nx_reasoning_paged.nx

nx_reasoning_paged.nx -- PREFIX-SHARED self-consistency: the paged-KV promotion of nx_reason_selfconsist. Prefill the ChatML prompt ONCE on a root sequence, then FORK per sample (refcounted page-table copy; the fork's first append copy-on-writes at most one tail block). MEASURED 1.48x vs the reset-loop at N=3 (nx_paged_fwd_gate, real model, bytes BIT-IDENTICAL per sample); the win grows with N since N-1 chunked prefills are eliminated. genealogy_id: wang_2023_self_consistency + kwon_2023_pagedattention lineage_id: substrate_reasoning_paged_v1