Current C++/Python/Java display code and labels, 900 pairs each. Four C++ replacements, reviewed Python/Java workload metadata, Java RSD<5%. Remove obsolete training/validation and prompt-expanded datasets.
Freeze base datasets with repaired unrenamed Python snapshot
Keep the three existing base dataset paths (900 pairs each). C++ and Java are byte-identical to the prior revision. Replace Python with the repaired unrenamed snapshot, preserving the existing filename. Historical measurement fields are retained; this candidate freeze does not certify full correctness or instruction-label revalidation. Training, validation, derived prompts, and other files are unchanged.
Rebuild python_ab_test from the latest published python_abdiff_test (same pair_id/direction), keeping AB template only. Ensures 900 base pairs with forward+reversed exactly once each and no embedded diff section.
Fix python_ab templates by remapping from matching python_abdiff pairs
Regenerate python_ab_train/val/test from the corresponding python_abdiff split for identical pairs and directions. This removes accidental embedded diff sections from python_ab inputs and enforces pure AB template structure.
Fix Python/Java test files to match canonical uniform benchmark
Replace the published Python/Java ab, abdiff, and adiff test files with versions rebuilt from the canonical final_benchmark_uniform sources. This fixes Python pair coverage drift, Java mixed-benchmark test generation, and the incorrect Java adiff header.
