nx_nofloat_mla_gate.nx

nx_nofloat_mla_gate.nx -- MLA no-float SERVE (the census's #1 remaining July-2026 gap, 2026-07-15). Proves DeepSeek's Multi-head Latent Attention runs in deterministic integer no-float AND that its KV-compression is LOSSLESS by construction. Two paths on the SAME weights/input: LATENT (cache = LC-dim c) vs FULL-KV (cache = 2*D). Teeth: T1 LOSSLESS COMPRESSION: latent-path output == full-KV-path output, EXACT (0 mismatch) -- caching the LC-dim latent loses NOTHING vs caching full K,V (they up-project deterministically from c) T2 COMPRESSION REAL + MEASURED: latent cache bytes < full-KV cache bytes; ratio = 2*D/LC (report it) T3 DETERMINISM: latent path byte-identical on repeat T4 REAL TRANSFORM: output differs from input on >half the cells T5 NEG-CONTROL: corrupt the down-projection Wdkv -> output CHANGES (the latent bottleneck is load-bearing, not a bypass) AND with LC>=2*D (no compression) it still matches full-KV (degenerate check sane) Synthetic Q16 weights (mechanism proof -- no small real MLA gguf exists; DeepSeek V4 = 671B). Runs in ms. expect_exit: 0 license_tier: ORIGINAL No hw writes (Rule 26).