Noisy long-tailed text classification is particularly challenging because mislabeled examples and clean tail-class examples often exhibit similarly high losses, despite requiring fundamentally different treatments. Consequently, conventional sample-selection methods may mistakenly discard informative minority-class instances, while uniform multi-teacher supervision can propagate conflicting learning signals. We propose GRIP (Gated Routing of Instance-level Preferences), a dual-teacher dual-gate framework that performs instance-level selective knowledge distillation. A noise-specialized teacher provides robust supervision for unreliable samples, whereas a tail-specialized teacher strengthens learning for under-represented classes. Two lightweight and independent gates use sample-level reliability, uncertainty, teacher-disagreement, and class-rarity signals to determine whether each teacher should supervise a given instance. The gates are optimized with reward-shaped weighted binary cross-entropy objectives, and the student combines the standard classification loss with selectively activated distillation losses. Experiments on eight text-classification benchmarks covering real-world and synthetic label noise, three noise types, and varying imbalance factors show that GRIP achieves the best performance across all reported synthetic noisy long-tailed settings and substantial gains on real-world noisy datasets, including a 9.76-point accuracy improvement on TREC over the strongest baseline. Ablation and routing analyses further show that the two teachers and gates learn complementary, non-redundant roles. Only the student is retained at inference time, preserving the model size and latency of a standard BERT classifier.