nx_pose_cnn.nx

nx_pose_cnn.nx -- T2 POSE FRONT-HALF: the sovereign software-f32 pose-CNN FORWARD PASS -- the heatmap PRODUCER that was the one remaining gap sitting UNDER nx_pose_keypoints. It composes the gated f32 vision substrate (nx_f32_conv2d_forward, itself only nx_f32_mul/nx_f32_add) into the real heatmap-pose-net SHAPE: a 3x3 stride-1 pad-1 spatial-conv BACKBONE stage (the ResBlock-class feature extractor) + a 1x1 conv HEAD that projects the feature channels into J per-joint confidence heatmaps, each followed by a ReLU that makes every heatmap value NON-NEGATIVE. KEY INVARIANT (load-bearing, why this closes the loop with NO glue): IEEE-754 non-negative floats are ORDER-ISOMORPHIC to their i64 bit-patterns -- for x,y >= +0.0, (x > y) as floats iff (bits(x) > bits(y)) as integers. So a ReLU'd heatmap can be consumed DIRECTLY by the integer, weights-free back-half nx_pose_keypoints (argmax + quarter-pixel subpixel + occlusion threshold) with NO f32->int conversion. The ReLU is REQUIRED: a negative activation carries the sign bit, whose pattern sorts ABOVE any positive peak, so the integer argmax would pick the wrong cell -- ReLU zeroing negatives to +0.0 is exactly what restores the isomorphism. This closes the pipeline image -> f32 conv backbone -> f32 1x1 head -> heatmaps -> integer keypoint back-half, end-to-end and sovereignly, on hand-set weights. Only TRAINED WEIGHTS + a deeper backbone remain (the T3 arc); the architecture + the arithmetic are proven here. NO libm, NO Q-scaling -- only nx_f32_conv2d/mul/add/gt. license_tier: ORIGINAL