Wildfire can spread from surface vegetation to the forest canopy through shrubs and small trees known as ladder fuels, making their mapping important for wildfire-risk assessment and forest management. Airborne LiDAR provides detailed 3D forest information, but the upper canopy blocks many laser pulses from reaching sub-canopy vegetation. When LiDAR point clouds are reduced for efficiency, methods such as voxel sampling, farthest-point sampling and token pruning may discard these poorly observed regions. Visibility-aware methods identify observed and unobserved regions, while forest-specific networks perform 3D segmentation, but they do not explicitly reason about hidden, uncertain or task-relevant regions. This study proposes an occlusion-aware 3D perception framework with an agentic reasoning module that combines LiDAR visibility, 3D features and prediction uncertainty to identify important regions and allocate computational resources. The study will support forest managers, wildfire risk analysts and fire agencies in wildfire-risk assessment, fuel-treatment planning and forest management.