Finite-lag transport tensor analysis

Builds a source-centered transport tensor from pairs of a recurrent network's own hidden states separated by a fixed lag (X_t, X_{t+delta}), decomposing it exactly into a conditional-spread trace and a coherent-displacement trace plus an antisymmetric coordinate-circulation statistic -- quantifying how much a real trained RNN's own state-transition dynamics locally expand/contract and rotate, as a function of lag and training phase.