Fix AA-Briefcase-Lite judge routing and subtitle transport - #3334

Fixes the existing Lite judging path and makes the native Gym benchmark run both scoring modes:

• Route each panel member through its own existing model adapter instead of silently sending every judge selection to Gemini. Match the published reasoning settings.
• Pass SRT subtitle contents through the shared bounded text reader.
• Request the required JSON on the first binary call using the existing format instruction; retain the fallback.
• Select all four tasks with binary and local pairwise judging by default. Document native preparation, execution and cached-artifact rejudging. Add configuration and pairwise adapter-routing tests; no new orchestration.

Validation: 104 focused tests passed, one optional-dependency skip; scoped pre-commit passed. Regression tests reproduce the routing/subtitle/format defects before their fixes. Real cached-artifact judging completed one four-task set with all 55 binary checks and 16 pairwise trials valid, including seven Claude pairwise trials. The documented run command also passes native CLI/config validation.

Still draft: the local 16,384-token Claude override is not a complete budget fix. A second four-task set finished with only one valid task: empty Claude answers are recorded in report analytical-quality and workbook-presentation pairwise checks, and presentation binary A-13. Persisted text does not establish their stop reason. Historical binary truncation also remains unresolved; one complete rerun does not prove it eliminated. No further cap, effort, timeout or brevity change is included.

These are public Lite checks plus a local pairwise approximation, not official AA-Briefcase Elo. No new policy generation was needed for this validation. Dataset preparation, sandbox, rubrics and judge sampling are unchanged. TEX/PDF recompilation is deferred. Private artifacts, credentials and deployment configuration are excluded.

References: AA methodology, subtitle rubric.