nx_f32_llm_live_load_test.nx

nx_f32_llm_live_load_test.nx -- LIVE-FIRE structural proof. Loads the FULL real Qwen2.5-0.5B-Instruct Q4_K_M GGUF + runs ONE forward pass with a 2-token prompt + verifies the resulting logits are finite (not NaN, not inf) and non-degenerate. Verifies that EVERY layer of the bits-up substrate works on real production data: sys_read_file -> 491 MB read nx_gguf_parse -> 291 tensors + 26 metadata entries nx_f32_llm_read_dims_from_gguf -> all 7 dims match Qwen2.5-0.5B spec nx_f32_llm_load_weights_v4_from_gguf -> 24 layers bound (with lazy Q4_K + eager Q5_0/Q6_K/Q8_0/F32 dispatch) nx_f32_kv_cache_alloc -> cache for real dims nx_f32_llm_forward_v4 -> ONE prefill pass through 24 transformer blocks (each with RMSNorm + attn + cache + FFN-SwiGLU) logits sanity check Skips the full autoregressive loop (which would take ~50 min in qemu due to 151936-column LM head matmul × 32 decode steps). This is the "loads + runs forward" structural proof, NOT a production tokens/sec bench.