Preserve shared scoring and native answer admission after extraction

[Maisiqi]

Change

Web Fetch answers can fail acceptance after harmless HTML/Markdown conversion, while metadata or a nearby table row can incorrectly supply a missing answer. This change ports Browser Eval PR51's shared scoring fixes into the benchmark package after Browser Eval's evaluator extraction.

• Match body and structured answers through the shared parsers, preserve table row ownership and native link titles, and distinguish declared exclusions from diagnostic hints.
• Make page identity an explicit task requirement. Preserve historical controls and add separate executable cases for revised policies.
• Validate declared answer fields, files, title and final URL against retained native bytes. The scorer uses the package's existing delivered-file acceptance; the execution harness retains evidence persistence and provider adapters.
• Bind the projection validator into the scoring source identity and exercise the same controls through standalone scoring.

Validation

make check passes all four task-suite definition checks, 1559 tests and 81 subtests. Controls accept a visible title paired with its own row's link and reject borrowed, hidden or mismatched titles; distinguish diagnostic text from an explicit task ban; and reject answer fields or result URLs that differ from retained native bytes.

The four task-suite definitions and frozen input hashes are unchanged. The synthetic scoring-control registry is SHA-256 159adaa11e5272adfe51555fddeba33283c5e5809a29a1d30bfc5ad911bcbe70. No measured outputs or provider calls are included. The existing Browser Eval PR51 will consume this pinned package revision and retain its saved report inputs.