nx_motion_neural_gate.nx

nx_motion_neural_gate.nx -- prove the neural-motion foundation is LEARNED + GENERATIVE (not the parametric keyword lookup): a codebook trained by VQ/k-means on a real motion corpus, and an n-gram token-LM that GENERATES a novel motion sequence, decoded back to drive the rig -> sovereign video. T1 the codebook LEARNS: trained distortion < initial distortion (k-means reduced quantization error) T2 tokens DISCRIMINATE: different moves -> different token sentences (Hamming > 0 across pairs) T3 round-trip PRESERVES motion: decode(encode(move)) per-frame error bounded (codec keeps the motion) T4 GENERATES novel motion: an LM sample is valid tokens, produces motion (decoded poses vary), and is NOT byte-identical to any single training move (real composition, not playback) Emits knowledge/synth_neuralmove.png (APNG of the GENERATED motion, rendered on the FK figure). license_tier: ORIGINAL expect_exit: 0