nx_f32_llm_e2e_test.nx

nx_f32_llm_e2e_test.nx -- end-to-end bits-up text-to-text smoke. Composes the entire f32 LLM pipeline on the tiny-Llama fixture: 1. Build fixture (12-tensor GGUF + BPE vocab + spec) 2. Allocate NxF32LlamaModel + populate dims from fixture 3. Load weights via nx_f32_llm_load_weights_from_gguf 4. Tokenize 'ab' via nx_bpe_encode -> [0, 1] 5. Allocate KV cache 6. Run nx_f32_llm_forward over the 2 prompt tokens 7. Argmax last token's logits -> next token id 8. Detokenize -> byte Zero-weight fixture means logits are all zero -> argmax returns 0 (the first token id 'a' = byte 0x61). This proves the entire pipeline executes without segfault and produces a deterministic known-good result.