Concrete crack segmentation in field imagery remains challenging because cracks are thin, elongated, low-contrast, and easily confused with stains, shadows, and heterogeneous textures. Existing lightweight networks reduce computation but often fragment long crack paths and blur fine boundaries. We present CrackLite, a lightweight topology-aware network for high-resolution concrete crack segmentation. CrackLite retains a hierarchical encoder–decoder backbone and introduces Direction-Guided Topology Aggregation (DGTA), which estimates local orientation priors and aggregates context along candidate crack directions through confidence-gated fusion. To improve thin-structure delineation, a Normal-Calibrated Local Geometry Refinement (NLGR) module sharpens crack boundaries and suppresses texture-induced false responses, while a training-only auxiliary branch regularizes centerline continuity and boundary localization without increasing inference-time cost. Experiments on a self-collected bridge crack dataset and two public benchmarks show that CrackLite achieves a favorable accuracy–efficiency trade-off for morphology-preserving concrete crack segmentation.