nx_llm_forward_profile.nx

nx_llm_forward_profile.nx -- clean per-decode-forward timing on the REAL model, to find where a decode token's time actually goes. The 34-token bench is too noisy on this host (WSL bounces + multi- session load) to attribute sub-20% changes. This times ONE forward_v4 decode call at a time (no sampler/detok/penalty/loop overhead), after a warmup prefill so the dequant cache + lm_head pool are hot. Combined with the isolated lm_head number (nx_matmul_t_pool_gate: ~143ms pooled at m=1,n=151936), it splits the decode forward into lm_head vs the 24-layer block stack. Prints each step's us (watch for growth = KV-cache-length effect) and the average. No vocab/bpe load (forward needs neither). expect_exit: 0