Hierarchical community structure appears naturally in many networks, but the multiscale edge heterogeneity it creates can make standard spectral perturbation arguments misleading. This talk studies spectral clustering based on the unnormalized graph Laplacian for hierarchical stochastic block models, with emphasis on recovering the first split of the hierarchy. We first characterize the population Fiedler vector and show that, under weak assortativity, its signs encode the root partition independently of lower-level connection probabilities. We then explain why dense small communities can help, rather than hurt, root-level recovery, even when within-leaf probabilities are much larger than root-scale probabilities. The main results establish weak and strong sign consistency for the empirical Fiedler vector under multiscale heterogeneity, including a refined balanced strong-leaf regime proved through direct nodewise and leave-one-out analysis. Synthetic and real-network experiments compare unnormalized, adjacency-based, and normalized spectral methods, illustrating when the Laplacian captures the top-level hierarchy and when degree heterogeneity changes the picture.