nx_paged_kv_gate.nx

nx_paged_kv_gate.nx -- MEASURED gate for the sovereign PAGED KV-cache (nx_kvcache + nx_f32_attn_paged). PURE (no model, fast): KAT pool alloc/refcount/free; fork refcounts; exhaustion fail-fast; truncate unref EQUIV paged attention == contiguous attention BIT-EXACT over 3 append rounds (5-token prefill, 1-token decode, 3-token chunk) x 2 layers x GQA heads FORK fork after prefill; diverge A/RA vs B/RB; copy-on-append fires (refcounts return to 1); BOTH forks bit-exact vs fresh contiguous replays of their histories REWIND truncate A back to the fork point, append RB -> bit-exact == B's outputs (the speculative-decode rewind contract) license_tier: ORIGINAL expect_exit: 0