nx_nofloat_mla.nx

nx_nofloat_mla.nx -- MLA (Multi-head Latent Attention, DeepSeek's marquee KV-compression) in no-float integer SERVE convention (2026-07-15). The 2026-frontier attention: instead of caching full K,V (2*kv_dim per token), x is DOWN-projected to a small latent c (LC-dim, LC << 2*kv_dim) which is the WHOLE KV cache; K,V are UP-projected from c at attention time. The KEY property (provable by construction): serving from the LC-dim latent cache produces the BYTE-IDENTICAL output as serving from a full-K,V cache -- LOSSLESS compression, ratio (2*kv_dim)/LC. This lib gives BOTH paths so a gate can differential them. NOPE variant (no RoPE) for rung 1 -- decoupled-RoPE (RoPE on a separate dim, since RoPE doesn't commute with the up-projection) is rung 2. Composes proven primitives (mm_out_in / fx_exp / qmul); Q16 throughout, deterministic (exact integer sums, positional tie-breaks). license_tier: ORIGINAL (lib: no main)