[lanyue-llk]

问题

Markdown 管道表格无法表达合并单元格、嵌套关系和单元格中的块内容；丢弃文内链接目标也会破坏导航。代码转 HTML 后备路径时，交互控件的排除规则还曾与普通代码路径不一致。有锚点的呈现表会错误退回数据表 HTML。

一图看懂

修改

• 简单表格输出 GFM。复杂表格有表头、标题、数据角色、表头关系或真实跨度时保留作者结构；只有嵌套布局和块内容的包装表展开为普通块。role=presentation/none 同样展开，包括带锚点的情况。

• 仅保留实际被引用的片段目标，保证每个目标只出现一次；覆盖表格、标题及 pre/code 根节点。按 ID 优先、精确匹配后再解码的顺序处理。无需默认禁用文内导航。

• 普通转换与 HTML 后备复用媒体、表单、数学和代码控件的语义判断；格式化与转义由各自输出路径完成。

• 修正后备路径的 Copy 按钮污染。带链接的预格式化代码保留 HTML 链接、语言和物理换行；普通包装中的标题仍输出 Markdown 标题。

• 按 HTML 非负整数规则判断单元格实际跨度：非法值与默认跨度继续使用 GFM；rowspan=0 和真实合并单元格保留结构。

• 空 poster 不生成空图片；已经有块级标签的元素不再增加重复 div，CSS 引入的块边界仍保留。

• 列表、表格和选项节点自行管理结构边界，CSS 补充边界不再重复分段；适配器请求的行内边界在两条输出路径中都用空格分隔。

• HTML 后备共享全局目标去重；图片保留被引用的 ID 与 title，别名放在链接开始标签之前。重新解析输出验证链接文字仍属于原链接。

• 原生 DOM 与浏览器 CLI 使用同一完整 HN 快照；CLI 实际输出严格比较，仅去除期望文件的末尾换行。

代码元素自身可以有补充 CSS 边界，其源码后代不应用这些边界。普通媒体名称优先使用 aria-label，再使用 title。列表的作者结构与计算盒分开处理：单个 CSS 块不改变列表松紧；角色匹配使用 HTML 空白，col/colgroup 不生成额外块。

列表先完成待输出正文，再判断松紧。嵌套列表位于末尾时保持紧凑；其后还有被空行分开的正文时，整个外层列表按松列表输出。

验证

• cargo test -p moli-html2md --locked：398 项通过；Clippy 通过。非法跨度和空 poster 的反例均先失败后通过；另有 CSS 块边界与实时 textarea 的两种输出路径验证。
• 普通代码、带链接代码和复杂表格中的同一组控件场景均通过；旧实现的后备路径会泄漏 Copy。
• 呈现表锚点反例在旧实现中生成 table，修复后保留标题、内容和可达目标，不生成数据表。
• 深层有锚点表格、嵌套标题的节点访问和文本复制量随输入线性增长；复杂布局表不逐层捕获整张表。没有片段链接时不建立 ID 清单。
• 验收覆盖行列、合并关系、内容归属和完整输出。没有表头也可能是合法的合并单元格表，因此保留其关系；不根据站点名称猜测用途。
• HTML 源码后备路径保留源字符、默认语言和开头空行，并通过重新解析核对实际代码文本。

复现

本 PR 的转换库测试可独立运行：cargo test -p moli-html2md --locked。浏览器接入在 #1337，以下 CLI 与冻结页复现使用完整 stack。

最小用例

跨两列的表头保持 colspan；文内链接与目标锚点同时保留。

fixture_html=$(cat <<'HTML'
<table><tr><th colspan="2">Total</th></tr><tr><td>A</td><td>B</td></tr></table><p><a href="#note">Note</a></p><p id="note">Details</p>
HTML
)
"$MOLI_BINARY" fetch "data:text/html;base64,$(printf '%s' "$fixture_html" | base64 | tr -d '\n')" --dump markdown

实际输出：

<table>
<tbody>
<tr>
<th colspan="2">Total</th>
</tr>
<tr>
<td>A</td>
<td>B</td>
</tr>
</tbody>
</table>

[Note](#note)

<a id="note"></a>

Details

真实 WebMainBench 页面

34b822a2… LinuxFocus 文章：保留嵌套表格内容与 colspan 关系。

七个相关页面通过 browser-eval 的现有 runner 共用一次定向运行，7/7 正常导出，并核对上述内容；不是 545 页质量重测。查看冻结输入、实际输出与核对结果。

• Moli：19bfc785；二进制 SHA-256：1c17bba8cb6b0a532144697411d13ec6a3872d1abad96d5872da0ded93d7b8b4。
• browser-eval：345ea093。
• 原始数据集 SHA-256：0efaa4b49a45e320a27fe6e5a0b6aad5b57259fc3321ac3448519cacc74c537e，runner 会校验。保留原网址，代理返回冻结 HTML，外部资源禁止访问，无点击或 DOM 改写。

环境准备与可复制命令

# Moli 仓库：构建完整 stack，供下面的 CLI 复现使用。
git fetch origin refs/pull/1337/head
git switch --detach 19bfc785f2e0114d7d88e17457de7c108a54069d
cargo build -p moli --bin moli --locked
export MOLI_BINARY="$PWD/target/debug/moli"

# 在单独目录准备 browser-eval。
git clone https://github.com/lexmount/browser-eval.git
cd browser-eval
git checkout 345ea09370a633e871670fde2de4ec25b53a4d05
uv sync --frozen
mkdir -p "$HOME/.browser-eval/public-benchmarks/webmainbench-545"
curl --fail --location \
  https://huggingface.co/datasets/opendatalab/WebMainBench/resolve/main/WebMainBench_545.jsonl \
  --output "$HOME/.browser-eval/public-benchmarks/webmainbench-545/WebMainBench_545.jsonl"

# 在 browser-eval 仓库运行；使用 macOS 的 loopback-only 沙箱。
# MOLI_BINARY 为上面构建的 Moli 的绝对路径。
export MOLI_BINARY
repro_root=$(mktemp -d)
cat > "$repro_root/cases.txt" <<'CASES'
75b0c1e9-d112-46d0-adc8-d1d09aa48731
bfaf0844-440c-466b-b55a-f82c88a54a4e
34b822a2-f7d6-4b69-b6af-133a1ca21326
8798f93a-240c-46d7-a8d1-a5fb011279fc
d80cf991-ba87-4068-80b4-3e9bd49ac8fd
72dd5ff7-8cb5-4660-8941-c1846a43a15d
ccb6033c-0a12-4f9c-8c68-794f26129841
CASES
uv run --frozen python scripts/run_webmainbench_benchmark.py \
  --provider moli --case-ids-file "$repro_root/cases.txt" \
  --preserve-original-origin --expected-version 1.1.19 \
  --timeout 90 --max-failure-retries 0 \
  --output "$repro_root/results"
printf '结果目录：%s\n' "$repro_root/results"

results.jsonl 记录成功状态、错误和输入/输出哈希，execution.json 记录二进制身份，artifacts/moli/ 保存实际 Markdown。定向运行只用于定位和回归，不进入正式榜单；报告中的历史成绩仍绑定原测量版本。

依赖

直接基于 #1335。#1334 → #1335 → #1336 → #1337 → #1380

[github-actions]

Spider Bench A/B

✅ All benchmark browser service runs completed; results are informational.

Public HEAD: 184 / 240 rows; 38 / 48 sites produced rows; 10 unexpected empty sites.

✅ Deterministic fixture contract is clean.

Common ancestor 9fc264a72080 → HEAD 803de4b5a824; benchmark order: head-first; workflow: success.

Public 48-site run · informational

Metric	Base	HEAD	Δ
Run status	success	success	—
Extracted rows / contract	184 / 240	184 / 240	0 (0.00%)
Contract fill	76.67%	76.67%	0.00 pp (0.00%)
Sites with rows	38 / 48	38 / 48	0 (0.00%)
Sites meeting row contract	36 / 48	36 / 48	0 (0.00%)
Unexpected empty sites	10	10	0 (0.00%)
Partial-row sites	2	2	0 (0.00%)
Site outcomes	extracted: 37; HTTP error + rows: 1; empty: 7; HTTP error + empty: 1; snapshot error + empty: 2	extracted: 36; HTTP error + rows: 2; empty: 8; HTTP error + empty: 1; snapshot error + empty: 1	—
Observed duration	46.54 s	42.84 s	-3.70 s (-7.95%)
Estimated CPU	40.57 s	36.92 s	-3.65 s (-8.99%)
Peak PSS	240.35 MiB	209.67 MiB	-30.69 MiB (-12.77%)
Peak RSS	243.99 MiB	213.26 MiB	-30.73 MiB (-12.59%)
Peak threads	26	27	+1 (+3.85%)
Mismatched sites	0	0	0
Timeout with items	0	0	0

Category rows	Base	HEAD	Δ
news	35 / 40	35 / 40	0
stocks	20 / 40	20 / 40	0
tech	35 / 40	35 / 40	0
sports	27 / 40	27 / 40	0
games	37 / 40	37 / 40	0
life	30 / 40	30 / 40	0

Public-site content, timing, and memory are noisy. This single A/B run reports evidence only; site outcome counts explain missing rows without treating them as a deterministic regression.

CPU and memory timelines

These are bounded 40-point views of the same complete process-tree samples used by the HTML report. CPU uses 100% per occupied logical core; memory lines are RSS first and PSS second when complete PSS samples are available.

Base

xychart-beta
    title "Base CPU"
    x-axis "Elapsed seconds" [1.036, 1.537, 2.538, 4.039, 5.04, 6.541, 8.043, 9.044, 10.547, 11.549, 13.052, 14.054, 15.556, 16.558, 17.058, 18.06, 19.561, 20.562, 22.066, 23.569, 24.57, 26.072, 27.073, 28.575, 29.576, 31.078, 32.58, 33.581, 34.583, 35.083, 36.086, 37.588, 39.088, 40.089, 41.59, 42.591, 44.094, 45.094]
    y-axis "CPU percent" 0 --> 200
    line [23.98, 131.89, 119.88, 141.96, 133.96, 109.98, 89.91, 117.77, 123.73, 103.87, 93.85, 75.82, 37.94, 0, 4, 4, 2, 35.94, 73.85, 125.66, 99.99, 101.86, 113.86, 131.91, 109.93, 99.93, 111.94, 5.99, 145.92, 79.95, 29.95, 91.78, 99.9, 100.01, 101.89, 121.77, 79.79, 105.97]

Loading

xychart-beta
    title "Base memory: RSS then PSS"
    x-axis "Elapsed seconds" [0.036, 1.036, 1.537, 2.538, 4.039, 5.04, 6.541, 8.043, 9.044, 10.547, 11.549, 13.052, 14.054, 15.556, 16.558, 17.058, 18.06, 19.561, 20.562, 22.066, 23.569, 24.57, 26.072, 27.073, 28.575, 29.576, 31.078, 32.58, 33.581, 34.583, 35.083, 36.086, 37.588, 39.088, 40.089, 41.59, 42.591, 44.094, 45.094, 46.538]
    y-axis "Memory MiB" 0 --> 500
    line [13.07, 81.93, 99.28, 103.37, 137.92, 152.79, 161.22, 128.25, 150.24, 163.59, 144.41, 135.17, 148.13, 125.74, 121.86, 120.74, 120.76, 120.39, 128.62, 120.38, 145.16, 243.99, 183.54, 208.8, 136.61, 168.27, 132.58, 148.04, 147.27, 153.18, 140.94, 140.68, 142.68, 161.92, 161.92, 161.93, 164.42, 140.66, 149.04, 0]
    line [9.62, 78.42, 95.66, 99.74, 134.3, 149.21, 157.6, 124.62, 146.61, 159.94, 140.76, 131.53, 144.9, 122.09, 118.21, 117.09, 117.11, 116.76, 124.92, 116.76, 140.78, 240.35, 180.17, 205.13, 131.68, 164.53, 128.8, 144.37, 138.79, 149.51, 137.33, 137.03, 139.03, 158.28, 158.28, 158.29, 160.78, 137.02, 145.63, 0]

Loading

HEAD

xychart-beta
    title "HEAD CPU"
    x-axis "Elapsed seconds" [1.035, 2.036, 2.537, 3.538, 4.539, 5.039, 5.539, 6.039, 7.04, 8.541, 9.543, 11.046, 12.046, 13.048, 14.551, 15.552, 16.554, 18.057, 19.059, 20.561, 21.563, 22.564, 24.066, 25.066, 26.568, 27.568, 28.568, 30.07, 31.071, 32.571, 33.573, 34.573, 36.075, 37.076, 38.078, 39.58, 40.582, 42.082]
    y-axis "CPU percent" 0 --> 200
    line [31.99, 95.89, 127.93, 93.94, 149.91, 122.03, 131.94, 124, 143.78, 63.97, 49.89, 111.97, 0, 31.95, 107.75, 87.96, 0, 125.76, 63.87, 93.79, 97.93, 123.86, 105.87, 132.01, 133.89, 110.06, 57.93, 27.92, 89.99, 19.99, 47.96, 123.9, 61.96, 5.99, 99.89, 147.85, 59.97, 135.85]

Loading

xychart-beta
    title "HEAD memory: RSS then PSS"
    x-axis "Elapsed seconds" [0.035, 1.035, 2.036, 2.537, 3.538, 4.539, 5.039, 5.539, 6.039, 7.04, 8.541, 9.543, 11.046, 12.046, 13.048, 14.551, 15.552, 16.554, 18.057, 19.059, 20.561, 21.563, 22.564, 24.066, 25.066, 26.568, 27.568, 28.568, 30.07, 31.071, 32.571, 33.573, 34.573, 36.075, 37.076, 38.078, 39.58, 40.582, 42.082, 42.837]
    y-axis "Memory MiB" 0 --> 500
    line [13.13, 84.43, 118.09, 114.46, 112.16, 131.06, 145.91, 147.95, 153.27, 144.38, 127.58, 137.21, 138.15, 111.17, 119.99, 152.31, 136.47, 127.55, 149.8, 118.35, 156.1, 167.92, 161.78, 213.26, 162.51, 147.89, 140.64, 128.52, 132.61, 138.43, 134.37, 142.11, 183.42, 138.56, 124.17, 146.55, 132.47, 126.68, 158.07, 0]
    line [9.76, 80.84, 114.52, 110.87, 108.57, 127.32, 142.76, 144.39, 149.66, 140.81, 124.07, 133.62, 134.31, 107.54, 116.36, 148.5, 132.89, 123.92, 146.9, 114.72, 150.35, 164.48, 158.13, 209.67, 158.87, 144.27, 136.07, 124.9, 128.96, 134.83, 130.73, 138.47, 179.23, 134.9, 120.5, 144.32, 128.81, 123.02, 154.48, 0]

Loading

Deterministic fixture · required contract

The fixture exercises 15 routes: 8 are expected to emit rows and 7 intentionally exercise empty, loading, or timeout behavior.

Metric	Base	HEAD	Δ
Run status	success	success	—
Extracted rows / contract	40 / 40	40 / 40	0 (0.00%)
Contract fill	100.00%	100.00%	0.00 pp (0.00%)
Sites with rows	8 / 15	8 / 15	0 (0.00%)
Sites meeting row contract	15 / 15	15 / 15	0 (0.00%)
Unexpected empty sites	0	0	0
Partial-row sites	0	0	0
Site outcomes	extracted: 7; HTTP error + rows: 1; empty: 3; timeout + empty: 4	extracted: 7; HTTP error + rows: 1; empty: 3; timeout + empty: 4	—
Observed duration	19.86 s	19.85 s	-0.00 s (0.00%)
Estimated CPU	0.42 s	0.42 s	0.00 s (0.00%)
Peak PSS	66.46 MiB	66.50 MiB	+0.04 MiB (+0.06%)
Peak RSS	69.89 MiB	70.14 MiB	+0.25 MiB (+0.36%)
Peak threads	18	18	0 (0.00%)
Mismatched sites	0	0	0
Timeout with items	0	0	0

Full HTML, JSON, CSV, logs, and page snapshots: workflow run and full spider-bench-results artifact.

The public-web A/B is informational. The exact HEAD fixture contract runs as a separate required CI check; timing and public-site content deltas remain non-blocking.

[github-actions]

CI Regression Report

Source CI run · source state at render: in_progress · artifacts: 6/6

Check	Status	Signal
Release regression	✅	HEAD/base failures 0/0; raw binary +0.022795%
Frontend differential	✅	1,020/1,020 cases matched; 0 issues
Agent episodes	✅	8/8 Moli episodes passed; 0 failures
Runtime/CDP contracts	✅	26 contract cases; 0 failures
CDP smoke	✅	53/53 groups passed; 625 scenarios
WebMainBench	✅	545/545 completed; 0 unexpected failures

Release regression — ✅ HEAD/base failures 0/0; raw binary +0.022795%

Package and image size

Metric	Base	HEAD	Delta	Delta %
Raw binary	150.25 MiB	150.28 MiB	+35.07 KiB	+0.022795%
Stripped binary	103.32 MiB	103.35 MiB	+32.00 KiB	+0.030246%
gzip binary	51.60 MiB	51.62 MiB	+19.76 KiB	+0.037397%
Rootfs	157.15 MiB	157.18 MiB	+30.00 KiB	+0.018643%
gzip rootfs	60.51 MiB	60.53 MiB	+19.97 KiB	+0.032233%

Startup latency and PSS

Case	Base p50	HEAD p50	Delta	Base PSS	HEAD PSS	Delta
Serve ready	26.79 ms	27.58 ms	+0.79 ms	5.31 MiB	4.54 MiB	−786.50 KiB
CDP first page	49.12 ms	51.87 ms	+2.74 ms	4.69 MiB	4.61 MiB	−78.50 KiB
CDP warm page	12.03 ms	13.52 ms	+1.48 ms	50.81 MiB	50.03 MiB	−799.00 KiB
CLI about:blank	15.60 ms	17.08 ms	+1.48 ms	4.70 MiB	3.40 MiB	−1.30 MiB
CLI local JS	24.65 ms	18.58 ms	−6.06 ms	4.40 MiB	4.92 MiB	+534.50 KiB

Concurrency matrix

Concurrency	Base p50	HEAD p50	Delta	Base PSS	HEAD PSS	Delta	HEAD failures	Unstable B/H
1	17.46 ms	17.37 ms	−0.09 ms	898.00 KiB	777.00 KiB	−121.00 KiB	0	0/0
5	20.68 ms	20.30 ms	−0.38 ms	28.18 MiB	14.09 MiB	−14.09 MiB	0	0/0
10	20.84 ms	21.09 ms	+0.25 ms	28.17 MiB	29.31 MiB	+1.14 MiB	0	0/0
25	20.28 ms	20.39 ms	+0.11 ms	13.85 MiB	28.33 MiB	+14.49 MiB	0	0/0
100	20.92 ms	19.72 ms	−1.21 ms	12.81 MiB	27.74 MiB	+14.93 MiB	0	0/0

Frontend differential — ✅ 1,020/1,020 cases matched; 0 issues

Match	DOM mismatch	Diagnostic mismatch	Moli error	Reference error	Infrastructure error	Mismatched frames	Duration
1,020	0	0	0	0	0	0	6m 32.20s

Agent episodes · Moli vs Chromium — ✅ 8/8 Moli episodes passed; 0 failures

Fast contract	Episodes	Passed	Assertions	Ready p50	Episode p50
Moli	8	8	47/47	2.76 ms	25.82 ms
Chromium	8	8	47/47	261.70 ms	84.94 ms

Operation p50	Chromium	Moli	Delta	Delta %
navigate	41.05 ms	9.02 ms	−32.03 ms	−78.02%
observe	0.91 ms	1.08 ms	+0.17 ms	+19.25%
fill	3.64 ms	3.39 ms	−0.26 ms	−7.01%
click	4.28 ms	4.07 ms	−0.21 ms	−4.85%

Idle-resource episode	Peak RSS	Peak PSS	Average CPU	Peak CPU	Peak processes	Sampler
Moli	66.18 MiB	63.58 MiB	0.71%	8.03%	1	✅
Chromium	1280.59 MiB	327.37 MiB	0.51%	8.04%	13	⚠️ PSS partial

Runtime and CDP session contracts — ✅ 26 contract cases; 0 failures

Suite	Cases	Failures	Median case p50	Median case p95	Median peak PSS p50
Synthetic fetch	13	0	17.81 ms	18.49 ms	837.00 KiB
Long-lived CDP session	13	0	18.01 ms	18.43 ms	—

CDP smoke — ✅ 53/53 groups passed; 625 scenarios

Workers: 4 · cumulative group time: 201.90 s · failed groups: 0

Slowest group	Scenarios	Duration	Status
scrollbar-visibility	22	29.10 s	✅
target-lifecycle	24	17.44 s	✅
dom-input	27	13.41 s	✅
multi-page	45	12.22 s	✅
chromium-cdp	56	11.26 s	✅

WebMainBench · 545 pages — ✅ 545/545 completed; 0 unexpected failures

Completed	Successful	Expected DNS failure	Unexpected failures	Panics	Crashes	Timeouts	Empty output
545/545	544	1	0	0	0	0	0

Frozen HTML · Linux · --wait done · 45 seconds per page · no retries. Content quality is evaluated separately.

Diagnostics: webmainbench-results — every page's Markdown, stderr, and exit status. Source run and artifacts

All artifact fields are parsed by the trusted default-branch renderer; missing or invalid inputs remain visible as unavailable.

[github-actions]

Sequential Navigation Soak A/B

✅ HEAD completed the 200-navigation resilience and memory observation.

One browser process, one target, and one CDP session navigate CSDN → SegmentFault → Huaban → example.com repeatedly. This run issued 200 Page.navigate commands per binary.

Common ancestor 9fc264a720 → HEAD 803de4b5a8; benchmark order: head-first; workflow: success.

Session resilience

Metric	Base	HEAD
Attempted / planned	200 / 200	200 / 200
Direct observable pass	195 / 200	197 / 200
Failures / recovered	5 / 5	3 / 3
Recovery failures	0	0
Lifecycle order violations	0	0
Network order violations	0	0
Wall time	612.0 s	578.2 s

A public-page failure is reported separately from an unrecoverable session. The soak requires all navigation attempts, zero failed recovery, zero lifecycle/network ordering violations, and complete resource evidence.

Process-tree memory

Metric	Base	HEAD	Δ
Boundary samples	200	200	0
Boundary RSS samples	200	200	0
Boundary PSS samples	200	200	0
Periodic peak RSS	267.6 MiB	261.9 MiB	-5.7 MiB
Periodic peak PSS	264.7 MiB	258.6 MiB	-6.1 MiB
RSS · first 10 avg	185.0 MiB	186.8 MiB	+1.9 MiB
RSS · last 10 avg	201.0 MiB	188.0 MiB	-13.0 MiB
RSS · last − first window	+16.1 MiB	+1.2 MiB	-14.9 MiB
RSS · warm slope / 100 nav	+5.7 MiB	+7.7 MiB	+2.0 MiB
PSS · first 10 avg	181.3 MiB	183.3 MiB	+2.0 MiB
PSS · last 10 avg	197.6 MiB	184.3 MiB	-13.4 MiB
PSS · last − first window	+16.4 MiB	+1.0 MiB	-15.4 MiB
PSS · warm slope / 100 nav	+5.6 MiB	+7.4 MiB	+1.9 MiB
Peak file descriptors	57	57	0
Peak threads	29	29	0

HEAD memory by 50-navigation quarter

HEAD range	Avg RSS	Final RSS	Peak RSS	Avg PSS	Final PSS	Peak FD
1–50	180.1 MiB	163.4 MiB	230.7 MiB	176.7 MiB	160.2 MiB	52
51–100	182.5 MiB	194.6 MiB	227.1 MiB	179.2 MiB	191.2 MiB	53
101–150	190.9 MiB	165.2 MiB	225.3 MiB	187.4 MiB	161.7 MiB	52
151–200	187.8 MiB	130.6 MiB	226.6 MiB	184.1 MiB	127.0 MiB	53

Raw reports and per-navigation boundary samples: workflow run and full sequential-navigation-soak-results artifact.

Public-site timing and memory are observational. Use the A/B deltas and trend shape as evidence, not as a deterministic performance threshold.

[lanyue-llk]

Closing after the consolidation audit. This belongs to an earlier experimental Markdown stack and was not part of the measured fixed build. Current failure evidence does not justify carrying its broad table and fragment expansion into #1501, so it is intentionally not merged.