@lark-apaas/coding-steering 0.1.32-beta.0 → 0.1.32-dev.5abff3b

package/steering/design-html/skills/pptx-style-extract/scripts/draft.py

@@ -5,26 +5,69 @@
 
 产出 <stage1-outdir>/l-out/：
     BRIEF.md          唯一必读简报：事实 + 草案依据 + 待判断清单
-    contact-sheet.png 候选图拼版（带编号，一次看完所有图）
-    manifest.yaml / frontmatter.yaml / layouts.yaml / body.md   四件草案，可直接进 package.py
+    vision-group-*.jpg 候选图与所在页语境拼版
+    asset-vision-groups.json 候选来源、位置与预算索引
+    manifest.yaml / frontmatter.yaml / layout-controls.yaml / layouts.yaml / body.md
+                                                        判断单与坐标事实，可直接进 package.py
 
 草案里所有数值都来自 extract.json；凡是需要「像人一样看」才能定的，写成 `TODO:` 行
 （package.py 见 TODO 即 FAIL），由 L 层改掉。
 """
 import argparse
+import copy
+import glob
 import json
 import os
 import re
 import shutil
+import subprocess
 import sys
 from collections import Counter, defaultdict
 
+sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+from ooxml import OFFICE_DEFAULT_FONTS              # noqa: E402
+from census import (ASSET_WARN_SINGLE, FULLSCREEN_COVERAGE, LUM_MID,   # noqa: E402
+                    REPEAT_MIN, SMALL_IMG_W_PCT, canvas_coverage)
+
+OPAQUE_ENOUGH = 128    # 能当背景的最低不透明度：低于半透明就遮不住底下的东西，
+                       # 那是叠加装饰不是背景
+FILL_MANY = 5          # 「被大量当填充铺开」的次数下限，用于区分卡片底与偶发用色
+BG_CONTENT_CAP = 5     # 内容页背景收几张：再多消费端也挑不过来，超出的写进 TODO 交人取舍
+SHEET_BATCH = 12       # 每张联系表最多 12 个候选；候选不截断，超出就继续生成下一张
+CONTEXT_BATCH = 8      # 每张整页语境表最多 8 页；同页只渲染一次
+# 与 studio_server_faas 的批量 vision 图数预算保持一致。单页大组允许 1 张整页图 +
+# 9 张候选；多个小页合组时，所有整页图和候选图合计最多 5 张。
+SINGLE_PAGE_IMAGE_BUDGET = 10
+MULTI_PAGE_IMAGE_BUDGET = 5
+# Skill 侧的总输入上限：限制模型需要读取的拼版数量和总视觉图数，而不是偷偷截断
+# 中间产物。首页、尾页优先；其余超限候选在 gaps 中显式说明。
+VISUAL_PACK_CAP = 5
+VISUAL_INPUT_CAP = 30
+VISUAL_PREVIEW_MAX_EDGE = 1200
+VISUAL_JPEG_QUALITY = 82
+
 HERE = os.path.dirname(os.path.abspath(__file__))
 SKILL_ROOT = os.path.dirname(HERE)
 SYS_FALLBACK = '"PingFang SC", "Microsoft YaHei", sans-serif'
 
 
 # ---------------------------------------------------------------- 小工具
+# 被名额截掉的东西统一记在这里，最后并进 gaps。截断本身是必要的（色板 40 个 token
+# 消费端挑不过来），但**不说**就成了「悄悄少了东西而产物看起来正常」——消费端会以为
+# 它拿到的就是全部。
+_TRUNCATED = []
+
+
+def note_truncation(kind, kept, total, advice='', where=''):
+    """记一条「这里按名额截断了」。kept >= total 时什么都不记。
+
+    按 kind 归并成一条 gap：同一类截断逐处各写一行会淹掉别的 gaps。
+    """
+    if total > kept:
+        _TRUNCATED.append((kind, kept, total, advice, where))
+    return kept
+
+
 def hex2rgb(h):
     h = h.lstrip('#')
     return tuple(int(h[i:i + 2], 16) for i in (0, 2, 4))
@@ -82,7 +125,64 @@ def q(v):
 
 
 # ---------------------------------------------------------------- 颜色
-def draft_colors(d):
+def bg_colors(d):
+    """页面/版式/母版的 `background` 声明里出现的底色，按声明次数排序。
+
+    「哪个色是底色」是直读事实（bgPr / bgRef），不用靠亮度猜：渐变里出现的浅色，
+    亮度可能比真底色更像底色。
+    """
+    cnt = Counter()
+    rows = (d.get('slides') or []) + (d.get('layouts') or []) \
+        + ((d.get('masters') or {}).get('masters') or [])
+    for row in rows:
+        bg = row.get('background')
+        if not isinstance(bg, dict):
+            continue
+        cols = []
+        if isinstance(bg.get('color'), dict):
+            cols.append(bg['color'])
+        for st in (bg.get('stops') or []):
+            if isinstance(st.get('color'), dict):
+                cols.append(st['color'])
+        for c in cols:
+            h = (c.get('hex') or '').upper()
+            if h:
+                cnt[h] += 1
+    return [h for h, _ in cnt.most_common()]
+
+
+def _gap_cut(vals, lo, hi):
+    """在排序后的值里找最大间隙，切点取间隙中点。
+
+    不用中位数：中位数会正好落在某个样本自己身上，它归哪边就只取决于写 >= 还是 >，
+    纯属任意。真正的分界在两族之间的空档里。夹在 [lo, hi] 内，避免整套同色的模板
+    把界推到极端。
+    """
+    v = sorted(vals)
+    if len(v) < 2:
+        return (lo + hi) / 2.0
+    _, mid = max((v[i + 1] - v[i], (v[i + 1] + v[i]) / 2.0) for i in range(len(v) - 1))
+    return min(max(mid, lo), hi)
+
+
+def palette_cuts(rows):
+    """「有彩 vs 中性」「深 vs 浅」的分界，按本模板自己的色分布切。
+
+    固定分界必然错一边：低饱和的莫兰迪配色整套都在低位，高饱和的品牌配色整套都在高位。
+    """
+    sat_cut = _gap_cut([r['sat'] for r in rows], 0.12, 0.45)
+    lums = sorted(r['lum'] for r in rows) or [0.0]
+    return sat_cut, lums[len(lums) // 2]
+
+
+def draft_colors(d, cusage=None):
+    """色板 token：名字按**实际用法**定，不只看亮度饱和度。
+
+    只看 lum/sat 会把「主要用来填色的纯黑」命名成 ink（文字色）、把「只出现在渐变里的
+    浅蓝」命名成 surface-alt。这里先看它在形状上主要干什么，再结合
+    亮度定名；用量太少的直接不进色板。
+    """
+    cusage = cusage or {}
     pool = [c for c in d['color_freq']
             if c.get('class') == 'design' and abs((c.get('alpha') or 100) - 100) < 0.1]
     seen, rows = set(), []
@@ -92,29 +192,63 @@ def draft_colors(d):
             continue
         seen.add(h)
         rgb = hex2rgb(h)
-        rows.append({'hex': h, 'n': c['n'], 'lum': lum(rgb), 'sat': satu(rgb)})
-    rows.sort(key=lambda r: -r['n'])
+        u = cusage.get(h) or Counter()
+        tot = sum(u.values())
+        main = u.most_common(1)[0][0] if tot else None
+        rows.append({'hex': h, 'n': c['n'], 'lum': lum(rgb), 'sat': satu(rgb),
+                     'use': u, 'use_n': tot, 'main': main})
+    # 用量只用来**命名**，不作准入门槛——color_usage 只数形状级的填充/描边/文字，
+    # 背景 p:bg 与主题色不在其中，拿它筛会把色板砍到只剩极少数几个。
+    strong = sorted(rows, key=lambda r: -r['n'])
+
+    SAT_CUT, LUM_CUT = palette_cuts(rows)
 
     tokens, used = [], set()
 
     def take(pred, names):
         for name in names:
-            for r in rows:
+            for r in strong:
                 if r['hex'] in used or not pred(r):
                     continue
                 used.add(r['hex'])
                 tokens.append((name, r))
                 break
 
-    take(lambda r: r['lum'] > 0.85 and r['sat'] < 0.15, ['surface', 'surface-alt'])
-    take(lambda r: r['lum'] < 0.32 and r['sat'] < 0.25, ['ink', 'ink-muted'])
-    take(lambda r: r['sat'] >= 0.35, ['primary', 'accent', 'accent-2', 'accent-3'])
-    take(lambda r: r['sat'] < 0.35, ['neutral'])
-    # 未取用但高频的留给 BRIEF 展示
-    rest = [r for r in rows if r['hex'] not in used][:6]
+    def kind(r):
+        if r['main'] == '文字':
+            return 'text'
+        if r['main'] in ('填充', '渐变', '描边'):
+            return 'paint'
+        return 'unknown'        # 形状层看不到用法，退回亮度/饱和度判断
+
+    # 墨色：主要用来写字（或看不出用法但本身是深中性色），且不是彩色
+    take(lambda r: r['sat'] < SAT_CUT and r['lum'] < min(LUM_CUT, LUM_MID)
+         and (kind(r) == 'text' or kind(r) == 'unknown'), ['ink', 'ink-muted'])
+    # 底色：直接取页面 background 声明里的色，按声明次数排
+    grounds = bg_colors(d)
+    for name in ('surface', 'surface-alt'):
+        for h in grounds:
+            r = next((x for x in strong if x['hex'] == h and x['hex'] not in used), None)
+            if r:
+                used.add(r['hex'])
+                tokens.append((name, r))
+                break
+
+    # 卡片/面板底：页面底色之外，真被大量当填充铺开的浅色（≥5 处才算）
+    take(lambda r: r['lum'] > max(LUM_CUT, 0.85) and r['sat'] < SAT_CUT
+         and (r['use'].get('填充') or 0) >= FILL_MANY, ['surface-raised'])
+    # 表达色：有彩度的按频次排
+    take(lambda r: r['sat'] >= SAT_CUT, ['primary', 'accent', 'accent-2', 'accent-3'])
+    # 其余低饱和色一律 neutral-N——它到底是卡片底、分隔线还是描边，数据分不出来，
+    # 就不要用名字去替消费方下结论；真实用法写在 Colors 表的用途列里。
+    take(lambda r: r['sat'] < SAT_CUT, ['neutral', 'neutral-2', 'neutral-3'])
+    spare = [r for r in rows if r['hex'] not in used]
+    note_truncation('设计色', 6, len(spare), '色板只收主要色，其余在联系表里看')
+    rest = spare[:6]
     return tokens, rest, rows
 
 
+
 # ---------------------------------------------------------------- 字体
 def parse_fallback_table():
     path = os.path.join(SKILL_ROOT, 'font-fallback.yaml')
@@ -142,6 +276,48 @@ def norm(s):
     return re.sub(r'[\s\-_]', '', s or '').lower()
 
 
+OFFICE_DEFAULT_FONTS_NORM = {norm(x) for x in OFFICE_DEFAULT_FONTS}
+
+
+def cover_slot_colors(tokens, archetypes, rows, cusage):
+    """slot CSS 里出现的每个色值都必须在色板里有名字。
+
+    Hard Rules 写「颜色只用 colors 里的 token」，而 slot CSS 的 color 是从模板直读的，
+    两者不对齐就等于产物自己违反自己的规则——slot 的色值直读自模板，未必都已进
+    色板。这里把缺的补进色板，按用法归族命名。
+    """
+    have = {r['hex'].upper() for _, r in tokens}
+    by_hex = {r['hex'].upper(): r for r in rows}
+    sat_cut, lum_cut = palette_cuts(rows)      # 与 draft_colors 同一套切点，别各切各的
+    used = [n for n, _ in tokens]
+
+    def nxt(fam):
+        if fam not in used:
+            return fam
+        i = 2
+        while '%s-%d' % (fam, i) in used:
+            i += 1
+        return '%s-%d' % (fam, i)
+
+    added = []
+    for a in archetypes:
+        for s in a['slots']:
+            h = (s.get('_color') or '').upper()
+            if not h.startswith('#') or h in have:
+                continue
+            have.add(h)
+            r = by_hex.get(h)
+            if r is None:            # 普查里没有这个色（理论上不该发生），跳过不编造
+                continue
+            fam = ('ink' if r['sat'] < sat_cut and r['lum'] < min(lum_cut, LUM_MID)
+                   else 'accent' if r['sat'] >= sat_cut else 'neutral')
+            name = nxt(fam)
+            used.append(name)
+            tokens.append((name, r))
+            added.append((name, h))
+    return added
+
+
 def draft_fonts(d):
     table = parse_fallback_table()
     groups = defaultdict(lambda: {'rendered': 0, 'weights': set(), 'names': set(), 'bold': 0})
@@ -167,6 +343,7 @@ def draft_fonts(d):
         return None
 
     out = []
+    note_truncation('字族', 4, len(ranked), '只报渲染量最大的几族')
     for key, g in ranked[:4]:
         fam = resolve(sorted(g['names'], key=len))
         stack = [sorted(g['names'], key=len)[0]]
@@ -201,35 +378,193 @@ def import_line(fonts):
     return "@import url('%s?%s&display=swap');" % (MIRROR, fam), webs
 
 
+def quant(hit, total):
+    """覆盖率决定量词——不到一半就不许说「一律/每页」。"""
+    if not total:
+        return None
+    r = hit / float(total)
+    if r >= 0.9:
+        return '一律'
+    if r >= 0.5:
+        return '多数'
+    return None
+
+
+def color_usage(shapes, d=None):
+    """每个色值在形状上的真实用法计数：填充 / 渐变 / 描边 / 文字。
+
+    用途列不能靠预设字典猜——同一个色在不同模板里的主用途完全不同。这里从 shapes
+    直接数，数不到就如实说数不到。
+    """
+    def hx(c):
+        return (c.get('hex') or '').upper() if isinstance(c, dict) else ''
+
+    def walk_text_colors(node, out):
+        """文本样式可能嵌在 lstStyle.lvlNpPr / defRPr / rPr 任一层——通用遍历，
+        别逐层枚举（枚举漏过 lvl2pPr，导致主色被写成「用途待确认」）。"""
+        if isinstance(node, dict):
+            if isinstance(node.get('color'), dict) and hx(node['color']):
+                out.append(hx(node['color']))
+            for v in node.values():
+                walk_text_colors(v, out)
+        elif isinstance(node, list):
+            for v in node:
+                walk_text_colors(v, out)
+
+    use = defaultdict(Counter)
+    for s in shapes:
+        f = s.get('fill') or {}
+        if f.get('type') == 'solid' and hx(f.get('color')):
+            use[hx(f['color'])]['填充'] += 1
+        for st in (f.get('stops') or []):
+            if hx(st.get('color')):
+                use[hx(st['color'])]['渐变'] += 1
+        ln = s.get('line') or {}
+        if hx(ln.get('color')):
+            use[hx(ln['color'])]['描边'] += 1
+        cols = []
+        walk_text_colors(s.get('text') or {}, cols)
+        for h in cols:
+            use[h]['文字'] += 1
+    for h in bg_colors(d or {}):
+        use[h]['页面背景'] += 1
+    # 主题 clrScheme：这类色常常只在主题里声明、页面上由 schemeClr 间接引用，
+    # 不记上就会在用途列写「未落在形状上」，看着像没人用。
+    for th in ((d or {}).get('themes') or []):
+        if not th.get('picked'):
+            continue
+        for slot, hexv in (th.get('clrScheme') or {}).items():
+            if isinstance(hexv, str) and hexv.startswith('#'):
+                use[hexv.upper()]['主题 ' + slot] += 1
+    return use
+
+
+def usage_phrase(counter):
+    """把用法计数写成一句话；主用法占六成以上就直接点名，否则并列前三。"""
+    if not counter:
+        return '普查里有声明，但未落在形状/背景/主题色上——用途待确认'
+    items = counter.most_common()
+    tot = sum(counter.values())
+    if items[0][1] >= tot * 0.6:
+        return '主要作%s（%d/%d 处）' % (items[0][0], items[0][1], tot)
+    return '、'.join('%s %d 处' % (k, v) for k, v in items[:3])
+
+
 def draft_anchors(d, tokens, fonts, roles, assets, archetypes):
-    names = [t[0] for t in tokens]
+    """anchors 只报测到的数，不下「这套风格是什么」的结论。
+
+    这一段在 design.md 里读起来像「设计总纲」，消费端会照它建全局样式。脚本写进去的
+    每一句解读都会被当成规则执行——实测把 1/8 覆盖率的 logo 描述成「跨页不动」，
+    消费端就建了全局 CSS 类，12 页全铺了 logo。所以这里只给覆盖率和计数，
+    「这是不是这套风格的特征」由看得到图的人判断。
+    """
+    A = []
+    n_arch = len(archetypes) or 1
+
+    # 1. 表达色：未取用的高频彩色要如实带上，不能说「其余全是中性」
+    names = [x[0] for x in tokens]
     chroma = [n for n in names if n.startswith(('primary', 'accent'))]
+    if chroma:
+        A.append((chroma[0] + '-led-palette', 'token',
+                  '有彩色 token 共 %d 个，用量最大的是 %s'
+                  % (len(chroma), '、'.join(chroma[:3]))))
+
+    # 2. 圆角：按普查占比
     radii = d.get('radii_census') or []
     zero = next((r for r in radii if r['px'] == 0), None)
-    total_r = sum(r['n'] for r in radii) or 1
-    geom = d.get('geom_census') or {}
+    tot_r = sum(r['n'] for r in radii) or 1
+    if zero:
+        q0 = quant(zero['n'], tot_r)
+        if q0:
+            A.append(('zero-radius', 'token',
+                      '圆角量为零的形状占 %d%%（普查 %d 个带圆角声明的形状）'
+                      % (round(100.0 * zero['n'] / tot_r), tot_r)))
+
+    # 3. 满屏底图：按有背景的页型占比
+    with_bg = sum(1 for a in archetypes if a.get('bg'))
+    qb = quant(with_bg, n_arch)
+    if qb:
+        A.append(('full-bleed-ground', 'pattern',
+                  '%d/%d 个页型声明了整幅铺满的底图' % (with_bg, n_arch)))
+
+    # 4. 标识：位置是不是真的固定，看有几个不同的 box
+    logo_slots = [s for a in archetypes for s in a['slots']
+                  if str(s.get('asset') or '').startswith(('logo', 'slogan'))]
+    logo_arch = sum(1 for a in archetypes
+                    if any(str(s.get('asset') or '').startswith(('logo', 'slogan'))
+                           for s in a['slots']))
+    boxes = {tuple(s['box']) for s in logo_slots}
+    # anchors 是「这套风格的定义性特征」，消费端读它来建全局样式。只在少数页型出现的
+    # 东西写进来，等于宣布它是全局元素——实测某模板 logo 只在 1/8 个页型上，anchor 仍
+    # 写成「跨页不动」，消费端据此建了个全局 CSS 类，12 页全铺了 logo。
+    # 所以这里和其他 anchor 用同一把尺：覆盖率不过半就不进 anchors。
+    ql = quant(logo_arch, n_arch)
+    if logo_arch and len(boxes) == 1 and ql:
+        A.append(('corner-locked-logo', 'component',
+                  '品牌标识出现在 %d/%d 个页型上，这些页型里它的 box 完全一致'
+                  % (logo_arch, n_arch)))
+    elif len(boxes) > 1:
+        A.append(('logo-moves-by-archetype', 'component',
+                  '品牌标识按页型换位换尺寸（共 %d 种摆法），必须按 layouts 里该页型的 box 放，'
+                  '不能沿用上一页' % len(boxes)))
+
+    # 5. 渐变：按普查计数
+    if (d.get('geom_census') or {}).get('gradient_fills'):
+        A.append(('gradient-accent', 'pattern',
+                  '全档共 %d 处渐变填充' % (d['geom_census']['gradient_fills'])))
+
+    # 6. 层级：字号跨度 + 字重是否单一（字重真单一才敢说「不靠字重」）
+    disp, body = roles.get('display'), roles.get('body')
+    if disp and body and disp['sz_px'] > body['sz_px']:
+        ws = {s.get('_font_weight') for a in archetypes for s in a['slots']
+              if s.get('_font_weight')}
+        tail = ('，字重只用 %s 一档' % list(ws)[0]) if len(ws) == 1 else ''
+        A.append(('size-driven-hierarchy', 'pattern',
+                  '最大字号档与正文档相差 %.1f 倍（见 typography）%s'
+                  % (disp['sz_px'] / body['sz_px'], tail)))
+
+    # 7. 阴影：只在描边极少时才敢说「不用描边分隔」
     eff = d.get('effects_census') or {}
-    A = []
-    if chroma:
-        A.append((chroma[0] + '-led-palette', 'token',
-                  '表达色只有 %s 这一组，其余全是中性底与墨色，见 colors' % '、'.join(chroma[:3])))
-    if zero and zero['n'] >= total_r * 0.7:
-        A.append(('zero-radius', 'token', '卡片、按钮、面板一律直角，圆角量在全 deck 压倒性为零'))
-    if any(a['kind'] == 'background' for a in assets):
-        A.append(('full-bleed-ground', 'pattern', '每页由整幅铺满的底图打底，元素浮在图上而不是浮在纯色块上'))
-    if any(a['kind'] == 'logo' for a in assets):
-        A.append(('corner-locked-logo', 'component', '品牌标识固定在同一角位，跨页不移动、不缩放'))
-    if geom.get('gradient_fills'):
-        A.append(('gradient-accent', 'pattern', '强调元素靠线性渐变承载，而不是纯色块'))
-    if len(fonts) >= 2:
-        A.append(('dual-family-typesetting', 'token', '中文与拉丁数字分属两套字族，同一行里混排'))
     if eff.get('outerShdw'):
-        A.append(('soft-shadow-card', 'component', '卡片靠极浅外阴影托起，不用描边分隔'))
-    disp, body = roles.get('display'), roles.get('body')
-    if disp and body and disp['sz_px'] >= body['sz_px'] * 3:
-        A.append(('size-driven-hierarchy', 'pattern', '层级靠字号跨度拉开而不是字重，展示档与正文档差出数倍'))
-    A.append(('archetype-reuse', 'pattern', '全 deck 只用 %d 种页型反复排列，版面骨架高度复用' % len(archetypes)))
-    A.append(('safe-area-discipline', 'token', '正文一律落在统一安全区内，不贴画布边'))
+        A.append(('soft-shadow-card', 'component',
+                  '全档 %d 处 outerShdw 外阴影' % eff['outerShdw']))
+
+    # 8. 双字族：只陈述分工存在，不断言「同一行混排」（普查没采集混排）
+    tot_r_font = sum(f['rendered'] for f in fonts) or 1
+    if len(fonts) >= 2 and fonts[1]['rendered']:
+        A.append(('dual-family-typesetting', 'token',
+                  '用了两套字族：%s 渲染 %d 处、%s 渲染 %d 处'
+                  % (fonts[0]['names'][0], fonts[0]['rendered'],
+                     fonts[1]['names'][0], fonts[1]['rendered'])))
+
+    # 9. 安全区：只在各页型正文左边界真的收敛时才写
+    # 「多宽算正文槽」按本包自己的槽宽分布定：固定 px 门槛在窄版心模板上会一个都不剩
+    widths = sorted(s['box'][2] for a in archetypes for s in a['slots'] if not s.get('asset'))
+    w_cut = widths[len(widths) // 2] if widths else 0
+    lefts = [s['box'][0] for a in archetypes for s in a['slots']
+             if not s.get('asset') and s['box'][2] >= w_cut]
+    if len(lefts) >= 4:
+        common = Counter(lefts).most_common(1)[0]
+        qs = quant(common[1], len(lefts))
+        if qs:
+            A.append(('shared-left-margin', 'token',
+                      '%d/%d 个正文槽的左边界落在同一个 x 上（坐标见 layouts）'
+                      % (common[1], len(lefts))))
+
+    # 10. 双主题：直读事实
+    themes = (d.get('theme_topology') or {}).get('themes') or []
+    if len(themes) > 1:
+        A.append(('dual-theme-masters', 'token',
+                  '模板声明了 %s 两套主题母版' % ' / '.join(themes)))
+
+    # 11. 画布：直读事实（兜底凑数也只用真事实）
+    cv = d['canvas']['px']
+    A.append(('fixed-canvas', 'token',
+              '画布 %d×%d，layouts 里的坐标都是这张画布上的绝对像素' % (cv[0], cv[1])))
+    if len(archetypes) >= 3:
+        A.append(('archetype-catalog', 'pattern',
+                  '归纳出 %d 种页型' % len(archetypes)))
+
     seen, out = set(), []
     for a in A:
         if a[0] in seen:
@@ -238,8 +573,6 @@ def draft_anchors(d, tokens, fonts, roles, assets, archetypes):
         out.append(a)
     return out[:8]
 
-
-# ---------------------------------------------------------------- 字号轴
 def draft_scale(d, archetypes=()):
     ts = [t for t in d['text_scale'] if t['sz_px'] >= 10]
     ts.sort(key=lambda t: -t['sz_px'])
@@ -251,8 +584,9 @@ def draft_scale(d, archetypes=()):
     display = by_px.get(max(title_sz)) if title_sz else None
     big = [t for t in ts if t['n'] >= 2] or ts
     display = display or big[0]
-    body_pool = [t for t in ts if t['sz_px'] <= 48]
-    body = max(body_pool, key=lambda t: t['n']) if body_pool else ts[-1]
+    # 正文档 = 渲染次数最多的那一档。不设「多大算正文」的上限：大字号排版的模板
+    # 正文本来就可能比别的模板的标题还大，预设上限会把它整档判错。
+    body = max([t for t in ts if t is not display] or ts, key=lambda t: t['n'])
     heading_pool = [t for t in ts if body['sz_px'] * 1.3 <= t['sz_px'] < display['sz_px']]
     heading = max(heading_pool, key=lambda t: t['n']) if heading_pool else None
     small_pool = [t for t in ts if t['sz_px'] < body['sz_px']]
@@ -279,7 +613,13 @@ def lh_of(t):
 
 # ---------------------------------------------------------------- 资产
 def probe_image(path):
-    info = {'w': None, 'h': None, 'alpha_mean': None, 'near_blank': False}
+    info = {
+        'w': None,
+        'h': None,
+        'alpha_mean': None,
+        'near_blank': False,
+        'near_white_ratio': None,
+    }
     try:
         from PIL import Image
     except Exception:
@@ -287,15 +627,119 @@ def probe_image(path):
     try:
         im = Image.open(path)
         info['w'], info['h'] = im.size
+        preview = im.convert('RGBA').resize((64, 64))
+        pixels = (preview.get_flattened_data()
+                  if hasattr(preview, 'get_flattened_data') else preview.getdata())
+        px = list(pixels)
+        alpha = [item[3] for item in px]
         if im.mode in ('RGBA', 'LA') or 'transparency' in im.info:
-            px = im.convert('RGBA').getchannel('A').resize((64, 64)).tobytes()
-            info['alpha_mean'] = sum(px) / len(px)
+            info['alpha_mean'] = sum(alpha) / len(alpha)
             info['near_blank'] = info['alpha_mean'] < 13  # <5% 不透明度
+        visible = [item for item in px if item[3] >= 32]
+        if visible:
+            near_white = [item for item in visible
+                          if item[0] >= 235 and item[1] >= 235 and item[2] >= 235]
+            info['near_white_ratio'] = round(len(near_white) / float(len(visible)), 3)
     except Exception:
         pass
     return info
 
 
+def needs_asset_judgment(candidate):
+    """局部图和半透明满屏叠加层需要看图定性；不透明满屏图按背景处理。"""
+    effective_alpha = candidate.get('effective_alpha_mean')
+    if ((candidate.get('probe') or {}).get('near_blank')
+            or (effective_alpha is not None and effective_alpha < 13)):
+        return False
+    if not candidate.get('fullscreen'):
+        return True
+    alpha = (effective_alpha if effective_alpha is not None
+             else (candidate.get('probe') or {}).get('alpha_mean'))
+    return alpha is not None and alpha < OPAQUE_ENOUGH
+
+
+def fullscreen_effective_alpha(data, outdir, shapes):
+    """满屏图片的实际平均 alpha，包含图片文件 alpha 与 OOXML 形状透明度。"""
+    media_out = {row.get('media'): row.get('out') for row in data.get('media') or []
+                 if row.get('media') and row.get('out')}
+    probed = {}
+    effective = {}
+    for shape in shapes:
+        media = shape.get('media')
+        if (shape.get('kind') != 'pic' or not media
+                or shape.get('w_pct', 0) < 95 or shape.get('h_pct', 0) < 95):
+            continue
+        if media not in probed:
+            out = media_out.get(media)
+            probe = probe_image(os.path.join(outdir, out)) if out else {}
+            probed[media] = probe.get('alpha_mean')
+        source_alpha = probed[media]
+        if source_alpha is None:
+            source_alpha = 255.0
+        try:
+            opacity = float(shape.get('opacity', 1.0))
+        except (TypeError, ValueError):
+            opacity = 1.0
+        alpha = source_alpha * max(0.0, min(opacity, 1.0))
+        effective[media] = min(effective.get(media, 255.0), alpha)
+    return effective
+
+
+def fullscreen_overlay_media(data, outdir, shapes):
+    """需要模型判断的满屏叠加层媒体。"""
+    return {
+        media for media, alpha in fullscreen_effective_alpha(data, outdir, shapes).items()
+        if 13 <= alpha < OPAQUE_ENOUGH
+    }
+
+
+def bg_busy_map(path, canvas, cells=12):
+    """把背景图切成网格，报每格的**局部对比度**（该格内亮度极差）。
+
+    「哪里不能压文字」的本质是「哪里花」。整幅渐变的底图各格对比度都低，说明没有
+    视觉主体；有山峰、人物、产品图的底图会在主体处出现明显更高的对比度。这里只出
+    客观数值和一个据此推出的草案，最终由看得到图的人定。
+    """
+    try:
+        from PIL import Image
+    except Exception:
+        return None
+    try:
+        im = Image.open(path).convert('L').resize((cells * 8, cells * 8))
+    except Exception:
+        return None
+    px = im.load()
+    grid = []
+    for gy in range(cells):
+        row = []
+        for gx in range(cells):
+            vals = [px[gx * 8 + x, gy * 8 + y] for y in range(8) for x in range(8)]
+            row.append(max(vals) - min(vals))
+        grid.append(row)
+    flat = sorted(v for row in grid for v in row)
+    if not flat:
+        return None
+    med = flat[len(flat) // 2]
+    hi = flat[int(len(flat) * 0.9)]
+    # 主体 = 对比度显著高于全图中位数的连片格子。阈值取「中位数与九分位的中点」，
+    # 由本图自己的分布定，不用固定值。
+    cut = (med + hi) / 2.0
+    cW, cH = canvas
+    hot = [(gx, gy) for gy in range(cells) for gx in range(cells) if grid[gy][gx] > cut]
+    if not hot:
+        return {'busy': None, 'median': med, 'p90': hi, 'why': '各处对比度一致，没有更花的区域'}
+    xs = [g[0] for g in hot]
+    ys = [g[1] for g in hot]
+    span = ((max(xs) - min(xs) + 1) * (max(ys) - min(ys) + 1)) / float(cells * cells)
+    if span > 0.5:
+        # 热格散落全图，外接矩形几乎覆盖整幅——圈出来等于没圈
+        return {'busy': None, 'median': med, 'p90': hi, 'why': '较花的格子散布全图，圈不出单一主体'}
+    box = [round(min(xs) * cW / cells), round(min(ys) * cH / cells),
+           round((max(xs) - min(xs) + 1) * cW / cells),
+           round((max(ys) - min(ys) + 1) * cH / cells)]
+    return {'busy': box, 'median': med, 'p90': hi, 'span': round(span, 2)}
+
+
 def copy_logo_candidates(outdir, logo_pool):
     if not logo_pool:
         return []
@@ -326,7 +770,8 @@ def copy_logo_candidates(outdir, logo_pool):
     return rows
 
 
-def draft_assets(d, outdir, bg_needed=(), cover_media=None, bg_under=None):
+def draft_assets(d, outdir, bg_needed=(), cover_media=None, bg_under=None,
+                 effective_alpha=None):
     imgs = {i['media']: i for i in d['images']}
     cluster_of = {}
     for c in d.get('media_clusters', []):
@@ -343,15 +788,35 @@ def draft_assets(d, outdir, bg_needed=(), cover_media=None, bg_under=None):
         boxes = img.get('boxes') or []
         top = max(boxes, key=lambda b: b.get('count', 0)) if boxes else {}
         parts = top.get('parts') or []
-        slides = sorted({slide_no(p) for p in parts if '/slides/' in p})
+        placements, seen_placements = [], set()
+        for cluster in boxes:
+            box = cluster.get('box') or {}
+            rounded = [round(box.get(key, 0)) for key in ('x', 'y', 'w', 'h')]
+            for part in cluster.get('parts') or []:
+                if '/slides/' in part:
+                    row = {'slide': slide_no(part), 'box': rounded}
+                elif '/slideLayouts/' in part:
+                    row = {'layout': os.path.basename(part), 'box': rounded}
+                else:
+                    continue
+                key = (row.get('slide'), row.get('layout'), tuple(rounded))
+                if key not in seen_placements:
+                    seen_placements.add(key)
+                    placements.append(row)
+        placements.sort(key=lambda row: (
+            row.get('slide', 9999), row.get('layout', ''), tuple(row['box'])))
+        slides = sorted({row['slide'] for row in placements if row.get('slide')})
         cands.append({
             'media': m['media'], 'file': os.path.basename(out_rel), 'out': out_rel,
             'bytes': m.get('bytes'), 'n': img.get('n', m.get('used_n', 0)),
+            'has_compressed': bool(m.get('compressed_out')),
             'fullscreen': bool(img.get('fullscreen')), 'w_pct': img.get('max_w_pct', 0),
             'box': top.get('box') or {}, 'slides': slides,
+            'placements': placements,
             'layer_only': bool(parts) and not slides,
             'repeat': bool(img.get('repeat_fixed')),
             'cluster': cluster_of.get(m['media']),
+            'effective_alpha_mean': (effective_alpha or {}).get(m['media']),
             'probe': probe, 'reasons': m.get('reasons', []),
         })
 
@@ -362,35 +827,62 @@ def draft_assets(d, outdir, bg_needed=(), cover_media=None, bg_under=None):
         if k not in best_of or c['n'] > best_of[k]['n']:
             best_of[k] = c
     kept = sorted(best_of.values(), key=lambda c: (-c['n'], -(c['bytes'] or 0)))
+    # 被同簇兄弟淘汰的 media 仍要能指到胜出者——封面底图常常是簇里 n 最小的那张
+    alias = {}
+    for c in cands:
+        w = best_of.get(c['cluster'] or c['media'])
+        if w and w['media'] != c['media']:
+            alias[c['media']] = w['media']
+    cover_media = alias.get(cover_media, cover_media)
+    bg_needed = {alias.get(m, m) for m in (bg_needed or ())}
 
     assets, rejected, todos = [], [], []
+    over_cap_bgs = []
     logo_pool = []
     bg_under = bg_under or {}
     bg_i = 0
-    canvas_w = d['canvas']['px'][0]
+    canvas_w, canvas_h = d['canvas']['px']
     for c in kept:
-        if c['probe'].get('near_blank'):
-            rejected.append((c, '近全透明（alpha 均值 %.0f/255），PPT 里看不见' % c['probe']['alpha_mean']))
+        effective_am = c.get('effective_alpha_mean')
+        if (c['probe'].get('near_blank')
+                or (effective_am is not None and effective_am < 13)):
+            rejected.append((c, '近全透明（alpha 均值 %.0f/255），PPT 里看不见'
+                             % (effective_am if effective_am is not None
+                                else c['probe']['alpha_mean'])))
+            continue
+        # 铺满 ≠ 能当背景。背景的定义性属性是**遮盖**：它得挡住底下的东西。一张大半透明
+        # 的图铺满整页也遮不住任何像素，它在 PPT 里是叠在幻灯片底色上的一层装饰（顶部
+        # 光晕之类），底色才是真背景。实测某模板一张 alpha 均值 30/255、72% 完全透明的
+        # 顶部光晕被当成满屏背景收进包，消费端每页铺它，顶部就多出一条原稿没有的浓色带。
+        am = effective_am if effective_am is not None else c['probe'].get('alpha_mean')
+        if c['fullscreen'] and am is not None and am < OPAQUE_ENOUGH:
+            rejected.append((c, 'alpha 均值只有 %.0f/255，遮不住底下的东西——'
+                             '它是叠在底色上的装饰层，不是背景' % am))
             continue
         if c['fullscreen']:
             if c['media'] == cover_media:
                 assets.append({'id': 'bg-cover', 'kind': 'background', 'role': 'cover',
-                               'src': c, 'use_full': True})
-            elif c['media'] in bg_needed and bg_i < 5:
+                               'src': c,
+                               # 只有真出了压缩版才能带原图；否则 path/full 指向同一
+                               # 文件，package.py 必 FAIL（封面不需要转码时就会踩到）
+                               'use_full': c['has_compressed']})
+            elif c['media'] in bg_needed and bg_i < BG_CONTENT_CAP:
                 bg_i += 1
                 assets.append({'id': 'bg-content-%d' % bg_i, 'kind': 'background',
                                'role': 'content', 'src': c, 'use_full': False})
+            elif c['media'] in bg_needed:
+                over_cap_bgs.append(c)
+                rejected.append((c, '有页型以它为主底，但内容页背景已收满 %d 张' % BG_CONTENT_CAP))
             else:
                 rejected.append((c, '满屏图但没有页面以它为主底（只在版式层备用）'))
-        elif c['w_pct'] < 30 and c['n'] >= 2:
+        elif c['w_pct'] < SMALL_IMG_W_PCT and c['n'] >= REPEAT_MIN:
+            # 品牌标识的共性是「小、重复出现、贴角」。这里只按贴角程度排序给出首选，
+            # 不设及格线——「多少分算 logo」没有客观依据，判断交 L 层，分项证据随 TODO 给出。
             b = c['box']
-            score = 0
-            score += 3 if b.get('y', 999) < 160 else (1 if b.get('y', 0) > canvas_w * 0.5 else 0)
-            score += 2 if b.get('x', 999) < 200 or b.get('x', 0) > canvas_w * 0.7 else 0
-            ar = (b.get('w') or 1) / max(b.get('h') or 1, 1)
-            score += 1 if 1.0 <= ar <= 8.0 else 0
-            score += 1 if c['n'] >= 2 else 0
-            logo_pool.append((score, c))
+            edge_x = min(b.get('x', 0), max(canvas_w - (b.get('x', 0) + (b.get('w') or 0)), 0))
+            edge_y = min(b.get('y', 0), max(canvas_h - (b.get('y', 0) + (b.get('h') or 0)), 0))
+            corner = (edge_x / canvas_w) + (edge_y / canvas_h)   # 越小越贴角
+            logo_pool.append((corner, c))
         else:
             rejected.append((c, '内容区图片（占宽 %.0f%%，出现 %d 次）' % (c['w_pct'], c['n'])))
 
@@ -404,46 +896,618 @@ def draft_assets(d, outdir, bg_needed=(), cover_media=None, bg_under=None):
             from PIL import Image
             im = Image.open(os.path.join(outdir, row['out'])).convert('RGB')
             b = c['box']
-            sx, sy = im.width / float(canvas_w), im.height / float(d['canvas']['px'][1])
+            sx, sy = im.width / float(canvas_w), im.height / float(canvas_h)
             crop = im.crop((int(b.get('x', 0) * sx), int(b.get('y', 0) * sy),
                             max(int((b.get('x', 0) + b.get('w', 1)) * sx), 1),
                             max(int((b.get('y', 0) + b.get('h', 1)) * sy), 1))).resize((16, 16))
             raw = crop.tobytes()
             px = [raw[i:i + 3] for i in range(0, len(raw), 3)]
-            return 'light' if sum(lum(p) for p in px) / len(px) > 0.55 else 'dark'
+            return 'light' if sum(lum(p) for p in px) / len(px) > LUM_MID else 'dark'
         except Exception:
             return None
 
-    logo_pool.sort(key=lambda kv: (-kv[0], -kv[1]['n']))
-    for i, (score, c) in enumerate(logo_pool):
+    # 贴角是品牌标识的定义性特征：离两边都超过画布 1/4 的重复小图，更可能是页内装饰。
+    # 这不是「多少分算 logo」那种凑出来的分数线——它直接来自「贴角」这个判据本身。
+    LOGO_CORNER_MAX = 0.5      # edge_x/W + edge_y/H，两边各 25% 即到上限
+    logo_pool.sort(key=lambda kv: (kv[0], -kv[1]['n']))
+    if logo_pool and logo_pool[0][0] > LOGO_CORNER_MAX:
+        todos.append('没有贴角的重复小图（最接近的一张离画布边 %.0f%%），本模板可能没有 logo；'
+                     '确认后要么从联系表挑一张补进 manifest，要么在 gaps 写明模板无品牌标识'
+                     % (logo_pool[0][0] * 50))
+        logo_pool = []
+    for i, (corner, c) in enumerate(logo_pool):
         b = c['box']
-        if i == 0 and score >= 5:
-            assets.append({'id': 'logo-primary', 'kind': 'logo', 'role': None, 'src': c,
-                           'use_full': False, 'on_bg': on_bg_of(c)})
-            todos.append('看联系表确认 `%s`（%.0fx%.0f @ %.0f,%.0f，出现 %d 次）真是品牌 logo；'
-                         '不是就把 manifest 的 logo-primary 换成别的候选或整条删掉'
-                         % (c['file'], b.get('w', 0), b.get('h', 0), b.get('x', 0), b.get('y', 0), c['n']))
+        if i == 0:
+            # 贴角、重复只能说明“像 logo”，不能替模型判定。比如一张产品功能角标也会
+            # 同时满足这些结构特征；先作为候选保留在联系表与图片槽中，由模型定为 logo
+            # 或 content，避免把内容图直接带进风格包。
+            rejected.append((c, '贴角重复小图候选（%.0fx%.0f @ %.0f,%.0f，出现 %d 次，'
+                                '离画布边 %.0f%%），结合样张判断 logo 或 content'
+                                % (b.get('w', 0), b.get('h', 0), b.get('x', 0), b.get('y', 0),
+                                   c['n'], corner * 50)))
         else:
-            rejected.append((c, '重复小图（%.0fx%.0f @ %.0f,%.0f），logo 相似度低于首选'
-                             % (b.get('w', 0), b.get('h', 0), b.get('x', 0), b.get('y', 0))))
+            rejected.append((c, '重复小图（%.0fx%.0f @ %.0f,%.0f），贴角程度 %.0f%% 不如首选'
+                             % (b.get('w', 0), b.get('h', 0), b.get('x', 0), b.get('y', 0),
+                                corner * 50)))
+
+    # 体量预算：包内资产总量超 20MB 直接 FAIL（V2-6）。`use_full` 的原图是唯一可能
+    # 单张爆预算的东西（未压缩的封面级大图可以单张达到数十 MB），所以在草案期就先丢 full，
+    # 不要留给 L 层去撞门禁再回修。
+    PACK_BUDGET = 20 * 1024 * 1024
+    est = sum(min(a['src'].get('bytes') or 0, ASSET_WARN_SINGLE) for a in assets)
+    for a in sorted([x for x in assets if x['use_full']],
+                    key=lambda x: -(x['src'].get('bytes') or 0)):
+        orig = a['src'].get('bytes') or 0
+        if est + orig > PACK_BUDGET * 0.9:
+            a['use_full'] = False
+            todos.append('`%s` 的原图 %.1fMB 会把包撑过 20MB 上限，草案已只保留压缩版；'
+                         '确实需要原图就改走 url 承载' % (a['id'], orig / 1024.0 / 1024))
+        else:
+            est += orig
 
+    if over_cap_bgs:
+        todos.append('模板有 %d 张内容页背景超出 %d 张上限（%s）；用到它们的页型在 layouts.md 里'
+                     '不会有 background，需要就手工补进 manifest 并删掉不重要的那几张'
+                     % (len(over_cap_bgs), BG_CONTENT_CAP,
+                        '、'.join(c['file'] for c in over_cap_bgs[:5])))
     if not any(a['role'] == 'cover' for a in assets):
         todos.append('没定出封面底图——从联系表挑一张补进 manifest（role: cover），或在 gaps 写明模板无封面主视觉')
     copy_logo_candidates(outdir, logo_pool)
-    return assets, rejected, todos
+    return assets, rejected, todos, alias, {c['media']: c for c in kept}
+
+
+def cover_background_media(archetypes):
+    """只从模板明确命名的 cover 页型读取封面背景。"""
+    return next((archetype['bg_raw'] for archetype in archetypes
+                 if archetype['name'] == 'cover'), None)
 
 
 # ---------------------------------------------------------------- 版式聚类
 DECOR_MIN = 40.0
 
+# 版式名 → role（模板自己按页型命名时直接用它，别再猜）。
+# 英文词按整词匹配：裸子串会让短词吃掉长词——`end` 一度把 `agenda`、`Appendix`、
+# `Trends Section` 全判成 closing，表里 `agenda -> section` 那条永远轮不到。
+ROLE_BY_WORD = [('封面', 'cover'), ('cover', 'cover'), ('首页', 'cover'),
+                ('title slide', 'cover'), ('标题幻灯片', 'cover'),
+                ('封底', 'closing'), ('尾页', 'closing'), ('结束', 'closing'),
+                ('致谢', 'closing'), ('谢谢', 'closing'), ('end', 'closing'),
+                ('thank you', 'closing'), ('closing', 'closing'),
+                ('章节', 'section'), ('目录', 'section'), ('过渡', 'section'),
+                ('section', 'section'), ('agenda', 'section'),
+                ('section header', 'section'), ('节标题', 'section'),
+                ('金句', 'quote'), ('问句', 'quote'), ('引言', 'quote'), ('quote', 'quote'),
+                ('空白', 'blank'), ('blank', 'blank')]
+PH_TO_TYPE = {'title': 'title', 'ctrTitle': 'title', 'subTitle': 'subtitle',
+              'body': 'body', 'pic': 'pic', 'clipArt': 'pic', 'tbl': 'table',
+              'chart': 'chart', 'media': 'media', 'dgm': 'pic',
+              'sldNum': 'slide-number', 'ftr': 'footer', 'dt': 'footer'}
+
+
+def role_of_name(name):
+    """版式名 → role。认不出返回 None，由调用方降置信度并留 TODO——不要静默当 content。
+
+    词表只覆盖中英文；换一种语言命名的模板会整份认不出。那时全落 content 且机检照过，
+    消费端拿到的是「每一页都是内容页」，封面/章节/结束页的语义整个丢掉且无处可查。
+    """
+    low = (name or '').lower()
+    for word, role in ROLE_BY_WORD:
+        if word.isascii():
+            if re.search(r'(?<![a-z])%s(?![a-z])' % re.escape(word), low):
+                return role
+        elif word in low:
+            return role
+    return None
+
+
+def clean_layout_name(name):
+    """`1_内容-左右排版（无副标题）` → `内容-左右排版（无副标题）`。"""
+    return re.sub(r'^\d+[_\-\s]*', '', (name or '').strip()) or '未命名版式'
+
+
+def is_bleed(s):
+    return (s.get('kind') == 'pic' and (s.get('w_pct') or 0) >= 95
+            and (s.get('h_pct') or 0) >= 95)
+
+
+def top_bleed_media(shapes):
+    """一串形状里最上层的满屏图。
+
+    OOXML 的 spTree 是绘制序，靠后的画在上面。一个版式常叠两张满屏图——通用底纹在
+    下、这一页的主视觉在上——所以看得见的是最后那张。取第一张会拿到底纹，实测让
+    章节页的深蓝主视觉被换成了另一张鲜蓝底纹，成品与原稿完全不是一个颜色。
+    """
+    out = None
+    for s in shapes:
+        if is_bleed(s) and s.get('media'):
+            out = s['media']
+    return out
+
+
+def slot_overlaps(slots):
+    """同一页型里坐标互相重叠的槽对。只报事实，不改坐标——坐标是从模板量的。"""
+    out = []
+    for i in range(len(slots)):
+        for j in range(i + 1, len(slots)):
+            a, b = slots[i].get('box'), slots[j].get('box')
+            if not (a and b):
+                continue
+            ox = min(a[0] + a[2], b[0] + b[2]) - max(a[0], b[0])
+            oy = min(a[1] + a[3], b[1] + b[3]) - max(a[1], b[1])
+            if ox > 0 and oy > 0:
+                out.append('%s×%s 叠 %dx%d' % (slots[i].get('role'), slots[j].get('role'),
+                                               round(ox), round(oy)))
+    return out
+
+
+def css_number(value, digits=3):
+    """CSS 数值稳定格式：整数不带小数，其余去掉无意义尾零。"""
+    number = round(float(value), digits)
+    if number == int(number):
+        return str(int(number))
+    return ('%.*f' % (digits, number)).rstrip('0').rstrip('.')
+
+
+def slot_style(s):
+    """占位符自带的排版样式，统一转成可直接写进 HTML style 的 CSS 声明串。
+
+    样式可能在三层：lstStyle.lvl1pPr（版式占位符常用）、段落 defRPr（Mac Office
+    导出把大量属性写在这一层）、段落 pPr（对齐）。逐层兜底，缺一层就往下取。
+
+    `box` 是布局几何，继续由 slot 独立承载；其余渲染属性不再泄漏成 size / color /
+    align / insets_px 等 PPTX 中间字段。下划线开头的键仅供 draft 内部统计，emit_layouts
+    不会写进消费者产物。
+    """
+    txt = s.get('text') or {}
+    inherited = dict((txt.get('lstStyle') or {}).get('lvl1pPr') or {})
+    ls = {}
+    # 四层逐级兜底，按 OOXML 的就近原则：run rPr → 段落 defRPr → 段落 pPr → lstStyle。
+    # 只枚举前几层会整份漏掉——有的导出器把字号全写在 run rPr 上，lstStyle 一个都没有。
+    for para in (txt.get('paragraphs') or []):
+        srcs = [r for r in (para.get('runs') or [])]
+        srcs.append(para.get('defRPr') or {})
+        srcs.append({k: v for k, v in para.items() if k not in ('runs', 'defRPr')})
+        for src in srcs:
+            for k, v in (src or {}).items():
+                if v is not None:
+                    ls.setdefault(k, v)
+    for k, v in inherited.items():
+        if v is not None:
+            ls.setdefault(k, v)
+    if not ls.get('sz_px'):
+        # 仍无声明：退到整形状里出现过的最大字号（generic walk），仍是文件里的值
+        anysz = shape_sz(s)
+        if anysz:
+            ls['sz_px'] = anysz
+    body = txt.get('bodyPr') or {}
+    css = []
+    out = {}
+    insets = body.get('insets_px') or {}
+    if insets:
+        css.append('box-sizing: border-box')
+        css.append('padding: %spx %spx %spx %spx' % (
+            css_number(insets.get('tIns', 0) or 0),
+            css_number(insets.get('rIns', 0) or 0),
+            css_number(insets.get('bIns', 0) or 0),
+            css_number(insets.get('lIns', 0) or 0),
+        ))
+    if ls.get('sz_px'):
+        # normAutofit 的 fontScale 是模板让大字装进小框的手段——不乘它，消费端拿到的是
+        # 未缩放字号，字比框高，渐变裁切会把溢出的底部切成透明。缺省 1.0（无 autofit / 无缩放）。
+        scale = body.get('font_scale')
+        raw = ls['sz_px'] * scale if scale else ls['sz_px']
+        size = round(raw)
+        css.append('font-size: %dpx' % size)
+        out['_font_size'] = size
+    typeface = ls.get('ea') or ls.get('latin') or ls.get('cs')
+    if typeface:
+        css.append('font-family: %s' % font_css([typeface]))
+    weight = ls.get('weight') or (700 if ls.get('bold') else None)
+    if weight:
+        css.append('font-weight: %s' % weight)
+        out['_font_weight'] = weight
+    if ls.get('italic'):
+        css.append('font-style: italic')
+    decorations = []
+    if ls.get('underline'):
+        decorations.append('underline')
+    if ls.get('strike'):
+        decorations.append('line-through')
+    if decorations:
+        css.append('text-decoration: %s' % ' '.join(decorations))
+    if ls.get('spc_px') is not None:
+        css.append('letter-spacing: %spx' % css_number(ls['spc_px']))
+    col = (ls.get('color') or {}).get('resolved')
+    if col:
+        css.append('color: %s' % col)
+        out['_color'] = col
+    else:
+        # 占位符的字色也可以是 gradFill（章节页的大号序号常这么做）。解析层已经把
+        # stops 和角度记全了，这里只取单色就会整条丢掉，消费端只能自己编一个平色。
+        # 与 decor 同一约定：css 是可直接写进 style 的声明串。
+        f = ls.get('fill') or {}
+        if f.get('type') == 'gradient':
+            g = _load_query()._css_gradient(f)
+            if g:
+                css += ['background-image: %s' % g, '-webkit-background-clip: text',
+                        'background-clip: text', 'color: transparent']
+    align = ls.get('algn')
+    if align:
+        css.append('text-align: %s' % {
+            'l': 'left', 'ctr': 'center', 'r': 'right', 'just': 'justify',
+        }.get(align, align))
+    line_spacing = ls.get('lnSpc') or {}
+    # normAutofit 的 lnSpcReduction 与 fontScale 同时把行距压缩，一起缩才装得进原框。
+    reduction = body.get('ln_spc_reduction') or 0
+    if line_spacing.get('mult'):
+        mult = line_spacing['mult'] * 1.2 * (1 - reduction)
+        css.append('line-height: %s' % css_number(mult))
+    elif line_spacing.get('px'):
+        css.append('line-height: %spx' % css_number(line_spacing['px'] * (1 - reduction)))
+    anchor = body.get('anchor')
+    if anchor in ('ctr', 'b'):
+        css += ['display: flex', 'flex-direction: column',
+                'justify-content: %s' % {'ctr': 'center', 'b': 'flex-end'}[anchor]]
+    if body.get('rot'):
+        try:
+            degrees = float(body['rot']) / 60000.0
+            css.append('rotate: %sdeg' % css_number(degrees))
+        except (TypeError, ValueError):
+            pass
+    if css:
+        out['css'] = '; '.join(css)
+    return out
+
+
+def instance_override(shapes, slide_part, slots, bgm, cW, cH, composites=None):
+    """实例页覆盖版式：版式是骨架，实例页才是设计师最终摆定的样子。
+
+    版式底图常是多个版式共用的通用底纹，实例页可能另铺主视觉大图；标题占位符的框高
+    也常被实例页放大以容纳多行。只读版式的包会让消费端拿到错的底图和装不下字的框，
+    只能自己缩字号。
+    """
+    ins = [s for s in shapes if s.get('part') == slide_part]
+    if not ins:
+        return slots, bgm
+    bgm = (composites or {}).get(slide_part) or top_bleed_media(ins) or bgm
+    texts = []
+    for s in ins:
+        b = s.get('box') or {}
+        if not (b.get('w') and b.get('h')) or not shape_text(s):
+            continue
+        texts.append({'sz': shape_sz(s), 'box': b, 'style': slot_style(s)})
+    texts.sort(key=lambda x: -x['sz'])
+    # 按字号大小依次顶替版式的文字槽（版式槽已按 y 排过，字号序更贴合语义层级）
+    tslots = [s for s in slots if s['type'] != 'pic']
+    for slot, ins_t in zip(sorted(tslots, key=lambda s: -(s.get('sz') or 0)), texts):
+        b = ins_t['box']
+        slot['box'] = [round(b.get('x', 0)), round(b.get('y', 0)),
+                       round(b.get('w', 0)), round(b.get('h', 0))]
+        slot['sz'] = ins_t['sz']
+        slot.update(ins_t['style'] or {})
+    return slots, bgm
+
+
+def layouts_from_template(d, shapes, cW, cH):
+    """form=3：模板自己用 slideLayout 声明了页型，直接读版式层。
+
+    拿样张聚类只能得到「样张数」个 archetype——模板往往只放 1-2 张样张，
+    真正的页型全在版式里。模板常见只放个位数样张却声明几十个语义版式，按样张聚类
+    只能得到「样张数」个 archetype，消费端搭页时大半无版式可抄，只能自己编。
+    """
+    by_part = defaultdict(list)
+    for s in shapes:
+        if (s.get('layer') == 'layout' and s.get('kind') == 'sp'
+                and (s.get('box') or {}).get('w') and (s.get('ph') or shape_text(s))):
+            by_part[s['part']].append(s)
+    bg_of_layout = {}
+    composites = d.get('background_composites') or {}
+    for s in shapes:
+        if s.get('layer') == 'layout' and is_bleed(s) and s.get('media'):
+            bg_of_layout[s['part']] = s['media']     # 靠后者在上层，最后一张才是看得见的
+    topo = d.get('theme_topology') or {}
+    theme_of_master = {m['master']: m.get('theme_label')
+                       for m in (topo.get('per_master') or [])}
+    master_of = (d.get('reference_graph') or {}).get('master_of_layout') or {}
+    # 只在版式恰好被 1 张实例页使用时才拿实例覆盖：多张实例共用一个版式时，
+    # 谁都不代表版式本身，硬挑一张会把别页的构图当成页型
+    lay_of_slide = (d.get('reference_graph') or {}).get('layout_of_slide') or {}
+    used_n = Counter(lay_of_slide.values())
+    slide_of_layout = {lp: sp for sp, lp in lay_of_slide.items() if used_n[lp] == 1}
+    sample_pages_of_layout = defaultdict(list)
+    for slide_part, layout_part in lay_of_slide.items():
+        sample_pages_of_layout[layout_part].append(slide_no(slide_part))
+    default_theme = topo.get('default')
+    multi = len(topo.get('themes') or []) > 1
+
+    rows = []
+    for l in d.get('layouts') or []:
+        phs = [s for s in by_part.get(l['part'], []) if (s.get('box') or {}).get('w')]
+        if not phs:
+            continue
+        theme = theme_of_master.get(master_of.get(l['part']))
+        phs.sort(key=lambda s: ((s['box'].get('y') or 0), (s['box'].get('x') or 0)))
+        slots, seen_kind = [], set()
+        for s in phs:
+            t = PH_TO_TYPE.get((s.get('ph') or {}).get('type'), 'body')
+            if t in ('slide-number', 'footer') and not shape_text(s):
+                continue                      # 空 chrome 占位符不是实际元素
+            b = s['box']
+            role = t if t in ('title', 'subtitle', 'footer', 'slide-number') else 'body'
+            if t == 'title' and 'title' in seen_kind:
+                role, t = 'subtitle', 'subtitle'
+            seen_kind.add(t)
+            row = {'role': role, 'type': t, 'sz': shape_sz(s),
+                   'box': [round(b.get('x', 0)), round(b.get('y', 0)),
+                           round(b.get('w', 0)), round(b.get('h', 0))],
+                   'txt': shape_text(s) or (s.get('name') or '')[:24]}
+            row.update(slot_style(s))
+            if t == 'body':
+                ph = s.get('ph') or {}
+                row.update({
+                    '_needs_role': True,
+                    '_source_layer': 'layout',
+                    '_placeholder': '%s/%s' % (
+                        ph.get('type') or '-', ph.get('idx') or '-'),
+                })
+            slots.append(row)
+        # 非满屏的图片元素（logo / 联名标 / 装饰）——它们逐版式换位置换尺寸，
+        # 必须按版式落进 slots，压成一条全局「固定位」规则就会撞标题。
+        bgm = composites.get(l['part']) or bg_of_layout.get(l['part'])
+        for s in shapes:
+            if s['part'] != l['part'] or s.get('kind') != 'pic' or not s.get('media'):
+                continue
+            if s['media'] == bgm or (s.get('w_pct', 0) >= 95 and s.get('h_pct', 0) >= 95):
+                continue
+            b = s.get('box') or {}
+            if not b.get('w'):
+                continue
+            slots.append({'role': 'logo', 'type': 'pic', 'sz': 0, 'txt': '',
+                          'media': s['media'],
+                          'box': [round(b.get('x', 0)), round(b.get('y', 0)),
+                                  round(b.get('w', 0)), round(b.get('h', 0))]})
+        if not slots:
+            continue
+        inst = slide_of_layout.get(l['part'])
+        if inst:
+            slots, bgm = instance_override(
+                shapes, inst, slots, bgm, cW, cH, composites)
+        taken = {tuple(s['box']) for s in slots}
+        decor = collect_decor(shapes, inst or l['part'], taken, (cW, cH))
+        named_role = role_of_name(l.get('name'))
+        rows.append({'zh': clean_layout_name(l.get('name')),
+                     'role': named_role or 'content', 'role_guessed': named_role is None,
+                     'slots': slots, 'decor': decor, 'bg_raw': bgm,
+                     'theme': theme, 'part': l['part'],
+                     'used': l.get('used_by_slides') or 0})
+
+    # 同名版式在 dark/light 两套 master 下各有一份——按名字归一，优先默认主题那份
+    best = {}
+    for r in rows:
+        k = r['zh']
+        cur = best.get(k)
+        if cur is None or (r['theme'] == default_theme and cur['theme'] != default_theme) \
+                or (r['used'] > cur['used']):
+            best[k] = r
+    picked = sorted(best.values(), key=lambda r: (
+        ['cover', 'section', 'quote', 'content', 'closing', 'blank'].index(r['role'])
+        if r['role'] in ('cover', 'section', 'quote', 'content', 'closing', 'blank') else 9,
+        -r['used'], r['part']))
+
+    used_key = Counter()
+    arch = []
+    for r in picked:
+        used_key[r['role']] += 1
+        n = used_key[r['role']]
+        key = r['role'] if n == 1 else '%s-%d' % (r['role'], n)
+        m_no = re.search(r'slideLayout(\d+)\.xml$', r['part'])
+        arch.append({'name': key, 'zh': r['zh'], 'role': r['role'], 'bg': None,
+                     'role_guessed': r.get('role_guessed'),
+                     'bg_raw': r['bg_raw'], 'slots': r['slots'],
+                     'decor': r.get('decor') or [], 'pages': [],
+                     'rep': None, 'rep_layout': int(m_no.group(1)) if m_no else None,
+                     # 版式名认不出 role 时不装作有把握：置信度降到 low，让 L 层看图定
+                     'pic_n': 0, 'confidence': 'low' if r.get('role_guessed') else 'high',
+                     'theme': r['theme'] if multi else None,
+                     '_layout_part': r['part'],
+                     '_sample_pages': sorted(sample_pages_of_layout.get(r['part']) or []),
+                     'source': 'layout:' + r['part'].split('/')[-1]})
+    return arch
+
+
+_QUERY = []
+
+
+def _load_query():
+    """复用 query.py 的 OOXML→CSS 渲染，不再写第二份。"""
+    if not _QUERY:
+        import importlib.util
+        spec = importlib.util.spec_from_file_location('_q', os.path.join(HERE, 'query.py'))
+        mod = importlib.util.module_from_spec(spec)
+        spec.loader.exec_module(mod)
+        _QUERY.append(mod)
+    return _QUERY[0]
+
+
+def collect_decor(shapes, part, taken_boxes, canvas, limit=10):
+    """页面上撑起版式骨架、但不含文字的形状（圆形图标托、卡片、分隔线）。
+
+    只给文字框的坐标，消费端看到的是「一段说明悬在半空、上方一片空白」，只能自己编
+    容器，编出来的形状与模板无关。这些形状必须进包。
+    """
+    q = _load_query()
+    cW, cH = canvas
+    out = []
+    for s in shapes:
+        if s.get('part') != part or s.get('kind') != 'sp':
+            continue
+        if any(r.get('text', '').strip()
+               for para in ((s.get('text') or {}).get('paragraphs') or [])
+               for r in (para.get('runs') or [])):
+            continue                       # 有文字的已经作为 slot 出过
+        b = s.get('box') or {}
+        w, h = b.get('w') or 0, b.get('h') or 0
+        if not (w or h):
+            continue                       # 零尺寸形状渲染不出任何东西
+        if canvas_coverage(b, cW, cH) >= FULLSCREEN_COVERAGE:
+            continue                       # 满屏底，属 background
+        box = [round(b.get('x', 0)), round(b.get('y', 0)), round(w), round(h)]
+        if tuple(box) in taken_boxes:
+            continue
+        css = q._recipe_css(s.get('fill'), s.get('line'),
+                            [s.get('radius_px')] if s.get('radius_px') else [], s.get('effects'))
+        # 声明要落成单行：含换行的声明会被下游的行式解析器从换行处截断，
+        # 且只记 PARSE-WARN 不 FAIL，整包照常出厂——带着半条渲染不出来的 CSS
+        css = [re.sub(r'\s*\n\s*', ' ', c.split('\x00')[0]).strip() for c in css if c]
+        if not css:
+            continue                       # 无填充无描边无阴影 = 看不见，不占篇幅
+        out.append({'box': box, 'geom': (s.get('geom') or {}).get('prst') or 'rect',
+                    'css': '; '.join(css), 'area': max(w * h, w, h)})
+    # 按面积降序取前 limit 条：撑起版式的结构性形状总在最前，零星噪点自然落在截断线外，
+    # 不需要再设一个「多小算噪点」的尺寸门槛（那种门槛会误杀 1px 分隔线）。
+    out.sort(key=lambda d: -d['area'])
+    note_truncation('装饰形状', limit, len(out), '按面积降序保留，剩下的多是零星小件',
+                    part.split('/')[-1])
+    return out[:limit]                      # 同款不同位置都要留，位置本身是版式信息
+
+
+def placeholder_key(shape):
+    ph = shape.get('ph') or {}
+    if not ph:
+        return None
+    return (ph.get('type') or 'body', str(ph.get('idx') or ''))
+
+
+def merge_dict(base, override):
+    """把实例页的非空声明叠到版式声明上；空实例占位符继续继承版式事实。"""
+    out = copy.deepcopy(base or {})
+    for key, value in (override or {}).items():
+        if value is None or value == []:
+            continue
+        if isinstance(value, dict) and isinstance(out.get(key), dict):
+            out[key] = merge_dict(out[key], value)
+        else:
+            out[key] = copy.deepcopy(value)
+    return out
+
+
+def inherited_text_shapes(layout_shapes, slide_shapes):
+    """返回实例页可用的文字形状，并补齐其引用版式中的占位符几何与样式。"""
+    layout_text = []
+    for shape in layout_shapes:
+        if shape.get('kind') != 'sp' or not (shape.get('box') or {}).get('w'):
+            continue
+        ph = shape.get('ph') or {}
+        ph_type = ph.get('type')
+        if shape_text(shape) or (ph and ph_type not in ('ftr', 'dt', 'sldNum')):
+            layout_text.append(shape)
+    by_placeholder = {placeholder_key(s): s for s in layout_text if placeholder_key(s)}
+    used = set()
+    out = []
+    for shape in slide_shapes:
+        if shape.get('kind') != 'sp':
+            continue
+        key = placeholder_key(shape)
+        base = by_placeholder.get(key)
+        if base:
+            merged = merge_dict(base, shape)
+            merged['text'] = merge_dict(base.get('text'), shape.get('text'))
+            if not shape_text(shape):
+                merged['text']['paragraphs'] = copy.deepcopy(
+                    (base.get('text') or {}).get('paragraphs') or [])
+            used.add(key)
+            out.append((merged, 'slide+layout'))
+        elif (shape.get('box') or {}).get('w') and shape_text(shape):
+            out.append((shape, 'slide'))
+    for shape in layout_text:
+        key = placeholder_key(shape)
+        if key not in used:
+            out.append((shape, 'layout'))
+    return out
+
+
+def slide_image_marks(data, included_fullscreen=()):
+    """从图片普查补齐形状图片填充；它们没有独立 pic 节点，但仍有媒体与坐标。"""
+    allowed_fullscreen = set(included_fullscreen)
+    out = defaultdict(list)
+    for image in data.get('images') or []:
+        media = image.get('media')
+        if not media or (image.get('fullscreen') and media not in allowed_fullscreen):
+            continue
+        for cluster in image.get('boxes') or []:
+            box = cluster.get('box')
+            if not box or not box.get('w'):
+                continue
+            for part in cluster.get('parts') or []:
+                if '/slides/' not in part and '/slideLayouts/' not in part:
+                    continue
+                out[part].append({'media': media, 'box': box})
+    return out
+
 
-def draft_layouts(d, outdir):
-    shapes = json.load(open(os.path.join(outdir, 'ref', 'shapes.json'), encoding='utf-8'))['shapes']
+def add_template_image_marks(archetypes, data, included_fullscreen=()):
+    """把版式和实例页的图片填充补进 form=3 页型。"""
+    marks_by_part = slide_image_marks(data, included_fullscreen)
+    layout_of_slide = (data.get('reference_graph') or {}).get('layout_of_slide') or {}
+    by_layout = {archetype.get('_layout_part'): archetype for archetype in archetypes}
+    for part, marks in marks_by_part.items():
+        layout_part = layout_of_slide.get(part, part)
+        archetype = by_layout.get(layout_part)
+        if not archetype:
+            continue
+        seen = {
+            (slot.get('media'), tuple(slot.get('box') or ()))
+            for slot in archetype.get('slots') or []
+            if slot.get('media')
+        }
+        for mark in marks:
+            box = mark['box']
+            rounded = [round(box.get(key, 0)) for key in ('x', 'y', 'w', 'h')]
+            key = (mark['media'], tuple(rounded))
+            if key in seen:
+                continue
+            seen.add(key)
+            archetype['slots'].append({
+                'role': 'logo',
+                'type': 'pic',
+                'sz': 0,
+                'txt': '',
+                'media': mark['media'],
+                'box': rounded,
+            })
+
+
+def preserve_image_bearing_groups(kept, ranked):
+    """有图片实例的孤例保留自己的页型，避免把资产绑定到近似但错误的版式。"""
+    return kept + [
+        group for group in ranked
+        if group not in kept and any(page.get('marks') for page in group[1])
+    ]
+
+
+def draft_layouts(d, outdir, effective_alpha=None):
+    with open(os.path.join(outdir, 'ref', 'shapes.json'), encoding='utf-8') as stream:
+        shapes = json.load(stream)['shapes']
     cW, cH = d['canvas']['px']
+    if effective_alpha is None:
+        effective_alpha = fullscreen_effective_alpha(d, outdir, shapes)
+    overlay_media = {
+        media for media, alpha in effective_alpha.items()
+        if 13 <= alpha < OPAQUE_ENOUGH
+    }
+    if (d.get('form_hint') or {}).get('form') == 3:
+        arch = layouts_from_template(d, shapes, cW, cH)
+        if len(arch) >= 3:
+            add_template_image_marks(arch, d, overlay_media)
+            return arch, [], []
     by_slide = defaultdict(list)
+    by_layout = defaultdict(list)
     for s in shapes:
         if s.get('layer') == 'slide':
             by_slide[s['part']].append(s)
+        elif s.get('layer') == 'layout':
+            by_layout[s['part']].append(s)
+    image_marks = slide_image_marks(d, overlay_media)
 
     bg_of_slide, layout_of_slide = {}, {}
     for s in d.get('slides', []):
@@ -451,94 +1515,200 @@ def draft_layouts(d, outdir):
         bg_of_slide[s['part']] = json.dumps(bg, sort_keys=True) if isinstance(bg, dict) else bg
         layout_of_slide[s['part']] = s.get('layout')
     # 版式层的满屏底图（form=2 常态：底图挂在 layout 上）
+    composites = d.get('background_composites') or {}
     bg_of_layout = {}
     for s in shapes:
-        if (s.get('layer') == 'layout' and s.get('kind') == 'pic'
-                and s.get('w_pct', 0) >= 95 and s.get('h_pct', 0) >= 95):
-            bg_of_layout.setdefault(s['part'], s.get('media'))
+        if s.get('layer') == 'layout' and is_bleed(s) and s.get('media'):
+            bg_of_layout[s['part']] = s['media']
 
     pages = []
     for part, sh in sorted(by_slide.items(), key=lambda kv: slide_no(kv[0])):
-        bg_media = None
-        for s in sh:
-            if s.get('kind') == 'pic' and s.get('w_pct', 0) >= 95 and s.get('h_pct', 0) >= 95:
-                bg_media = s.get('media')
-                break
+        layout_part = layout_of_slide.get(part)
+        layout_shapes = by_layout.get(layout_part) or []
+        bg_media = top_bleed_media(sh)
         if bg_media is None:
-            bg_media = bg_of_layout.get(layout_of_slide.get(part))
+            bg_media = bg_of_layout.get(layout_part)
+        rendered_bg = (composites.get(part)
+                       or composites.get(layout_part)
+                       or bg_media)
         texts = []
-        for s in sh:
-            if s.get('kind') != 'sp':
-                continue
-            txt = shape_text(s)
-            if not txt:
-                continue
+        for s, source_layer in inherited_text_shapes(layout_shapes, sh):
+            txt = shape_text(s) or (s.get('name') or '')[:24]
             b = s.get('box') or {}
             if b.get('w', 0) < DECOR_MIN or b.get('h', 0) < 16:
                 continue
-            texts.append({'sz': shape_sz(s), 'box': b, 'txt': txt})
+            ph = s.get('ph') or {}
+            ph_type = ph.get('type')
+            direct_type = PH_TO_TYPE.get(ph_type, 'body')
+            texts.append({
+                'sz': shape_sz(s),
+                'box': b,
+                'txt': txt,
+                'style': slot_style(s),
+                'direct_type': direct_type,
+                'needs_role': direct_type == 'body',
+                'source_layer': source_layer,
+                'placeholder': '%s/%s' % (ph_type or '-', ph.get('idx') or '-'),
+            })
         texts.sort(key=lambda t: (-t['sz'], t['box'].get('y', 0)))
-        pics = [s for s in sh if s.get('kind') == 'pic' and s.get('w_pct', 0) < 95]
+        visible_shapes = layout_shapes + sh
+        pics = []
+        for shape in visible_shapes:
+            if shape.get('kind') != 'pic':
+                continue
+            if shape.get('w_pct', 0) < 95 or shape.get('media') in overlay_media:
+                pics.append(shape)
+        # 小图元素（logo / 角标 / 装饰）逐页记位置，供 archetype 落 slots
+        marks = [{'media': s['media'], 'box': s['box']} for s in pics
+                 if s.get('media') and (s.get('box') or {}).get('w')]
+        seen_marks = {
+            (mark['media'], round(mark['box'].get('x', 0)), round(mark['box'].get('y', 0)))
+            for mark in marks
+        }
+        for mark in image_marks.get(part) or []:
+            key = (mark['media'], round(mark['box'].get('x', 0)),
+                   round(mark['box'].get('y', 0)))
+            if key not in seen_marks:
+                seen_marks.add(key)
+                marks.append(mark)
         pages.append({'part': part, 'no': slide_no(part), 'bg_media': bg_media,
+                      'rendered_bg': rendered_bg,
                       'bg_color': bg_of_slide.get(part), 'texts': texts, 'pic_n': len(pics),
-                      'shape_n': len(sh)})
+                      'marks': marks, 'shape_n': len(visible_shapes), 'layout': layout_part})
 
-    def kind_of(p):
+    # 页型的**角色**（封面 / 章节页 / 内容页……）不在这里判：那是看图才能下的结论，
+    # 交给读得到重建图的模型。脚本只做客观归并——同一张底图 + 文字块数量相近的页
+    # 归成一组，档位按本 deck 自己的分布切，不用「字号 ≥60 就是章节页」这类固定数。
+    ns = sorted(len(p['texts']) for p in pages) or [0]
+    q1, q2 = ns[len(ns) // 3], ns[len(ns) * 2 // 3]
+
+    def density_band(p):
         n = len(p['texts'])
-        top = p['texts'][0]['sz'] if p['texts'] else 0
-        if p['no'] == 1:
-            return 'cover'
-        if n <= 3 and top >= 60:
-            return 'section'
-        if n >= 8 or p['pic_n'] >= 4:
-            return 'content-dense'
-        return 'content'
+        return 0 if n <= q1 else (1 if n <= q2 else 2)
 
+    last_page_no = max((p['no'] for p in pages), default=None)
     groups = defaultdict(list)
     for p in pages:
-        groups[(p['bg_media'] or p['bg_color'] or 'none', kind_of(p))].append(p)
+        if p['no'] == 1:
+            # 首页单独成组：它是 deck 唯一的入口页，版面通常和后面任何一页都不同，
+            # 并进别的组就会被代表页顶掉、坐标全丢。这只是不合并，不代表它是封面。
+            groups[('__first__', -1)] = [p]
+            continue
+        if p['no'] == last_page_no:
+            # 末页也单独保留完整结构：它可能是封底，也可能只是最后一张内容页，脚本
+            # 不替模型下结论。和首页一样，拆组只避免它被聚类代表页吞掉。
+            groups[('__last__', -2)] = [p]
+            continue
+        groups[(p['bg_media'] or p['bg_color'] or 'none', density_band(p))].append(p)
 
     ranked = sorted(groups.items(), key=lambda kv: (-len(kv[1]), kv[1][0]['no']))
-    kept = [g for g in ranked if len(g[1]) >= 2 or g[0][1] == 'cover'][:8]
+    # 首页所在的组一定收——deck 的第一页是模板的门面，孤例也不能被名额挤掉。
+    # 这只保证它进包，它是不是封面由看图的人定。
+    first = [g for g in ranked if g[0][0] == '__first__']
+    last = [g for g in ranked if g[0][0] == '__last__']
+    kept = first + last + [
+        g for g in ranked
+        if g not in first and g not in last and len(g[1]) >= 2
+    ][:max(0, 8 - len(first) - len(last))]
     for g in ranked:                      # 名额没用满就把最大的孤例页也收进来
         if len(kept) >= 8:
             break
         if g not in kept:
             kept.append(g)
+    # 图片用途必须与它实际所在的版式绑定。若把图片孤例并到“最接近”页型，装饰会被
+    # 绑定到错误布局；是否为内容图、logo 墙或装饰由后续模型看图判断，不按图片数量猜。
+    kept = preserve_image_bearing_groups(kept, ranked)
     leftover = sorted(p['no'] for g in ranked if g not in kept for p in g[1])
 
     archetypes = []
-    used = Counter()
-    for (bg_raw, kind), ps in kept:
+    for gi, ((bg_raw, _band), ps) in enumerate(kept, 1):
         rep = max(ps, key=lambda p: len(p['texts']))
-        used[kind] += 1
-        name = kind if used[kind] == 1 else '%s-%d' % (kind, used[kind])
-        # 标题按「位置 + 跨度」认，不按字号——巨号数值（21%、7,869）常比标题还大
-        band = [t for t in rep['texts']
-                if t['box'].get('y', 1e9) < cH * 0.28 and t['box'].get('w', 0) >= cW * 0.25]
-        title = max(band, key=lambda t: t['sz']) if band else (
+        if bg_raw == '__first__':
+            bg_raw = rep['bg_media'] or rep['bg_color'] or 'none'
+        elif bg_raw == '__last__':
+            bg_raw = rep['bg_media'] or rep['bg_color'] or 'none'
+        rendered_bg = rep.get('rendered_bg')
+        if rendered_bg:
+            bg_raw = rendered_bg
+        name = 'layout-%d' % gi
+        # 标题按「位置 + 跨度」认，不按字号——big-number 类的巨号数值常比标题还大
+        # 标题 = 该页最靠上的那批文本里最宽的一块。不按「画布前 28%」这类固定比例切：
+        # 版心靠下的模板会整页认不出标题。以该页自身的文本框分布定「靠上」。
+        ys = sorted(t['box'].get('y', 0) for t in rep['texts'])
+        y_cut = ys[max(len(ys) // 4, 0)] if ys else 0
+        band = [t for t in rep['texts'] if t['box'].get('y', 1e9) <= y_cut]
+        title = max(band, key=lambda t: (t['box'].get('w', 0), t['sz'])) if band else (
             max(rep['texts'], key=lambda t: t['sz']) if rep['texts'] else None)
         rest = [t for t in rep['texts'] if t is not title]
         rest.sort(key=lambda t: (t['box'].get('y', 0), t['box'].get('x', 0)))
         ordered = ([title] if title else []) + rest
         slots = []
-        for i, t in enumerate(ordered[:6]):
+        for i, t in enumerate(ordered):
             b = t['box']
-            if t is title:
+            if t.get('needs_role'):
+                role = typ = 'body'
+            elif t.get('direct_type') in ('title', 'subtitle', 'footer', 'slide-number'):
+                role = typ = t['direct_type']
+            elif t is title:
                 role = typ = 'title'
-            elif (title and i == 1 and t['sz'] >= 28
-                  and abs(b.get('x', 0) - title['box'].get('x', 0)) < 120
+            elif (title and i == 1
+                  # 副标题 = 紧跟在标题下方、与标题左对齐的那一块。三个量都相对标题
+                  # 自身：绝对 px 门槛在大字号排版的模板上会整片认不出来。
+                  and abs(b.get('x', 0) - title['box'].get('x', 0)) <= title['box'].get('h', 0)
                   and 0 <= b.get('y', 0) - (title['box'].get('y', 0)
-                                            + title['box'].get('h', 0)) < 220):
+                                            + title['box'].get('h', 0))
+                  <= title['box'].get('h', 0) * 2):
                 role = typ = 'subtitle'
             else:
                 role = typ = 'body'
-            slots.append({'role': role, 'box': [round(b.get('x', 0)), round(b.get('y', 0)),
-                                                round(b.get('w', 0)), round(b.get('h', 0))],
-                          'type': typ, 'sz': t['sz'], 'txt': t['txt']})
+            row = {'role': role, 'box': [round(b.get('x', 0)), round(b.get('y', 0)),
+                                        round(b.get('w', 0)), round(b.get('h', 0))],
+                   'type': typ, 'sz': t['sz'], 'txt': t['txt']}
+            row.update(t.get('style') or {})
+            if t.get('needs_role'):
+                row.update({
+                    '_needs_role': True,
+                    '_source_layer': t.get('source_layer'),
+                    '_placeholder': t.get('placeholder'),
+                })
+            slots.append(row)
+        # 同组页面上的图片元素按素材+位置去重后落候选 slots。内容图去掉具体资产引用，
+        # 保留通用图片槽；装饰图绑定资产，避免非代表页上的装饰没有进入 layouts。
+        seen_mark = set()
+        for page in ps:
+            for mk in page.get('marks') or []:
+                b = mk['box']
+                key = (mk['media'], round(b.get('x', 0)), round(b.get('y', 0)))
+                if key in seen_mark:
+                    continue
+                seen_mark.add(key)
+                slots.append({'role': 'logo', 'type': 'pic', 'sz': 0, 'txt': '',
+                              'media': mk['media'],
+                              'box': [round(b.get('x', 0)), round(b.get('y', 0)),
+                                      round(b.get('w', 0)), round(b.get('h', 0))]})
+        taken = {tuple(s['box']) for s in slots}
+        decor = []
+        seen_decor = set()
+        for source_part in (rep.get('layout'), rep['part']):
+            for item in collect_decor(shapes, source_part, taken, (cW, cH)):
+                key = (tuple(item['box']), item['geom'], item['css'])
+                if key not in seen_decor:
+                    seen_decor.add(key)
+                    decor.append(item)
         archetypes.append({'name': name, 'bg': None, 'bg_raw': bg_raw, 'slots': slots,
+                           'decor': decor,
                            'pages': sorted(p['no'] for p in ps), 'rep': rep['no'],
                            'pic_n': rep['pic_n'],
+                           '_source_layouts': sorted({
+                               p['layout'] for p in ps if p.get('layout')
+                           }),
+                           '_source_backgrounds': sorted({
+                               p.get('rendered_bg') or p.get('bg_media') or p.get('bg_color')
+                               for p in ps
+                               if p.get('rendered_bg') or p.get('bg_media') or p.get('bg_color')
+                           }),
+                           '_text_n': len(rep['texts']),
+                           '_last_page_candidate': rep['no'] == last_page_no,
                            'confidence': 'high' if len(ps) >= 3 else
                                          ('medium' if len(ps) == 2 else 'low')})
     return archetypes, pages, leftover
@@ -547,15 +1717,24 @@ def draft_layouts(d, outdir):
 # ---------------------------------------------------------------- 联系表
 def layout_sheet(outdir, archetypes, path):
     """把各 archetype 的代表页光栅出来拼成一张——版式命名得看得见页面。"""
-    reps = [a['rep'] for a in archetypes]
+    use_layout = all(a.get('rep') is None for a in archetypes)
+    reps = [a.get('rep_layout') if use_layout else a.get('rep') for a in archetypes]
+    reps = [x for x in reps if x is not None]
     if not reps:
         return None
-    import subprocess
-    r = subprocess.run([sys.executable, os.path.join(HERE, 'render_pages.py'), outdir,
-                        '--pages', 'slides', '--only', ','.join(map(str, reps)), '--no-html'],
-                       capture_output=True, text=True)
     png_dir = os.path.join(outdir, 'ref', 'rebuild', 'png')
-    if r.returncode or not os.path.isdir(png_dir):
+    kind = 'layout' if use_layout else 'slide'
+    missing = [no for no in reps
+               if not os.path.exists(os.path.join(png_dir, '%s-%s.png' % (kind, no)))]
+    if missing:
+        import subprocess
+        r = subprocess.run([sys.executable, os.path.join(HERE, 'render_pages.py'), outdir,
+                            '--pages', 'layouts' if use_layout else 'slides',
+                            '--only', ','.join(map(str, missing)), '--no-html'],
+                           capture_output=True, text=True)
+        if r.returncode:
+            return None
+    if not os.path.isdir(png_dir):
         return None
     try:
         from PIL import Image, ImageDraw
@@ -570,26 +1749,31 @@ def layout_sheet(outdir, archetypes, path):
     for i, a in enumerate(archetypes):
         x = pad + (i % cols) * (cw + pad)
         y = pad + (i // cols) * (ch + pad + lab)
-        f = os.path.join(png_dir, 'slide-%d.png' % a['rep'])
+        no = a.get('rep_layout') if use_layout else a.get('rep')
+        f = os.path.join(png_dir, '%s-%s.png' % (kind, no))
         if os.path.exists(f):
             im = Image.open(f).convert('RGB')
             im.thumbnail((cw, ch))
             sheet.paste(im, (x, y))
         dr.rectangle([x, y, x + cw, y + ch], outline=(120, 120, 128))
-        dr.text((x + 2, y + ch + 5), '[%s]  slide %d  x%d pages  bg=%s'
-                % (a['name'], a['rep'], len(a['pages']), a.get('bg') or '-'),
+        # 标注只写 ASCII——Pillow 默认字体没有 CJK 字形，中文会渲染成方框
+        dr.text((x + 2, y + ch + 5), '[%s]  %s  bg=%s'
+                % (a['name'],
+                   ('layout %s' % a.get('rep_layout')) if use_layout
+                   else ('slide %s x%d pages' % (a.get('rep'), len(a['pages']))),
+                   a.get('bg') or '-'),
                 fill=(20, 20, 24))
     sheet.save(path, optimize=True)
     return path
 
 
-def contact_sheet(outdir, cands, path):
+def contact_sheet(outdir, cands, path, start_index=1):
     try:
         from PIL import Image, ImageDraw
     except Exception:
         return None
     cell, pad, cols = 220, 20, 4
-    items = cands[:12]
+    items = cands            # 上限由调用方定，编号与 BRIEF 表格一一对应
     if not items:
         return None
     rows = (len(items) + cols - 1) // cols
@@ -613,19 +1797,548 @@ def contact_sheet(outdir, cands, path):
             dr.text((x + 8, y + 8), 'unreadable', fill=(200, 0, 0))
         dr.rectangle([x, y, x + cell, y + cell], outline=(120, 120, 128))
         dr.text((x + 2, y + cell + 4), '[%d] %s  %dx%d  used=%d'
-                % (idx + 1, c['file'], c['probe'].get('w') or 0, c['probe'].get('h') or 0, c['n']),
+                % (c.get('_candidate_index', start_index + idx), c['file'],
+                   c['probe'].get('w') or 0,
+                   c['probe'].get('h') or 0, c['n']),
                 fill=(20, 20, 24))
     sheet.save(path, optimize=True)
     return path
 
 
+def contact_sheets(outdir, cands, ldir):
+    paths = []
+    legacy = os.path.join(ldir, 'contact-sheet.png')
+    if os.path.exists(legacy):
+        os.remove(legacy)
+    for start in range(0, len(cands), SHEET_BATCH):
+        batch = cands[start:start + SHEET_BATCH]
+        path = os.path.join(ldir, 'contact-sheet-%d.png' % (start // SHEET_BATCH + 1))
+        if contact_sheet(outdir, batch, path, start + 1):
+            paths.append(path)
+    if paths:
+        shutil.copy2(paths[0], legacy)
+    return paths
+
+
+def asset_vision_contexts(candidates):
+    """把同一素材的每个归纳页型实例放进对应语境，不只展示最早出现的页面。"""
+    contexts = []
+    for candidate in candidates:
+        placements = candidate.get('placements') or []
+        instance_placements = [row for row in placements if row.get('slide')]
+        if not instance_placements:
+            row = dict(candidate)
+            row['source_placements'] = placements
+            contexts.append(row)
+            continue
+        seen = set()
+        id_counts = Counter()
+        for placement in instance_placements:
+            box = tuple(placement['box'])
+            layout = placement.get('archetype')
+            key = (layout, box) if layout else (placement['slide'], box)
+            if key in seen:
+                continue
+            seen.add(key)
+            row = dict(candidate)
+            base_id = '%s-s%d' % (candidate['id'], placement['slide'])
+            id_counts[base_id] += 1
+            row['id'] = (base_id if id_counts[base_id] == 1
+                         else '%s-%d' % (base_id, id_counts[base_id]))
+            row['placements'] = [placement]
+            row['slides'] = [placement['slide']]
+            row['layout'] = layout
+            row['source_placements'] = placements
+            contexts.append(row)
+    return contexts
+
+
+def _group_input_count(group):
+    # 与 FaaS 一致：每页桶预留一张页面语境图；无实例页的版式候选也占一个图位。
+    page_count = len([page for page in group['pages'] if page > 0]) or 1
+    return len(group['candidates']) + page_count
+
+
+def build_asset_vision_groups(candidates):
+    """按 FaaS 的 10/5 图数预算，把候选按所在页组合成视觉判断批次。"""
+    buckets = defaultdict(list)
+    for candidate in asset_vision_contexts(candidates):
+        page = next((row['slide'] for row in candidate.get('placements') or []
+                     if row.get('slide')), 0)
+        buckets[page].append(candidate)
+
+    groups, queued = [], []
+
+    def flush_small_pages():
+        if not queued:
+            return
+        current, current_pages, inputs = [], [], 0
+        for page, page_candidates in queued:
+            page_inputs = 1 + len(page_candidates)
+            if current and inputs + page_inputs > MULTI_PAGE_IMAGE_BUDGET:
+                groups.append({'pages': current_pages, 'candidates': current})
+                current, current_pages, inputs = [], [], 0
+            current.extend(page_candidates)
+            current_pages.append(page)
+            inputs += page_inputs
+        if current:
+            groups.append({'pages': current_pages, 'candidates': current})
+        del queued[:]
+
+    for page in sorted(buckets):
+        page_candidates = buckets[page]
+        page_inputs = 1 + len(page_candidates)
+        if page_inputs > MULTI_PAGE_IMAGE_BUDGET:
+            flush_small_pages()
+            per_group = SINGLE_PAGE_IMAGE_BUDGET - 1
+            for start in range(0, len(page_candidates), per_group):
+                groups.append({
+                    'pages': [page],
+                    'candidates': page_candidates[start:start + per_group],
+                })
+            continue
+        queued.append((page, page_candidates))
+    flush_small_pages()
+
+    for index, group in enumerate(groups, 1):
+        group['id'] = 'vision-%d' % index
+        group['input_count'] = _group_input_count(group)
+    return groups
+
+
+def select_asset_vision_groups(groups, slide_count):
+    """受总预算约束选择视觉批次：首页和尾页的所有可容纳分批优先于中间页。"""
+    if not groups:
+        return [], []
+    first_page = 1
+    last_page = slide_count or max(
+        (page for group in groups for page in group['pages'] if page > 0), default=0)
+    selected, selected_ids, inputs = [], set(), 0
+
+    def add(group):
+        nonlocal inputs
+        if (group['id'] in selected_ids or len(selected) >= VISUAL_PACK_CAP
+                or inputs + group['input_count'] > VISUAL_INPUT_CAP):
+            return False
+        selected.append(group)
+        selected_ids.add(group['id'])
+        inputs += group['input_count']
+        return True
+
+    # 首尾页先于中间页保留全部可容纳分批。交错加入避免首页多批先占满总预算，尾页
+    # 连首批都进不去；单页 deck 不重复扫描。
+    priority_batches = [
+        [group for group in groups if page in group['pages']]
+        for page in dict.fromkeys((first_page, last_page))
+    ]
+    for batch_index in range(max(map(len, priority_batches), default=0)):
+        for batches in priority_batches:
+            if batch_index < len(batches):
+                add(batches[batch_index])
+
+    # 首尾的第一个批次已经保证；剩余按页码保留前段内容，优先丢弃尾页之前的后段。
+    remainder = sorted(
+        (group for group in groups if group['id'] not in selected_ids),
+        key=lambda group: (
+            min((page for page in group['pages'] if page > 0), default=999999),
+            group['id'],
+        ),
+    )
+    for group in remainder:
+        add(group)
+
+    selected.sort(key=lambda group: (
+        min((page for page in group['pages'] if page > 0), default=999999), group['id']))
+    omitted = [group for group in groups if group['id'] not in selected_ids]
+    return selected, omitted
+
+
+def _safe_remove(pattern):
+    for path in glob.glob(pattern):
+        try:
+            os.remove(path)
+        except OSError:
+            pass
+
+
+def _fit_image(image, width, height):
+    copy = image.copy()
+    copy.thumbnail((width, height))
+    return copy
+
+
+def _draw_checkerboard(draw, box, size=14):
+    x, y, w, h = box
+    for row in range(0, h, size):
+        for col in range(0, w, size):
+            if (row // size + col // size) % 2 == 0:
+                draw.rectangle([x + col, y + row, x + col + size - 1, y + row + size - 1],
+                               fill=(214, 214, 218))
+
+
+def _candidate_is_visual_risk(candidate):
+    probe = candidate.get('probe') or {}
+    alpha = candidate.get('effective_alpha_mean')
+    if alpha is None:
+        alpha = probe.get('alpha_mean')
+    return ((alpha is not None and alpha < 230)
+            or (probe.get('near_white_ratio') or 0) >= 0.7)
+
+
+def _paste_candidate_preview(sheet, draw, image, box, dark=False):
+    x, y, w, h = box
+    if dark:
+        draw.rectangle([x, y, x + w, y + h], fill=(54, 54, 58))
+    else:
+        _draw_checkerboard(draw, box)
+    preview = _fit_image(image.convert('RGBA'), w - 8, h - 8)
+    px = x + (w - preview.width) // 2
+    py = y + (h - preview.height) // 2
+    sheet.paste(preview, (px, py), preview)
+
+
+def _placement_text(candidate):
+    rows = []
+    for placement in candidate.get('placements') or []:
+        box = placement['box']
+        if placement.get('slide'):
+            rows.append('s%d@%d,%d,%d,%d' % (
+                placement['slide'], box[0], box[1], box[2], box[3]))
+        else:
+            rows.append('%s@%d,%d,%d,%d' % (
+                placement.get('layout') or 'layout', box[0], box[1], box[2], box[3]))
+    return ';'.join(rows)
+
+
+def _save_visual_sheet(sheet, path):
+    if max(sheet.size) > VISUAL_PREVIEW_MAX_EDGE:
+        ratio = VISUAL_PREVIEW_MAX_EDGE / float(max(sheet.size))
+        sheet = sheet.resize((max(1, round(sheet.width * ratio)),
+                              max(1, round(sheet.height * ratio))))
+    sheet.save(path, 'JPEG', quality=VISUAL_JPEG_QUALITY, optimize=True, progressive=True)
+
+
+def render_asset_vision_pages(outdir, pages):
+    """视觉判断必须有页面语境；截图失败时中止草案而非让模型盲判。"""
+    if not pages:
+        return None
+    result = subprocess.run(
+        [sys.executable, os.path.join(HERE, 'render_pages.py'), outdir,
+         '--pages', 'slides', '--only', ','.join(map(str, pages)), '--no-html'],
+        capture_output=True, text=True,
+    )
+    png_dir = os.path.join(outdir, 'ref', 'rebuild', 'png')
+    missing = [
+        page for page in pages
+        if not os.path.isfile(os.path.join(png_dir, 'slide-%d.png' % page))
+    ]
+    if result.returncode or missing:
+        detail = (result.stderr or result.stdout or '').strip().splitlines()
+        raise RuntimeError(
+            '视觉判断所需页面截图生成失败%s%s' % (
+                '（缺第%s页）' % '、'.join(map(str, missing)) if missing else '',
+                '：' + detail[-1] if detail else '',
+            )
+        )
+    return png_dir
+
+
+class VisionContextError(RuntimeError):
+    pass
+
+
+def has_pillow():
+    try:
+        from PIL import Image  # noqa: F401
+    except Exception:
+        return False
+    return True
+
+
+def asset_vision_group_sheet(outdir, group, png_dir, path):
+    """把一组整页语境和候选图做成可索引拼版；每张候选保持独立卡片。"""
+    try:
+        from PIL import Image, ImageDraw
+    except Exception:
+        return None
+
+    candidates = group['candidates']
+    page_count = len([page for page in group['pages'] if page > 0])
+    page_w, page_h = 440, 248
+    cell, pad, label_h = 220, 16, 42
+    asset_cols = 3
+    asset_rows = max(1, (len(candidates) + asset_cols - 1) // asset_cols)
+    page_rows = max(1, (page_count + 1) // 2) if page_count else 0
+    page_cols = min(2, page_count) if page_count else 0
+    width = max(2 * (page_w + pad) + pad if page_cols else 0,
+                asset_cols * (cell + pad) + pad)
+    header_h = 28
+    page_area_h = (page_rows * (page_h + label_h + pad) + pad) if page_rows else 0
+    asset_top = header_h + page_area_h
+    height = asset_top + asset_rows * (cell + label_h + pad) + pad
+    sheet = Image.new('RGB', (width, height), (245, 245, 247))
+    draw = ImageDraw.Draw(sheet)
+    draw.text((pad, 7), '%s  inputs=%d  pages=%s' % (
+        group['id'], group['input_count'],
+        ','.join(map(str, group['pages'])) or 'layout'), fill=(20, 20, 24))
+
+    for index, page in enumerate([page for page in group['pages'] if page > 0]):
+        x = pad + (index % 2) * (page_w + pad)
+        y = header_h + (index // 2) * (page_h + label_h + pad)
+        source = (os.path.join(png_dir, 'slide-%d.png' % page)
+                  if png_dir else None)
+        if source and os.path.exists(source):
+            try:
+                image = Image.open(source).convert('RGB')
+                image = _fit_image(image, page_w, page_h)
+                sheet.paste(image, (x + (page_w - image.width) // 2,
+                                    y + (page_h - image.height) // 2))
+            except Exception as exc:
+                raise VisionContextError('视觉判断所需页面截图不可读取：第%d页' % page) from exc
+        else:
+            raise VisionContextError('视觉判断所需页面截图缺失：第%d页' % page)
+        draw.rectangle([x, y, x + page_w, y + page_h], outline=(120, 120, 128))
+        draw.text((x + 2, y + page_h + 5), '[page %d] context for candidates below' % page,
+                  fill=(20, 20, 24))
+
+    for index, candidate in enumerate(candidates):
+        x = pad + (index % asset_cols) * (cell + pad)
+        y = asset_top + (index // asset_cols) * (cell + label_h + pad)
+        image_path = os.path.join(outdir, candidate['out'])
+        try:
+            image = Image.open(image_path)
+            if _candidate_is_visual_risk(candidate):
+                half = (cell - 3) // 2
+                _paste_candidate_preview(sheet, draw, image, (x, y, half, cell))
+                _paste_candidate_preview(sheet, draw, image, (x + half + 3, y, cell - half - 3, cell),
+                                         dark=True)
+            else:
+                _paste_candidate_preview(sheet, draw, image, (x, y, cell, cell))
+        except Exception:
+            draw.text((x + 8, y + 8), 'unreadable', fill=(200, 0, 0))
+        draw.rectangle([x, y, x + cell, y + cell], outline=(120, 120, 128))
+        probe = candidate.get('probe') or {}
+        alpha = candidate.get('effective_alpha_mean')
+        if alpha is None:
+            alpha = probe.get('alpha_mean')
+        risk = (' a=%s w=%s' % (
+            '?' if alpha is None else round(alpha),
+            '?' if probe.get('near_white_ratio') is None
+            else round(probe['near_white_ratio'] * 100),
+        )) if _candidate_is_visual_risk(candidate) else ''
+        draw.text((x + 2, y + cell + 3), '[%s] %s %dx%d%s' % (
+            candidate['id'], candidate['file'], probe.get('w') or 0, probe.get('h') or 0, risk),
+            fill=(20, 20, 24))
+        first = (candidate.get('placements') or [{}])[0]
+        total_placements = len(candidate.get('source_placements') or
+                               candidate.get('placements') or [])
+        if first.get('slide'):
+            box = first['box']
+            draw.text((x + 2, y + cell + 18), 's%d @%d,%d %dx%d seen=%d' % (
+                first['slide'], box[0], box[1], box[2], box[3],
+                total_placements), fill=(20, 20, 24))
+        else:
+            draw.text((x + 2, y + cell + 18), 'layout x%d' % len(candidate.get('placements') or []),
+                      fill=(20, 20, 24))
+    _save_visual_sheet(sheet, path)
+    return path
+
+
+def emit_asset_vision_groups(outdir, candidates, slide_count, ldir):
+    """生成受预算约束的拼版和结构化索引，返回已选/未选组。"""
+    groups = build_asset_vision_groups(candidates)
+    selected, omitted = select_asset_vision_groups(groups, slide_count)
+    _safe_remove(os.path.join(ldir, 'vision-group-*.jpg'))
+    _safe_remove(os.path.join(ldir, 'contact-sheet-*.png'))
+    _safe_remove(os.path.join(ldir, 'contact-sheet.png'))
+    _safe_remove(os.path.join(ldir, 'asset-context-sheet-*.png'))
+
+    paths = []
+    if not has_pillow():
+        omitted = groups
+        selected = []
+    else:
+        pages = sorted({page for group in selected for page in group['pages'] if page > 0})
+        png_dir = render_asset_vision_pages(outdir, pages)
+        try:
+            for index, group in enumerate(selected, 1):
+                path = os.path.join(ldir, 'vision-group-%d.jpg' % index)
+                if asset_vision_group_sheet(outdir, group, png_dir, path):
+                    paths.append(path)
+                    group['sheet'] = os.path.basename(path)
+            if selected and len(paths) != len(selected):
+                raise RuntimeError('视觉判断拼版生成失败')
+        except VisionContextError:
+            raise
+        except Exception:
+            _safe_remove(os.path.join(ldir, 'vision-group-*.jpg'))
+            selected, omitted, paths = [], groups, []
+    if paths:
+        # 旧流程只认 contact-sheet.png；保留首个视觉组的 PNG 别名，新的 BRIEF 不再要求读它。
+        try:
+            from PIL import Image
+            legacy = os.path.join(ldir, 'contact-sheet-1.png')
+            Image.open(paths[0]).convert('RGB').save(legacy, 'PNG', optimize=True)
+            shutil.copy2(legacy, os.path.join(ldir, 'contact-sheet.png'))
+        except Exception:
+            pass
+
+    def serialize(group):
+        return {
+            'id': group['id'],
+            'sheet': group.get('sheet'),
+            'pages': group['pages'],
+            'input_count': group['input_count'],
+            'candidates': [{
+                'id': candidate['id'],
+                'source_media': candidate['file'],
+                'source_px': [candidate['probe'].get('w'), candidate['probe'].get('h')],
+                'bytes': candidate.get('bytes'),
+                'repeat_count': candidate.get('n'),
+                'fullscreen': candidate.get('fullscreen'),
+                'effective_alpha_mean': candidate.get('effective_alpha_mean'),
+                'near_white_ratio': candidate['probe'].get('near_white_ratio'),
+                'placements': candidate.get('placements') or [],
+                'source_placements': candidate.get('source_placements') or
+                                     candidate.get('placements') or [],
+            } for candidate in group['candidates']],
+        }
+
+    index = {
+        'version': 2,
+        'limits': {
+            'single_page_image_budget': SINGLE_PAGE_IMAGE_BUDGET,
+            'multi_page_image_budget': MULTI_PAGE_IMAGE_BUDGET,
+            'pack_cap': VISUAL_PACK_CAP,
+            'input_cap': VISUAL_INPUT_CAP,
+        },
+        'selected': [serialize(group) for group in selected],
+        'omitted': [serialize(group) for group in omitted],
+    }
+    with open(os.path.join(ldir, 'asset-vision-groups.json'), 'w', encoding='utf-8') as stream:
+        json.dump(index, stream, ensure_ascii=False, indent=2)
+        stream.write('\n')
+    return selected, omitted, paths
+
+
+def asset_context_sheets(outdir, cands, ldir):
+    """按候选主所在页去重拼整页语境，供模型识别 logo 墙和装饰用途。"""
+    reviewed = [c for c in cands if needs_asset_judgment(c)]
+    pages = []
+    seen = set()
+    for c in reviewed:
+        page = next((no for no in c.get('slides') or [] if no and no != 9999), None)
+        if page is not None and page not in seen:
+            seen.add(page)
+            pages.append(page)
+    if not pages:
+        return []
+    import subprocess
+    result = subprocess.run(
+        [sys.executable, os.path.join(HERE, 'render_pages.py'), outdir,
+         '--pages', 'slides', '--only', ','.join(map(str, pages)), '--no-html'],
+        capture_output=True, text=True,
+    )
+    png_dir = os.path.join(outdir, 'ref', 'rebuild', 'png')
+    if result.returncode or not os.path.isdir(png_dir):
+        return []
+    try:
+        from PIL import Image, ImageDraw
+    except Exception:
+        return []
+    paths = []
+    candidate_ids = defaultdict(list)
+    for index, c in enumerate(cands, 1):
+        if not needs_asset_judgment(c):
+            continue
+        for page in c.get('slides') or []:
+            if page in seen:
+                candidate_ids[page].append(index)
+    for start in range(0, len(pages), CONTEXT_BATCH):
+        batch = pages[start:start + CONTEXT_BATCH]
+        cols, cw, ch, pad, lab = 2, 480, 270, 16, 22
+        rows = (len(batch) + cols - 1) // cols
+        sheet = Image.new('RGB', (cols * (cw + pad) + pad,
+                                  rows * (ch + pad + lab) + pad), (245, 245, 247))
+        draw = ImageDraw.Draw(sheet)
+        for offset, page in enumerate(batch):
+            x = pad + (offset % cols) * (cw + pad)
+            y = pad + (offset // cols) * (ch + pad + lab)
+            source = os.path.join(png_dir, 'slide-%d.png' % page)
+            if os.path.exists(source):
+                image = Image.open(source).convert('RGB')
+                image.thumbnail((cw, ch))
+                sheet.paste(image, (x, y))
+            draw.rectangle([x, y, x + cw, y + ch], outline=(120, 120, 128))
+            draw.text((x + 2, y + ch + 5), 'slide %d  candidates=%s'
+                      % (page, ','.join(map(str, candidate_ids[page]))),
+                      fill=(20, 20, 24))
+        path = os.path.join(ldir, 'asset-context-sheet-%d.png'
+                            % (start // CONTEXT_BATCH + 1))
+        sheet.save(path, optimize=True)
+        paths.append(path)
+    return paths
+
+
 # ---------------------------------------------------------------- 落盘
 def write(p, s):
     with open(p, 'w', encoding='utf-8') as f:
         f.write(s)
 
 
-def emit_manifest(d, assets, ldir):
+def bound_visual_candidates(candidates, archetypes):
+    """只把最终有图片槽的候选交给判断单；其余仍留在联系表供视觉核对。"""
+    bound_files = {
+        s.get('source_media')
+        for a in archetypes
+        for s in a.get('slots') or []
+        if s.get('source_media')
+    }
+    return [c for c in candidates if c.get('file') in bound_files]
+
+
+def visual_slot_candidates(candidates, archetypes):
+    """把候选绑定到最终槽位；页面截图只作该槽位的视觉语境。"""
+    source_candidates = {
+        candidate.get('file'): candidate
+        for candidate in bound_visual_candidates(candidates, archetypes)
+    }
+    rows, seen = [], set()
+    for archetype in archetypes:
+        pages = set(archetype.get('pages') or archetype.get('_sample_pages') or [])
+        for slot in archetype.get('slots') or []:
+            source = slot.get('source_media')
+            raw_box = slot.get('box')
+            candidate = source_candidates.get(source)
+            if not pages or not candidate or not raw_box or not needs_asset_judgment(candidate):
+                continue
+            box = [round(value) for value in raw_box]
+            key = (source, archetype['name'], tuple(box))
+            if key in seen:
+                continue
+            seen.add(key)
+            source_placements = candidate.get('placements') or []
+            matching = [
+                placement for placement in source_placements
+                if tuple(placement.get('box') or ()) == tuple(box)
+                and placement.get('slide') in pages
+            ]
+            if not matching:
+                continue
+            slide = matching[0]['slide']
+            row = dict(candidate)
+            row['placements'] = [{
+                'slide': slide,
+                'box': box,
+                'archetype': archetype['name'],
+            }]
+            row['slides'] = [slide] if slide else []
+            row['source_placements'] = source_placements
+            rows.append(row)
+    return rows
+
+
+def emit_manifest(d, assets, vision_groups, ldir):
     L = ['version: alpha',
          'name: TODO-style-name        # 英文 kebab，体现气质，不要用文件名',
          'name_zh: TODO中文名',
@@ -646,6 +2359,28 @@ def emit_manifest(d, assets, ldir):
                 L.append('    on-bg: %s' % (a.get('on_bg') or 'light'))
             if a['use_full']:
                 L.append('    use_full: true')
+    if vision_groups:
+        L += [
+            'asset_vision_groups:',
+            '  # 每项对应拼版中的一个候选实例；同源图在不同页型/位置可分别定性。',
+            '  # 取值与 FaaS 对齐：logo|slogan|background|texture|icon|decorative|illustration|photo|chart|screenshot|footer-copyright|page-number|watermark|content-image|unknown。',
+        ]
+        for group in vision_groups:
+            for candidate in group['candidates']:
+                placement = (candidate.get('placements') or [{}])[0]
+                L.append('  - id: %s' % candidate['id'])
+                L.append('    source_media: %s' % q(candidate['file']))
+                if placement.get('box'):
+                    L.append('    box: %s' % placement['box'])
+                L.append('    visual_kind: TODO-visual-kind-%s   # %s；视觉组 %s'
+                         % (candidate['id'], candidate['id'], group['id']))
+        L += [
+            'asset_decisions:',
+            '  # 仅在不在视觉预算内的图片、或需要覆盖已有判断时追加。',
+            '  # 位置例外写 box；同图同坐标跨页型不同，再补 layout。',
+            '  # - {source_media: example.png, visual_kind: chart}',
+            '  # - {source_media: example.png, layout: layout-2, box: [0, 0, 100, 100], visual_kind: decorative}',
+        ]
     write(os.path.join(ldir, 'manifest.yaml'), '\n'.join(L) + '\n')
 
 
@@ -669,20 +2404,32 @@ def emit_frontmatter(d, tokens, fonts, roles, anchors, gaps, ldir):
     edge = {}
     for p in pads:
         edge.setdefault(p['edge'], p['px'])
-    if edge:
+    # 四边都测出来才写 spacing / safe-area。缺一边就整段不写，并在 gaps 说明——
+    # 拿另一套模板的边距当默认值，会让消费端按一个从没在本模板出现过的网格排版。
+    edges_full = all(edge.get(k) is not None for k in ('top', 'right', 'bottom', 'left'))
+    if edges_full:
         L.append('spacing:')
         L.append('  page-padding: {top: %s, right: %s, bottom: %s, left: %s}'
-                 % (edge.get('top', 73), edge.get('right', 90),
-                    edge.get('bottom', 95), edge.get('left', 90)))
-    radii = [r for r in (d.get('radii_census') or []) if r['px'] >= 2 and r['n'] >= 6]
-    if radii:
-        top = max(radii, key=lambda r: r['n'])
+                 % (edge['top'], edge['right'], edge['bottom'], edge['left']))
+    # rounded.card 是全局 token，只能表达全档共同的一档圆角。多个非零档位或零/非零
+    # 混用时，圆角属于 layouts.md 里的局部形状事实，压成一个值会把直角容器也圆角化。
+    radii = d.get('radii_census') or []
+    if len(radii) == 1 and radii[0]['px'] >= 1:
+        top = radii[0]
         L.append('rounded:')
         L.append('  card: %dpx' % round(top['px']))
-    L.append('safe-area:')
-    L.append('  content: {top: %s, right: %s, bottom: %s, left: %s, applies-to: [content]}'
-             % (edge.get('top', 73), edge.get('right', 90), edge.get('bottom', 95), edge.get('left', 90)))
-    L.append('  confidence: medium')
+    if edges_full:
+        L.append('safe-area:')
+        L.append('  content: {top: %s, right: %s, bottom: %s, left: %s, applies-to: [content]}'
+                 % (edge['top'], edge['right'], edge['bottom'], edge['left']))
+        L.append('  confidence: medium')
+    else:
+        gaps = list(gaps) + ['本模板没测出四边都稳定的页边距（普查到 %s），'
+                             '因此不给 spacing / safe-area：按各页型 slot 的实际坐标排版，'
+                             '不要自造统一边距。'
+                             % ('、'.join('%s=%s' % (k, edge[k]) for k in
+                                          ('top', 'right', 'bottom', 'left') if edge.get(k) is not None)
+                                or '一边都没有')]
     L.append('anchors:')
     for aid, typ, desc in anchors:
         L.append('  - {id: %s, type: %s, desc: "%s"}' % (aid, typ, desc))
@@ -692,131 +2439,638 @@ def emit_frontmatter(d, tokens, fonts, roles, anchors, gaps, ldir):
     write(os.path.join(ldir, 'frontmatter.yaml'), '\n'.join(L) + '\n')
 
 
-def emit_layouts(archetypes, ldir):
-    L = ['# 只改 names 这一段：给每个页型起表意的中文名（看 layout-sheet.png）。下面 layouts 段不要动。',
-         'names:']
+def draft_flow(a, facts, canvas):
+    """从结构事实推出「区带」草案：一页 = 若干竖直区带，高度由内容决定。
+
+    绝对坐标只能表达「模板样张那份内容摆在哪」。真实内容长度不同，上面的区带一变高，
+    下面的就该整体下移——这件事在一张坐标表里表达不出来，只能靠消费端自己算，而它
+    算错的方向有两个：估小了压穿下一块，估大了留一片空。
+
+    这里只出草案，最终用绝对还是流式由看得到重建图的人定。
+    """
+    cW, cH = canvas
+    # 装饰件也算进来：很多模板的版式层只有几个占位符，真正撑起版面的是卡片容器
+    # （在 decor 里）。只看 slots 会把一页的主体结构整个漏掉。
+    slots = [s for s in a['slots'] if s.get('box')]
+    fixed_roles = {'logo', 'slide-number', 'page-number', 'header', 'footer'}
+    fixed = [s for s in slots if s.get('role') in fixed_roles]
+    content_slots = [s for s in slots if s.get('role') not in fixed_roles]
+    containers = [{'role': 'container', 'type': 'decor', 'box': dcr['box'],
+                   'css': dcr.get('css')} for dcr in (a.get('decor') or [])]
+    items = group_flow_cards(content_slots, containers)
+    if len(items) < 2:
+        return None
+    items.sort(key=lambda s: (s['box'][1], s['box'][0]))
+    gaps = [items[i + 1]['box'][1] - (items[i]['box'][1] + items[i]['box'][3])
+            for i in range(len(items) - 1)]
+    pos = [g for g in gaps if g > 0]
+    if not pos:
+        return None
+    # 区带边界 = 间距分布里的最大空档。同一区带内部的间距（网格行距之类）总是明显
+    # 小于区带之间的间距，用本页自己的分布切，不设固定阈值。
+    cut = _gap_cut(pos, min(pos), max(pos)) if len(pos) > 1 else pos[0]
+    regions, cur = [], [items[0]]
+    for i, g in enumerate(gaps):
+        if g >= cut:
+            regions.append(cur)
+            cur = []
+        cur.append(items[i + 1])
+    regions.append(cur)
+
+    # 整页左右边距 = 所有内容的横向外包络，作为各区带的缺省。
+    lefts = [s['box'][0] for s in items]
+    rights = [s['box'][0] + s['box'][2] for s in items]
+    page_margin = [min(lefts), cW - max(rights)]
+
+    out = []
+    for reg in regions:
+        if not reg:
+            continue
+        # 同一区带里 y 接近的算一行；每行元素数一致且 >1 就是网格
+        rows, cr = [], [reg[0]]
+        for s in reg[1:]:
+            if abs(s['box'][1] - cr[-1]['box'][1]) <= max(s['box'][3], 1) * 0.5:
+                cr.append(s)
+            else:
+                rows.append(cr)
+                cr = [s]
+        rows.append(cr)
+        widths = {len(r) for r in rows}
+        if len(rows) >= 1 and widths == {len(rows[0])} and len(rows[0]) > 1:
+            cols = len(rows[0])
+            xs = sorted(s['box'][0] for s in rows[0])
+            col_gap = round((xs[1] - xs[0]) - rows[0][0]['box'][2]) if cols > 1 else 0
+            row_gap = 0
+            if len(rows) > 1:
+                row_gap = round(rows[1][0]['box'][1]
+                                - (rows[0][0]['box'][1] + rows[0][0]['box'][3]))
+            region = {'kind': 'grid', 'cols': cols, 'gap': [max(col_gap, 0), max(row_gap, 0)],
+                      'items': rows[0]}
+            # 卡片组的横向范围常和整页不同（标题贴左、卡片居中）。整页边距是所有元素的
+            # 外包络，直接套给居中卡片组会把它拉偏成左对齐。区带范围和整页明显不一致时，
+            # 落这个区带自己的左右边距，消费端把网格放进它再填 1fr。按落盘的整数比较，
+            # 亚像素噪声不触发多余的区带边距。
+            reg_margin = [min(s['box'][0] for s in rows[0]),
+                          cW - max(s['box'][0] + s['box'][2] for s in rows[0])]
+            if [int(reg_margin[0]), int(reg_margin[1])] != [int(page_margin[0]), int(page_margin[1])]:
+                region['margin'] = reg_margin
+            out.append(region)
+        elif len(rows) == len(reg):
+            # 每行一个元素 = 真的竖着排
+            inner = 0
+            if len(reg) > 1:
+                inner = round(reg[1]['box'][1] - (reg[0]['box'][1] + reg[0]['box'][3]))
+            out.append({'kind': 'stack', 'gap': max(inner, 0), 'items': reg})
+        else:
+            # 每行元素数不一致（比如左列两张、右列一张跨两行）。硬说成 stack 会让消费端
+            # 以为它们是竖排的，比不给还糟。如实说这块推不出规整结构，按坐标摆。
+            out.append({'kind': 'free', 'items': reg})
+    if fixed:
+        out.append({'kind': 'free', 'items': fixed})
+    if len(out) < 2:
+        return None
+    return {'top': items[0]['box'][1], 'margin': page_margin,
+            'gap': round(cut), 'regions': out}
+
+
+def box_contains(outer, inner):
+    return (outer[0] <= inner[0] and outer[1] <= inner[1]
+            and outer[0] + outer[2] >= inner[0] + inner[2]
+            and outer[1] + outer[3] >= inner[1] + inner[3])
+
+
+def boxes_overlap(a, b):
+    return (min(a[0] + a[2], b[0] + b[2]) > max(a[0], b[0])
+            and min(a[1] + a[3], b[1] + b[3]) > max(a[1], b[1]))
+
+
+def overlap_ratio(outer, inner):
+    width = min(outer[0] + outer[2], inner[0] + inner[2]) - max(outer[0], inner[0])
+    height = min(outer[1] + outer[3], inner[1] + inner[3]) - max(outer[1], inner[1])
+    if width <= 0 or height <= 0 or inner[2] <= 0 or inner[3] <= 0:
+        return 0
+    return width * height / (inner[2] * inner[3])
+
+
+def group_flow_cards(slots, containers):
+    """把并列卡片容器及其文字组成一层 group，避免拍平成多列元素。"""
+    candidates = []
+    for container in containers:
+        children = [slot for slot in slots if box_contains(container['box'], slot['box'])]
+        if len(children) >= 2:
+            candidates.append((container, children))
+    selected = []
+    for container, children in sorted(
+            candidates, key=lambda pair: pair[0]['box'][2] * pair[0]['box'][3]):
+        if not any(boxes_overlap(container['box'], other['box']) for other, _ in selected):
+            selected.append((container, children))
+    if len(selected) < 2:
+        return slots + containers
+
+    grouped_slots = {id(slot) for _, children in selected for slot in children}
+    nested_by_container = {}
+    for container, _ in selected:
+        nested_by_container[id(container)] = [
+            other for other in containers
+            if other is not container and overlap_ratio(container['box'], other['box']) >= 0.9
+        ]
+    grouped_containers = {
+        id(container)
+        for container, _ in selected
+        for container in [container] + nested_by_container[id(container)]
+    }
+    out = [slot for slot in slots if id(slot) not in grouped_slots]
+    out += [container for container in containers if id(container) not in grouped_containers]
+    for container, children in selected:
+        children = children + nested_by_container[id(container)]
+        children = sorted(children, key=lambda slot: (slot['box'][1], slot['box'][0]))
+        gaps = [children[i + 1]['box'][1]
+                - (children[i]['box'][1] + children[i]['box'][3])
+                for i in range(len(children) - 1)]
+        outer = container['box']
+        insets = [
+            min(child['box'][1] - outer[1] for child in children),
+            min(outer[0] + outer[2] - child['box'][0] - child['box'][2] for child in children),
+            min(outer[1] + outer[3] - child['box'][1] - child['box'][3] for child in children),
+            min(child['box'][0] - outer[0] for child in children),
+        ]
+        padding = max(0, round(min(insets)))
+        css = container.get('css') or ''
+        if padding:
+            css = '; '.join(part for part in (
+                css.rstrip('; '), 'box-sizing: border-box', 'padding: %dpx' % padding) if part)
+        out.append({
+            'role': 'group',
+            'type': 'group',
+            'box': outer,
+            'css': css,
+            'gap': max(0, round(min(gaps))) if gaps else 0,
+            'items': children,
+        })
+    return out
+
+
+def structure_facts(archetypes, d, shapes):
+    """每个页型的**结构事实**：栅格、垂直间距序列、容器样式配方、样张里的实际字数。
+
+    这些是判「该用绝对坐标还是流式」的依据，脚本只测不判：
+    - 栅格拟合好不好，决定这页是不是一个规整的多列区带
+    - 垂直间距序列里的突变点，就是区带的边界（网格内部 24、区带之间 110）
+    - 样张字数说明这个框是按几行内容设计的——框高本身看不出这件事
+    """
+    q = _load_query()
+    by_part = defaultdict(list)
+    for s in shapes:
+        by_part[s.get('part')].append(s)
+
+    # 容器样式配方：跨全档聚类一次，记出现次数与跨页数，供判断「哪些是共性风格」
+    groups = {}
+    for s in shapes:
+        fill, line, fx = s.get('fill'), s.get('line'), s.get('effects')
+        if not fill and not line and not fx:
+            continue
+        if isinstance(fill, dict) and fill.get('type') == 'image':
+            continue
+        k = q._sig(fill, line, fx)
+        if k[0] == 'none' and k[1] == 'none' and not k[2]:
+            continue
+        g = groups.setdefault(k, {'n': 0, 'parts': set(), 'radii': [],
+                                  'fill': fill, 'line': line, 'fx': fx, 'shapes': set()})
+        g['n'] += 1
+        g['parts'].add(s.get('part'))
+        g['radii'].append(s.get('radius_px') or 0)
+        g['shapes'].add(id(s))
+    ranked = sorted(groups.values(), key=lambda g: -g['n'])
+    recipe_id = {}
+    recipes = []
+    for i, g in enumerate(ranked, 1):
+        rid = 'r%d' % i
+        css = [re.sub(r'\s*\n\s*', ' ', c.split('\x00')[0]).strip()
+               for c in q._recipe_css(g['fill'], g['line'], g['radii'], g['fx']) if c]
+        recipes.append({'id': rid, 'n': g['n'], 'pages': len(g['parts']),
+                        'css': '; '.join(css)})
+        for sid in g['shapes']:
+            recipe_id[sid] = rid
+
+    grids = (d.get('spacing_candidates') or {}).get('grids') or []
+    grid_by_part = defaultdict(list)
+    for gd in grids:
+        grid_by_part[gd.get('part')].append(gd)
+
+    out = {}
+    for a in archetypes:
+        part = None
+        if a.get('source', '').startswith('layout:'):
+            part = 'ppt/slideLayouts/' + a['source'].split(':', 1)[1]
+        elif a.get('rep'):
+            part = 'ppt/slides/slide%d.xml' % a['rep']
+        boxes = [s['box'] for s in a['slots']] + [x['box'] for x in (a.get('decor') or [])]
+        boxes.sort(key=lambda b: b[1])
+        gaps = [boxes[i + 1][1] - (boxes[i][1] + boxes[i][3]) for i in range(len(boxes) - 1)]
+        chars = [(s['box'], len(s.get('txt') or '')) for s in a['slots'] if s.get('txt')]
+        used = []
+        for s in by_part.get(part, []):
+            rid = recipe_id.get(id(s))
+            if rid and rid not in used:
+                used.append(rid)
+        out[a['name']] = {'grids': grid_by_part.get(part) or [], 'gaps': gaps,
+                          'chars': chars, 'recipes': used}
+    return out, recipes
+
+
+LAYOUT_CONTROL_KEYS = {
+    'names', 'roles', 'text_roles', 'layout_modes', 'bg_rules',
+}
+
+
+def emit_layout_controls(layout_lines, ldir, text_role_candidates, flow_archetypes):
+    """从兼容用的 layouts.yaml 分出模型只需编辑的判断区。
+
+    旧判断单把控制项和每个 slot 的坐标正文混在一起。模型为补一个角色读取整份文件，
+    在版式很多的模板上会把时间耗在无须判断的数值上。仍保留旧文件给既有调用方；
+    新文件只承载最终可覆盖它的五个顶层判断区。
+    """
+    blocks, current = {}, None
+    for line in layout_lines:
+        match = re.match(r'^([A-Za-z_][\w-]*):', line)
+        if match:
+            current = match.group(1)
+            if current in LAYOUT_CONTROL_KEYS:
+                blocks[current] = [line]
+            continue
+        if current in blocks:
+            blocks[current].append(line)
+
+    controls = [
+        '# 版式判断控制区 —— 只读并编辑本文件；不要打开或修改 layouts.yaml。',
+        '# package.py 会用本文件覆盖 layouts.yaml 的同名判断区，后者仅保留坐标事实与兼容输入。',
+        '# 可编辑顶层键仅为 names / roles / text_roles / layout_modes / bg_rules。',
+    ]
+    for key in ('names', 'roles'):
+        if key in blocks:
+            controls.extend([''] + blocks[key])
+    if text_role_candidates:
+        controls += [
+            '',
+            'text_roles:',
+            '# 默认文字槽都是 body；仅把确认属于 title|subtitle|header|footer 的例外填为',
+            '#   <id>: title（不要给普通正文补 body）。候选对应的原始槽位在下列注释中。',
+        ]
+        for role_id, slot in text_role_candidates:
+            controls.append('# %s：%s' % (
+                role_id, (slot.get('txt') or '（无样本文字）')[:60]))
+    if flow_archetypes:
+        controls += [
+            '',
+            'layout_modes:',
+            '# 默认 slots。只有样张明确需要内容随高度重排时，取消注释并填 `<页型>: flow`。',
+        ]
+        controls.extend('# %s: flow' % archetype['name'] for archetype in flow_archetypes)
+    if 'bg_rules' in blocks:
+        controls.extend([''] + blocks['bg_rules'])
+    write(os.path.join(ldir, 'layout-controls.yaml'), '\n'.join(controls) + '\n')
+
+
+def emit_layouts(archetypes, ldir, busy_hints=None, facts=None, recipes=None):
+    sampled_archetypes = {
+        id(archetype) for archetype in archetypes
+        if (archetype.get('rep') is not None
+            or archetype.get('pages')
+            or archetype.get('_sample_pages'))
+    }
+    prefilled = sum(1 for a in archetypes if a.get('zh'))
+    L = ['# 判断单草案 —— package.py 读它产出 layouts.md，deck 的版式坐标从 layouts.md 读。',
+         '# 只改 names / roles / text_roles / layout_modes / bg_rules 五段（都是扁平键值，'
+         '改完 package.py 自动并回各页型）。',
+         '# 下面 layouts 段是普查数值，一个字都不要动——改它容易连带删掉 slots/confidence。']
+    if prefilled:
+        L.append('# names 已按模板自带的版式名填好 %d 条，读一遍确认表意即可，通常不用改。' % prefilled)
+    if recipes:
+        L.append('# 容器样式配方（按出现次数排；跨页数多 = 共性风格，只在一处出现的多半不是）：')
+        for r in recipes[:8]:
+            L.append('#   %s  出现 %d 次 / 跨 %d 处   %s' % (r['id'], r['n'], r['pages'], r['css']))
+    L.append('names:')
+    for a in archetypes:
+        if a.get('zh'):
+            # 模板自己给版式起了名（form=3），直接用——比看图起名准，也省掉一轮判断
+            L.append('  %s: %s' % (a['name'], q(a['zh'])))
+        else:
+            L.append('  %s: TODO中文名（代表页 %s，共 %d 页）'
+                     % (a['name'], a['rep'], len(a['pages'])))
+    # 角色（封面 / 章节页 / 内容页……）是看图才能下的结论，脚本不猜。模板自己按页型
+    # 命名时用它的标注，否则连同客观事实一起摆出来，由看得到重建图的你来定。
+    need_role = [a for a in archetypes if not a.get('role')]
+    if need_role:
+        L.append('roles:   # 取值 cover|section|content|quote|closing|blank|custom')
+        for a in need_role:
+            szs = sorted({round(s['sz']) for s in a['slots'] if s.get('sz')}, reverse=True)
+            L.append('  %s: TODO角色   # 代表页 %s，共 %d 页；文字块 %d 个，字号 %s；'
+                     '图片 %d 张%s%s'
+                     % (a['name'], a['rep'], len(a['pages']),
+                        len([s for s in a['slots'] if not s.get('asset')]),
+                        '/'.join(str(x) for x in szs[:5]) or '未声明',
+                        a.get('pic_n') or 0, '；有满屏底图' if a.get('bg_raw') else '',
+                        '；末页候选，结合样张判断 closing 或实际角色'
+                        if a.get('_last_page_candidate') else ''))
+    # 普通正文先保持 body：它是安全且可消费的默认值。标题/页眉/页脚的少量例外依然
+    # 要由模型看样张后写入 text_roles；把每一个正文槽都做成 TODO 会迫使模型逐行复述
+    # 近百个显然的 body，挤占真正的视觉判断时间。
+    text_role_ids = {}
+    for a in archetypes:
+        index = 0
+        for slot in a.get('slots') or []:
+            if not slot.get('_needs_role'):
+                continue
+            index += 1
+            text_role_ids[id(slot)] = '%s-text-%d' % (a['name'], index)
+    if text_role_ids:
+        L.append('# 文字槽默认均为 body。看样张后，只把确实属于 title|subtitle|header|footer 的'
+                 '例外追加到 text_roles:；不要为普通正文逐条补 body。')
+    flow_archetypes = [a for a in archetypes if a.get('flow')]
+    if flow_archetypes:
+        L.append('# 同时有 flow 与 slots 时默认保留 slots，保证固定构图可消费。'
+                 '只有样张明确需要内容随高度重排时，才在 layout_modes: 中写 <页型>: flow。')
+    # 禁放区是**背景图**的属性，不是页型的属性——按背景资产分组，页型再多也不涨
+    bgs = []
     for a in archetypes:
-        L.append('  %s: TODO中文名（代表页 %s，共 %d 页）' % (a['name'], a['rep'], len(a['pages'])))
+        if a['bg'] and a['bg'] not in bgs:
+            bgs.append(a['bg'])
+    if bgs:
+        L.append('bg_rules:')
+        for bg in bgs:
+            users = [a['name'] for a in archetypes if a['bg'] == bg]
+            L.append('  %s:   # 用它的页型：%s' % (bg, ', '.join(users)))
+            hint = (busy_hints or {}).get(bg)
+            if hint:
+                L.append('    # 图像局部对比度：中位 %s、九分位 %s；%s'
+                         % (hint['median'], hint['p90'],
+                            ('更花的一片在 %s' % hint['busy']) if hint.get('busy')
+                            else hint.get('why', '')))
+            # text_safe 不是判断题：模板自己已经把文字放在哪儿写死了。取用这张背景的
+            # 所有页型的槽与装饰件的外接并集即可——让人看图猜只会猜得更松，把模板从不
+            # 放字的区域也划进安全区，这个字段就白设了。
+            boxes = [s['box'] for a in archetypes if a['bg'] == bg for s in a['slots']] + \
+                    [dcr['box'] for a in archetypes if a['bg'] == bg for dcr in (a.get('decor') or [])]
+            if boxes:
+                x0 = min(b[0] for b in boxes)
+                y0 = min(b[1] for b in boxes)
+                x1 = max(b[0] + b[2] for b in boxes)
+                y1 = max(b[1] + b[3] for b in boxes)
+                L.append('    text_safe: [%d, %d, %d, %d]   # 由该背景各页型的槽位并集算出'
+                         % (x0, y0, x1 - x0, y1 - y0))
+            elif any(id(archetype) in sampled_archetypes
+                     for archetype in archetypes if archetype['bg'] == bg):
+                L.append('    text_safe: TODO安全文字区[x,y,w,h]（该背景下没有任何槽位可依据）')
+            else:
+                L.append('    text_safe: [0, 0, 0, 0]   # 未见对应样张，没有可依据的文字区')
+            if any(id(archetype) in sampled_archetypes
+                   for archetype in archetypes if archetype['bg'] == bg):
+                L.append('    avoid: TODO禁放区列表；无禁放区写 []，有则写 [{box: [x,y,w,h], reason: "..."}]')
+                L.append('    pairing_rule: "TODO这张背景上标题/正文/图表要避让哪些区域"')
+            else:
+                L.append('    avoid: []   # 未见对应样张，不额外推断禁放区')
+                L.append('    pairing_rule: "未见对应样张；沿用该页型已有槽位"')
     L.append('layouts:')
     for a in archetypes:
+        fx = (facts or {}).get(a['name']) or {}
+        if fx:
+            # 结构事实：判「这页该用绝对坐标还是流式」的依据。脚本只测不判。
+            for gd in (fx.get('grids') or [])[:2]:
+                c, r = gd.get('cols') or {}, gd.get('rows') or {}
+                L.append('  # 栅格：%s 列%s%s' % (
+                    c.get('n'), ' @%gpx 步距方差 %.2f' % (c.get('pitch') or 0, c.get('sd') or 0)
+                    if c.get('regular') else '（列不规整）',
+                    '，行 %s' % (('%d @%gpx' % (r.get('n') or 0, r.get('pitch') or 0))
+                                if r.get('regular') else '不规整')))
+            if fx.get('gaps'):
+                L.append('  # 垂直间距：%s（突变处即区带边界）'
+                         % '、'.join(str(int(g)) for g in fx['gaps'][:10]))
+            if fx.get('chars'):
+                L.append('  # 样张字数：%s'
+                         % '、'.join('%s=%d字' % (b, n) for b, n in fx['chars'][:6]))
+            if fx.get('recipes'):
+                L.append('  # 命中配方：%s' % '、'.join(fx['recipes'][:4]))
+        # 槽与槽在坐标上重叠：PPT 里占位符互相压是常态（文字 valign 居中、样张只有一行，
+        # 看不出来），照抄坐标做成 HTML 后内容一变长就撞。实测封面 title 框比 subtitle
+        # 的顶还低 41px，两行标题直接压在副标题上。这里只报事实，怎么让开由你定。
+        ov = slot_overlaps(a.get('slots') or [])
+        if ov:
+            L.append('  # 槽位重叠：%s（模板里靠文字居中不显形，内容变长会撞）'
+                     % '、'.join(ov[:3]))
         L.append('  %s:' % a['name'])
-        L.append('    role: %s' % a['name'].split('-')[0])
+        if a.get('role'):
+            L.append('    role: %s' % a['role'])
         if a['bg']:
             L.append('    background: %s' % a['bg'])
-            L.append('    text_safe: TODO安全文字区[x,y,w,h]，按背景主体避让后填写')
-            L.append('    avoid: TODO禁放区列表；无禁放区写 []，有则写 [{box: [x,y,w,h], reason: "..."}]')
-            L.append('    pairing_rule: "TODO说明该页型必须配这张背景时，标题/正文/图表需要避让哪些区域"')
+        fl = a.get('flow')
+        if fl:
+            L.append('    flow:')
+            L.append('      top: %d' % fl['top'])
+            L.append('      margin: [%d, %d]' % tuple(fl['margin']))
+            L.append('      gap: %d' % fl['gap'])
+            L.append('      regions:')
+            for r in fl['regions']:
+                if r['kind'] == 'grid':
+                    L.append('        - kind: grid')
+                    L.append('          cols: %d' % r['cols'])
+                    L.append('          gap: [%d, %d]' % tuple(r['gap']))
+                    if r.get('margin'):
+                        L.append('          margin: [%d, %d]  # 本区带自己的左右边距，'
+                                 '和整页 margin 不同（居中卡片组不跟标题的左边距）'
+                                 % tuple(r['margin']))
+                elif r['kind'] == 'free':
+                    L.append('        - kind: free   # 推不出规整结构，按 slots 的坐标摆')
+                else:
+                    L.append('        - kind: stack')
+                    L.append('          gap: %d' % r['gap'])
+                L.append('          items:')
+                for s in r['items']:
+                    if s.get('type') == 'group':
+                        L.append('            - role: group')
+                        L.append('              gap: %d' % s['gap'])
+                        if s.get('css'):
+                            L.append('              css: "%s"'
+                                     % str(s['css']).replace('"', "'"))
+                        L.append('              items:')
+                        for child in s['items']:
+                            role_id = text_role_ids.get(id(child))
+                            if role_id:
+                                L.append('                # text-role: %s' % role_id)
+                            if child.get('type') == 'decor':
+                                L.append('                - {role: container, css: "%s"}'
+                                         % str(child.get('css') or '').replace('"', "'"))
+                                continue
+                            extra = ''
+                            if child.get('css') is not None:
+                                extra += ', css: "%s"' % str(child['css']).replace('"', "'")
+                            if child.get('asset'):
+                                extra += ', asset: %s' % child['asset']
+                            if child.get('source_media'):
+                                extra += ', source_media: %s' % child['source_media']
+                                extra += ', source_box: %s' % child['box']
+                            L.append('                - {role: %s, type: %s%s}'
+                                     % (child['role'], child['type'], extra))
+                        continue
+                    role_id = text_role_ids.get(id(s))
+                    if role_id:
+                        L.append('            # text-role: %s' % role_id)
+                    # free 区带按坐标摆，而 slots 会被删掉，所以坐标必须写在这里
+                    bx = ', box: %s' % s['box'] if r['kind'] == 'free' else ''
+                    if s.get('type') == 'decor':
+                        L.append('            - {role: container%s, css: "%s"}'
+                                 % (bx, (s.get('css') or '').replace('"', "'")))
+                        continue
+                    extra = bx
+                    if s.get('css') is not None:
+                        # CSS 串一律加引号：里面的逗号/冒号在 flow map 里是分隔符
+                        extra += ', css: "%s"' % str(s['css']).replace('"', "'")
+                    if s.get('asset'):
+                        extra += ', asset: %s' % s['asset']
+                    if s.get('source_media'):
+                        extra += ', source_media: %s' % s['source_media']
+                        if r['kind'] != 'free':
+                            extra += ', source_box: %s' % s['box']
+                    L.append('            - {role: %s, type: %s%s}' % (s['role'], s['type'], extra))
         L.append('    slots:')
         for s in a['slots']:
-            L.append('      - {role: %s, box: %s, type: %s}' % (s['role'], s['box'], s['type']))
+            role_id = text_role_ids.get(id(s))
+            if role_id:
+                L.append('      # text-role: %s' % role_id)
+            extra = ''
+            if s.get('asset'):
+                extra += ', asset: %s' % s['asset']
+            if s.get('source_media'):
+                extra += ', source_media: %s' % s['source_media']
+            if s.get('css') is not None:
+                extra += ', css: "%s"' % str(s['css']).replace('"', "'")
+            L.append('      - {role: %s, box: %s, type: %s%s}'
+                     % (s['role'], s['box'], s['type'], extra))
+        if a.get('decor'):
+            L.append('    decor:')
+            for dcr in a['decor']:
+                L.append('      - {box: %s, geom: %s, css: "%s"}'
+                         % (dcr['box'], dcr['geom'], dcr['css'].replace('"', "'")))
         L.append('    confidence: %s' % a.get('confidence', 'medium'))
     write(os.path.join(ldir, 'layouts.yaml'), '\n'.join(L) + '\n')
-
-
-def emit_body(d, tokens, fonts, roles, assets, archetypes, exceptions, ldir):
+    emit_layout_controls(L, ldir, [
+        (text_role_ids[id(slot)], slot)
+        for archetype in archetypes
+        for slot in archetype.get('slots') or []
+        if id(archetype) in sampled_archetypes and id(slot) in text_role_ids
+    ], [
+        archetype for archetype in flow_archetypes
+        if id(archetype) in sampled_archetypes
+    ])
+
+
+def emit_body(d, tokens, fonts, roles, assets, archetypes, exceptions, cusage, ldir,
+              has_asset_candidates=False):
+    """design.md 正文。
+
+    每条规则只出现一次——同一条散在 Fast Path / Usage / Background Safety /
+    Hard Rules 各写一遍时措辞必然漂移，消费端无法判断哪份权威。
+    坐标、字号、色值、资产位置的权威都在 layouts.md；本文件只给色板、字体栈与纪律。
+    """
     canvas = d['canvas']['px']
     cover = next((a for a in assets if a['id'] == 'bg-cover'), None)
-    logo = next((a for a in assets if a['kind'] == 'logo'), None)
-    content_bgs = [a['id'] for a in assets
-                   if a.get('kind') == 'background' and a.get('role') == 'content']
-    key_assets = []
-    for a in assets[:8]:
-        role = a.get('role') or a.get('kind')
-        key_assets.append('`%s` -> `%s` (%s)' % (a['id'], a.get('out_rel') or a['id'], role))
-
-    L = ['## Agent Fast Path', '',
-         '消费本风格时先读这一节；它是给生成 Agent 的短路径，目标是把风格理解控制在 1 分钟内，避免把审计材料重新理解一遍。',
-         '',
-         '- **时间预算**：风格导入最多做 1 次 `read_file design.md`；如需坐标，最多再读 1 次 `layouts.md`。完成这两步后必须直接开始生成，不要再探索风格包。',
-         '- **只读入口**：常规生成只需要 `design.md`；需要坐标时再打开 `layouts.md`。',
-         '- **附件/zip 兜底**：如果当前内容来自 zip 附件的文本摘要，直接使用摘要中 `design.md` / `layouts.md` 的文本；不要尝试修复 zip、解析二进制、搜索附件目录或重建压缩包。',
-         '- **禁止动作**：不要读取 `ref/color-freq-raw.json`、`ref/font-clusters.json`、`ref/extract.json`、`ref/rebuild/`、`ref/rebuild/png/*`；不要 summarize / view 参考页；不要重新统计颜色、字体或版式。',
-         '- **信息来源优先级**：本节 > `## Usage` > frontmatter `assets` / `colors` / `typography` > `layouts.md`。除此之外的文件只用于人工审计，不用于生成。',
-         '- **缺信息时降级**：如果某个细节本节没有写，用 frontmatter token 和最接近的 `layouts.md` archetype 推断；不要打开审计文件补证。',
-         '- **画布**：所有坐标按 `%dx%d` 绝对像素理解。' % (canvas[0], canvas[1])]
-    if cover:
-        L.append('- **封面背景**：优先使用 `%s`；整幅铺满画布，禁止自造渐变替代。' % cover['id'])
-    if content_bgs:
-        L.append('- **内容页背景**：按 `layouts.md` 中 archetype 的 `background` 字段取；常用内容背景为 %s。'
-                 % '、'.join('`%s`' % x for x in content_bgs[:4]))
-    if cover or content_bgs:
-        L.append('- **背景安全区**：背景图和版式必须配对。按 archetype 的 `background`、`text_safe`、`avoid` 一起放文字和卡片；标题、正文、关键数字、图表、卡片、时间线及其容器的外接矩形都不得压到背景视觉主体、强光斑、深色透明区上，透明容器也不能跨进禁放区。')
-    if logo:
-        L.append('- **标识资产**：只使用 `%s`；不得重画、不得改比例。' % logo['id'])
-    L += ['- **配色与字体**：颜色只取 frontmatter `colors`；字体/字号只取 frontmatter `typography`；内容主题不得引入新色相。',
-          '- **版式**：优先使用下面内联版式索引；需要更多 slot 时才读 `layouts.md`；不要用 `ref/rebuild/png` 反推坐标。']
-    if archetypes:
-        L += ['', '内联版式索引（先用这里，不要为了选页型再读文件）：']
-        for a in archetypes:
-            parts = ['`%s`' % a['name']]
-            if a.get('bg'):
-                parts.append('背景 `%s`' % a['bg'])
-            slots = []
-            for s in a['slots'][:4]:
-                slots.append('%s/%s %s' % (s['role'], s['type'], s['box']))
-            L.append('- %s：%s' % (' / '.join(parts), '；'.join(slots) or '按最接近用途套用'))
-    if key_assets:
-        L += ['', '关键资产：'] + ['- ' + a for a in key_assets]
-
-    L += ['', '## Overview', '',
-         'TODO: 两三句话讲清这套模板的性格与适用场景——看过联系表和页面事实之后再写，不要套话。', '',
-         '画布 %dx%d px，%d 页样张归纳出 %d 种页型；版式坐标全部放在 `layouts.md`。'
-         % (canvas[0], canvas[1], d['counts']['slides'], len(archetypes)), '',
-         '## Usage', '',
-         '搭一页 PPT 就三步：', '',
-         '1. **挑版式** —— 打开 `layouts.md`，按用途选一个 archetype，`slots[].box` 的 `[x, y, w, h]` 是 %dx%d 画布上的绝对像素，直接照搬，不要自己排版；同时遵守该 archetype 的 `text_safe` 和 `avoid`。'
-         % (canvas[0], canvas[1])]
-    if assets:
-        L.append('2. **铺资产** —— 背景和标识只用 `assets/` 里的文件，见下表；不要自造渐变或重画 logo。背景必须和版式配对，文字必须留在安全区内。')
-    else:
-        L.append('2. **铺背景** —— 本包无图片资产，背景用 `colors` 里的底色。')
-    L += ['3. **填色与字** —— 颜色只从 frontmatter 的 `colors` 取，字号只从 `typography` 取。', '']
-    if assets:
-        L += ['{{ASSET_TABLE}}', '']
-        L += ['## Background Safety', '',
-              '真实背景不是纯装饰底。填写本包时必须根据 `contact-sheet.png` 和 `layout-sheet.png` 补充每个 archetype 的 `text_safe`、`avoid`、`pairing_rule`；消费时文字、图表、卡片、表格、时间线、标题容器、正文容器和宽透明容器的外接矩形不得与 `avoid` 区域重叠。内容量超出安全区时换页型或拆页，不要扩大文字区域压住背景主体。', '']
-    L += ['**色板纪律**：整套页面只用 `colors` 里的 %d 个 token，内容主题（咖啡、医疗、金融……）不改变配色；'
-          '需要强调时用 `primary`，不要引入模板外的色相。' % len(tokens), '',
+    imp, webs = import_line(fonts)
+    sidecar = '`layouts.md`'
+
+    L = ['## Overview', '',
+         'TODO: 两三句话讲清这套模板的性格与适用场景——看过联系表和页面重建图之后再写。', '']
+    L.append(('模板自带 %d 种版式，页型、坐标和 CSS 样式都直读自版式层。'
+              % len(archetypes)) if (d.get('form_hint') or {}).get('form') == 3 else
+             ('%d 页样张归纳出 %d 种页型。' % (d['counts']['slides'], len(archetypes))))
+    L += ['', '## Usage', '',
+          '搭一页 PPT 六步，中间四步的数据都在 %s：' % sidecar, '']
+    L += ['1. **定画布** —— 舞台按 `layouts.md` 的 `canvas` 设成 %d×%d，'
+          '别套用默认尺寸：源模板的长宽比不一定是 16:9，套错了整页坐标全偏。'
+          '舞台尺寸改不了时，整体等比缩放 `min(舞台宽/%d, 舞台高/%d)` 后居中留白——'
+          '逐轴拉伸会把圆压成椭圆、把字挤扁。' % (canvas[0], canvas[1], canvas[0], canvas[1]),
+          '2. **挑页型** —— 在 %s 里按用途选一个 archetype（清单见下面 Layouts 段）。'
+          '页数多于页型时，挑最接近的一个原样套用它的 slot：用不到的槽删掉，'
+          '内容比槽多就按同类槽的间距等距加，**坐标一律沿用该页型给的那套，不要自己另起网格**。'
+          % sidecar,
+          '3. **按页型给的形态落元素** —— 页型给 `flow` 就用流式，给 `slots` 就用绝对，'
+          '两者只会出现一个。'
+          '**flow**：整块用一个纵向 flex 容器，`top` 是它的起始 y，`margin` 是整块的左右边距，'
+          '`gap` 是区带之间的间距；`regions` 从上往下依次排，**每个区带的高度由它自己的'
+          '内容决定，不要写死高度**——上面的区带内容变多时，下面的自然被推下去，这正是'
+          '这套表达要解决的事。区带内部：`kind: grid` 用 `grid-template-columns: repeat(cols, 1fr)` '
+          '配 `gap: [行间距, 列间距]`；`kind: stack` 用纵向 flex 配 `gap`；`kind: free` '
+          '按 item 自带的 `box` 绝对定位。区带自带 `margin: [左, 右]` 时用它的、'
+          '覆盖整块的 `margin`（模板里居中的卡片组和贴左的标题横向范围本就不同）；'
+          '没带就用整块的 `margin`。`grid` 在自己这份左右边距里再 `repeat(cols, 1fr)`。'
+          '`grid` 里的 `role: group` 是一张卡片：'
+          'group 的 `css` 用于外层容器，内部 `items` 按顺序纵向排布并使用 group 的 `gap`。'
+          '每个 `role: container` 的项是容器，把它的 `css` 逐项原样写进 style，内容放进去；'
+          '其中没有 `border-radius` 就按 `0`，不得自行补圆角。',
+          '4. **按 slot 落元素（页型给的是 slots 时）** —— 每个 slot 渲染成一个绝对定位元素：`box` 是 '
+          '`[x, y, w, h]`（%dx%d 画布上的绝对像素），机械展开成 `left/top/width/height`；'
+          'slot 的 `css` 是模板排版属性已转译好的声明串，原样写进 style，不要另选字号、'
+          '内边距、颜色或对齐。'
+          '带 `asset` 的 slot 是图片元素（logo、角标），把该资产放在它自己的 `box` 里；'
+          '这个页型没有 `asset` 槽，这一页就不出现该资产。' % (canvas[0], canvas[1]),
+          '5. **铺装饰几何** —— 页型的 `decor` 是这一页的图形骨架（图标托底的圆、'
+          '卡片、分隔线）：每条渲染成一个绝对定位空元素，`box` 给位置，`css` 逐项原样写进 '
+          'style；没有 `border-radius` 就按 `0`。只有 `geom: ellipse` 另加 '
+          '`border-radius: 50%`。它们压在背景之上、slot 之下，'
+          '落在 slot 上的图标正是靠它们托住。',
+          '6. **落实全局设计** —— `design.md` frontmatter 的 `colors`、`typography`、'
+          '`spacing`、`rounded`、`components` 是全局 token；用 CSS variables、类名或内联'
+          '样式承载。局部 slot / decor 的 `css` 优先，不能再解释成另一套视觉系统。'
+          '字体使用 Typography 的完整栈与降级，不在运行时安装字体或依赖。',
+          '7. **保持标题结构** —— 有合适页型可参考时，沿用该页型已有的标题层级与局部 '
+          '`css`；只渲染该页型已有的文字槽，背景中已经可见的固定标题不再创建文本，'
+          '页型没有 `subtitle` 槽就不新增副标题。没有合适参考时，按本包整体视觉组织标题。']
+    if assets or has_asset_candidates:
+        L += ['', '资产文件（背景由页型的 `background` 字段指定，'
+              '图片资产的位置由该页型 `slots` 里带 `asset` 的槽给出）：', '',
+              '{{ASSET_TABLE}}', '',
+              '将包内 `assets/` 复制到项目内相对目录，再引用复制后的路径；最终 HTML 不引用'
+              '抽取工作目录或本机绝对路径。附件只提供 `assetRoot` / `assetPaths` 时，把'
+              '`assetRoot` 当作不透明前缀，只拼接清单中声明的相对路径。']
+    L += ['', '文字与容器的外接矩形落在该页型 `background` 对应的 `text_safe` 内，'
+          '避开 `avoid` 列出的区域（两者都在 %s 的 `backgrounds` 段）。内容装不下时换页型或拆页。'
+          % sidecar, '',
           '## Colors', '', '| token | 值 | 用途 |', '|---|---|---|']
-    USE = {'surface': '页面与卡片主底色', 'surface-alt': '次级底色，分区/强调区块的浅底',
-           'ink': '正文与标题文字色', 'ink-muted': '次级文字色，说明与标签',
-           'primary': '主强调色：图表主序列、关键数字、行动点',
-           'accent': '副强调色，多与 primary 组成渐变',
-           'accent-2': '渐变与图表的第二落点色', 'accent-3': '渐变收尾色，用量最少',
-           'neutral': '中性弱化色：分隔线、次要标签'}
     for name, r in tokens:
-        L.append('| `%s` | `%s` | %s |' % (name, r['hex'], USE.get(name, '按 token 名对应的角色使用')))
-    imp, webs = import_line(fonts)
+        L.append('| `%s` | `%s` | %s |'
+                 % (name, r['hex'], usage_phrase(cusage.get(r['hex'].upper()))))
     L += ['', '## Typography', '']
     for f in fonts[:2]:
         L.append('- **%s** —— 栈 `%s`%s' % (
             f['names'][0], font_css(f['stack']),
             '，源为商业/内部字体无 web 分发源，按气质降级到 %s' % f['stack'][1]
             if len(f['stack']) > 1 else ''))
-    L += ['', '字号轴：' + '、'.join('%s %dpx' % (k, round(v['sz_px'])) for k, v in roles.items()), '',
+    L += ['', '字号轴：' + '、'.join('%s %dpx' % (k, round(v['sz_px'])) for k, v in roles.items())
+          + '。slot 自带 `css` 时以其中的 `font-size` 为准；没有 slot CSS 的新增层级，'
+          '复用轴上最接近的一档。', '',
           '字体加载（**HARD REQUIREMENT：下面这行 @import 原样写入全局样式首行，禁止替换为 '
           'fonts.googleapis.com 或其他域**）：', '', '```', imp, '```', '',
-          '镜像只保证 wght 400 一档，更粗的字重由浏览器合成，不要把字重当唯一区分手段；'
-          '系统字体 PingFang SC / Microsoft YaHei 置于栈末保底，中文场景负字距一律清零。', '',
-          '## Layouts', '',
-          '**搭页前先读 `layouts.md`**——%d 个 archetype 的 slot 坐标都在那里，本文件不重复。'
-          % len(archetypes), '', '{{LAYOUT_LIST}}']
-    L += ['', '## Hard Rules', '']
+          '镜像只保证 wght 400 一档，更粗的字重由浏览器合成，字重不能作为唯一区分手段；'
+          '系统字体 PingFang SC / Microsoft YaHei 置于栈末保底，中文场景负字距清零。', '',
+          '## Layouts', '', '页型清单如下，每个页型的 slots、background、'
+          '禁放区都在 %s：' % sidecar, '', '{{LAYOUT_LIST}}',
+          '', '## Hard Rules', '']
     if cover:
-        L.append('- 封面页背景必须铺 `bg-cover`（文件见 Usage 表），整幅铺满 %dx%d，不要自造渐变或换图。'
-                 % (canvas[0], canvas[1]))
+        L.append('- 封面页铺满 `bg-cover`，整幅覆盖 %dx%d 画布。' % (canvas[0], canvas[1]))
     if any(a['role'] == 'content' for a in assets):
-        L.append('- 内容页背景整幅铺满，用哪一张按 `layouts.md` 里该 archetype 的 `background` 字段取，不要混用。')
-    if logo:
-        b, pages = logo['src']['box'], logo['src']['slides']
-        where = ('只出现在源第 %s 页' % '、'.join(map(str, pages))
-                 if pages and len(pages) <= 4 else '每页放一次')
-        L.append('- `%s` 放在 (%d, %d)，尺寸 %dx%d px，%s；不要重画、不要替换成文字、不要改比例。'
-                 % (logo['id'], b.get('x', 0), b.get('y', 0), b.get('w', 0), b.get('h', 0), where))
-    L += ['- 版式坐标只从 `layouts.md` 的 `slots[].box` 取。',
-          '- 颜色只用 `colors` 里的 token；字号只用 `typography` 里的档位。',
-          '- `@import` 行原样写入，禁止替换域名。',
-          '- TODO: 补 1-2 条这套模板特有的硬规则（看过联系表之后写，例如主色只许用在哪类元素）。',
+        L.append('- 内容页的背景由该页型的 `background` 字段指定，整幅铺满。')
+    L.append('{{LOGO_RULES}}')
+    L += ['- 坐标、字号、色值、资产位置以 %s 为准；本文件的 Colors / Typography 是可用值的清单。'
+          % sidecar,
+          '- 强调色族以 Colors 和 %s 的 slot CSS 为主；必要时可以使用 Colors 之外的颜色，'
+          '但不能形成与模板主色竞争的第二强调色。' % sidecar,
+          '- 新增颜色应与模板整体的色相、明度和饱和度关系协调。允许新增中性色、低彩度辅助色'
+          '或局部语义色表达正负、风险、警告、状态、图表序列，但保持辅助层级；'
+          '只要新色通过高饱和、高对比、大面积或跨页重复获得主视觉权重，'
+          '或被用于标题、关键数字、图表主序列、卡片底色或渐变，就属于新的强调色，改用模板'
+          '强调色族的深浅、透明度，或改用线型、纹理、标签区分。',
+          '- 交付前逐页检查：色板、字体、版式、背景、资产和本段规则均来自本风格包；'
+          '页面无资源加载失败、内容溢出或画幅裁切。',
+          '- 本包里的数值就是普查结果，照用即可，无需重新统计颜色、字体或版式。',
+          '- 风格包以文本形式（zip 摘要等）到手时，直接用摘要里 design.md / layouts.md 的文本。',
           '', '## Exceptions', '']
     if exceptions:
         L += ['- ' + e for e in exceptions]
@@ -827,32 +3081,82 @@ def emit_body(d, tokens, fonts, roles, assets, archetypes, exceptions, ldir):
 
 
 def emit_brief(d, ctx, ldir):
-    (tokens, rest, fonts, roles, assets, rejected, todos, archetypes, cands, sheet,
-     leftover, lsheet) = ctx
+    (tokens, rest, fonts, roles, assets, rejected, todos, archetypes, cands, sheets,
+     selected_vision_groups, omitted_vision_groups, leftover, lsheet) = ctx
     canvas = d['canvas']['px']
-    L = ['# 抽取简报（草案已生成，读完这一页就能改）', '',
+    def sample_pages(archetype):
+        return archetype.get('pages') or archetype.get('_sample_pages') or []
+
+    sampled_archetypes = [
+        archetype for archetype in archetypes
+        if archetype.get('rep') is not None or sample_pages(archetype)
+    ]
+    template_only_archetypes = [
+        archetype for archetype in archetypes
+        if archetype not in sampled_archetypes
+    ]
+    L = ['# 抽取简报（第 1/3 步产物；改完草案跑 package.py 出包）', '',
          '源：`%s`  画布 %dx%d  %d 页 / %d 版式  主题 %s  form=%s'
          % (d['source']['filename'], canvas[0], canvas[1], d['counts']['slides'],
             d['counts']['layouts'], d['theme_topology']['themes'],
             d['form_hint']['form']), '',
          '## 待判断（草案里已标 TODO，逐条改掉）', '']
-    base_todos = ['给风格起名：`manifest.yaml` 的 name / name_zh / description（看两张图定气质）',
-                  '`layouts.yaml` 顶部 `names:` 一段填 %d 个中文页型名（看 layout-sheet.png，'
-                  '一次改完；下面 layouts 段不要动）' % len(archetypes),
-                  '`body.md` 的 Overview 与 Hard Rules 末条（Colors 用途列草案已填好，觉得不对再改）']
-    for t in base_todos + todos:
+    # 待判断清单从草案实时扫 TODO 生成，不写死：写死的清单会和草案对不上——
+    # 既漏掉后加的段（模型读到一半才发现还有活），又在草案已预填时还催人去填。
+    HINT = {'manifest.yaml': '看两张图定气质',
+            'layout-controls.yaml': '看 layout-sheet.png；只改这个控制区',
+            'body.md': 'Colors 用途列草案已填好，觉得不对再改'}
+    for fn in ('manifest.yaml', 'body.md', 'layout-controls.yaml', 'frontmatter.yaml'):
+        path = os.path.join(ldir, fn)
+        if not os.path.exists(path):
+            continue
+        keys = []
+        for line in open(path, encoding='utf-8'):
+            if 'TODO' not in line:
+                continue
+            m = re.match(r'\s*[-#]?\s*([\w-]+):', line)
+            keys.append(m.group(1) if m else line.strip()[:24])
+        if not keys:
+            continue
+        seen, uniq = set(), []
+        for k in keys:
+            if k not in seen:
+                seen.add(k)
+                uniq.append(k)
+        hint = HINT.get(fn)
+        L.append('- `%s` %d 处：%s%s'
+                 % (fn, len(keys), '、'.join(uniq[:6]) + ('…' if len(uniq) > 6 else ''),
+                    '（%s）' % hint if hint else ''))
+    for t in todos:
         L.append('- ' + t)
-    L += ['', '## 联系表（一次看完所有候选图）', '',
-          '`l-out/contact-sheet.png`  —— 编号对应下表；看完再决定 logo / 封面归属。' if sheet
-          else '（Pillow 不可用，未生成联系表；逐张看 `media-out/`）', '',
-          '| # | 文件 | 尺寸 | 出现 | 满屏 | 页 | 草案判定 |', '|---|---|---|---|---|---|---|']
+    L += ['', '## 资产判断（按视觉组一次看完）', '',
+          ('视觉判断拼版：%s。每张都含候选独立卡与所在页截图；只读这些拼版，不逐张打开素材。'
+           % '、'.join('`l-out/%s`' % os.path.basename(path) for path in sheets))
+          if sheets else '（未生成视觉拼版；不要给图片候选定性，已按内容图保留位置并在 gaps 说明。）',
+          '`l-out/asset-vision-groups.json` 记录每张候选的原图尺寸、所有页内位置和尺寸；'
+          '透明/近白候选在拼版中同时给棋盘格和深灰底预览。',
+          '按每个候选实例填 `asset_vision_groups.visual_kind`；同源图在不同页型/位置可不同。'
+          '第三方 logo 墙属于 `content-image`，不是 deck 的 `logo`。',
+          '',
+          '| ID | 文件 | 原图 | 出现 | 页 | 所有位置 |', '|---|---|---|---|---|---|']
     decided = {a['src']['file']: a['id'] for a in assets}
     why = {c['file']: r for c, r in rejected}
-    for i, c in enumerate(cands[:12], 1):
-        L.append('| %d | `%s` | %sx%s | %d | %s | %s | %s |' % (
-            i, c['file'], c['probe'].get('w') or '?', c['probe'].get('h') or '?', c['n'],
-            'Y' if c['fullscreen'] else '', ','.join(map(str, c['slides'][:6])) or 'layout',
-            decided.get(c['file']) or ('✗ ' + why.get(c['file'], '未采纳'))))
+    selected_candidates = {
+        candidate['file']: candidate
+        for group in selected_vision_groups
+        for candidate in group['candidates']
+    }
+    for c in selected_candidates.values():
+        L.append('| `%s` | `%s` | %sx%s | %d | %s | `%s` |' % (
+            c['id'], c['file'], c['probe'].get('w') or '?', c['probe'].get('h') or '?',
+            c['n'], ','.join(map(str, c['slides'])) or 'layout', _placement_text(c)))
+    if omitted_vision_groups:
+        omitted_pages = sorted({
+            page for group in omitted_vision_groups for page in group['pages'] if page > 0
+        })
+        L.append('')
+        L.append('未进视觉预算：%s；对应 slot 默认保留内容图片位置，不会自动升为风格资产。'
+                 % ('第%s页' % '、'.join(map(str, omitted_pages)) if omitted_pages else '版式候选'))
     L += ['', '## 颜色（草案 token 已写进 frontmatter.yaml）', '',
           '| token | hex | 出现 |', '|---|---|---|']
     for name, r in tokens:
@@ -868,24 +3172,55 @@ def emit_brief(d, ctx, ldir):
     L.append('')
     L.append('字号轴：' + '、'.join('%s=%dpx(n=%d)' % (k, round(v['sz_px']), v['n'])
                                  for k, v in roles.items()))
-    L += ['', '## 版式聚类（草案已写进 layouts.yaml）', '',
-          '`l-out/layout-sheet.png` 是各页型代表页的重建图——**看它给页型起名**，'
-          '不用再逐页查 shapes。' if lsheet else '（未生成版式图，按下面的 slot 原文命名）', '',
+    L += ['', '## 版式聚类（判断项在 layout-controls.yaml，坐标事实在 layouts.yaml）', '',
+          ('`l-out/layout-sheet.png` 是各页型代表页的重建图——**看它给页型起名**，'
+           '不用再逐页查 shapes。' if sampled_archetypes else
+           '`l-out/layout-sheet.png` 是模板版式层的重建图；用它看整体视觉即可，'
+           '没有对应样张的版式已按模板名称预填，不逐项改名或判角色。')
+          if lsheet else '（未生成版式图，按下面的 slot 原文命名）', '',
           '| archetype | 页数 | 代表页 | 背景 | slot 数 |', '|---|---|---|---|---|']
-    for a in archetypes:
+    for a in sampled_archetypes:
+        pages = sample_pages(a)
+        representative = a.get('rep') or (pages[0] if pages else None)
         L.append('| `%s` | %d | %s | %s | %d |' % (
-            a['name'], len(a['pages']), a['rep'], a['bg'] or '（无资产底图）', len(a['slots'])))
+            a['name'], len(pages), representative, a['bg'] or '（无资产底图）', len(a['slots'])))
+    if template_only_archetypes:
+        L.append('')
+        L.append('另有 %d 个模板声明版式没有对应样张：名称、角色和坐标已预填并会进入最终包；'
+                 '除非当前样张直接证明不对，不需要逐项判断。'
+                 % len(template_only_archetypes))
+    sampled_pages = {
+        page: archetype
+        for archetype in sampled_archetypes
+        for page in sample_pages(archetype)
+    }
+    first_page = 1
+    last_page = d['counts']['slides']
+    if first_page in sampled_pages:
+        first_archetype = sampled_pages[first_page]
+        L.append('')
+        L.append('第 1 页实际使用页型：`%s`。若样张确为封面，只在 `roles.%s` 填 `cover`，'
+                 '不要按页型名称猜。'
+                 % (first_archetype['name'], first_archetype['name']))
+    if last_page != first_page and last_page in sampled_pages:
+        last_archetype = sampled_pages[last_page]
+        L.append('第 %d 页实际使用页型：`%s`。若样张确为封底，只在 `roles.%s` 填 `closing`，'
+                 '不要按页型名称猜。'
+                 % (last_page, last_archetype['name'], last_archetype['name']))
     if leftover:
         L += ['', '未归入 archetype 的页：%s —— 都是单页孤例，需要就自己补一个 archetype。'
               % ', '.join(map(str, leftover))]
-    L += ['', '各 archetype 的 slot 原文（据此起中文页型名、改 role）：', '']
-    for a in archetypes:
-        L.append('- `%s`（第 %s 页，覆盖 %s）' % (a['name'], a['rep'], a['pages']))
+    L += ['', '有样张页型的 slot 原文（据此起中文页型名，并在 text_roles 判断文本角色）：', '']
+    for a in sampled_archetypes:
+        pages = sample_pages(a)
+        representative = a.get('rep') or (pages[0] if pages else None)
+        L.append('- `%s`（第 %s 页，覆盖 %s）' % (a['name'], representative, pages))
         for s in a['slots']:
             L.append('  - %s %spx 「%s」' % (s['role'], round(s['sz']), s['txt']))
     L += ['', '## 下一步', '',
-          '1. 看 `contact-sheet.png` 和 `layout-sheet.png`；'
-          '2. 用一次批量编辑/patch 改掉四份草案里的 TODO；3. 跑 `package.py`。']
+          '1. 并行看全部 `vision-group-*.jpg` 和 `layout-sheet.png`；'
+          '2. 先填 asset_vision_groups，再用少量 asset_decisions 写例外，最后一次批量改完其它 TODO；'
+          '3. 只改 `layout-controls.yaml` 的版式判断项，再跑 `package.py`。']
     write(os.path.join(ldir, 'BRIEF.md'), '\n'.join(L) + '\n')
 
 
@@ -898,46 +3233,158 @@ def main(argv=None):
     ldir = os.path.join(outdir, 'l-out')
     os.makedirs(ldir, exist_ok=True)
 
-    tokens, rest, _ = draft_colors(d)
+    all_shapes = json.load(open(os.path.join(outdir, 'ref', 'shapes.json'),
+                                encoding='utf-8'))['shapes']
+    cusage = color_usage(all_shapes, d)
+    tokens, rest, rows = draft_colors(d, cusage)
     fonts = draft_fonts(d)
-    archetypes, pages, leftover = draft_layouts(d, outdir)
-    cover_media = next((a['bg_raw'] for a in archetypes if a['name'] == 'cover'), None)
-    bg_needed = {a['bg_raw'] for a in archetypes if a['bg_raw'] and a['bg_raw'].startswith('ppt/media')}
-    bg_under = {p['no']: p['bg_media'] for p in pages}
-    assets, rejected, todos = draft_assets(d, outdir, bg_needed, cover_media, bg_under)
+    effective_alpha = fullscreen_effective_alpha(d, outdir, all_shapes)
+    archetypes, pages, leftover = draft_layouts(d, outdir, effective_alpha)
+    # 只有模板已声明 cover 页型时才能直读它的封面背景。首页和末页会单独保留样张，
+    # 但它们的角色仍由模型看图判断，不能因为页码就自动升格为 cover / closing。
+    cover_media = cover_background_media(archetypes)
+    exported_media = {m['media'] for m in d.get('media', []) if m.get('exported')}
+    bg_needed = {a['bg_raw'] for a in archetypes if a['bg_raw'] in exported_media}
+    bg_under = {p['no']: p.get('rendered_bg') or p['bg_media'] for p in pages}
+    assets, rejected, todos, alias, pool = draft_assets(
+        d, outdir, bg_needed, cover_media, bg_under, effective_alpha)
     media_to_asset = {a['src']['media']: a['id'] for a in assets}
+    for m, w in (alias or {}).items():
+        if w in media_to_asset:
+            media_to_asset.setdefault(m, media_to_asset[w])
+
+    cW, cH = d['canvas']['px']
     for a in archetypes:
         a['bg'] = media_to_asset.get(a['bg_raw'])
+        # 版式自带的图片元素：映射到资产 id。映射不到时**保留槽位但不写 asset**——
+        # 删掉整条槽，消费端看到的是一个没有图标的托底圆，和图标不进包是同一个失败模式，
+        # 而且它连「这里本来有东西」都不知道。
+        keep = []
+        for s in a['slots']:
+            if s.get('media'):
+                aid = media_to_asset.get(s['media'])
+                if not aid:
+                    c = pool.get(alias.get(s['media'], s['media'])) or pool.get(s['media'])
+                    s['role'] = 'asset-candidate'
+                    if c:
+                        s['source_media'] = c['file']
+                    s.pop('media', None)
+                    keep.append(s)
+                    continue
+                s['asset'] = aid
+                c = pool.get(alias.get(s['media'], s['media'])) or pool.get(s['media'])
+                if c:
+                    s['source_media'] = c['file']
+                # role 跟着资产走：图标槽写成 logo 会让消费端把它当品牌标识，每页都摆一个
+                s['role'] = next((x['kind'] for x in assets if x['id'] == aid), s['role'])
+            keep.append(s)
+        a['slots'] = keep
     roles = draft_scale(d, archetypes)
-    cands = sorted([c for c in [a['src'] for a in assets]] +
-                   [c for c, _ in rejected], key=lambda c: (-c['n'], c['file']))
-    sheet = contact_sheet(outdir, cands, os.path.join(ldir, 'contact-sheet.png'))
+    slot_added = cover_slot_colors(tokens, archetypes, rows, cusage)
+    # 局部图和半透明满屏叠加层必须结合页面语境定性。候选在本阶段按图片槽过滤：
+    # 没有最终槽位的媒体无需让模型判断；有槽位但超出视觉预算的则保留通用 pic 槽。
+    decided_c = sorted([a['src'] for a in assets], key=lambda c: (-c['n'], c['file']))
+    other_c = sorted([c for c, _ in rejected], key=lambda c: (-c['n'], c['file']))
+    cands, seen_file = [], set()
+    for c in decided_c + other_c:            # 同一张图可能有多条候选记录（不同位置各一条）
+        if c['file'] not in seen_file:
+            seen_file.add(c['file'])
+            cands.append(c)
+    review_candidates = []
+    for candidate_index, candidate in enumerate(
+            visual_slot_candidates(cands, archetypes), 1):
+        row = dict(candidate)
+        row['id'] = 'asset-%d' % candidate_index
+        review_candidates.append(row)
+    selected_vision_groups, omitted_vision_groups, sheets = emit_asset_vision_groups(
+        outdir, review_candidates, d['counts']['slides'], ldir)
     lsheet = layout_sheet(outdir, archetypes, os.path.join(ldir, 'layout-sheet.png'))
 
     anchors = draft_anchors(d, tokens, fonts, roles, assets, archetypes)
     gaps, exceptions = [], []
+    if omitted_vision_groups:
+        omitted_pages = sorted({
+            page for group in omitted_vision_groups for page in group['pages'] if page > 0
+        })
+        if omitted_pages:
+            gaps.append('视觉略过：%s页' % '/'.join(map(str, omitted_pages)))
+        else:
+            gaps.append('视觉判断超预算，未覆盖版式候选')
+    if review_candidates and not sheets:
+        gaps.append('视觉拼版不可用，图片候选按内容图保留，未做风格定性。')
     for c, why in rejected:
         if '近全透明' in why:
             gaps.append('母版/版式里的 %s 是%s，不是设计资产，任何情况下不要当背景用。' % (c['file'], why))
-    for f in fonts[:2]:
-        if len(f['stack']) > 1:
+        elif '不是背景' in why:
+            gaps.append('%s 在模板里铺满整页，但%s；那几页的真实背景是幻灯片自身的底色，'
+                        '需要时按 Colors 里的 surface 铺纯色。' % (c['file'], why))
+    by_kind = {}
+    for kind, kept, total, advice, where in _TRUNCATED:
+        e = by_kind.setdefault(kind, {'kept': 0, 'total': 0, 'advice': advice, 'where': []})
+        e['kept'] += kept
+        e['total'] += total
+        if where:
+            e['where'].append(where)
+    for kind, e in by_kind.items():
+        at = ('（%s）' % '、'.join(e['where'][:6])) if e['where'] else ''
+        gaps.append('%s%s按名额截断：普查到 %d 个，包内留了 %d 个%s。'
+                    % (kind, at, e['total'], e['kept'],
+                       '；' + e['advice'] if e['advice'] else ''))
+    # 「没命中映射表」不等于「装不上」：降级目标本身（Noto Sans SC 之类）和 Office 出厂体
+    # 都不在 match 列里，但它们本来就可用。真正危险的是**既没命中、又不是已知可用字体**的
+    # 那种——design.md 的字体栈里留着一个消费端装不上的商业字体名，且没有任何降级说明。
+    web_ok = {norm(x) for fam in parse_fallback_table() for x in fam['fallback']}
+    web_ok |= {norm(x.strip().strip('"')) for x in SYS_FALLBACK.split(',')}
+    for f in fonts:
+        if f.get('mapped'):
             gaps.append('源字体 %s 无 web 授权源，已按 font-fallback 表降级到 %s；字形细节与原稿有差异。'
                         % (f['names'][0], f['stack'][1]))
+        elif norm(f['names'][0]) in OFFICE_DEFAULT_FONTS_NORM:
+            gaps.append('%s 是 Office 出厂字体，多半是模板里没清干净的残留而非设计选型；'
+                        '按正文/标题的实际气质挑替代体，不要照抄它。' % f['names'][0])
+        elif norm(f['names'][0]) not in web_ok:
+            gaps.append('源字体 %s 不在 font-fallback 表里，字体栈只有原名，消费端很可能装不上；'
+                        '按气质挑一个有 web 分发源的近似体补进栈，不要照抄原名。' % f['names'][0])
+    nosize = [(a['name'], s['box']) for a in archetypes for s in a['slots']
+              if not s.get('asset') and not s.get('_font_size')]
+    if nosize:
+        gaps.append('这些文字槽在源文件任何层级都没有字号声明（都不是占位符，是普通文本框，'
+                    '继承源是 presentation.xml 的 defaultTextStyle，本抽取按约定不解继承链）：'
+                    '%s。用 typography 里最接近的档位，不要自造新档。'
+                    % '、'.join('%s %s' % (n, b) for n, b in nosize[:6]))
+
     if leftover:
         exceptions.append('源 deck 第 %s 页是单页孤例，没有归纳成 archetype；需要类似构图时按最接近的页型改。'
                           % '、'.join(map(str, leftover)))
 
-    emit_manifest(d, assets, ldir)
+    emit_manifest(d, assets, selected_vision_groups, ldir)
     emit_frontmatter(d, tokens, fonts, roles, anchors, gaps, ldir)
-    emit_layouts(archetypes, ldir)
-    emit_body(d, tokens, fonts, roles, assets, archetypes, exceptions, ldir)
-    emit_brief(d, (tokens, rest, fonts, roles, assets, rejected, todos, archetypes, cands, sheet,
-                   leftover, lsheet), ldir)
-
-    print('草案就绪 -> %s' % ldir)
-    print('  资产 %d（%s）  版式 %d  色 %d  字体 %d'
+    # 每张背景量一次局部对比度，作为「哪里不能压文字」的客观依据摆进判断单。
+    # 只报测到的数，不替人填 avoid——哪块算主体、要不要避让，是看图才能定的。
+    busy_hints = {}
+    for a in assets:
+        if a['kind'] != 'background' or not a['src'].get('out'):
+            continue
+        r = bg_busy_map(os.path.join(outdir, a['src']['out']), (cW, cH))
+        if r:
+            busy_hints[a['id']] = r
+    facts, recipes = structure_facts(archetypes, d, all_shapes)
+    for a in archetypes:
+        a['flow'] = draft_flow(a, facts.get(a['name']) or {}, (cW, cH))
+    emit_layouts(archetypes, ldir, busy_hints, facts, recipes)
+    emit_body(d, tokens, fonts, roles, assets, archetypes, exceptions, cusage, ldir,
+              has_asset_candidates=any(needs_asset_judgment(c) for c in cands))
+    emit_brief(d, (tokens, rest, fonts, roles, assets, rejected, todos, archetypes, cands, sheets,
+                   selected_vision_groups, omitted_vision_groups, leftover, lsheet), ldir)
+
+    # 这几行落在模型判断「skill 是不是做完了」的那一刻。只报数就会被读成「包已生成」，
+    # 于是判断和打包整段被跳过，deck 拿不到任何版式坐标。所以这里报进度与下一条命令。
+    print('第 1/3 步完成，判断单草案 -> %s' % ldir)
+    print('  待你确认：资产 %d（%s）  版式 %d  色 %d  字体 %d'
           % (len(assets), ', '.join(x['id'] for x in assets), len(archetypes), len(tokens), len(fonts)))
-    print('  先读 l-out/BRIEF.md，再看 l-out/contact-sheet.png')
+    print('  第 2 步 读 l-out/BRIEF.md，并行看视觉组拼版与版式图；版式判断只改 layout-controls.yaml')
+    print('  第 3 步 package.py 产出 design.md + layouts.md —— deck 的版式坐标只从这两份读')
+    sys.stdout.flush()
     return 0