Deep learning has become a cornerstone of Network Intrusion Detection Systems (NIDS), but its susceptibility to adversarial attacks reveals a major gap between reported accuracy and actual robustness. To close this gap, we introduce the Robustness-oriented Intrusion Detection Framework (R-IDF), which combines CTGAN-based data balancing, adversarial training strategies, diverse attack models, and a robustness metric suite. Evaluated on the CSE-CICIDS2018 dataset with LSTM-based models, R-IDF shows that although all models achieve nearly 99% clean accuracy, their retained accuracy under adaptive adversarial attacks often falls below 50%. This discrepancy, formalized as the Accuracy Fallacy Gap (AFG) and summarized by AUCRA, demonstrates that clean accuracy alone severely overestimates resilience. While adversarial training and defensive distillation improve resistance to transfer-based attacks, both fail against stronger adaptive adversaries. These results highlight the urgent need for robustness-oriented evaluation. By exposing hidden vulnerabilities and standardizing multi-attack assessment, R-IDF provides a reproducible benchmark that shifts NIDS evaluation beyond accuracy toward a more realistic measure of security.