@topmindspace/tms-skills 0.1.1 → 0.1.3

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

@@ -3,10 +3,12 @@
 """
 TopPPT HTML· HTML 报告质量校验（交付闭环）
 用法:
-    python validate_report.py <报告.html> [--strict] [--json]
+    python validate_report.py <报告.html> [--strict] [--layout-qa] [--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）。
+Mode A / presentation + --strict：自动启用 --layout-qa（V 契约/截断/溢出未拆页/半空卡/对齐）；
+research / architecture 仍须显式传 --layout-qa（不强制）。
 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），
@@ -218,7 +220,12 @@ 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）。"""
+    """图表与版式多样性（阈值单源 charts.variety；判定逻辑 checks_html）。
+
+    多样性仍是特性：全部 data-chart（含 advanced）计入下限——用 advanced 不罚。
+    preferCoreFirst：核图多样性不足却靠 advanced 撑场时 WARN（勿凑下限）。
+    禁反模式由 layout-qa / sizeByComplexity / CHART_SKEW 另检。
+    """
     v = CHART_VARIETY
     if not v:
         return
@@ -230,8 +237,14 @@ def _check_chart_variety(txt, chk, mode):
     n_chart_pages = len([1 for ts in per_page if ts])
 
     floor = checks_html.variety_floor(mode, n_chart_pages, v)
+    # advanced 与核图一并计入 distinct（不因用 advanced 受罚）
     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}")
+    # preferCoreFirst WARN 仅 A/B：C 架构以结构/甘特等 advanced 为主属正常，不噪音
+    if mode in ('presentation', 'research'):
+        pref = checks_html.variety_core_preference(used, v)
+        if pref:
+            chk("图表多样性·核图优先（preferCoreFirst）", False, pref, level="WARN")
 
     if v.get('noRepeatAdjacent'):
         adj = checks_html.adjacent_same_type(per_page)
@@ -323,7 +336,7 @@ def _check_bands(txt, chk, mode="presentation"):
         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）")
+        ("; ".join(est_over) + "（按「重构承载→拆页/分章→换形态→有限 fontShrink」；禁静默截断；或长结构页加 band--flow）")
         if est_over else "")
 
     bad_grids = 0
@@ -888,6 +901,45 @@ def _check_v9_hard_gates(txt, chk, model):
         not oversize, "; ".join(oversize[:4]) if oversize else "")
 
 
+
+def _class_tokens(tag_or_html: str) -> set[str]:
+    """Extract HTML class tokens from class="..." attributes."""
+    out: set[str] = set()
+    for m in re.finditer(r'class="([^"]*)"', tag_or_html):
+        out.update(m.group(1).split())
+    return out
+
+
+def _mixed_grid_needs_align(band: str) -> bool:
+    """True when a grid/g-* region pairs media with cards/lists (token-exact; no t-metric / data-chart false hits)."""
+    def has_media(classes: set[str], chunk: str) -> bool:
+        if classes & {'fig', 'media', 'chart'}:
+            return True
+        if any(c.startswith('media') or c.startswith('fig') for c in classes):
+            return True
+        return 'data-chart=' in chunk
+
+    def has_cards(classes: set[str]) -> bool:
+        if classes & {'card', 'ul', 'metric'}:
+            return True
+        return any(
+            c.startswith('ul--') or c.startswith('card') or c.startswith('metric__')
+            for c in classes)
+
+    # Find each grid opening and inspect following chunk
+    for m in re.finditer(r'<div class="([^"]*)"', band):
+        classes_on_grid = set(m.group(1).split())
+        if not (classes_on_grid & {'grid'} or any(re.fullmatch(r'g-\d+', c) or c.startswith('g-') for c in classes_on_grid)):
+            # allow g-hero / g-side / g-2 etc
+            if not any(c == 'grid' or c.startswith('g-') for c in classes_on_grid):
+                continue
+        chunk = band[m.start(): m.start() + 4500]
+        classes = _class_tokens(chunk)
+        if has_media(classes, chunk) and has_cards(classes):
+            return True
+    return False
+
+
 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 {}
@@ -920,7 +972,10 @@ def _check_layout_grammar(txt, chk, mode):
 
     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')):
+        # intentional whitespace：封面/章节幕/金句/收尾等休止页不参与 FILL 门禁
+        if any(f'id="{x}' in head for x in (
+                'refs', 'appendix', 'cover', 'agenda', 'quote',
+                'closing', 'section', 'chapter', 'next')):
             continue
         if 'band--flow' in head:
             continue
@@ -956,11 +1011,10 @@ def _check_layout_grammar(txt, chk, mode):
         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))
+        # ③ 混排对齐（grid 内同时有图/媒体与卡/列表；token 精确）
+        mixed = _mixed_grid_needs_align(b)
         if mixed and mixed_need:
-            if not any(t in b for t in align_tokens):
+            if not any(re.search(r'\b' + re.escape(tok) + r'\b', b) for tok 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）
@@ -1161,6 +1215,33 @@ def _check_media(txt, chk, model):
         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")
 
+    # 配图页 caption / so-what（廉价 WARN）；近邻主张/导语亦可
+    bands = re.split(r'(?=<section\b)', txt)
+    miss_cap = []
+    near_empty = []
+    for i, b in enumerate(bands):
+        if not re.search(r'class="[^"]*\bmedia\b|<img\b|class="[^"]*media--', b):
+            continue
+        if len(b) < 80:
+            continue
+        has_cap = bool(re.search(
+            r'class="[^"]*(?:fig__cap|media__cap|media__ph|caption|exhibit__src|footnote|so-what|lead)', b)
+            or re.search(r'<figcaption\b|class="[^"]*\bsoWhat\b', b))
+        if not has_cap:
+            miss_cap.append(f'band[{i}]')
+        # 近图过空：有 media 但正文文本极少且无要点/指标（WARN）
+        textish = re.sub(r'<script[\s\S]*?</script>|<style[\s\S]*?</style>|<[^>]+>', ' ', b)
+        textish = re.sub(r'\s+', ' ', textish).strip()
+        has_points = bool(re.search(r'class="[^"]*(?:points|bullets|kpi|metrics|card)', b))
+        if len(textish) < 40 and not has_points and not re.search(r'media--ph', b):
+            near_empty.append(f'band[{i}]')
+    chk("配图页含 caption/图注（IMAGE_CAPTION）", not miss_cap,
+        f"{len(miss_cap)} 处配图区缺 .fig__cap/.media__cap/.footnote/so-what 等图注" if miss_cap else "",
+        level="WARN")
+    chk("配图页近邻过空（IMAGE_NEAR_EMPTY）", not near_empty,
+        f"{len(near_empty)} 处配图页几乎无注解/要点（补 caption 或要点条）" if near_empty else "",
+        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」，
@@ -1233,11 +1314,227 @@ def _print_fix_guide(results, strict, width):
             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("  ① 重构承载（列表/卡/表/图）② 拆页/分章 ③ 换布局形态 "
+              "④ 有限缩字号 ⑤ 禁止静默截断/砍 so-what")
     print("  完整失败模式库与错误解释纠正表: references/failure-modes.md")
 
 
+
+def _check_layout_qa(txt, chk, mode, model):
+    """布局 QA（--layout-qa；presentation+--strict 自动开）：
+    V 契约 / 极偏环图 / 骨架连用 / 缺 data-skel / 简单全幅 /
+    截断迹象 / 溢出未拆页 / 半空卡 / 列对齐节奏。
+    """
+    bands = _bands(txt)
+    skip_ids = {'cover', 'agenda', 'refs', 'appendix', 'next', 'quote'}
+    content = []
+    for i, b in enumerate(bands, 1):
+        head = b[:240]
+        if any(f'id="{sid}"' in head for sid in skip_ids):
+            continue
+        if 'band--deep' in head or 'band--accent' in head:
+            # 金句/强调带允许无骨架
+            if 'data-skel=' not in b and not re.search(r'class="[^"]*\bP\d+\b', b):
+                continue
+        content.append((i, b, head))
+
+    # ① 缺 data-skel：若全文已出现 data-skel（scaffold/render 管线），则内容页必须都有
+    has_any_skel = 'data-skel="' in txt
+    missing = []
+    if has_any_skel:
+        for i, b, head in content:
+            if 'data-skel="' not in b and not re.search(r'class="[^"]*\bP\d+\b', b):
+                missing.append(f"第{i}页")
+    chk("LAYOUT_QA_MISSING_SKEL 内容页须带 data-skel（管线产物）",
+        not missing,
+        ("缺骨架: " + "; ".join(missing[:5])) if missing else
+        ("（全文无 data-skel，跳过——手写示例豁免；scaffold/render 会写入）" if not has_any_skel else ""))
+
+    # ② 连续 ≥3 页同一 data-skel（或同一 layout 签名）
+    skels = []
+    for i, b, head in content:
+        m = re.search(r'data-skel="(P\d+)"', b)
+        if m:
+            skels.append((i, m.group(1)))
+        else:
+            sig = _layout_sig(b)
+            if sig:
+                skels.append((i, f"sig:{sig}"))
+    streak_hits = []
+    run_v, run_s, run_n = None, 0, 0
+    for i, v in skels:
+        if v == run_v:
+            run_n += 1
+        else:
+            if run_v and run_n >= 3:
+                streak_hits.append(f"{run_v}×{run_n}（起第{run_s}页）")
+            run_v, run_s, run_n = v, i, 1
+    if run_v and run_n >= 3:
+        streak_hits.append(f"{run_v}×{run_n}（起第{run_s}页）")
+    chk("LAYOUT_QA_SKEL_STREAK 同一 data-skel/版式签名连续 <3 页",
+        not streak_hits, "; ".join(streak_hits[:3]) if streak_hits else "")
+
+    # ③ 极偏仍用 donut/pie（模型侧；与 CHART_SKEW 互补，专打 layout-qa 关键字）
+    skew_hits = []
+    sections = (model or {}).get('sections') or []
+    for si, sec in enumerate(sections, 1):
+        if not isinstance(sec, dict):
+            continue
+        ch = sec.get('chart') if isinstance(sec.get('chart'), dict) else {}
+        ctype = str(ch.get('type') or (sec.get('type') if sec.get('type') in
+                                         ('donut', 'pie', 'multidonut') else '') or '').lower()
+        if sec.get('type') == 'donut':
+            ctype = 'donut'
+        if ctype not in ('donut', 'pie', 'multidonut'):
+            continue
+        vals = [float(v) for v in (ch.get('values') or []) if isinstance(v, (int, float))]
+        vals = [v for v in vals if v >= 0]
+        if len(vals) < 2:
+            continue
+        total = sum(vals) or 1.0
+        pcts = [v / total * 100 for v in vals]
+        mn, mx = min(pcts), max(pcts)
+        ratio = (mx / mn) if mn > 0 else 999
+        if mn < 5.0 or ratio > 20:
+            skew_hits.append(f"sections[{si}] {ctype} 最小{mn:.1f}% 比{ratio:.0f}:1 → 改 V3/KPI")
+    chk("LAYOUT_QA_SKEW_DONUT 极偏占比禁用 donut/pie（改 V3 KPI）",
+        not skew_hits, "; ".join(skew_hits[:3]) if skew_hits else "")
+
+    # ④ 演示模式：简单图（≤2 类）却近全幅（svg height≥320 或 width:100% 无注解带）
+    bleed_hits = []
+    if mode == 'presentation':
+        for i, b, head in content:
+            for m in re.finditer(
+                    r'<svg\b[^>]*data-chart="([^"]+)"[^>]*style="([^"]*)"[^>]*>', b):
+                ctype, style = m.group(1), m.group(2)
+                svg_end = b.find('</svg>', m.end())
+                block = b[m.start():svg_end if svg_end > 0 else m.end() + 600]
+                n_lab = len(re.findall(r'<text\b', block))
+                n_pts = max(n_lab, len(re.findall(r'<rect\b', block)),
+                            len(re.findall(r'<circle\b', block)))
+                hm = re.search(r'height:\s*(\d+(?:\.\d+)?)px', style)
+                h_px = float(hm.group(1)) if hm else 0
+                simple = n_lab <= 2 or n_pts <= 3
+                full = h_px >= 320 or ('width:100%' in style and h_px >= 240)
+                has_anno = bool(re.search(r'class="[^"]*(?:sowhat|anno|points|callout)', b))
+                if simple and full and not has_anno:
+                    bleed_hits.append(f"第{i}页 {ctype} 简单全幅无注解")
+    chk("LAYOUT_QA_SIMPLE_FULLBLEED 演示页简单图禁全幅无注解（用 V1–V4）",
+        not bleed_hits, "; ".join(bleed_hits[:3]) if bleed_hits else "")
+
+    # ⑤ V 契约：演示页 data-v / data-skel 与内容复杂度粗检
+    v_hits = []
+    if mode == 'presentation' and sections:
+        for si, sec in enumerate(sections, 1):
+            if not isinstance(sec, dict):
+                continue
+            pt = str(sec.get('type') or '')
+            if pt in ('cover', 'agenda', 'closing', 'quote'):
+                continue
+            ch = sec.get('chart') if isinstance(sec.get('chart'), dict) else {}
+            vals = ch.get('values') or []
+            n = len(vals) if vals else len(sec.get('points') or sec.get('items') or [])
+            skel = str(sec.get('layoutPreset') or '')
+            # 极偏却仍 donut
+            if pt == 'donut' or ch.get('type') in ('donut', 'pie'):
+                if vals:
+                    nums = [float(v) for v in vals if isinstance(v, (int, float)) and v >= 0]
+                    if len(nums) >= 2:
+                        total = sum(nums) or 1
+                        pcts = [v / total * 100 for v in nums]
+                        if min(pcts) < 5:
+                            v_hits.append(f"sections[{si}] 极偏仍 {pt or ch.get('type')}（应 V3/kpi）")
+            # 简单 1–2 点却声明 V1 全幅主视觉（layoutPreset P1 + 简单）
+            if skel == 'P1' and n <= 2 and (pt in ('bar', 'donut') or ch.get('type')):
+                v_hits.append(f"sections[{si}] P1+简单{n}点（应 V3/P3 或加注解）")
+    chk("LAYOUT_QA_V_CONTRACT 演示 V1–V4 与复杂度匹配",
+        not v_hits, "; ".join(v_hits[:3]) if v_hits else "")
+
+    # ⑥ 截断迹象：正文/列表项以省略号截断充数（antiTruncation）
+    at = (LC.get('qualityGates') or {}).get('antiTruncation') or {}
+    trunc_hits = []
+    if at.get('forbidEllipsisTruncate', True):
+        patterns = list(at.get('ellipsisPatterns') or ['…', '...', '……'])
+        for i, b, head in content:
+            plain_parts = re.findall(r'<(?:li|p)[^>]*>([\s\S]*?)</(?:li|p)>', b)
+            plain_parts += re.findall(
+                r'class="[^"]*(?:card__b|sowhat__v|point)[^"]*"[^>]*>([\s\S]*?)</',
+                b)
+            for raw in plain_parts:
+                plain = _plain(raw).strip()
+                if len(plain) < 8:
+                    continue
+                for pat in patterns:
+                    if plain.endswith(pat) or plain.endswith(pat + '。'):
+                        # 排除「等…」短收口
+                        if re.search(r'等[…\.]{1,3}$', plain) and len(plain) <= 16:
+                            continue
+                        trunc_hits.append(f"第{i}页「{plain[:18]}」")
+                        break
+    chk("LAYOUT_QA_TRUNCATION 禁静默截断（列表/卡/结论勿以省略号砍义）",
+        not trunc_hits, "; ".join(trunc_hits[:4]) if trunc_hits else "")
+
+    # ⑦ 溢出却无拆页策略：页高估算超预算，且无 band--flow / 续页标记 / 多部分 title
+    overflow_hits = []
+    if at.get('overflowNoSplitFail', True):
+        B = MODE_BUDGETS.get(mode, MODE_BUDGETS['presentation'])
+        fs_px = int(B.get('bodyPx') or 17)
+        wrap_px = int(B.get('wrap') or 1400) - WRAP_INSET_PX
+        cpl = max(10, int(wrap_px / fs_px))
+        for i, b, head in content:
+            if 'band--flow' in head:
+                continue
+            # 已有拆页/续页信号则豁免
+            if re.search(r'(续|续表|01[ab]|02[ab]|part\s*[12]|跟进|详见下页)', b, re.I):
+                continue
+            b_clean = re.sub(r'<script\b[\s\S]*?</script>', ' ', b)
+            b_clean = re.sub(r'<style\b[\s\S]*?</style>', ' ', b_clean)
+            plen = len(_plain(b_clean).strip())
+            u = _band_heavy_units(b)
+            est_h = plen / cpl * fs_px * LINE_FACTOR + u * UNIT_PX + SHEAD_PX
+            if est_h > SCREEN_BUDGET_PX * 1.08:
+                overflow_hits.append(f"第{i}页≈{est_h:.0f}px 无拆页/换形态信号")
+    chk("LAYOUT_QA_OVERFLOW_NO_SPLIT 溢出须拆页/换形态（禁硬塞）",
+        not overflow_hits, "; ".join(overflow_hits[:3]) if overflow_hits else "")
+
+    # ⑧ 半空卡 vs 塞爆：同页多卡时过半卡正文过短
+    he = (LC.get('qualityGates') or {}).get('halfEmpty') or {}
+    half_hits = []
+    min_cards = int(he.get('minCards') or 3)
+    short_n = int(he.get('shortPlainChars') or 12)
+    max_ratio = float(he.get('maxShortRatio') or 0.5)
+    for i, b, head in content:
+        cards = re.findall(r'class="[^"]*\bcard\b[^"]*"[^>]*>([\s\S]*?)(?=<div class="[^"]*\bcard\b|</section>|$)', b)
+        if len(cards) < min_cards:
+            # also count .card blocks via simpler split
+            cards = re.split(r'class="[^"]*\bcard\b', b)[1:]
+        if len(cards) < min_cards:
+            continue
+        shorts = 0
+        for c in cards:
+            plen = len(_plain(c[:800]).strip())
+            if plen < short_n:
+                shorts += 1
+        if shorts / max(len(cards), 1) > max_ratio and shorts >= 2:
+            half_hits.append(f"第{i}页半空卡 {shorts}/{len(cards)}")
+    chk("LAYOUT_QA_HALF_EMPTY 半空卡过多（与塞爆同样不合格）",
+        not half_hits, "; ".join(half_hits[:3]) if half_hits else "",
+        level="WARN")
+
+    # ⑨ 列对齐节奏：同页多个 grid 混排却完全无对齐 token（加强 LAYOUT_ALIGN）
+    align_cfg = (LC.get('layoutSystem') or {}).get('align') or {}
+    rhythm_hits = []
+    if align_cfg.get('requireColumnRhythm', True):
+        tokens = set(align_cfg.get('mixedGridClasses') or ['a-start', 'a-c', 'a-end'])
+        for i, b, head in content:
+            mixed = _mixed_grid_needs_align(b)
+            has_align = any(re.search(r'\b' + re.escape(tok) + r'\b', b) for tok in tokens)
+            if mixed and not has_align:
+                rhythm_hits.append(f"第{i}页混排缺对齐类")
+    chk("LAYOUT_QA_ALIGN_RHYTHM 混排列节奏/共享对齐类",
+        not rhythm_hits, "; ".join(rhythm_hits[:3]) if rhythm_hits else "")
+
+
 def main():
     if len(sys.argv) < 2:
         print(__doc__)
@@ -1245,6 +1542,9 @@ def main():
     path = Path(sys.argv[1])
     strict = '--strict' in sys.argv
     as_json = '--json' in sys.argv
+    layout_qa = '--layout-qa' in sys.argv
+    # Mode A / presentation：--strict 隐含 --layout-qa（正式演示交付不漏 V 契约）
+    # research/architecture 不自动开启，避免污染密排/架构路径
     if not path.exists():
         print(f"文件不存在: {path}")
         return 2
@@ -1278,6 +1578,8 @@ def main():
         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'
+    if strict and mode == 'presentation' and not layout_qa:
+        layout_qa = True  # presentation + --strict → 自动 layout-qa
 
     # ── 宽屏与页面高度模型 ──
     B0 = MODE_BUDGETS[mode]
@@ -1400,9 +1702,33 @@ def main():
         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")
 
+    if mode == 'presentation':
+        # Mode A：主张/行动句标题（action title）；纯话题标签 WARN（硬 FAIL 过脆）
+        STRUCT_A = ('报告大纲', '大纲', '议程', 'Agenda', '参考资料', '下一步', '结论',
+                    '封面', '目录', '附录', '谢谢', 'Thank', 'Q&A', '问答')
+        TOPIC_ONLY = (
+            '现状分析', '市场格局', '风险与挑战', '背景介绍', '项目概述', '总结',
+            '概览', '概述', '简介', '背景', '方案', '规划', '进展', '回顾',
+            '分析', '对比', '数据', '附录', '下一步计划', '内容', '主题',
+        )
+        h1s_a = re.findall(r'<h2 class="t-h1 shead__title"[^>]*>(.*?)</h2>', txt)
+        titles_a = [re.sub(r'<[^>]+>', '', h).strip() for h in h1s_a]
+        topic_hits = []
+        for h in titles_a:
+            if not h or any(h.startswith(s) or h == s for s in STRUCT_A):
+                continue
+            # 纯话题：命中话题词表，或极短且无判断/数字/动词痕迹
+            if h in TOPIC_ONLY or (len(h) <= 6 and not re.search(
+                    r'\d|是|应|须|将|已|要|可|能|达|超|降|升|破|卡|成|未|无|有', h)):
+                topic_hits.append(h)
+        chk("presentation 主张/行动标题（禁纯话题标签）", not topic_hits,
+            f"话题式: {topic_hits[:4]}" if topic_hits else "", level="WARN")
+
     _check_content_quality(txt, chk, mode, model)
     _check_v9_hard_gates(txt, chk, model)
     _check_layout_grammar(txt, chk, mode)
+    if layout_qa:
+        _check_layout_qa(txt, chk, mode, model)
 
     # ── 去AI味（词表来自单源） ──
     body_plain = _plain(body_txt)