Predicting pedestrian crossing intention is vital for enabling safe and anticipatory behaviour in autonomous vehicles (AVs), particularly in complex and visually degraded urban settings. Existing approaches often rely on high-quality visual inputs, dense semantic maps, and computationally intensive architectures, limiting their generalizability and real-time applicability. To address these limitations, we propose an end-to-end framework that unifies visual appearance, motion dynamics, and symbolic scene attributes for robust intention forecasting. Our method integrates a MobileViT-S backbone for visual encoding, a GRU-based trajectory encoder for motion, and a template-driven semantic context encoder whose outputs are embedded via MiniLM-L6-v2. These heterogeneous cues are fused through an Attention-Gated Cross-Modal Fusion Block that adaptively aligns modalities via cross-attention and gating mechanisms. Evaluated on the PIE and JAAD datasets, our model achieves competitive results (AUC: 0.92 on PIE, 0.89 on JAAD) using only ~0.99M trainable parameters.