nx_nofloat_kvcache_int8_gate.nx

nx_nofloat_kvcache_int8_gate.nx -- CAP-NF-KVCACHE-INT8: a QUANTIZED KV-cache (store cached K/V at int8 = 8x less cache memory, the standard long-context inference technique), pure integer Q16. Single-block causal attention (random weights -- about cache fidelity, not model quality). Decode storing K/V at a configurable precision: full (Q16) vs int8 (127 levels) vs ternary (teeth). T1 int8 cache PRESERVES DECISIONS: argmax(int8-cache) == argmax(full-cache) at every position. T2 MEMORY: int8 = 8 bits/value vs i64 = 64 bits = 8x smaller cache (the point of quantizing the cache). T3 TEETH: a ternary cache FLIPS some decisions -> precision matters (int8 is the quality-preserving choice). Sovereign: nx_nofloat_autograd (qmul/isqrt/sin/cos/exp) + nx_syscalls. expect_exit: 0