@topmindspace/tms-skills 0.1.2 → 0.1.4

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

@@ -87,6 +87,7 @@ INCLUDE = [
     'references/*',
     'scripts/*',
     'evals/*',
+    'agents/*',
 ]
 
 # 出包规则（命中即跳过；与 .gitignore 口径一致）
@@ -120,6 +121,9 @@ REQUIRED = [
     'references/layout-grammar.md',
     'references/default-surface.md',
     'references/presentation-craft.md',
+    'references/illustration-layout.md',
+    'references/page-type-matrix.md',
+    'references/chart-decision-tree.md',
     'references/modes.md', 'references/outline-design.md', 'references/design-system.md',
     'references/styles.md', 'references/content-rules.md', 'references/components.md',
     'references/charts.md',
@@ -145,13 +149,15 @@ REQUIRED = [
     'references/infographics-stats.md', 'references/infographics-structure.md',
     'evals/prompts.csv', 'evals/run_evals.py', 'evals/rubric.schema.json',
     'evals/trace.example.json',
+    'evals/trigger-queries.json',
+    'scripts/check_triggers.py',
 ]
 
 # 必须达到最小数量的集合（名称含日期，故按数量校验）
 MIN_COUNTS = {
     'assets/examples/*.html': 3,  # Batch3 每模式 1 份黄金样张
     'assets/examples/*.model.json': 3,
-    'references/*.md': 22,  # craft 轮后含 default-surface + presentation-craft；≥22 防误删
+    'references/*.md': 25,  # + page-type-matrix + chart-decision-tree；≥25 防误删
 }
 
 
@@ -255,6 +261,16 @@ def main():
         return 1
     print('  校验通过：必需文件齐全 · frontmatter 合规 · 无临时/缓存路径')
 
+    # P0-1 trigger heuristic gate（廉价、无 LLM）
+    trig = ROOT / 'scripts' / 'check_triggers.py'
+    if trig.exists():
+        import subprocess
+        r = subprocess.run([sys.executable, str(trig)], cwd=str(ROOT))
+        if r.returncode != 0:
+            print('  ✗ trigger coverage 失败（evals/trigger-queries.json）')
+            return 1
+        print('  ✓ trigger coverage 通过')
+
     if check_only:
         print('\n(--check 模式：仅校验，不打包)')
         return 0

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

@@ -16,7 +16,7 @@
     ② 无 LLM 的 rubric 启发式（content/layout/chart/infographic/tone 五维）
     ③ --deliver：按 SKILL.md「交付说明」七要素（路径/字节数/模式/风格/篇幅/格式/校验+引用）
        生成结构化交付说明——七要素缺一即 gate FAIL，杜绝手拼遗漏
-    交付时一键跑完，避免「只过了结构、内容质量靠自觉」。
+    交付时一键跑完；互不依赖的子进程门禁（HTML / PPTX / evals）并行执行以缩短墙钟。
 """
 from __future__ import annotations
 
@@ -26,6 +26,7 @@ import re
 import subprocess
 import sys
 import time
+from concurrent.futures import ThreadPoolExecutor, as_completed
 from pathlib import Path
 
 try:
@@ -242,33 +243,43 @@ def main() -> int:
     print(f'交付质量门禁 · {html.name}')
     print('-' * 60)
 
-    # ① HTML strict（Mode A / presentation 自动附带 --layout-qa）
+    # ①–③ 互不依赖子进程并行（HTML strict / PPTX strict / evals）
     head_snip = html.read_text(encoding='utf-8')[:2500]
     mode_m = re.search(r'data-mode="([^"]+)"', head_snip)
     mode = mode_m.group(1) if mode_m else 'presentation'
+    jobs: list[tuple[str, list[str], str | None]] = []
+    # (name, cmd, fixed_note) — fixed_note 非空则覆盖尾部摘要
     vcmd = [sys.executable, str(ROOT / 'scripts' / 'validate_report.py'), str(html), '--strict']
     if mode == 'presentation':
         vcmd.append('--layout-qa')
-    rc, out, el = run(vcmd)
-    gate_name = 'validate_report --strict' + (' --layout-qa' if mode == 'presentation' else '')
-    tail = [ln for ln in out.splitlines() if ln.startswith('PASS ') or ln.startswith('结论')]
-    gate(gate_name, rc == 0, tail[-1] if tail else f'exit {rc}', el)
-
-    # ② PPTX strict（可选）
+    gate_html = 'validate_report --strict' + (' --layout-qa' if mode == 'presentation' else '')
+    jobs.append((gate_html, vcmd, None))
     if args.pptx:
-        cmd = [sys.executable, str(ROOT / 'scripts' / 'validate_pptx.py'), args.pptx, '--strict']
+        pcmd = [sys.executable, str(ROOT / 'scripts' / 'validate_pptx.py'), args.pptx, '--strict']
         if args.model:
-            cmd += ['--model=' + args.model]
-        rc, out, el = run(cmd)
-        tail = [ln for ln in out.splitlines() if ln.startswith('PASS ') or ln.startswith('结论')]
-        gate('validate_pptx --strict', rc == 0, tail[-1] if tail else f'exit {rc}', el)
-
-    # ③ evals 确定性
-    cmd = [sys.executable, str(ROOT / 'evals' / 'run_evals.py'), '--score', str(html)]
+            pcmd += ['--model=' + args.model]
+        jobs.append(('validate_pptx --strict', pcmd, None))
+    ecmd = [sys.executable, str(ROOT / 'evals' / 'run_evals.py'), '--score', str(html)]
     if args.trace:
-        cmd += ['--trace', args.trace]
-    rc, out, el = run(cmd)
-    gate('evals --score', rc == 0, '确定性 + 效率检查', el)
+        ecmd += ['--trace', args.trace]
+    jobs.append(('evals --score', ecmd, '确定性 + 效率检查'))
+
+    results: dict[str, tuple[int, str, float, str | None]] = {}
+    with ThreadPoolExecutor(max_workers=max(1, len(jobs))) as pool:
+        futs = {pool.submit(run, cmd): (name, note) for name, cmd, note in jobs}
+        for fut in as_completed(futs):
+            name, note = futs[fut]
+            rc, out, el = fut.result()
+            results[name] = (rc, out, el, note)
+
+    # 按 jobs 声明顺序输出（稳定可读；墙钟已并行）
+    for name, _cmd, _note in jobs:
+        rc, out, el, note = results[name]
+        if note is not None:
+            gate(name, rc == 0, note, el)
+        else:
+            tail = [ln for ln in out.splitlines() if ln.startswith('PASS ') or ln.startswith('结论')]
+            gate(name, rc == 0, tail[-1] if tail else f'exit {rc}', el)
 
     # ④ rubric 启发式
     rubric = heuristic_rubric(html)

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

@@ -0,0 +1,180 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""Regression gates for usage-feedback defects (validator + static engine checks)."""
+from __future__ import annotations
+
+import re
+import sys
+from pathlib import Path
+from xml.etree import ElementTree as ET
+
+ROOT = Path(__file__).resolve().parent
+sys.path.insert(0, str(ROOT))
+import validate_pptx as V  # noqa: E402
+
+EMU = 914400
+NS = V.NS
+fails: list[str] = []
+
+
+def ok(name: str, cond: bool, detail: str = "") -> None:
+    status = "PASS" if cond else "FAIL"
+    print(f"  [{status}] {name}" + (f" — {detail}" if detail and not cond else ""))
+    if not cond:
+        fails.append(name)
+
+
+def test_shape_bounds_reads_p_xfrm() -> None:
+    xml = (
+        '<?xml version="1.0" encoding="UTF-8" standalone="yes"?>'
+        f'<p:graphicFrame xmlns:a="{NS["a"]}" xmlns:p="{NS["p"]}">'
+        "<p:xfrm>"
+        f'<a:off x="{int(0.6 * EMU)}" y="{int(2.5 * EMU)}"/>'
+        f'<a:ext cx="{int(8.5 * EMU)}" cy="{int(3.5 * EMU)}"/>'
+        "</p:xfrm>"
+        "</p:graphicFrame>"
+    )
+    el = ET.fromstring(xml)
+    box = V.shape_bounds(el)
+    ok("1 shape_bounds reads p:xfrm", box is not None and box[2] > 0, f"got {box}")
+
+
+def test_shape_bounds_still_reads_a_xfrm() -> None:
+    xml = (
+        '<?xml version="1.0" encoding="UTF-8" standalone="yes"?>'
+        f'<p:sp xmlns:a="{NS["a"]}" xmlns:p="{NS["p"]}">'
+        "<p:spPr><a:xfrm>"
+        f'<a:off x="{EMU}" y="{EMU}"/>'
+        f'<a:ext cx="{2 * EMU}" cy="{EMU}"/>'
+        "</a:xfrm></p:spPr></p:sp>"
+    )
+    el = ET.fromstring(xml)
+    box = V.shape_bounds(el)
+    ok("1b shape_bounds reads a:xfrm", box is not None and box[0] == EMU)
+
+
+def _text_shape(paras, x=0.6, y=2.5, w=4.0, h=0.6):
+    parts = []
+    for text, pt in paras:
+        parts.append(
+            f'<a:p><a:r><a:rPr sz="{int(pt * 100)}"/><a:t>{text}</a:t></a:r></a:p>'
+        )
+    xml = (
+        '<?xml version="1.0" encoding="UTF-8" standalone="yes"?>'
+        f'<p:sp xmlns:a="{NS["a"]}" xmlns:p="{NS["p"]}">'
+        "<p:spPr><a:xfrm>"
+        f'<a:off x="{int(x * EMU)}" y="{int(y * EMU)}"/>'
+        f'<a:ext cx="{int(w * EMU)}" cy="{int(h * EMU)}"/>'
+        "</a:xfrm></p:spPr>"
+        f'<p:txBody>{"".join(parts)}</p:txBody></p:sp>'
+    )
+    return ET.fromstring(xml)
+
+
+def test_text_overflow_vertical() -> None:
+    paras = [(f"这是第{i}段足够长的中文内容用来累计行高", 12.0) for i in range(8)]
+    shape = _text_shape(paras, h=0.55)
+    issues = V.text_overflow_vertical_check(shape, 1)
+    codes = [i["code"] for i in issues]
+    ok(
+        "2 TEXT_OVERFLOW_VERTICAL fires on multi-segment cumulative",
+        "TEXT_OVERFLOW_VERTICAL" in codes,
+        f"codes={codes}",
+    )
+
+
+def test_annotation_band_overlap() -> None:
+    xml = (
+        '<?xml version="1.0" encoding="UTF-8" standalone="yes"?>'
+        f'<p:sp xmlns:a="{NS["a"]}" xmlns:p="{NS["p"]}">'
+        "<p:spPr><a:xfrm>"
+        f'<a:off x="{int(0.6 * EMU)}" y="{int(5.5 * EMU)}"/>'
+        f'<a:ext cx="{int(4 * EMU)}" cy="{int(1.4 * EMU)}"/>'
+        "</a:xfrm></p:spPr></p:sp>"
+    )
+    el = ET.fromstring(xml)
+    issues = V.annotation_band_overlap_check([el], 1, int(7.5 * EMU))
+    codes = [i["code"] for i in issues]
+    ok(
+        "3 ANNOTATION_BAND_OVERLAP fires when content crushes band",
+        "ANNOTATION_BAND_OVERLAP" in codes,
+        f"codes={codes}",
+    )
+
+
+def test_font_size_not_snapped() -> None:
+    xml = (
+        '<?xml version="1.0" encoding="UTF-8" standalone="yes"?>'
+        f'<p:sld xmlns:a="{NS["a"]}" xmlns:p="{NS["p"]}">'
+        "<p:cSld><p:spTree><p:sp><p:txBody>"
+        '<a:p><a:r><a:rPr sz="1130"/><a:t>OffLadderSampleTextHere</a:t></a:r></a:p>'
+        "</p:txBody></p:sp></p:spTree></p:cSld></p:sld>"
+    )
+    root = ET.fromstring(xml)
+    issues = V.font_size_snap_check(root, 1)
+    codes = [i["code"] for i in issues]
+    ok(
+        "4 FONT_SIZE_NOT_SNAPPED fires on off-ladder size",
+        "FONT_SIZE_NOT_SNAPPED" in codes,
+        f"codes={codes}",
+    )
+
+
+def test_engine_static() -> None:
+    export = (ROOT.parent / "assets" / "pptx-export.js").read_text(encoding="utf-8")
+    build = (ROOT.parent / "scripts" / "build_pptx.js").read_text(encoding="utf-8")
+
+    m = re.search(r"function vbarShapes\([\s\S]*?\n\}", export)
+    body = m.group(0) if m else ""
+    ok("B vbarShapes handles ec.series", "ec.series" in body and "series.forEach" in body)
+
+    ok(
+        "C A-channel streamgraph reserves legend inside plot",
+        "Reserve legend INSIDE" in export and "plotBottom" in export,
+    )
+    ok(
+        "C B-channel infoStreamgraph reserves legend inside plot",
+        "Reserve legend INSIDE" in build and "plotBottom" in build,
+    )
+
+    ok(
+        "D A-channel caption uses capYImg from imageLayoutShapes",
+        "capYImg" in export
+        and re.search(r"capYImg\s*=\s*imageLayoutShapes", export) is not None,
+    )
+    ok(
+        "D B-channel caption uses capYImg from imageLayoutShapes",
+        "capYImg" in build
+        and re.search(r"capYImg\s*=\s*imageLayoutShapes", build) is not None,
+    )
+
+    cards = re.search(
+        r"else if \(type === 'cards'\) \{[\s\S]*?else if \(type === 'split'\)",
+        export,
+    )
+    cards_body = cards.group(0) if cards else ""
+    ok(
+        "A A-channel cards uses sz: N object props (not bare sz())",
+        "sz: 12" in cards_body
+        and not re.search(r"(?<![\w.])sz\s*\(\s*12\s*\)", cards_body),
+    )
+    ok(
+        "A modeSize applied in txSp serialization",
+        "function modeSize" in export and "modeSize(" in export,
+    )
+
+
+def main() -> int:
+    print("feedback-gates regression")
+    test_shape_bounds_reads_p_xfrm()
+    test_shape_bounds_still_reads_a_xfrm()
+    test_text_overflow_vertical()
+    test_annotation_band_overlap()
+    test_font_size_not_snapped()
+    test_engine_static()
+    print(f"\n{len(fails)} failed" if fails else "\nAll feedback gates PASS")
+    return 1 if fails else 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

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

@@ -108,6 +108,9 @@ STRICT_FAILURE_CODES = {
     "UNBALANCED_EMPTY_SPACE",
     "UNJUSTIFIED_LARGE_IMAGE",
     "TEXT_OVERFLOW_ESTIMATE",
+    "TEXT_OVERFLOW_VERTICAL",
+    "ANNOTATION_BAND_OVERLAP",
+    "FONT_SIZE_NOT_SNAPPED",
     "TEXT_INCOMPLETE",
     "CHART_SKEW_INVALID",
     "CHART_OVERSIZE",
@@ -207,9 +210,17 @@ def slide_chart_text(archive: zipfile.ZipFile, slide_name: str) -> str:
 
 
 def shape_bounds(element: ET.Element) -> tuple[int, int, int, int] | None:
+    """Read shape geometry from DrawingML ``a:xfrm`` **or** PresentationML ``p:xfrm``.
+
+    Native charts live in ``p:graphicFrame``, which carries ``p:xfrm`` (not ``a:xfrm``).
+    Skipping ``p:xfrm`` made chart overflow / overlap gates blind.
+    """
     xfrm = element.find(".//a:xfrm", NS)
+    if xfrm is None:
+        xfrm = element.find(".//p:xfrm", NS)
     if xfrm is None:
         return None
+    # p:xfrm and a:xfrm both nest a:off / a:ext
     offset = xfrm.find("a:off", NS)
     extent = xfrm.find("a:ext", NS)
     if offset is None or extent is None:
@@ -298,6 +309,147 @@ def text_overflow_check(shape: ET.Element, slide_no: int) -> list[dict[str, Any]
     return out
 
 
+def text_overflow_vertical_check(shape: ET.Element, slide_no: int) -> list[dict[str, Any]]:
+    """Multi-segment cumulative vertical overflow (TEXT_OVERFLOW_VERTICAL).
+
+    Per-paragraph area checks miss the case where each line fits individually but
+    the *sum* of estimated line heights exceeds the text-frame height.
+    """
+    out: list[dict[str, Any]] = []
+    box = shape_bounds(shape)
+    if box is None:
+        return out
+    _, _, cx, cy = box
+    w_in, h_in = cx / 914400.0, cy / 914400.0
+    if w_in <= 0 or h_in <= 0:
+        return out
+    cfg = _containers_cfg()
+    lf = float(cfg.get("lineFactor") or 1.35)
+    total_h = 0.0
+    segments = 0
+    for para in shape.findall(".//a:p", NS):
+        line = "".join((t.text or "") for t in para.findall(".//a:t", NS))
+        if not line.strip():
+            continue
+        sizes = font_sizes_pt(para)
+        fz = min(sizes) if sizes else 12.0
+        est_w = est_text_width_in(line, fz)
+        lines_needed = max(1, int((est_w / max(w_in, 0.01)) + 0.999))
+        total_h += lines_needed * (fz / 72.0) * lf
+        segments += 1
+    # Single-segment cases already covered by TEXT_OVERFLOW_ESTIMATE; this gate
+    # targets multi-para cumulative overflow (tolerance mirrors horizontal gate).
+    if segments >= 2 and total_h > h_in * TEXT_OVERFLOW_TOLERANCE:
+        out.append(issue(
+            "TEXT_OVERFLOW_VERTICAL",
+            f"多段文本累计高度估算 {total_h:.2f}in > 文本框 {h_in:.2f}in"
+            f"（{segments} 段，容差 {TEXT_OVERFLOW_TOLERANCE:.0%}）——"
+            "请拆段/缩字号阶梯/换页，禁止静默截断。",
+            slide=slide_no,
+        ))
+    return out
+
+
+def annotation_band_overlap_check(
+    elements: list[ET.Element],
+    slide_no: int,
+    slide_height_emu: int,
+) -> list[dict[str, Any]]:
+    """Detect primary content crushing into the annotation band.
+
+    Annotation band ≈ [contentBottomWithNote, contentBottom] (default 6.4–6.9in).
+    Shapes that start in the body and extend into the band (e.g. streamgraph
+    legends stacked below the plot) fire ANNOTATION_BAND_OVERLAP.
+    """
+    out: list[dict[str, Any]] = []
+    try:
+        lc_path = Path(__file__).resolve().parent / "layout-constants.json"
+        lc = json.loads(lc_path.read_text(encoding="utf-8"))
+        lay = ((lc.get("pageTypes") or {}).get("layout") or {})
+        band_top_in = float(lay.get("contentBottomWithNote") or 6.4)
+        band_bot_in = float(lay.get("contentBottom") or 6.9)
+    except (OSError, json.JSONDecodeError, TypeError, ValueError):
+        band_top_in, band_bot_in = 6.4, 6.9
+    band_top = int(band_top_in * 914400)
+    band_bot = int(band_bot_in * 914400)
+    # Ignore footer/page-number chrome near the very bottom
+    pager_floor = int(min(band_bot_in + 0.05, slide_height_emu / 914400.0 - 0.05) * 914400)
+    min_overlap = int(0.08 * 914400)  # 0.08in
+    for el in elements:
+        box = shape_bounds(el)
+        if box is None:
+            continue
+        x, y, cx, cy = box
+        if cx <= 0 or cy <= 0:
+            continue
+        bottom = y + cy
+        # Legitimate annotation/note content sits entirely inside the band
+        if y >= band_top - int(0.02 * 914400):
+            continue
+        # Page chrome (pager) — tiny shapes near bottom edge
+        if y >= pager_floor:
+            continue
+        overlap = min(bottom, band_bot) - max(y, band_top)
+        if overlap >= min_overlap and bottom > band_top:
+            out.append(issue(
+                "ANNOTATION_BAND_OVERLAP",
+                f"主内容侵入注释带（元素底边 {bottom/914400:.2f}in 越过注释带顶 "
+                f"{band_top_in:.2f}in，重叠 {overlap/914400:.2f}in）——"
+                "图例/系列请收入主图区或压缩系列数，禁止压进 so-what/来源行。",
+                slide=slide_no,
+            ))
+            # One finding per slide is enough to gate
+            break
+    return out
+
+
+def font_size_snap_check(root: ET.Element, slide_no: int) -> list[dict[str, Any]]:
+    """Font sizes must sit on the fontShrink ladder (0.5pt snap grid).
+
+    fitFont already selects from the ladder; this gate catches callers that
+    arithmetic-shift sizes (e.g. fz-1) or hardcode off-ladder values without
+    re-snapping through the ladder / modeSize path.
+    """
+    out: list[dict[str, Any]] = []
+    try:
+        lc_path = Path(__file__).resolve().parent / "layout-constants.json"
+        lc = json.loads(lc_path.read_text(encoding="utf-8"))
+        ladder = ((lc.get("containers") or {}).get("fontShrink") or {}).get("ladder") or []
+        allowed = {round(float(v), 2) for v in ladder}
+    except (OSError, json.JSONDecodeError, TypeError, ValueError):
+        allowed = {15, 14, 13.5, 13, 12.5, 12, 11.5, 11, 10.5, 10, 9.5, 9, 8.5}
+    # Also allow common display sizes used by cover/hero (not on shrink ladder)
+    allowed |= {44, 36, 30, 28, 24, 22, 20, 18, 16, 8.0, 7.5}
+    alien: list[float] = []
+    for run in root.findall(".//a:r", NS):
+        text = "".join(t.text or "" for t in run.findall("a:t", NS))
+        if not text.strip():
+            continue
+        pr = run.find("a:rPr", NS)
+        raw = pr.get("sz") if pr is not None else None
+        if raw is None:
+            continue
+        try:
+            pt = int(raw) / 100.0
+        except (TypeError, ValueError):
+            continue
+        if round(pt, 2) not in allowed and abs(pt * 2 - round(pt * 2)) > 0.01:
+            alien.append(pt)
+        elif round(pt, 2) not in allowed:
+            # On 0.5 grid but not on declared ladder / display set — still flag
+            # when far from any allowed value (>0.26pt)
+            if min(abs(pt - a) for a in allowed) > 0.26:
+                alien.append(pt)
+    if alien:
+        out.append(issue(
+            "FONT_SIZE_NOT_SNAPPED",
+            f"页内出现未对齐字号阶梯的字号 {sorted(set(round(v,2) for v in alien))[:8]}——"
+            "fitFont/modeSize 须回落到 containers.fontShrink.ladder（或封面展示档）。",
+            slide=slide_no,
+        ))
+    return out
+
+
 @lru_cache(maxsize=1)
 def _containers_cfg() -> dict[str, Any]:
     """容器内边距与锚点容差（单源 scripts/layout-constants.json 的 containers / anchorTolerance）。"""
@@ -628,8 +780,13 @@ def inspect_slide(
 
     for shape in shapes:
         warnings.extend(text_overflow_check(shape, slide_number))
+        warnings.extend(text_overflow_vertical_check(shape, slide_number))
         warnings.extend(container_overflow_check(shape, slide_number))
 
+    warnings.extend(annotation_band_overlap_check(
+        [*shapes, *pictures, *graphic_frames], slide_number, height))
+    warnings.extend(font_size_snap_check(root, slide_number))
+
     for table in tables:
         warnings.extend(table_checks(table, slide_number))
 

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

@@ -1215,6 +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」，