Agent/examples/process_pipeline/script/search_eval/procedure-dsl/spec/tools/renderer.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
renderer.py — Procedure DSL 可视化共享模板.

每个 case 提供 case_data (一个 dict, 见 build_html 的 docstring),
import 本模块并调用 build_html(case_data) → HTML 字符串.

spec 引用: spec.md §12 (.html 可视化结构规范).
"""
import html
import json
import re
from pathlib import Path


# ── 数据 loaders (从 spec/taxonomy 单一来源加载) ────────────────────
def _load_json(rel_path: str) -> dict:
    """从 spec/ 子路径加载 JSON. rel_path 是相对 spec/ 的路径 (e.g. 'taxonomy/type.json')."""
    return json.loads((Path(__file__).resolve().parent.parent / rel_path).read_text(encoding='utf-8'))


def _load_stdlib_types() -> dict:
    """type stdlib 渲染元信息. 新结构 type.json 无 type_metadata → 返回 {};
    stdlib 叶子的 in_tree 标记在 build_html 里按 type.json 的 $leaves 补齐."""
    return {}


_DRAWER_TITLES = {
    'effect': ('作用 (chip 上 data-prefix="作用")',
               '这一步在 AIGC 生产工序链中的位置 + 作用; 从 §A.1 字典树取 L3 叶子'),
    'action': ('动作 (chip 上 data-prefix="动作")',
               '这一步的动作动词; 从 §A.2 字典树取路径'),
    'type':   ('类型 (chip 上 data-type)',
               '领域语义类型 (按功能角色分类: 程序控制/数据复用/内容/知识); 从 §A.3 字典树取叶子'),
}


def _add_label_suffix(node, depth=0):
    """递归把 dict 的 key 加 (L1)/(L2) 显示后缀, 让 drawer UI 区分层级.
    leaf string 保留原状; $* 元数据 key 跳过."""
    if not isinstance(node, dict):
        return node
    result = {}
    for k, v in node.items():
        if k.startswith('$'):
            continue
        new_k = k
        if isinstance(v, dict):
            if depth == 0:
                new_k = f'{k} (L1)'
            elif depth == 1:
                new_k = f'{k} (L2)'
        result[new_k] = _add_label_suffix(v, depth + 1)
    return result


def _tree_to_drawer(nodes):
    """新结构 最终分类树 (list of {分类名称, 分类说明, 子分类}) → drawer 嵌套 dict.
    叶子 → {名称: 分类说明 字符串}; 非叶 → {名称: 子树 dict}. 层级即 根→叶 路径."""
    out = {}
    for n in nodes or []:
        name = n.get('分类名称', '?')
        kids = n.get('子分类') or []
        out[name] = _tree_to_drawer(kids) if kids else (n.get('分类说明', '') or '')
    return out


def _load_drawer_tree(dim: str) -> dict:
    """spec/taxonomy/{dim}.json → 重塑为 drawer UI 需要的 {title, desc, tree} 格式.
    新结构读 最终分类树 (中文键); 自动加 (L1)/(L2) 层级标签. dim ∈ {'effect', 'action', 'type'}."""
    raw = _load_json(f'taxonomy/{dim}.json')
    title, desc = _DRAWER_TITLES[dim]
    nested = _tree_to_drawer(raw.get('最终分类树', []))
    return {
        'title': title,
        'desc': desc,
        'tree': _add_label_suffix(nested, depth=0),
    }


# =============================================================================
# STDLIB · type registry (字典树 §A.3 叶子的 in_tree 标记基础; 现为空, build_html 按 $leaves 补齐)
# =============================================================================
STDLIB_TYPE_REGISTRY = _load_stdlib_types()


# =============================================================================
# 字典树 (spec §A.1 作用 / §A.2 动作 / §A.3 类型)
# =============================================================================
EFFECT_TREE = _load_drawer_tree("effect")

ACTION_TREE = _load_drawer_tree("action")

TYPE_TREE = _load_drawer_tree("type")


def _build_type_paths() -> dict:
    """type.json 最终分类树 → {叶子名: '根/.../叶子' 路径}, 供 chip 显示完整路径."""
    raw = _load_json('taxonomy/type.json')
    out: dict = {}
    def walk(nodes, prefix):
        for n in nodes:
            path = prefix + [n.get('分类名称', '')]
            kids = n.get('子分类') or []
            if kids:
                walk(kids, path)
            else:
                out[n.get('分类名称', '')] = '/'.join(path)
    walk(raw.get('最终分类树', []), [])
    return out


TYPE_PATHS = _build_type_paths()

FEATURE_TAXONOMY = _load_json("taxonomy/feature.json")

EXTERNAL_TAXONOMIES = {
    '实质': {'title': '实质 (内容是什么)',  'desc': '理念 / 表象 — 911 路径',
            'source': 'external', 'file': '分类库导出_实质_*.json'},
    '形式': {'title': '形式 (内容怎么呈现)', 'desc': '呈现 / 架构 — 565 路径',
            'source': 'external', 'file': '分类库导出_形式_*.json'},
}


# =============================================================================
# Render helpers
# =============================================================================
def he(s):
    """HTML escape, 默认 quote=True 把 `"` 转成 `"`.
    这点关键: title / data-* / class 等属性值如果含 `"` 会提前终止属性,
    导致 tooltip 显示不完整 (典型 bug: '原文方法 3 只说"自己写动作"' 在 title 中只显示到 "只说" 就被截断).
    """
    if s is None:
        return ''
    return html.escape(str(s))


def render_intent(text):
    """目的列: 简短自然语言句, **尽量** 把其他列里的结构化值都做成 {kind:value} token.

    每个 token 底色对应其引用的列, 让人一眼看出该值来自哪里. 合法 kind:
      effect          → 作用列 (灰, 需求组)
      via             → 外部工具列 (浅绿 + 等宽字体, 实现组)
      act             → 动作列 (绿, 实现组)
      control         → 逻辑控制列 (浅青, 实现组) — 并行/遍历/分支/请求/等待
      in-type         → 输入·类型 (黄圆胶囊)         out-type → 输出·类型 (蓝圆胶囊)
      in-sub          → 输入·实质 (黄矩形 tag)        out-sub  → 输出·实质 (蓝矩形 tag)
      in-form         → 输入·形式 (黄矩形 tag 斜体)   out-form → 输出·形式 (蓝矩形 tag 斜体)

    **特性列 (feature) 不允许在 intent 中引用** — feature 是内部执行特征 (随机/幂等/人工/读写外部),
    不出现在面向使用者的描述. 写成 `ik-other` 灰色显示作为 lint 警告.

    严禁变量名 token (no `{in:X}` / `{out:X}`); 严禁 dataflow 公式 / case-specific 简写.
    """
    def sub(m):
        kind = m.group(1)
        val = m.group(2)
        kc = {
            'effect': 'ik-effect',
            'via': 'ik-via',
            'act': 'ik-act',
            'in-type': 'ik-in-type', 'out-type': 'ik-out-type',
        }.get(kind, 'ik-other')
        return f'<span class="intent-tok {kc}">{he(val)}</span>'
    return re.sub(r'\{([\w-]+):([^}]+)\}', sub, text or '')


def render_chip(type_name):
    if not type_name:
        return ''
    if type_name == '-':
        return he(type_name)
    # 显示完整路径 (e.g. 程序控制类型/指令/提示词); data-type 仍存叶子名, drawer 查找不受影响.
    # case-specific 类型 (不在字典树) 显示原名.
    disp = TYPE_PATHS.get(type_name, type_name)
    return f'<span class="chip" data-type="{he(type_name)}">{he(disp)}</span>'


def render_path(prefix, value):
    if not value:
        return ''
    if isinstance(value, list):
        spans = []
        for val in value:
            if val:
                spans.append(f'<span data-prefix="{prefix}" data-value="{he(val)}">{he(val)}</span>')
        return '\n'.join(spans)
    if isinstance(value, str):
        if '+' in value:
            parts = [p.strip() for p in value.split('+') if p.strip()]
            spans = []
            for val in parts:
                spans.append(f'<span data-prefix="{prefix}" data-value="{he(val)}">{he(val)}</span>')
            return '\n'.join(spans)
        return f'<span data-prefix="{prefix}" data-value="{he(value)}">{he(value)}</span>'
    return ''


_VALUE_DESC_RE = re.compile(r'^<(.+)>$', re.DOTALL)


def render_value(vl):
    """值列渲染:
      - 若整段以 `<...>` 括起 → 渲染为斜体浅灰背景, 表示"这是对内容的描述, 不是内容本身"
        (适用于无法在 cell 中直接嵌入的非文本数据: 视频/图像/音频).
      - 否则 → 渲染为普通文本 (适用于文本数据本身, 如 prompt 全文).
    """
    if vl is None:
        return ''
    s = str(vl).strip()
    m = _VALUE_DESC_RE.match(s)
    if m:
        inner = m.group(1)
        return f'<span class="value-desc">&lt;{he(inner)}&gt;</span>'
    return f'<span class="natural">{he(s)}</span>'


def render_focus_class(cell_key, focus_list):
    return ' row-focus' if cell_key in (focus_list or []) else ''


def cell_attrs(field_key, focus, io_reason=None, is_empty=False):
    """组合 cell 的额外 class 和属性 (focus + 推断补全).

    field_key 例: 'action', 'in-value-0', 'out-type-1', etc.
    io_reason: 当前 cell 所属 IO item 整体被标 inferred 时, 传入 reason 字符串.
    is_empty: 该 cell 内容是否为空 (留空的推断 cell 角标变 推?).
    返回 (class_suffix_str, extra_attrs_str).
    """
    cls = render_focus_class(field_key, focus)
    extra = ''
    if io_reason:
        cls += ' is-inferred'
        if is_empty:
            cls += ' is-low-confidence'
        extra = f' title="推断补全: {he(io_reason)}"'
    return cls, extra


def render_io_value(item):
    """IO 值列: output 的 id 作为小标签贴在值前 (供 anchor 1:1 引用对照); input 无 id."""
    iid = item.get('id')
    id_tag = f'<span class="io-id" data-io-id="{he(iid)}">{he(iid)}</span> ' if iid else ''
    return id_tag + render_value(item.get('value'))


_OUT_ID_RE = re.compile(r'^s\d+(?:\.\d+)?o\d+$')
_OUT_PATH_RE = re.compile(r's(\d+(?:\.\d+)?)\.outputs\[(\d+)\]')


def _ref_output_id(anchor):
    """把一条输入 anchor (来源 ←) 归一到它引用的**输出编号**, 取不到返回 None.

    认: ← s1o1 / ← s5o1[-1] (去索引) / ← p1.s1.outputs[0] (JSON 路径式也救, 归一成 s1o1)。
    不认 (返回 None): ← 工序输入 / ← s1.名字 (按名引用) / 字面量。供"去处"自动反推用。
    """
    if not anchor:
        return None
    s = str(anchor).strip()
    if s.startswith('←'):
        s = s[1:].strip()
    m = _OUT_PATH_RE.search(s)                       # JSON 路径式 → sNo(idx+1)
    if m:
        return f's{m.group(1)}o{int(m.group(2)) + 1}'
    base = s.split('[')[0].strip()                   # 去掉 [i] / [-1] 索引
    return base if _OUT_ID_RE.match(base) else None


def build_output_consumers(proc):
    """扫一个工序内所有输入的来源 ← anchor, 反推 {输出编号: [消费它的步骤 id]}。

    "去处"(输出 →)其实是"来源"(输入 ←)的逆: 谁的输入引用了本输出, 谁就是去处。
    模型没填 → 时, 渲染层据此自动补出去处, 不必逼模型冗余地正反都填。
    """
    consumers: dict = {}
    for st in proc.get('steps', []) or []:
        sid = st.get('id')
        for io in st.get('inputs') or []:
            oid = _ref_output_id(io.get('anchor'))
            if oid:
                lst = consumers.setdefault(oid, [])
                if sid and sid not in lst:
                    lst.append(sid)
    return consumers


def render_step_row(step, idx_label, type_reg=None, consumers=None):
    """渲染一个 step (kind: step / block / nested) 为一组 tr 行 (14 列).

    需求组 (rowspan): # / 目的 / 作用 / 实质 / 形式 (后两者 step 级).
    输入/输出 (逐 IO): 类型 / 值 / 来源(去处). 实现组 (rowspan): 外部工具 / 动作 / 指令.
    type_reg: 已合并 STDLIB + case-specific 的类型注册表 (用于 chip in_tree 标记).
    consumers: {输出编号: [消费步骤 id]} 反向索引; 输出 → 没填时据此自动补"去处"。
    """
    inputs = step.get('inputs', [])
    outputs = step.get('outputs', [])
    N = max(len(inputs), len(outputs), 1)
    focus = step.get('focus', [])

    is_nested = step['kind'] == 'nested'
    is_block = step['kind'] == 'block'

    main_cls = 'step step-main'
    if is_block:
        main_cls = 'step block-header'
    if is_nested:
        main_cls = 'step step-main step-nested'
    sub_cls = 'step step-sub'

    data_step = step['id']
    data_group = step.get('group', '')

    rows = []
    for k in range(N):
        tr_cls = main_cls if k == 0 else sub_cls
        attrs = f' data-step="{data_step}"'
        if data_group:
            attrs += f' data-group="{data_group}"'
        attrs += f' data-focus="{",".join(focus)}"'
        if k == 0:
            attrs = f' id="{data_step}"' + attrs

        cells = []
        if k == 0:
            rs = f' rowspan="{N}"' if N > 1 else ''
            arrow = '<span class="arrow">▼</span> ' if is_block else ''
            indent = '  └ ' if is_nested else ''
            cells.append(f'<td class="idx{render_focus_class("idx", focus)}"{rs}>{arrow}{indent}{he(idx_label)}</td>')
            intent_html = render_intent(step.get('intent', ''))
            c, a = cell_attrs('intent', focus)
            cells.append(f'<td class="intent{c}"{a}{rs}><div class="intent-text">{intent_html}</div></td>')
            c, a = cell_attrs('effect', focus)
            cells.append(f'<td class="effect{c}"{a}{rs}>{render_path("作用", step.get("effect", ""))}</td>')
            c, a = cell_attrs('substance', focus)
            cells.append(f'<td class="substance{c}"{a}{rs}>{render_path("实质", step.get("substance")) if step.get("substance") else ""}</td>')
            c, a = cell_attrs('form', focus)
            cells.append(f'<td class="form{c}"{a}{rs}>{render_path("形式", step.get("form")) if step.get("form") else ""}</td>')

        in_item = inputs[k] if k < len(inputs) else None
        if in_item:
            tp = in_item.get('type'); vl = in_item.get('value'); ac = in_item.get('anchor')
            io_inf = in_item.get('inferred_reason') if in_item.get('inferred') else None
            c, a = cell_attrs(f'in-type-{k}', focus, io_inf, is_empty=not tp)
            cells.append(f'<td class="in-type{c}"{a}>{render_chip(tp)}</td>')
            c, a = cell_attrs(f'in-value-{k}', focus, io_inf, is_empty=not vl)
            cells.append(f'<td class="in-value{c}"{a}>{render_io_value(in_item)}</td>')
            c, a = cell_attrs(f'in-anchor-{k}', focus, io_inf, is_empty=not ac)
            cells.append(f'<td class="in-anchor{c}"{a}><span class="flow">{he(ac)}</span></td>')
        else:
            cells += ['<td class="in-type"></td>', '<td class="in-value"></td>', '<td class="in-anchor"></td>']

        if k == 0:
            rs = f' rowspan="{N}"' if N > 1 else ''
            c, a = cell_attrs('via', focus)
            cells.append(f'<td class="via{c}"{a}{rs}>{he(step.get("via", ""))}</td>')
            c, a = cell_attrs('action', focus)
            cells.append(f'<td class="action{c}"{a}{rs}>{render_path("动作", step.get("action", ""))}</td>')
            c, a = cell_attrs('directive', focus)
            directive = step.get('directive', '') or ''
            inner = f'<span class="natural">{he(directive)}</span>' if directive else ''
            cells.append(f'<td class="directive{c}"{a}{rs}>{inner}</td>')

        out_item = outputs[k] if k < len(outputs) else None
        if out_item:
            tp = out_item.get('type'); vl = out_item.get('value'); ac = (out_item.get('anchor') or '')
            io_inf = out_item.get('inferred_reason') if out_item.get('inferred') else None
            # 去处: 模型填了就用原值; 没填则从下游 ← 反推 (谁的输入引用了本输出)。
            derived = ''
            if not ac.strip() and consumers and out_item.get('id'):
                cs = consumers.get(out_item['id'])
                if cs:
                    derived = '→ ' + ', '.join(cs)
            flow_txt = ac if ac.strip() else derived
            flow_cls = 'flow' if ac.strip() else ('flow derived' if derived else 'flow')
            flow_attr = '' if ac.strip() else (' title="自动推导自下游来源 (模型未填 →)"' if derived else '')
            c, a = cell_attrs(f'out-type-{k}', focus, io_inf, is_empty=not tp)
            cells.append(f'<td class="out-type{c}"{a}>{render_chip(tp)}</td>')
            c, a = cell_attrs(f'out-value-{k}', focus, io_inf, is_empty=not vl)
            cells.append(f'<td class="out-value{c}"{a}>{render_io_value(out_item)}</td>')
            c, a = cell_attrs(f'out-anchor-{k}', focus, io_inf, is_empty=not flow_txt)
            cells.append(f'<td class="out-anchor{c}"{a}><span class="{flow_cls}"{flow_attr}>{he(flow_txt)}</span></td>')
        else:
            cells += ['<td class="out-type"></td>', '<td class="out-value"></td>', '<td class="out-anchor"></td>']

        rows.append(f'<tr class="{tr_cls}"{attrs}>{"".join(cells)}</tr>')
    return '\n      '.join(rows)


def render_declarations(case_data, procedure=None):
    """工序头部 + declare 块 (输入/资源/返回), 可折叠.

    procedure: 该工序的 dict (新 schema 下 case_data.procedures[i]). 不传时退化到
    老 case_data.procedure (兼容老 case 直到全部迁移完).
    """
    proc = procedure if procedure is not None else case_data.get('procedure', {})
    decls = proc.get('declarations') if procedure is not None else case_data.get('declarations', {})
    decls = decls or {}
    parts = []
    parts.append('<details class="declarations" open>')
    parts.append('<summary class="decl-summary">')
    parts.append(f'  <span class="kw">工序</span> <b class="proc-name">{he(proc.get("name", ""))}</b>')
    if proc.get('purpose'):
        parts.append(f'  <span class="decl-purpose">#目的: {he(proc["purpose"])}</span>')
    if proc.get('category'):
        parts.append(f'  <span class="tag-mini">类别: {he(proc["category"])}</span>')
    meta_bits = []
    for label, key in [('平台', 'platform'), ('作者', 'author')]:
        if proc.get(key):
            meta_bits.append(f'#{label}: {he(proc[key])}')
    if case_data.get('case_id') is not None:
        meta_bits.append(f'case: {he(case_data["case_id"])}')
    if meta_bits:
        parts.append(f'  <span class="decl-meta">{" · ".join(meta_bits)}</span>')
    parts.append('</summary>')
    parts.append('<div class="decl-body">')

    def section(label, items, renderer):
        out = [f'<div class="decl-section"><div class="decl-label">{label}</div>']
        for it in items:
            out.append(renderer(it))
        out.append('</div>')
        return '\n'.join(out)

    def render_io(it):
        chip = render_chip(it.get('type', ''))
        name = he(it.get('name', ''))
        default = it.get('default')
        line = f'<div class="decl-row">{chip} <span class="name" data-var="{name}">{name}</span>'
        if default:
            line += f'  <span class="decl-default">= {he(default)}</span>'
        if it.get('desc'):
            line += f'  <span class="decl-desc">— {he(it["desc"])}</span>'
        return line + '</div>'

    if decls.get('inputs'):
        parts.append(section('输入', decls['inputs'], render_io))
    if decls.get('resources'):
        parts.append(section('资源 (跨 case 长期资产)', decls['resources'], render_io))
    if decls.get('returns'):
        ret = decls['returns']
        chip = render_chip(ret.get('type', ''))
        line = f'<div class="decl-row">{chip}'
        if ret.get('note'):
            line += f'  <span class="decl-desc">— {he(ret["note"])}</span>'
        line += '</div>'
        parts.append(f'<div class="decl-section"><div class="decl-label">返回</div>{line}</div>')

    parts.append('</div>')
    parts.append('</details>')
    return '\n'.join(parts)


_INLINE_MEDIA_RE = re.compile(r'\[(image|video):\s*(https?://[^\s\]]+)\]')


def render_source_body(body_text):
    """把 body_text 渲染为 HTML: [image:URL] → <img>, [video:URL] → <a>, 其他按 \\n 分段."""
    if not body_text:
        return ''
    pieces = []
    last = 0
    for m in _INLINE_MEDIA_RE.finditer(body_text):
        seg = body_text[last:m.start()]
        if seg.strip():
            for para in seg.split('\n'):
                p = para.strip()
                if p:
                    pieces.append(f'<p class="src-p">{he(p)}</p>')
        kind, url = m.group(1), m.group(2)
        if kind == 'image':
            pieces.append(f'<div class="src-img-wrap"><img class="src-img" src="{he(url)}" loading="lazy" referrerpolicy="no-referrer" alt=""></div>')
        else:
            pieces.append(f'<div class="src-video"><a href="{he(url)}" target="_blank" rel="noopener">▶ 视频</a></div>')
        last = m.end()
    tail = body_text[last:]
    if tail.strip():
        for para in tail.split('\n'):
            p = para.strip()
            if p:
                pieces.append(f'<p class="src-p">{he(p)}</p>')
    return '\n'.join(pieces)


def render_source(case_data):
    """原文 折叠块: 元信息 + body_text 完整正文 (含内嵌图片/视频)."""
    src = case_data.get('source')
    if not src:
        return ''
    url = src.get('url', '')
    title = src.get('title', '')
    excerpt = src.get('excerpt', '')
    body_text = src.get('body_text', '')
    cover = src.get('cover_image', '')
    meta_bits = [b for b in [src.get('platform'), src.get('author'), src.get('date')] if b]
    parts = ['<details class="source-block">']
    parts.append(f'<summary>原文: <b>{he(title or "(无标题)")}</b></summary>')
    parts.append('<div class="source-body">')
    if url:
        parts.append(f'<div class="source-url"><a href="{he(url)}" target="_blank" rel="noopener">{he(url)}</a></div>')
    if meta_bits:
        parts.append(f'<div class="source-meta">{" · ".join(he(b) for b in meta_bits)}</div>')
    if excerpt:
        parts.append(f'<p class="source-excerpt"><b>摘要</b>: {he(excerpt)}</p>')
    if cover:
        parts.append(f'<div class="src-cover"><img class="src-img" src="{he(cover)}" loading="lazy" referrerpolicy="no-referrer" alt=""></div>')
    if body_text:
        parts.append('<hr class="src-divider">')
        parts.append('<div class="source-full">')
        parts.append(render_source_body(body_text))
        parts.append('</div>')
    parts.append('</div></details>')
    return '\n'.join(parts)


def render_thead():
    return '''<thead>
  <tr>
    <th colspan="5" class="col-group-demand">需求</th>
    <th colspan="3" class="col-group-input">输入</th>
    <th colspan="3" class="col-group-impl">实现</th>
    <th colspan="3" class="col-group-output">输出</th>
  </tr>
  <tr>
    <th class="col-idx">#</th>
    <th class="col-intent">目的</th>
    <th class="col-effect">作用</th>
    <th class="col-substance">实质</th>
    <th class="col-form">形式</th>
    <th class="col-in-type">类型</th>
    <th class="col-in-value">值</th>
    <th class="col-in-anchor">来源</th>
    <th class="col-via">外部工具</th>
    <th class="col-action">动作</th>
    <th class="col-directive">指令</th>
    <th class="col-out-type">类型</th>
    <th class="col-out-value">值</th>
    <th class="col-out-anchor">去处</th>
  </tr>
</thead>'''


def render_legend():
    return '''<div class="legend">
    <div class="group gray">
      <span class="gh">需求</span>
      <span class="col-toggle" data-col="idx">#</span>
      <span class="col-toggle" data-col="intent">目的</span>
      <span class="col-toggle" data-col="effect">作用</span>
      <span class="col-toggle" data-col="substance">实质</span>
      <span class="col-toggle" data-col="form">形式</span>
    </div>
    <div class="group yellow">
      <span class="gh">输入</span>
      <span class="col-toggle" data-col="in-type">类型</span>
      <span class="col-toggle" data-col="in-value">值</span>
      <span class="col-toggle" data-col="in-anchor">来源</span>
    </div>
    <div class="group green">
      <span class="gh">实现</span>
      <span class="col-toggle" data-col="via">外部工具</span>
      <span class="col-toggle" data-col="action">动作</span>
      <span class="col-toggle" data-col="directive">指令</span>
    </div>
    <div class="group blue">
      <span class="gh">输出</span>
      <span class="col-toggle" data-col="out-type">类型</span>
      <span class="col-toggle" data-col="out-value">值</span>
      <span class="col-toggle" data-col="out-anchor">去处</span>
    </div>
    <span class="inferred-toggle" id="inferred-toggle" title="高亮所有标为推断补全的 cell">高亮推断</span>
    <span class="legend-hint">点击列名 ↔ 显示/隐藏 · 点击组名 ↔ 整组切换 · 「推」角标 hover 看推断理由</span>
  </div>'''


# =============================================================================
# CSS
# =============================================================================
CSS = (Path(__file__).resolve().parent / "styles.css").read_text(encoding="utf-8")

# =============================================================================
# JS
# =============================================================================
JS_TEMPLATE = (Path(__file__).resolve().parent / "script.js").read_text(encoding="utf-8")


# =============================================================================
# build_html — 主入口
# =============================================================================
def build_html(case_data: dict) -> str:
    """从 case_data 构建完整 HTML.

    case_data 结构 (新 schema):
        page_title:    str
        case_id:       int | str | None
        source:        {platform, author, date, url, title, excerpt, body_text?, cover_image?}
        procedures: [
            {
                id:            str (e.g. 'p1-simple' / 'p1')
                name:          str
                purpose:       str
                category:      str
                platform:      str
                author:        str
                declarations:  {inputs[], resources[], returns{}}
                type_registry: dict (该工序的 case-specific 类型, 跟 STDLIB 合并)
                steps:         [{id, kind, effect, via, action, ...}]
                return_row:    {arrow, text}
            },
            ...
        ]

    """
    # type_registry 合并: STDLIB + 所有 procedures.type_registry (跨工序合并)
    type_reg = dict(STDLIB_TYPE_REGISTRY)
    for proc in case_data.get('procedures', []):
        type_reg.update(proc.get('type_registry') or {})

    # 补齐 type.json $leaves: 两类问题都要修
    #   (a) 缺 entry: STDLIB 跟 type.json 不同步, 21 个叶子在字典树但 STDLIB 没条目
    #       → 补 stub {in_tree: True}
    #   (b) 有 entry 但缺 in_tree 标记: e.g. type.json.type_metadata 里"提示词" 有 metadata 但没 in_tree,
    #       同时它又是 $leaves 之一. STDLIB 直接复制 type_metadata 段就没带 in_tree.
    #       → in-place 加 in_tree=True
    # 这是 script.js isInTypeTree() 正确返 true 的硬要求.
    try:
        leaves_file = Path(__file__).resolve().parent.parent / 'taxonomy' / 'type.json'
        if leaves_file.exists():
            leaves_data = json.loads(leaves_file.read_text(encoding='utf-8'))
            for leaf in leaves_data.get('$leaves', []):
                if leaf not in type_reg:
                    type_reg[leaf] = {'in_tree': True}
                elif isinstance(type_reg[leaf], dict) and not type_reg[leaf].get('in_tree'):
                    type_reg[leaf] = {**type_reg[leaf], 'in_tree': True}
    except Exception:
        pass    # 兜底失败不阻塞渲染

    # build thead 一次 (24 列结构跨 procedures 共享)
    thead = render_thead()

    # 逐 procedure 渲染: 每个 procedure 出一段 (declarations 折叠 + 工序表)
    procedure_blocks: list[str] = []
    for proc in case_data.get('procedures', []):
        proc_decl_html = render_declarations(case_data, procedure=proc)

        consumers = build_output_consumers(proc)   # {输出编号: [消费步骤]} → 自动补"去处"
        body_rows = []
        for step in proc.get('steps', []):
            body_rows.append(render_step_row(step, step['id'], type_reg, consumers))
        rr = proc.get('return_row') or {}
        if rr:
            body_rows.append(f'<tr class="return-row"><td class="idx">{rr.get("arrow", "↩")}</td><td colspan="13">{rr.get("text", "")}</td></tr>')
        proc_tbody = '<tbody>\n      ' + '\n      '.join(body_rows) + '\n    </tbody>'

        proc_id_attr = he(proc.get('id', ''))
        procedure_blocks.append(f'''<section class="procedure" id="proc-{proc_id_attr}">
    {proc_decl_html}
    <div class="table-wrap">
      <table class="proc">
        {thead}
        {proc_tbody}
      </table>
    </div>
  </section>''')

    procedures_html = '\n  '.join(procedure_blocks)

    # taxonomy for JS
    taxonomy_js = {
        '作用': EFFECT_TREE,
        '动作': ACTION_TREE,
        '类型': TYPE_TREE,
        '实质': EXTERNAL_TAXONOMIES['实质'],
        '形式': EXTERNAL_TAXONOMIES['形式'],
    }

    js = (JS_TEMPLATE
          .replace('__TYPE_REGISTRY__', json.dumps(type_reg, ensure_ascii=False, indent=2))
          .replace('__TAXONOMY__', json.dumps(taxonomy_js, ensure_ascii=False, indent=2)))

    page_title = case_data.get('page_title', 'Procedure DSL 可视化')
    source_html = render_source(case_data)
    legend_html = render_legend()

    return f'''<!doctype html>
<html lang="zh">
<head>
  <meta charset="utf-8">
  <title>{he(page_title)}</title>
  <style>{CSS}</style>
</head>
<body>
  <div class="scroll-area">
    <header class="page-header">
      <h1>{he(page_title)}</h1>
    </header>
    {source_html}
    {legend_html}
    {procedures_html}
  </div>
  <div class="drawer-overlay" id="drawer-overlay"></div>
  <aside class="drawer" id="drawer" aria-hidden="true">
    <header>
      <h2>详情</h2>
      <button class="close" aria-label="关闭">×</button>
    </header>
    <div class="content"></div>
  </aside>
  <script>{js}</script>
</body>
</html>
'''