Learning-augmented navigation filters introduce data-driven process-noise adaptation into recursive state estimation, yet erroneous learned parameters can compromise measurement acceptance and uncertainty reporting. This work investigates runtime guardrails that constrain learned process-noise proposals in a hybrid inertial navigation system with ultra-wideband aiding (INS/UWB). Local analysis characterizes how variations in process-noise parameters influence innovation-based measurement gating, while experiments examine magnitude, rate, and projection restrictions within an error-state extended Kalman filter. Across eight held-out quadrotor recordings, magnitude restrictions suppress two severe Ridge-based failures, reducing recording-level position root-mean-square error (RMSE) from over 650 m to below 0.5 m. Under a substantial upward perturbation of the learned process-noise proposals, magnitude restriction reduces mean position RMSE by 11.9%, with improvement across all eight recordings. Channel-specific tests identify effective angular-rate process-noise bounds, while discrete projection reveals that lower error may result from suppressed adaptation. However, improved accuracy does not necessarily imply calibrated uncertainty or recovery following estimator divergence. These findings characterize the benefits and limitations of runtime guardrails for learning-based navigation estimation.