nx_pose_backprop_gate.nx

nx_pose_backprop_gate.nx -- proves the MULTI-LAYER pose-net backprop COMPOSES correctly (the "scale build-2" de-risk): a 2-layer conv net conv1(1->2) -> ReLU -> conv2(2->1), then the loss gradient is backpropagated through conv2 (nx_f32_conv2d_backward) -> ReLU backward (f32_relu_bwd) -> conv1 (nx_f32_conv2d_backward). Inputs/weights are integers chosen so every pre-activation is POSITIVE (ReLU = identity) -> the net is locally linear -> the chained gradients are exact small integers, verified BIT-EXACT (no tolerance). This shows the training loop already gated for one layer (nx_f32_conv_train_gate) chains correctly through DEPTH -- so a real backbone+head pose net trains; only SCALE + labeled data remain. 1x1 convs at H=W=1 keep the arithmetic hand-checkable. expect_exit: 0