel="noopener")[^>]*>', txt)
+    chk('.ref-link 均带 target="_blank" rel="noopener"', not bad_links,
+        f"{len(bad_links)} 处" if bad_links else "")
+    _check_anchors(txt, chk)
+
+    # ── 单一强调色（色值表来自 layout-constants.json styleAccents 单源） ──
+    style_m = re.search(r'<html[^>]*data-style="([^"]+)"', txt)
+    style = style_m.group(1) if style_m else 'business-blue'
+    body_txt = txt[txt.find('</style>'):] if '</style>' in txt else txt
+    body_txt = re.sub(r'<script\b[\s\S]*?</script>', ' ', body_txt)
+    if STYLE_ACCENTS:
+        own = STYLE_ACCENTS.get(style, set())
+        others = ALL_ACCENTS - own
+        bad_hues = [h for h in others if h in body_txt.lower()]
+        if style == 'spectrum':
+            chk("彩色模块边界（spectrum：数据色限 c1–c5，无其它风格强调色）", not bad_hues,
+                f"残留 {bad_hues}" if bad_hues else "")
+        else:
+            chk(f"单一强调色（{style}，正文无第二色相）", not bad_hues,
+                f"残留 {bad_hues}" if bad_hues else "")
+
+    hard_hex = re.findall(r'style="[^"]*(?:color|background|border-color|fill|stroke)\s*:\s*#',
+                          body_txt)
+    chk("正文无内联写死 hex 色（用 var()/语义类）", not hard_hex,
+        f"{len(hard_hex)} 处" if hard_hex else "", level="WARN")
+
+    # ── 内容与密度 ──
+    paras = re.findall(r'<p class="t-body"[^>]*>(.*?)</p>', txt, re.S)
+    long_paras = [x for x in paras if len(re.sub(r'<[^>]+>', '', x).strip()) > 110]
+    chk("卡片内无超 110 字纯段落（应转列表）", not long_paras,
+        f"{len(long_paras)} 段" if long_paras else "")
+    _check_bands(txt, chk, mode)
+    _check_table_rows(txt, chk, mode)
+    _check_mode_layouts(txt, chk, mode)
+    if mode != 'architecture':
+        _check_icons(txt, chk)
+    _check_pptx_export(txt, chk)
+    model = _check_model_consistency(txt, chk, mode, style)
+    _check_type_features(txt, chk, model)
+    _check_exhibits(txt, chk, mode)
+    _check_research_extras(txt, chk, mode)
+    _check_emphasis(txt, chk)
+    _check_annotations(txt, chk, model)
+    _check_media(txt, chk, model)
+
+    if mode == 'research':
+        STRUCT = ('报告大纲', '大纲', '参考资料', '数据来源', '下一步', '结论', '全文核心', 'agenda', '附录')
+        h1s = re.findall(r'<h2 class="t-h1 shead__title"[^>]*>(.*?)</h2>', txt)
+        short = [re.sub(r'<[^>]+>', '', h).strip() for h in h1s]
+        short = [h for h in short if 0 < len(h) < 12 and not h.lower().startswith(STRUCT)]
+        chk("research 行动标题（章节主标题 ≥12 字，标题即结论）", not short,
+            f"过短: {short[:3]}" if short else "", level="WARN")
+
+    _check_content_quality(txt, chk, mode, model)
+    _check_v9_hard_gates(txt, chk, model)
+    _check_layout_grammar(txt, chk, mode)
+
+    # ── 去AI味（词表来自单源） ──
+    body_plain = _plain(body_txt)
+    hits = [w for w in AI_FLAVOR if re.search(w, body_plain)]
+    chk("去AI味（无高危 AI 腔词汇）", not hits, f"命中 {hits}" if hits else "", level="WARN")
+
+    # ── 图表 ──
+    _check_charts(txt, chk)
+    _check_chart_variety(txt, chk, mode)
+
+    # ── 汇总 ──
+    n_fail = sum(1 for lv, _, ok, _ in results if not ok and lv == "FAIL")
+    n_warn = sum(1 for lv, _, ok, _ in results if not ok and lv == "WARN")
+    n_pass = sum(1 for _, _, ok, _ in results if ok)
+    verdict = "通过，可交付" if n_fail == 0 and (not strict or n_warn == 0) else \
+              ("有警告" if n_fail == 0 else "不通过，需修复后重跑")
+    exit_code = 1 if (n_fail > 0 or (strict and n_warn > 0)) else 0
+
+    if as_json:
+        print(json.dumps({
+            "file": str(path),
+            "strict": strict,
+            "pass": n_pass, "warn": n_warn, "fail": n_fail,
+            "verdict": verdict,
+            "checks": [{"level": lv, "name": nm, "ok": ok, "note": nt}
+                       for lv, nm, ok, nt in results],
+        }, ensure_ascii=False, indent=2))
+        return exit_code
+
+    W = 68
+    print("=" * W)
+    print(f"TopPPT HTML· 质量校验 {LC.get('version', '')}  {'[STRICT]' if strict else ''}")
+    print(f"文件: {path.name}  ({len(txt.encode('utf-8'))/1024:.1f} KB)")
+    print("=" * W)
+    for level, name, ok, note in results:
+        tag = "PASS" if ok else ("WARN" if level == "WARN" else "FAIL")
+        line = f"[{tag}] {name}"
+        if note and not ok:
+            line += f"  -> {note}"
+        print(line)
+    print("-" * W)
+    print(f"PASS {n_pass}  WARN {n_warn}  FAIL {n_fail}")
+    print(f"结论: {verdict}")
+    if exit_code != 0:
+        _print_fix_guide(results, strict, W)
+    print("=" * W)
+    return exit_code
+
+
+if __name__ == '__main__':
+    sys.exit(main())