nx_nofloat_batched_forward_gate.nx

nx_nofloat_batched_forward_gate.nx -- THE DECISIVE SPECULATIVE-DECODING FOUNDATION EXPERIMENT. Speculative decoding's speedup on MEMORY-BOUND decode = a BATCHED verify reads each weight ONCE and applies it to K token positions, amortizing the weight-read that dominates m=1 decode. Q: does batching actually amortize the weight-read on this box, or is it compute-bound (no gain)? This settles the whole speculative lever BEFORE any core-engine build. Same total work (M tokens x N outputs), two loop orders on the real FFN shape: m=1 (today's decode): M passes over the weight array (each token re-reads every weight) m=8 (batched verify): 1 pass over the weight array (each weight read once, reused for 8 token activations) If m=8 per-token time << m=1 -> weight-read amortizes -> speculative pays. If ~equal -> compute-bound -> it won't. Isolated gate (i8 weights, the current fast mode); touches NO serve code. license_tier: ORIGINAL expect_exit: 0