gpt-oss 20B Heretic-v2

User reports gpt-oss-20b-heretic-v2 generating at 119 tok/s on an M3 Ultra after a monkey-patched fix. Setup is MLX (mlx-lm 0.31.3) with MXFP4 g32 experts, converted from a BF16 checkpoint and loaded in mlx and oMLX. The post is a bug report: sanitize() in mlx_lm/models/gpt_oss.py fails on BF16 checkpoints in the [E, H, 2F] layout, reporting 72 parameters not in model. The fix was verified end-to-end through convert, load and coherent generation.