Deep learning models for automated knee osteoarthritis (OA) grading achieve high in-distribution accuracy but frequently exploit spurious image-acquisition cues — a phenomenon termed shortcut learning — rather than learning disease-relevant anatomical features. We present a systematic three-phase experimental framework to detect, characterise, and mitigate shortcut learning in Kellgren–Lawrence (KL) grade classification using a disentangled ResNet-18 architecture extended with a Gradient Reversal Layer (GRL). Phase 1 (baseline) achieves a macro one-vs-rest AUC of 0.989 while simultaneously exhibiting a domain sensitivity probe AUC of 0.887, corresponding to an estimated mutual information I(Z; D) of 1.04 nats between backbone features Z and scanner-site domain labels D. Phase 2 (fisheye geometric augmentation) disrupts spatial shortcuts, reducing disease AUC by 17.6 pp and domain sensitivity AUC by 26.6 pp. Phase 3 (hybrid GRL, λ = 1.0) recovers disease AUC to 0.847 while suppressing domain sensitivity AUC to 0.541 and reducing I(Z; D) to approximately 0.09 nats. Integrated Gradients attribution analysis — verified against the completeness and sensitivity axioms — confirms an anatomical shift in model focus from non-diagnostic image borders to the medial tibio-femoral joint space. An information-theoretic decomposition of feature variance shows that the GRL reduces domain-explained variance from 44% to 13% of total representation capacity. Collectively, these findings establish that benchmark AUC alone is insufficient for validating clinical AI, and provide a mathematically grounded framework for shortcut characterisation and adversarial mitigation in musculoskeletal diagnostics.