Indoor wireless coverage has emerged as a persistent constraint in modern 5G deployments, particularly in high-rise urban buildings where signal attenuation, pilot contamination, and architectural obstructions undermine network reliability and user experience. While Unmanned Aerial Vehicles (UAVs) provide a flexible means of acquiring external RF measurements, inferring the interior signal landscape solely from these sparse and non-invasive observations remains an open and pressing challenge. We propose SpiderGNN (Spatial-aware Predictive Inference with Dynamic Edge Reasoning Graph Neural Network), a topology-adaptive GNN framework for reconstructing indoor 5G signal quality from sparse UAV-based external observations. Unlike conventional distance-based graphs, SpiderGNN builds sparse structures dynamically using attention-derived feature affinity, enabling better modeling of the non-Euclidean nature of signal propagation in complex building environments. Built upon this learned topology, a residual Graph Attention Network (GAT) encoder propagates multi-hop signal semantics, followed by dual decoding branches tailored to the distinct physical characteristics of RSSI and CQI. To enhance stability and generalization, an ensemble-based fusion layer aggregates predictions across multiple stochastic runs. Empirical evaluation on a real-world, multi-floor 5G dataset demonstrates that SpiderGNN achieves substantial improvements over state-of-the-art location-aware baselines, reducing RMSE by 76% for RSSI and 89% for CQI, while attaining coefficient of determination (R \(^2\) ) scores exceeding 0.99 on both metrics. By uniting sparse UAV sensing, non-Euclidean graph modeling, and spatially contextual inference, SpiderGNN establishes a scalable and non-intrusive paradigm for indoor signal estimation. The framework offers a technically grounded and practically deployable solution to urban-scale wireless planning, bridging the longstanding gap between outdoor sensing and indoor coverage prediction.