nx_f32_llama_v4p.nx

nx_f32_llama_v4p.nx -- PAGED-KV twins of the v4 forward chain: block / stack / model-forward over *NxPagedSeq instead of the contiguous *NxF32KVCache. FAITHFUL copies of nx_f32_llama_block_v4 / nx_f32_llama_stack_v4 / nx_f32_llm_forward_v4 with EXACTLY these substitutions (the codebase twin convention -- v2/v3/v4 precedent): block: cache -> pseq; RoPE cache_before = pseq.seq_len; nx_f32_attn_with_cache -> nx_f32_attn_with_paged stack: nx_pkv_ensure_append ONCE before the layer loop (blocks + copy-on-append privatization must precede ALL layers' appends); nx_f32_kv_cache_advance -> nx_pkv_advance forward: cache -> pseq; stack_v4 -> stack_v4p (embed/rmsnorm/lm_head identical) Debug dumps (blk4_dump8/ffnmax "log2mag" stdout spam) omitted -- numerically irrelevant. bp_* phase profiling kept (parity with the contiguous profiler). Bit-exactness vs the contiguous chain is gated end-to-end by nx_paged_fwd_gate on the REAL model. This is what makes the paged pool PAY: prefill a prompt ONCE, fork N sequences (refcounted, copy-on-append) -- the prefix-shared self-consistency / best-of-N serving pattern. genealogy_id: kwon_2023_pagedattention + standard_transformer_stack lineage_id: substrate_f32_llama_v4p_v1