Dataset Splits, Vision Encoder, and Hyper-PELT Implementation Details

Benchmarks span MRPC→GQA; text splits follow prior work, images downsampled to a 7×7 grid, visual encoder is frozen for fair param counts.