Qwen 3.6 vs 3.5: Same 37 tok/s on RTX 4070, +43% on Frontend Generation

Same RTX 4070, same 35B-A3B footprint. Generation speed barely moved (38.76 vs 36.7 tok/s). But WebBench jumped +43% and Terminal-Bench +27%. AIME barely moved. The generation gap lives on the agent axis, not the tok/s axis.