Finding 9: add MoE result, correct the prediction (MoE loses harder, not less)

Add MoE DFlash2 A/B: loses harder than dense (0.89x code), contradicting the prediction

Add finding 9: DFlash2 also a net loss on Metal despite 76% draft acceptance

Add DFlash2 speculative decoding A/B on dense 27B (net loss despite 76% acceptance)

Add section 8: oMLX (paged SSD KV cache) dense 27B at 8k, link the two upstream bugs it turned up

Add oMLX Qwen3.8-27B dense 4bit at 8k ctx (2 sessions, GPU idle both times)

Upload README.md

Add M2 Max 32GB benchmarks: MoE vs dense, MTP on Metal, KV quant, MLX vs llama.cpp

initial commit