Fix MMR-Bench data integrity and optional joint input splitting

[AIGNLAI]

Summary

Fix production MMR-Bench V1 loader issues found during the full-7 reproduction, without changing router algorithms or metrics.

• Parse mixed boolean/string/numeric correctness labels into numeric values; reject malformed/out-of-range labels and invalid monetary costs. Preserve original costs, including valid zeros.
• Make candidate coverage explicit. The shipped config selects complete-models: on the audited release, exclude GPT-5-nano (missing 3,040 MathVision observations), keep nine candidates in canonical order and all 10,370 queries. No missing-cell imputation or query dropping. Strict error policy is available and is the fallback for programmatic configs without a policy.
• Expose actual released answers as model_i_response, retaining model_i_prediction as an identical backward-compatible alias. This allows EA-RAM to train its ex-post evaluator instead of silently skipping response-dependent training.
• Support archive-native and prepared image layouts, with explicit errors for missing/ambiguous paths.
• Add opt-in joint-grouped in-domain splitting using transitive components of identical decoded RGB images OR whitespace-normalized query text. The default remains the historical legacy row-random split, with an overlap warning.
• Add regression tests, a complete example config and protocol documentation.

Protocol caveats

The corrected nine-model candidate pool is not the historical ten-model evaluation. Joint grouping is optional and is not the untouched default protocol: fractions are approximate, grouping is not per-task stratified, and an entire task can fall in test. It prevents split input overlap, not external pretrained-model exposure.

Validation

• Full local offline/unit smoke suite: 135 tests passed, including 17 new MMR-Bench regression tests.
• Regression coverage includes EA-RAM actually fitting its evaluator on released-response aliases; original costs and model-column identity; strict missing-data behavior; transitive image/text grouping; duplicate image encodings; image failures; and exact legacy-split compatibility.
• Complete official raw-data validation through the production loader with the joint-grouped protocol: 2,059 train / 8,311 test, nine complete candidates, all 10,370 queries retained. Every ID, prompt, image path, answer, task, quality, cost and model response matches the audited frozen reproduction inputs.
• Python compilation and git diff --check passed.

Scope

Only the loader, benchmark/example configs, documentation and regression tests are included. No experiment runners, cache machinery, datasets, results, weights, NIRT gateway changes, credentials or machine-specific paths are committed.