nx_nofloat_rope_gate.nx

nx_nofloat_rope_gate.nx -- HARD-EVIDENCE gate for RoPE (rotary position embedding) backward + RoPE composed into the attention path. Advances CAP-NF-ATTN-FULL: with RoPE the attention is positionally-aware AND still fully differentiable in pure integer Q16. RoPE is parameter-free (orthogonal rotation), so its backward is the inverse rotation -- here proven correct by finite differences AND inside the full attention block. A1 RoPE gradcheck : loss=mse(RoPE(x), t); tape grad dL/dx == central finite difference. A2 attention+RoPE gradcheck: X->{Q,K,V}; RoPE(Q),RoPE(K); S=QK^T; scale; causal softmax; O=A.V; mse; gradcheck dL/dWq -- the gradient flowing through RoPE + softmax + both matmuls (the real RoPE-attn backward). D neg-control teeth : a wrong RoPE grad is rejected. C bit-exact : the attention+RoPE gradient is identical across two runs (determinism). Evidence -> knowledge/status/nofloat_rope.log. Sovereign: imports nx_nofloat_autograd (pure integer) + nx_syscalls. HONEST scope: single-head; output-projection + multi-head/GQA remain (bookkeeping over matmul+vadd). expect_exit: 0