Vehicle re-identification (Re-ID) aims to match images of the same vehicleacross non-overlapping camera views, a task central to multi-camera visualsurveillance and large-scale video analytics.The core difficulty stems from near-identical inter-instance appearancesand substantial intra-class variation induced by viewpoint and illuminationchanges.Current methods face three fundamental limitations: geometric reasoning relies on manually annotated keypoints, relational graph construction employs deterministic similarity metrics that propagate errors under appearance ambiguity, and inference cost is treated as a fixed quantity regardless of input complexity.In this paper, we propose Keypoint-Aware Graph Transformer Network (KAGT-Net), an identity-supervised graph transformer with uncertainty-calibrated edges, organized as a four-stage hierarchical framework for vehicle Re-ID. We discover geometric structure through an identity-supervised keypoint prediction module that leverages Procrustes consistency across identity-matched pairs as the sole geometric supervision signal,eliminating the need for manual part annotations.We construct a sparse relational graph using uncertainty-calibrated edge probabilities derived from heteroscedastic estimation, suppressing spurious associations under appearance ambiguity.We align cross-camera feature distributions via differentiable optimal transport, coupled with a temporal memory module for identity-consistent embedding aggregation.An adaptive budget controller dynamically adjusts graph sparsity to balanceretrieval accuracy against latency constraints.Experiments on VeRi-776, VehicleID, and VERI-Wild demonstrate consistentimprovements over recent state-of-the-art methods, confirming theeffectiveness of uncertainty-calibrated relational reasoning for fine-grainedvehicle retrieval.