greghavens/Qwen3.8-27B-bnb-4bit

Restore the 15 MTP tensors dropped by the NF4 conversion

Add NoveltyBench output-diversity benchmark (null result, 3-arm)

Add BFCL tool-calling benchmark: -1.04pp overall, isolated to live_irrelevance over-calling

Add three-way vision VQA benchmark; bf16 tower shows no measurable benefit

Add paired IFEval benchmark vs bf16 base (no significant degradation, CI +/-2.7pp) + eval scripts and raw generations

NF4 (bnb) quant of Qwen3.8-27B; vision tower + lm_head kept bf16

initial commit