nx_gpu_q4k_gate.nx

nx_gpu_q4k_gate.nx -- proves a SOVEREIGN organ can drive the GPU. WHY THIS EXISTS (2026-08-01): Q4_K matvec was measured correct on the RTX 5080 at 362 GB/s DRAM-resident (=> ~42 tok/s for a 14B, against 1.06 tok/s for today's 0.5B CPU seat). But NishiLang emits STATIC ELF with syscalls only -- it has no dlopen, so it cannot link libcuda. The GPU therefore lives behind nx_gpu_q4kd (a small C daemon holding weights resident in VRAM) and we reach it over TCP, which this language already speaks. This gate is the proof that the sovereign side can actually drive it -- not that the daemon works in isolation. ★ THE COMPARISON IS BIT-EXACT, ON PURPOSE. GPU and CPU reduce in different orders, so a float comparison would normally need a tolerance -- and a tolerance is a place for a wrong answer to hide. Instead the test data is chosen so EVERY product is a small integer: d=1.0, dmin=0, all sub-block scales=1, x integral. Integer-valued f32 addition is exact in ANY order, so the GPU result must equal the expected value BIT FOR BIT. No tolerance, nothing to tune. EXIT CONTRACT: 0 = all teeth pass. 1 = a tooth failed. 2 = could not reach the daemon (refuses to report GREEN when it never measured -- absence of findings and absence of measurement must never render identically). nx_gpu_q4k_gate [port] default 18140 expect_exit: 0 license_tier: ORIGINAL