[GDPVal] Add per-provider judge transports with deterministic eligibility replay

[Kh4L]

Ports the per-provider judge transport layer, executed and validated in a full 220-task GDPVal evaluation campaign, onto the production-hardening branch.

Changes

• Per-judge media transport modes: native_pdf, images_and_text, and native_pdf_overflow_images (native PDFs kept; over-cap PDFs rasterized at a pinned DPI with every page retained).
• Provider-specific request limits as judge-panel config: serialized-request bytes, native-PDF page/document/byte caps (aggregate and per-document), video count, raster DPI tiers.
• Per-document native-PDF cap (max_native_pdf_bytes_per_document) forces lossless full rasterization of any PDF above the provider's per-file limit; page-count overflow rasterizes only the required prefix, keeping the suffix native.
• Deterministic transport eligibility preflight before any billable call, with seeded schedule replay; exclusions recorded per row (media_routing_exclusions, transport_by_judge).
• Overflow eligibility also enforces the judge's aggregate document/byte caps against the post-overflow native remainder, and a render page cap so a plan the renderer cannot satisfy is excluded at preflight instead of failing at dispatch.
• New transport_assignment.py: provider-free task/reference reassignment repair using exact artifact footprints against provider caps.
• transport_assignment_repair settings enter the multistage fingerprint, so changing them invalidates journals instead of silently replaying stale plans on resume.
• Final verdict reminder block appended to judge messages; the text budget's framing reserve includes its serialized upper bound so the request-size invariant holds at any configured cap.

Validation

• resources_servers/gdpval/tests: 379 passed, 3 skipped, including 16 new tests in test_transport_modes.py (overflow planning/application, eligibility receipts, seeded replay, assignment solver capacity contract, fingerprint invalidation).
• ruff check / ruff format clean.
• Same code path that produced a strict 220/220-task, 880/880-trial judging PASS in the validating campaign, rebased onto this branch.

First of four planned child PRs (media caps/tiling, provider timeout/retry hardening, transport receipts follow). Draft until the example three-provider judge panel config lands.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.