Add format-tax macro-F1 vs accuracy analysis

Fix: keep zero-padded design-class codes (consistent with 2022 + eval mapping)

Add corpus-wide design-class backfill (2023-2026, from BigQuery uspc)

Add design_class_difficulty_perpatent.json

Add design_class_difficulty.json

Add design-class analysis artifact: design_class_2023_2025_eval.json