[Maisiqi]

Summary

Web Fetch's retained ten-method comparison now has a reusable configured cohort. Select --providers configured --provider-config configs/providers.example.json through the existing validation, diagnostic, benchmark or certification entry point. The shared per-call default changes from 45 to 30 seconds; the existing default of one failure retry remains. Each retry receives a fresh budget, retains its physical evidence, and selects the final response rather than the best response.

• Add one shared Markdown adapter for Bright Data, Browser Use Fetch, Browserbase Fetch, DeepSeek Fetch API and Crawl4AI Cloud, with raw-response projection, redacted caller diagnostics and bilingual method/account documentation through the existing catalog template mechanism.
• Configure Moli, Firecrawl, Cloudflare Markdown, Tavily, Exa and the five hosted methods with fixed native parameters. Exa uniformly requests up to 100 rich links. Remove Moli's per-host private-network exceptions; retain its security checks and pinned binary identity.
• Reject per-case retrieval assistance before transport. Use native URL routes without helper document downloads. Preserve the 20 MB HTTP response/Moli output ceiling without imposing a Moli subresource-download cap.
• Stop fixed-native timing before scorer-input preparation. Preserve retry cost and call-sequence time separately from final-call latency. Timed-out live workers remain quarantined.
• Keep the supported diagnostic runner and its --measure-only option. No new --collect-only mode or hosted JSON/schema/structured output is included. Historical unscored collections still cannot silently become scored reports.

Validation

• 561 targeted offline tests and 31 subtests passed. The 15 CLI tests requiring a local loopback server were rerun outside the restricted sandbox; no real provider APIs were used.
• Repository lint checks and git diff --check passed.
• Existing fixture and installed-package checks passed.
• No credentials, measured outputs, raw responses, logs or report-instance profiles are included in Git.

Earlier development did include live smoke checks: archived runs contain five cases each for Bright Data, Browser Use Fetch, Browserbase Fetch and DeepSeek Fetch API, plus one Crawl4AI Cloud case. Those inspected runs used a 45-second budget. They are diagnostic evidence, not validation of the current 30-second configuration or a certified comparison. This revision made no new live provider calls and did not alter historical evidence.

Scope and compatibility

The configured method list is reusable configuration, not a report instance. Publishable cohorts still require explicit frozen task selection and the existing certified pipeline. CLI timeout/retry overrides remain available and recorded; explicit provider lists keep their meaning. Historical adapters remain selectable and the no-argument CLI remains an offline fixture.

Hosted adapters currently implement webpage body/structure tasks only. Their output contract is Markdown; other output formats are rejected before transport. Crawl4AI Cloud retains proxy=none; native paid proxies remain configurable but must be selected before a run. Cloud and local Crawl4AI evidence are distinct.

Local execution still requires usable public DNS. The inspected host resolved one benchmark domain to a private address while public DNS returned public addresses. Keep the existing preflight rejection rather than restoring a private-host exception; resolve the environment difference before executing an affected public cohort. This is not a provider capability failure.