When Latent Linearity Helps: Contact-Aware Evaluation of LeWorldModel Planning

World models are often evaluated by short-horizon prediction or visual fidelity, while a physical agent must preserve actionable structure through contact transitions and long-horizon control. We study this gap in LeWorldModel (LeWM), a pixel-based JEPA-style world model with SIGReg regularization, an inverse dynamics model (IDM), latent predictor rollouts, and cross-entropy method (CEM) planning. We evaluate a label-free tail-error reweighting intervention (TailReg) together with representation geometry, contact-stratified prediction error, action decoding, and downstream planning across four simulated control environments. At a frame skip of five, latent straightness is positive for PushT ($0.5170\pm0.0964$) and Cube ($0.4120\pm0.0461$), but negative for Reacher ($-0.0728\pm0.0839$) and TwoRoom ($-0.3651\pm0.0892$). On PushT, direct IDM control reaches $35.5\%$ mean success, whereas the strongest tested CEM configuration reaches $21.5\%$. TailReg reduces contact-conditioned latent prediction MSE from $0.428$ to $0.247$, reduces its p95 from $0.605$ to $0.333$, and increases the mean Pearson correlation of five IDM gripper dimensions from $0.450$ to $0.577$. However, no successful Cube rollout is observed in 50 episodes for any evaluated strategy. These findings suggest that latent linearity and physical representation metrics should be reported together with action-conditioned, long-horizon decision metrics.