We reformulate world model validity as a spatially-resolved control problem. We introduce Vectorial Resonance (R in [0,1]^d), a per-dimension signal comparing learned model predictions against a physics prior, with quantile-calibrated sensitivity. Combined with an L_inf decision rule and Dual-Horizon sign-change analysis for noise rejection, this monitor localizes which state dimensions have structurally failed. We couple this detection with a physically-factored architecture (shared trunk, independent heads per dimension) and a masked loss function enabling targeted partial reconstruction with zero gradient leakage to healthy parameters. On a CartPole environment with 7 shift scenarios across 10 seeds, vectorial monitoring achieves perfect spatial precision (1.00) with zero false positives under noise bursts, while detecting micro-shifts (18% gravity change) that scalar monitoring misses entirely due to geometric dilution. The framework also guarantees graceful degradation: incorrect physics priors are automatically neutralized through calibration scaling. These results demonstrate that spatializing model validity enables surgical model repair without catastrophic forgetting.