[lanyue-llk]

The Web Fetch harness could not fairly run and explain a browser-native comparison across seven task families. Native Markdown, rendered HTML, embedded resources, direct files, retries, scoring, and failure review were handled by separate paths. The first implementation also stored one report's profile, rerun scope, release prose, and participant identities in Git.

This change makes those boundaries reusable:

• Route webpage and document-body tasks through each browser's declared native content interface; Chromium alone uses the fixed HTML-to-Markdown converter. Embedded-media tasks inspect addresses already present in that output, while direct-media tasks verify original bytes. The harness does not switch interfaces after a failure or add an external document parser.
• Keep dynamic completion at task-family or registered-contract scope. Per-case selectors, answer-dependent waits, and provider-specific scoring branches remain rejected.
• Preserve visible record ownership through HTML and Markdown normalization. Image alternative text cannot replace a required visible name, and fields from adjacent cards, lists, wrappers, or asides cannot complete another record.
• Normalize bounded main-navigation transport failures before the shared retry decision, preserve native DNS context for attachment calls, retain explicit loopback fixtures, and keep experiment-specific DNS settings out of portable defaults.
• Build the seven-family, 430-case dataset from hash-bound reusable sources and corrections. A report profile is an explicit external input; the repository does not store a report's profile or rerun receipt.
• Read report release history and pricing overrides from hash-bound report-assets when supplied. The root release registry is a legacy fallback, and the shared pricing catalog contains product facts rather than one run's wording.
• Use generic Moli release/candidate configuration slots. Actual versions, commits, executable hashes, review payloads, prose, and publication state are bound by external evidence.

The published report remains on Artifact Site, with only a compact version/profile-hash archive binding in Git: https://artifact-site.lexmount.cn/v/IqGpMUYHWat-dw4n8usQttTptoPzboAo

Reproduce the 430-case definition

Use the report profile from the evidence package. It fixes the participants, ordered case inventory, run policy, executable identities, and report output without putting one report instance back into Git.

uv run --frozen python scripts/assemble_webfetch_dataset.py \
  --assembly evaluation/webfetch-seven-family-430/dataset-assembly.json \
  --profile <report-profile.json> \
  --output /tmp/webfetch-430.json

Assembly rejects a missing profile, stale source hash, mismatched case ownership, unregistered correction control, or report-local data embedded in the reusable authority.

Reproduce representative Browser Eval cases

These are real cases from the compiled 430-case definition, rather than fixtures made for this PR:

• document-01: direct PDF body delivery. A method passes only when it returns readable document content.
• dynamic-104-mui-data-grid: native dynamic-page completion and structured content.
• replacement-ko-yonhap-sports: a live dynamic page that exercises the declared failure-retry path.
• risk-diverse-134: record ownership and body acceptance on a risk-controlled page.

Run both Moli builds in alternating order so they see the same case sequence. Replace the two external files with the hash-bound files from the evidence package and point their generic Moli slots at the binaries under review.

uv run --frozen python scripts/run_certified_benchmark.py \
  --dataset /tmp/webfetch-430.json \
  --output /tmp/webfetch-badcases-run \
  --work-root /tmp/webfetch-badcases-work \
  --providers moli-release,moli-candidate \
  --case-ids document-01,dynamic-104-mui-data-grid,replacement-ko-yonhap-sports,risk-diverse-134 \
  --provider-config <providers.json> \
  --workers 1 \
  --call-order case-alternating \
  --retrieval-mode native

The certified output retains every physical call and always uses the final declared attempt. The comparison audit checks the two binary identities, serial call order, native route, diagnostic log, output hash, and complete case matrix.

Validation

At merge, the focused architecture, scoring, report-content, pricing, and participant tests passed (83 tests and 14 subtests), all 240 evaluation/research/configuration JSON files parsed, Ruff F checks passed, and git diff --check passed.

Post-merge verification on main at 28534e24 reran the affected evidence, native-route, report-interface, and method tests: 42 tests and 45 subtests passed. The same code rescored the four saved Browser Eval cases above without changing their verdicts:

• document-01: fail — Moli does not support the required readable PDF output.
• dynamic-104-mui-data-grid: pass.
• replacement-ko-yonhap-sports: fail — page loading/extraction timed out.
• risk-diverse-134: pass.

This verification used saved, hash-bound outputs. It made no provider calls and did not alter the published results.

[github-actions]

Claude encountered an error after 3s —— View job

---

I'll analyze this and get back to you.

[github-actions]

Claude encountered an error after 3s —— View job

---

I'll analyze this and get back to you.

[github-actions]

Claude encountered an error after 3s —— View job

---

I'll analyze this and get back to you.

[github-actions]

Claude encountered an error after 3s —— View job

---

I'll analyze this and get back to you.

[github-actions]

Claude encountered an error after 3s —— View job

---

I'll analyze this and get back to you.

[github-actions]

Claude encountered an error after 3s —— View job

---

I'll analyze this and get back to you.

[github-actions]

Claude encountered an error after 3s —— View job

---

I'll analyze this and get back to you.

[github-actions]

Claude encountered an error after 3s —— View job

---

I'll analyze this and get back to you.

[github-actions]

Claude encountered an error after 3s —— View job

---

I'll analyze this and get back to you.

[github-actions]

Claude encountered an error after 3s —— View job

---

I'll analyze this and get back to you.

[github-actions]

Claude encountered an error after 3s —— View job

---

I'll analyze this and get back to you.

[github-actions]

Claude encountered an error after 3s —— View job

---

I'll analyze this and get back to you.

[lanyue-llk]

审查结论：建议修正下列问题后再合入。

本 PR 将 Web Fetch 扩展为五浏览器、七类任务、每浏览器 430 题，并统一原生接口、正文评分和发布校验。共享任务投影与统一路由选择器是合适的抽象；保留动态任务分类也合理，但调用校验没有同步迁移。资源投影还把单个 URL 当成 srcset 列表解析，导致合法资源地址损坏。

确认两处功能回退，均做了基线对照：目标分支通过，PR 失败；在隔离副本中验证最小修复后，相关测试 63 项通过。现有相关测试分两组共 478 项通过、13 项跳过；启用四种真实本地浏览器后，资源、取消和文件测试 24 项通过、1 项失败，失败原因见第三条。进程清理失败后的隔离行为符合现有安全合同，不列为缺陷。

这次审查覆盖代码、调用合同及相关测试，不改变报告中独立逐结果复核 0/2,150 的范围。

[github-actions]

Claude encountered an error after 3s —— View job

---

I'll analyze this and get back to you.