We study symbol detection in Multiple-Input Multiple-Output Orthogonal Time Frequency Space (MIMO-OTFS) systems over geometry-based, time-varying multipath channels, where multipath propagation induces spatial and fractional Doppler–Delay (DD) coupling and leads to non-uniform scaling across symbol components. We analytically characterize this non-uniform diagonal structure of the MIMO-OTFS Gram matrix, showing that all diagonal variation arises exclusively from inter-path coupling. Motivated by this result, we adopt a deep-unfolded Conjugate Gradient (CG) framework and introduce a geometry-aware coordinate-wise regularizer directly into the underlying linear system, with weights learned in a data-driven manner. Simulation results under both low- and high-mobility scenarios show that, for moderately separated propagation paths, the proposed detector outperforms low-complexity baselines and approaches the performance of more complex neural architectures at medium and high Signal to Noise Ratio (SNR).