Linear-solve ensemble controller

Integrate this at the residual-vector-field operation of a small neural ODE, latent SSM, or continuous-depth residual block. Let the base dynamics be f_theta(t,y) and add F_interp(t,y). Choose N anchor examples or task-support states z_i(t) and desired vector-field values v_i(t). These values can come from a teacher trajectory, a target latent trajectory, or finite-difference derivatives of labeled target states. At each ODE solver evaluation: (1) compute the anchor states and construct V from the desired derivatives; (2) form K_ij = G_0(z_i,z_j); (3) symmetrize K as (K + K^T)/2; (4) solve (K + lambda I) A = V using a Cholesky or QR solve; and (5) evaluate F_interp(t,y) = sum_i A_i G_0(y,z_i), returning f_theta(t,y) + F_interp(t,y). The paper's exact interpolation is the lambda = 0 case with nonsingular K. Use a small ridge lambda as a numerical safeguard and log the smallest eigenvalue and condition number of K. If conditioning is poor, merge similar anchors or increase lambda. First test on two-dimensional spiral and sine trajectories, then on a two-task MNIST latent ODE. Compare against a shared neural ODE trained only by backpropagation at equal parameter count and equal wall-clock budget. Measure anchor derivative error, final target-state error, gradient norm variance, and task accuracy. Success means near-zero simultaneous anchor error, fewer optimization steps, or better two-task accuracy without increasing inference cost excessively.