Indoor positioning using commodity Wi-Fi has gained significant attention; however, achieving sub-meter accuracy across diverse layouts remains challenging due to multipath fading and non-line-of-sight (NLOS) effects. In this work, we propose a hybrid Graph–Temporal Convolutional Network (GTCN) model that incorporates access point (AP) geometry through graph convolutions while capturing temporal signal dynamics via dilated temporal convolutional networks. The proposed model adaptively learns per-AP importance using a lightweight gating mechanism and jointly exploits Wi-Fi Received Signal Strength (RSS) and Round Trip Time (RTT) features for enhanced robustness.It is evaluated across four experimental areas such as lecture theatre, office, corridor, and building floor covering areas from 15 × 14.5 m2 to 92 × 15 m2. We further analyze the sensitivity of the model to AP density under both LOS and NLOS conditions, demonstrating that positioning accuracy systematically improves with denser AP deployment, particularly in large-scale mixed environments. Despite its high accuracy, the proposed GTCN remains computationally lightweight, requiring fewer than 105 trainable parameters and only tens of MFLOPs per inference, enabling real-time operation on embedded and edge devices.