Heisenberg latent upsampler

Integrate this as an optional sequence upsampler between temporal resolutions in a world model, RNN, SSM, or graph trajectory encoder. Store every latent token as three channels (x,y,z), or project a larger latent vector into several independent Heisenberg triples plus ordinary residual channels. For each four-token window, compute horizontal insertion coordinates a_i and b_i with the four-point rule: a_i=-omega*x[i-1]+(0.5+omega)*(x[i]+x[i+1])-omega*x[i+2], and similarly for b_i, using omega=1/16 initially. Compute the inserted central coordinate with z_new=-omega*z[i-1]+(0.5+omega)*(z[i]+z[i+1])-omega*z[i+2]+0.5*(x[i]*b_i-y[i]*a_i). Copy even-indexed input states unchanged, concatenate the refined sequence, and pass it through a learned linear projection or MLP if the downstream block expects a different width. The group formula is used only in the central correction; all other channels can use ordinary interpolation. Compare against nearest-neighbor, linear, and ordinary four-point interpolation on a synthetic task where a 2D control trajectory determines a target signed-area or winding label, followed by a small temporal predictor. Train at low temporal resolution and evaluate after upsampling. Measure prediction error, latent trajectory consistency under composition, and gradient norms. Success means lower area-sensitive prediction error at the same latent width, with no instability or excessive growth in z; also test reversed trajectories, for which the signed-area term should change sign.