arXiv:2611.22176v1 Announce Type: cross Abstract: Autonomous vehicles plan time-optimal routes through flow fields that are known only through reduced, finite-resolution models of an underlying multiscale flow. Classical Hamilton-Jacobi homogenization concerns convergence of the value function as the fine scale vanishes; but what a vehicle commits to is a route, and value accuracy need not imply route accuracy. We study this gap for a fixed-resolution reduction rather than in the homogenization limit. The committed route's dependence on the model-reduction error splits into two regimes. When the optimal route is a unique, nondegenerate minimizer with a regular value-function neighborhood, it depends Lipschitz-continuously on the reduction error, while the extra travel time from acting on the reduced model grows only quadratically with that error. When two routes are tied on the cut locus of the value function, the value remains stable, but an arbitrarily small change in the flow can switch the selected route, making route selection discontinuous in the reduction error. This dichotomy identifies where a reduced model must be accurate: not where the flow error is largest, but where the route choice is fragile. In the regular regime it yields a resolution-allocation principle, while an oracle marginal-regret benchmark illustrates the potential benefit of decision-aware resolution allocation relative to refinement based on flow-error magnitude.