nx_f32_llm_special_tokens.nx

nx_f32_llm_special_tokens.nx -- read BOS/EOS/PAD token ids from GGUF. Composes with nx_f32_llm_run: the caller passes the eos_token_id arg. Pre-this-brick, callers had to hardcode it. Post: read it from the GGUF and pass it through. Production runs no longer loop forever on models that emit EOS to terminate generation. Metadata keys (GGUF v3, gpt2-style tokenizers + Llama tokenizers): tokenizer.ggml.bos_token_id (u32) beginning-of-sequence tokenizer.ggml.eos_token_id (u32) end-of-sequence tokenizer.ggml.padding_token_id (u32) optional padding token All three return -1 if not present in metadata. genealogy_id: ggml_tokenizer_special_token_keys lineage_id: substrate_f32_llm_special_tokens_v1