nx_nofloat_kvcache_gate.nx

nx_nofloat_kvcache_gate.nx -- CAP-NF-KVCACHE: a KV-CACHE for O(T) autoregressive decode (vs O(T^2) full-recompute), pure integer Q16. Single-block causal attention model (random weights -- this is about inference CORRECTNESS + EFFICIENCY, not model quality, so no training needed). Two decoders share the SAME primitives: full-recompute re-derives every past K/V each step; cached computes each K/V once and reuses it. T1 CORRECT: cached logits == full-recompute logits at every position (bit-exact, by construction). T2 CHEAPER: cached does far fewer K/V projections (full = T(T+1)/2, cached = T) -- measured counts. T3 TEETH: a CORRUPTED cache (one stale entry) yields DIFFERENT logits -> the cache content is really used. Sovereign: nx_nofloat_autograd (qmul/isqrt/sin/cos/exp) + nx_syscalls. expect_exit: 0