Adaptive Reliability-Aware Nonlinear State Estimation Framework for GNSS/INS Integrated Navigation with Bounded Position Correction

State estimation in tightly coupled nonlinear GNSS/INS navigation is complicated by persistent variation in measurement quality, deteriorating satellite geometry, and occasional abnormal updates that produce large position corrections. This paper develops an adaptive reliability-aware framework that addresses these effects at the measurement, geometry, and state-feedback levels. Redundant measurement noise covariance estimation (RMNCE)-inspired residual statistics are converted into a bounded long-term reliability indicator for satellite-wise covariance adaptation without mandatory rejection. The same reliability information is incorporated into a quality-aware weighted geometric dilution of precision criterion for satellite selection. An event-triggered bounded correction then limits excessive position feedback while leaving nominal updates and velocity correction unchanged. Under the experimental setting $C_p=C_h$ and $C_s<C_p$, this generalized mapping reduces to event-triggered hard norm clipping; the experiments therefore evaluate the triggering policy and active saturation level, not the inactive smooth-transition or geometry-dependent terms. Theoretical analysis establishes positive-definite adaptive covariance, finite weighted-geometry evaluation, and bounded position correction under the stated conditions. These structural properties do not imply global convergence, mean-square stability, estimator consistency, or formal integrity guarantees for the complete nonlinear estimator. Vehicle-mounted GNSS/IMU experiments under nominal, persistent multipath-like, and low-satellite-visibility conditions show that the reliability-aware front end provides the dominant improvement. A matched protection-policy ablation further shows that event-triggered clipping improves 2D RMSE and the 95th-percentile horizontal error relative to both no additional position protection and unconditional clipping, although the most extreme horizontal errors remain unchanged. The results support coordinated reliability adaptation, geometric selection, and layered state-feedback safeguards.