{% extends "base.html" %}

{% block htmltitle %}
{# Pages with a frontmatter title (home + conf/area index pages) use it verbatim,
   dropping Material's " - {site_name}" suffix that otherwise eats ~17 width units
   and truncates the SEO hook in Google SERPs. Paper pages have no frontmatter
   title and keep the default suffix via super(). #}
{%- if page and page.meta and page.meta.title %}
  <title>{{ page.meta.title }}</title>
{%- elif page and page.is_homepage %}
  <title>{{ config.site_name }}</title>
{%- else %}
  {{ super() }}
{%- endif %}
{% endblock %}

{% block extrahead %}
{% if config.extra.bing_site_verification %}
<meta name="msvalidate.01" content="{{ config.extra.bing_site_verification }}" />
{% endif %}

{# Cloudflare Web Analytics — cookieless, no consent banner needed.
   Load only on the canonical host so legacy GitHub Pages copies and previews
   cannot contaminate production analytics. #}
{% if config.extra.cloudflare_analytics_token %}
<script data-site-url="{{ config.site_url | e }}"
        data-analytics-token="{{ config.extra.cloudflare_analytics_token | e }}">
(function() {
  var analyticsConfig = document.currentScript.dataset;
  var canonicalHost = new URL(analyticsConfig.siteUrl).hostname;
  if (window.location.hostname !== canonicalHost) return;

  var beacon = document.createElement('script');
  beacon.src = 'https://static.cloudflareinsights.com/beacon.min.js';
  beacon.dataset.cfBeacon = JSON.stringify({
    token: analyticsConfig.analyticsToken
  });
  document.head.appendChild(beacon);
})();
</script>
{% endif %}

{% if page and page.url == 'search/' %}
<meta name="robots" content="noindex,follow" />
{% else %}
<meta name="robots" content="index,follow,max-snippet:180,max-image-preview:large" />
<meta name="googlebot" content="index,follow,max-snippet:180,max-image-preview:large" />
{% endif %}

<script>
(function() {
  function markChromeNoSnippet() {
    var selectors = [
      '.md-header', '.md-tabs', '.md-sidebar', '.md-search',
      '.md-footer', '.md-source', '.md-top'
    ];
    selectors.forEach(function(selector) {
      document.querySelectorAll(selector).forEach(function(node) {
        node.setAttribute('data-nosnippet', '');
      });
    });
  }
  if (document.readyState === 'loading') {
    document.addEventListener('DOMContentLoaded', markChromeNoSnippet);
  } else {
    markChromeNoSnippet();
  }
})();
</script>

{# hreflang: bilingual alternate URLs for zh / en sites.
   Only emitted for pages that exist on BOTH sites — the set is loaded by
   hooks/hreflang.py from .hreflang_pairs.json (generated by
   src/gen_hreflang_pairs.py in the parent Auto-Research repo).
   Avoids 404 hreflang clusters which Google penalises. #}
{% set _alt_pages = config.extra.hreflang_pairs %}
{% if _alt_pages and page and page.url in _alt_pages %}
<link rel="alternate" hreflang="zh-CN" href="https://papernotes.org/{{ page.url }}" />
<link rel="alternate" hreflang="en" href="https://en.papernotes.org/{{ page.url }}" />
<link rel="alternate" hreflang="x-default" href="https://papernotes.org/{{ page.url }}" />
{% endif %}

{% if page %}

{# Detect page type — frontmatter title already carries any '[论文解读]' prefix
   from src/add_seo.py, so the template never adds another prefix. #}
{% set _src = (page.file.src_uri if page and page.file else '') %}
{% set _is_index = _src.endswith('index.md') %}
{% set _is_paper = page and not page.is_homepage and not _is_index %}

{# Prefer frontmatter `title:` (SEO-tuned) over nav entry / H1 (page.title). #}
{# _seo_title_raw is the unescaped form for JSON-LD (uses | tojson). #}
{# _seo_title is HTML-escaped for use in OG / Twitter meta attribute values. #}
{% set _seo_title_raw = (page.meta.title if page and page.meta and page.meta.title else page.title) %}
{% set _seo_title = _seo_title_raw | e %}
{% set _desc_raw = (page.meta.description if page and page.meta and page.meta.description else (config.site_description | default(''))) %}
{% set _desc = _desc_raw | e %}
{% set _tags = (page.meta.tags if page and page.meta and page.meta.tags else []) %}

{% if _tags %}
<meta name="keywords" content="{{ _tags | join(', ') | e }}" />
{% if _is_paper %}
{% for tag in _tags %}
<meta property="article:tag" content="{{ tag | e }}" />
{% endfor %}
{% endif %}
{% endif %}

<!-- OpenGraph & Twitter Card -->
<meta property="og:site_name" content="{{ config.site_name }}" />
<meta property="og:type" content="{% if _is_paper %}article{% else %}website{% endif %}" />
<meta property="og:url" content="{{ page.canonical_url }}" />
<meta property="og:title" content="{{ _seo_title }}" />
<meta property="og:description" content="{{ _desc }}" />
<meta name="twitter:card" content="summary" />
<meta name="twitter:title" content="{{ _seo_title }}" />
<meta name="twitter:description" content="{{ _desc }}" />

{# All JSON-LD string values use | tojson (NOT | e) to avoid GSC "Bad escape
   sequence in string" errors. HTML-escape produces &amp; / &quot; etc. which
   pollute JSON output, and never escapes the literal \ that crashes the JSON
   parser when titles/descriptions contain LaTeX (\theta, \boldsymbol, ...) #}
<!-- JSON-LD: Organization (site-wide entity; referenced by @id elsewhere) -->
<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Organization",
  "@id": "https://papernotes.org/#organization",
  "name": "AI Paper Notes",
  "alternateName": "论文笔记",
  "url": "https://papernotes.org/",
  "sameAs": [
    "https://github.com/zhaoyang97/Paper-Notes",
    "https://en.papernotes.org/"
  ]
}
</script>
{% if page.is_homepage %}
<!-- JSON-LD: WebSite (homepage) -->
<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "WebSite",
  "name": {{ config.site_name | tojson }},
  "url": {{ page.canonical_url | tojson }},
  "description": {{ _desc_raw | tojson }},
  "inLanguage": "zh-CN",
  "publisher": { "@id": "https://papernotes.org/#organization" },
  "author": { "@id": "https://papernotes.org/#organization" }
}
</script>
{% elif _is_index %}
<!-- JSON-LD: CollectionPage (conference / area index) -->
<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "CollectionPage",
  "name": {{ _seo_title_raw | tojson }},
  "description": {{ _desc_raw | tojson }},
  "url": {{ page.canonical_url | tojson }},
  "inLanguage": "zh-CN",
  "dateModified": {{ (page.meta.date_modified | default(config.extra.build_date | default(''))) | string | tojson }},
  {% if _tags %}
  "keywords": {{ (_tags | join(', ')) | tojson }},
  "about": [
    {%- for tag in _tags[:10] %}
    { "@type": "Thing", "name": {{ tag | tojson }} }{% if not loop.last %},{% endif %}
    {%- endfor %}
  ],
  {% endif %}
  "publisher": { "@id": "https://papernotes.org/#organization" },
  "isPartOf": {
    "@type": "WebSite",
    "name": {{ config.site_name | tojson }},
    "url": {{ config.site_url | tojson }}
  }{% if page.meta and page.meta.item_list %},
  "mainEntity": {
    "@type": "ItemList",
    "numberOfItems": {{ page.meta.item_total | default(page.meta.item_list | length) }},
    "itemListElement": [
    {%- for it in page.meta.item_list %}
      {
        "@type": "ListItem",
        "position": {{ loop.index }},
        "name": {{ it.t | tojson }},
        "url": {{ (page.canonical_url ~ it.u) | tojson }}
      }{% if not loop.last %},{% endif %}
    {%- endfor %}
    ]
  }{% endif %}
}
</script>
{% else %}
<!-- JSON-LD: ScholarlyArticle (paper page) -->
<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "ScholarlyArticle",
  "name": {{ _seo_title_raw | tojson }},
  "headline": {{ _seo_title_raw | tojson }},
  "description": {{ _desc_raw | tojson }},
  "url": {{ page.canonical_url | tojson }},
  "mainEntityOfPage": {{ page.canonical_url | tojson }},
  "inLanguage": "zh-CN",
  "isAccessibleForFree": true,
  "learningResourceType": "AI paper note",
  "genre": "论文解读",
  {% if _tags %}
  "keywords": {{ (_tags | join(', ')) | tojson }},
  "about": [
    {%- for tag in _tags[:10] %}
    { "@type": "Thing", "name": {{ tag | tojson }} }{% if not loop.last %},{% endif %}
    {%- endfor %}
  ],
  {% endif %}
  "datePublished": {{ (page.meta.date_published | default(config.extra.build_date | default(''))) | string | tojson }},
  "dateModified": {{ (page.meta.date_modified | default(page.meta.date_published | default(config.extra.build_date | default('')))) | string | tojson }},
  "author": { "@id": "https://papernotes.org/#organization" },
  "publisher": { "@id": "https://papernotes.org/#organization" }
}
</script>
{% endif %}

<!-- JSON-LD: BreadcrumbList -->
{# Build crumbs: every ListItem MUST have an "item" URL (Google requirement). #}
{# Derive ancestor URLs from canonical_url segments, dedupe against current page. #}
{% if page and not page.is_homepage %}
{% set _site = config.site_url.rstrip('/') %}
{% set _canon = page.canonical_url.rstrip('/') ~ '/' %}
{% set _rel = _canon[(_site|length + 1):].rstrip('/') %}
{% set _segs = _rel.split('/') if _rel else [] %}
{% set _ancestors = page.ancestors | reverse | list %}
{% set ns = namespace(items=[{'name': 'AI Paper Notes', 'url': _site ~ '/'}]) %}
{% for anc in _ancestors %}
  {% set _depth = loop.index %}
  {% if _depth <= _segs|length %}
    {% set _url = _site ~ '/' ~ (_segs[:_depth] | join('/')) ~ '/' %}
    {% if _url != _canon %}
      {% set ns.items = ns.items + [{'name': anc.title, 'url': _url}] %}
    {% endif %}
  {% endif %}
{% endfor %}
{% set ns.items = ns.items + [{'name': page.title, 'url': _canon}] %}
<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "BreadcrumbList",
  "itemListElement": [
  {%- for c in ns.items %}
    {
      "@type": "ListItem",
      "position": {{ loop.index }},
      "name": {{ c.name | tojson }},
      "item": {{ c.url | tojson }}
    }{% if not loop.last %},{% endif %}
  {%- endfor %}
  ]
}
</script>
{% endif %}
{% endif %}

<!-- Lazy-load search worker: defer search_index.json download until user opens search -->
<!-- Index is title-only (~7MB, slimmed by hooks/slim_search_index.py post-build) -->
<!-- Mobile: redirect to Bing site search (no search index download at all) -->
<!-- Desktop: create search worker lazily only when user focuses on search input -->
<script>
(function() {
  var isMobile = /Android|iPhone|iPad|iPod/i.test(navigator.userAgent);

  // ── Bilingual query expansion mapping (中↔英) ──
  var P = [
    "上下文学习|in-context learning","持续学习|continual learning","课程学习|curriculum learning",
    "小样本学习|few-shot learning","零样本|zero-shot","元学习|meta-learning",
    "模仿学习|imitation learning","主动学习|active learning","自监督学习|self-supervised learning",
    "多任务学习|multi-task learning","迁移学习|transfer learning","强化学习|reinforcement learning",
    "对比学习|contrastive learning","知识蒸馏|knowledge distillation","联邦学习|federated learning",
    "微调|fine-tuning","预训练|pre-training","指令微调|instruction tuning",
    "合成数据|synthetic data","数据增强|data augmentation",
    "推理|reasoning","空间推理|spatial reasoning","视觉推理|visual reasoning",
    "幻觉|hallucination","越狱攻击|jailbreak","对齐|alignment","奖励模型|reward model",
    "偏好优化|preference optimization","推测解码|speculative decoding","长上下文|long context",
    "知识编辑|knowledge editing","知识图谱|knowledge graph","代码生成|code generation",
    "多智能体|multi-agent","具身智能|embodied ai","视觉定位|visual grounding",
    "文生图|text-to-image","视频生成|video generation","新视角合成|novel view synthesis",
    "运动生成|motion generation","鲁棒性|robustness","对抗攻击|adversarial attack",
    "差分隐私|differential privacy","可解释性|interpretability","泛化|generalization",
    "不确定性|uncertainty","异常检测|anomaly detection","因果推理|causal inference",
    "基础模型|foundation model","世界模型|world model","扩散模型|diffusion model",
    "目标检测|object detection","语义分割|semantic segmentation","实例分割|instance segmentation",
    "图像分割|image segmentation","点云|point cloud","深度估计|depth estimation",
    "光流|optical flow","姿态估计|pose estimation","图像复原|image restoration",
    "图像生成|image generation","图像编辑|image editing","超分辨率|super-resolution",
    "视频理解|video understanding","动作识别|action recognition","视觉问答|visual question answering",
    "文生视频|text-to-video","语音识别|speech recognition","自动驾驶|autonomous driving",
    "遥感|remote sensing","医学影像|medical imaging","模型压缩|model compression",
    "量化|quantization","多模态|multi-modal","跨模态|cross-modal",
    "图神经网络|graph neural network","注意力机制|attention mechanism",
    "域适应|domain adaptation","领域泛化|domain generalization","分布外|out-of-distribution",
    "开放词汇|open-vocabulary","人脸识别|face recognition","机器翻译|machine translation",
    "情感分析|sentiment analysis","文档理解|document understanding","问答|question answering",
    "轨迹预测|trajectory prediction","图像修复|image inpainting","去噪|denoising",
    "模型剪枝|model pruning","参数高效微调|parameter-efficient fine-tuning",
    "灾难性遗忘|catastrophic forgetting","3D目标检测|3d object detection",
  ];
  var M = {};
  P.forEach(function(p) { var s = p.split('|'); M[s[0]] = s[1]; M[s[1]] = s[0]; });

  function expand(q) {
    var ql = q.toLowerCase().trim();
    if (M[ql]) return q + ' ' + M[ql];
    var extra = [], seen = {};
    for (var t in M) {
      if (ql.indexOf(t) !== -1 && !seen[M[t]]) { seen[M[t]] = true; extra.push(M[t]); }
    }
    return extra.length ? q + ' ' + extra.join(' ') : q;
  }
  window.__pnExpand = expand;

  // ── Press Enter in the header search box → open the full /search/ results page ──
  document.addEventListener('keydown', function(e) {
    if (e.key !== 'Enter') return;
    var t = e.target;
    if (!t || !t.classList || !t.classList.contains('md-search__input')) return;
    e.preventDefault(); e.stopPropagation();
    var i = location.pathname.indexOf('/search/');
    var b = i >= 0 ? location.pathname.slice(0, i + 1) : '/';
    var v = t.value.trim();
    location.href = b + 'search/' + (v ? '?q=' + encodeURIComponent(v) : '');
  }, true);

  // ── Mobile: disable search worker entirely, use Bing ──
  if (isMobile) {
    if (window.Worker) {
      var origWorkerMobile = window.Worker;
      window.Worker = function(url) {
        if (typeof url === 'string' && url.indexOf('search') !== -1) {
          return { postMessage: function(){}, terminate: function(){}, onmessage: null };
        }
        return new origWorkerMobile(url);
      };
    }
    document.addEventListener('click', function(e) {
      var el = e.target.closest('label[for="__search"], .md-search__icon, .md-header__button[data-md-component="search"]');
      if (!el) return;
      e.preventDefault(); e.stopPropagation();
      var q = prompt('搜索论文笔记');
      if (q && q.trim()) {
        window.open('https://www.bing.com/search?q=site:papernotes.org+' + encodeURIComponent(q.trim()), '_blank');
      }
    }, true);
    return;
  }

  // ── Desktop: lazy-load search worker (only download index when user opens search) ──
  if (window.Worker) {
    var OrigWorker = window.Worker;
    window.Worker = function(url, opts) {
      // Non-search workers: create immediately
      if (typeof url !== 'string' || url.indexOf('search') === -1) {
        return new OrigWorker(url, opts);
      }

      // Search worker: create a lazy proxy
      var realWorker = null;
      var queued = [];
      var proxy = {
        onmessage: null,
        onerror: null,
        terminate: function() { if (realWorker) realWorker.terminate(); },
        postMessage: function(msg) {
          // Sanitize + bilingual-expand the query before it reaches lunr.
          if (msg && typeof msg.data === 'string' && msg.data.length > 0 && msg.data.length < 200) {
            // Strip lunr query-syntax metacharacters. A literal ':' (field
            // separator), '~' (fuzzy), '^' (boost) or '*' (wildcard) in the raw
            // text makes lunr throw QueryParseError -> 0 results. This is why
            // pasting a paper's full title like "Cog-RAG: Cognitive-Inspired ..."
            // returned "no results": the ':' broke the query parser.
            var cleaned = msg.data.replace(/[:~^*]/g, ' ').replace(/\s+/g, ' ').trim();
            msg = Object.assign({}, msg, { data: expand(cleaned) });
          }
          if (realWorker) {
            realWorker.postMessage(msg);
          } else {
            queued.push(msg);
          }
        },
        addEventListener: function(type, fn) {
          if (type === 'message') proxy.onmessage = fn;
          else if (type === 'error') proxy.onerror = fn;
        },
        removeEventListener: function() {}
      };

      // Materialize the real worker (called once when user opens search)
      function materialize() {
        if (realWorker) return;
        realWorker = new OrigWorker(url, opts);
        realWorker.onmessage = function(e) {
          if (typeof proxy.onmessage === 'function') proxy.onmessage(e);
        };
        realWorker.onerror = function(e) {
          if (typeof proxy.onerror === 'function') proxy.onerror(e);
        };
        // Flush queued messages
        for (var i = 0; i < queued.length; i++) realWorker.postMessage(queued[i]);
        queued = [];
      }

      // Trigger worker creation when user opens search
      function onSearchOpen() {
        materialize();
        // Clean up listeners after first trigger
        document.removeEventListener('focusin', onFocusIn, true);
        var cb = document.getElementById('__search');
        if (cb) cb.removeEventListener('change', onSearchOpen);
      }

      function onFocusIn(e) {
        if (e.target && e.target.classList && e.target.classList.contains('md-search__input')) {
          onSearchOpen();
        }
      }

      // Listen for search input focus or search checkbox toggle
      document.addEventListener('focusin', onFocusIn, true);
      var searchCheckbox = document.getElementById('__search');
      if (searchCheckbox) {
        searchCheckbox.addEventListener('change', onSearchOpen);
      }

      return proxy;
    };
    window.Worker.prototype = OrigWorker.prototype;
  }
})();
</script>

<!-- Add deep-search fallback link in search results (both mobile and desktop) -->
<script>
(function() {
  // Inject a "全文深度搜索 via Bing" link at the bottom of search results.
  //
  // IMPORTANT: we must NOT use a MutationObserver on `.md-search` with
  // subtree:true and then appendChild into `.md-search-result`, because the
  // append mutates the observed subtree → triggers the observer again → infinite
  // loop → tab OOM ("喔唷，崩溃啦 / Out of Memory"). Instead we listen to the
  // search input's `input` event (which does NOT fire from our own DOM
  // mutations) and debounce updates to the fallback link.
  var lastQuery = null;
  var scheduled = false;

  function updateLink() {
    scheduled = false;
    var input = document.querySelector('.md-search__input');
    var output = document.querySelector('.md-search-result');
    if (!input || !output) return;

    var query = input.value.trim();

    // Only touch the DOM if query actually changed — avoids churn while user
    // is typing and prevents redundant re-renders.
    if (query === lastQuery) return;
    lastQuery = query;

    var existing = output.querySelector('.deep-search-link');
    if (!query) {
      if (existing) existing.remove();
      return;
    }

    var encoded = encodeURIComponent(query);
    var site = 'site:papernotes.org+';
    var html =
      '<a href="https://www.bing.com/search?q=' + site + encoded +
      '" target="_blank" rel="noopener" style="color:var(--md-accent-fg-color);font-size:0.85rem;">' +
      '🔍 全文深度搜索「' + query.replace(/</g, '&lt;').replace(/>/g, '&gt;') + '」(via Bing)</a>' +
      '<span style="display:block;font-size:0.75rem;color:var(--md-default-fg-color--light);margin-top:2px;">' +
      '站内搜索仅匹配标题、标签和关键词，点此搜索全文内容</span>';

    if (existing) {
      existing.innerHTML = html;
    } else {
      var link = document.createElement('div');
      link.className = 'deep-search-link';
      link.style.cssText = 'padding:12px 16px;text-align:center;border-top:1px solid var(--md-default-fg-color--lightest);';
      link.innerHTML = html;
      output.appendChild(link);
    }
  }

  function schedule() {
    if (scheduled) return;
    scheduled = true;
    // Defer to rAF so we merge rapid events and run after Material renders
    // its own results.
    (window.requestAnimationFrame || function(fn){setTimeout(fn,16);})(updateLink);
  }

  document.addEventListener('DOMContentLoaded', function() {
    var input = document.querySelector('.md-search__input');
    if (!input) return;
    input.addEventListener('input', schedule);
    input.addEventListener('focus', schedule);
  });
})();
</script>

<!-- Promote near-exact title matches to the top of search results -->
<script>
(function() {
  // Why: lunr (BM25) length-normalization penalizes long titles, so pasting a
  // paper's FULL title can rank that paper BELOW shorter titles that merely
  // share one common token (e.g. "hypergraph"). We re-rank the *rendered*
  // results client-side: when one result's title covers almost all of the
  // query's words, we move it to the top. We never use a MutationObserver
  // (it would loop → OOM); we re-apply on the input event with a few short
  // delayed passes so it survives Material's async re-render.
  var SEP = /[\s\u200b\-\.\,\!\?\;\:\(\)\[\]\{\}\/\|]+/;
  function toks(s) {
    return (s || '').toLowerCase().split(SEP).filter(function(t){ return t.length >= 2; });
  }
  function promote() {
    var input = document.querySelector('.md-search__input');
    var list = document.querySelector('.md-search-result__list');
    if (!input || !list) return;
    var q = input.value.trim();
    if (q.length < 8) return;                 // only long, title-like queries
    var qToks = toks(q);
    if (qToks.length < 3) return;             // leave short keyword searches alone
    var qSet = {}; qToks.forEach(function(t){ qSet[t] = true; });
    var items = [].slice.call(list.children).filter(function(el){
      return el.classList && el.classList.contains('md-search-result__item');
    });
    if (items.length < 2) return;
    var best = null, bestScore = 0;
    items.forEach(function(it) {
      var h = it.querySelector('h1, .md-search-result__title');
      var tToks = toks(h ? h.textContent : '');
      if (!tToks.length) return;
      var tSet = {}; tToks.forEach(function(t){ tSet[t] = true; });
      var hit = 0; qToks.forEach(function(t){ if (tSet[t]) hit++; });
      var coverage = hit / qToks.length;      // share of query words found in title
      if (coverage > bestScore) { bestScore = coverage; best = it; }
    });
    // Only reorder when the best title clearly IS what the user typed.
    if (best && bestScore >= 0.8 && best !== list.firstElementChild) {
      list.insertBefore(best, list.firstElementChild);
    }
  }
  function schedule() {
    [180, 420, 800].forEach(function(d){ setTimeout(promote, d); });
  }
  document.addEventListener('DOMContentLoaded', function() {
    var input = document.querySelector('.md-search__input');
    if (input) input.addEventListener('input', schedule);
  });
})();
</script>
{% endblock %}

{% block site_nav %}
  {% if nav %}
    {% set hidden = "" %}
    {% if page.meta and page.meta.hide %}
      {% set hidden = "hidden" if "navigation" in page.meta.hide %}
    {% endif %}
    <div class="md-sidebar md-sidebar--primary" data-md-component="sidebar" data-md-type="navigation" data-nosnippet {{ hidden }}>
      <div class="md-sidebar__scrollwrap">
        <div class="md-sidebar__inner">
          {% include "partials/nav.html" %}
        </div>
      </div>
    </div>
  {% endif %}
  {% if "toc.integrate" not in features %}
    {% set hidden = "" %}
    {% if page.meta and page.meta.hide %}
      {% set hidden = "hidden" if "toc" in page.meta.hide %}
    {% endif %}
    <div class="md-sidebar md-sidebar--secondary" data-md-component="sidebar" data-md-type="toc" data-nosnippet {{ hidden }}>
      <div class="md-sidebar__scrollwrap">
        <div class="md-sidebar__inner">
          {% include "partials/toc.html" %}
        </div>
      </div>
    </div>
  {% endif %}
{% endblock %}

{# B3: MathJax library is loaded only on paper pages. The config
   (javascripts/mathjax.js) stays global via extra_javascript and is rendered
   by super() before this script, so window.MathJax is defined first; index /
   home pages skip the ~1 MB tex-mml-chtml download entirely. #}
{% block scripts %}
{{ super() }}
{% set _src = (page.file.src_uri if page and page.file else '') %}
{% if page and not page.is_homepage and not _src.endswith('index.md') %}
<script src="https://cdnjs.cloudflare.com/ajax/libs/mathjax/3.2.2/es5/tex-mml-chtml.min.js"></script>
{% endif %}
{% endblock %}

{% block content %}
{% if page %}
  {% set _src = (page.file.src_uri if page.file else '') %}
  {% set _is_index = _src.endswith('index.md') %}
  {% set _is_paper = not page.is_homepage and not _is_index %}
  {% if _is_paper and page.meta and page.meta.description %}
<aside class="paper-seo-summary" aria-label="论文摘要">
  <p class="paper-seo-summary__desc">{{ page.meta.description }}</p>
  {% if page.meta.tags %}
  <p class="paper-seo-summary__tags">标签：{{ page.meta.tags | join(' · ') }}</p>
  {% endif %}
</aside>
  {% endif %}
{% endif %}
{{ super() }}
{% endblock %}