Individual tree segmentation from LiDAR is central to automated inventory, yet many evaluations cover only one forest type or one methodological family. It also remains unclear whether point-level and instance-level metrics rank methods in the same order, which complicates operational choices. We therefore harmonise preprocessing, a binary tree versus non-tree semantic task, and a joint evaluation protocol. Representative unsupervised (graph-based TreeIso; region-growing TreeX) and deep learning (vote-based TreeLearn; mask-based ForestFormer3D) pipelines are compared on both public benchmarks and a fused unmanned aerial and mobile laser scanning benchmark from rugged subtropical Hong Kong woodland called HKTrees. Performance is strongest on regular coniferous plots. Broadleaved heterogeneity and crown overlap lower instance recall relative to semantic scores. HKTrees is the hardest regime and shows pronounced domain shift. Point-level and instance-level rankings are not always aligned. We discuss trade-offs in annotation cost, generalisation, and interpretability and outline expansion of HKTrees dataset for larger-scale training.