The growing backlog of the Brazilian Judiciary demands scalable semantic organization beyond keyword-based retrieval. This study investigates the trade-off between topic coherence and thematic granularity in neural topic modeling for legal texts. We cluster 1,831 Brazilian labor summaries using BERTopic with LegalBERT-pt embeddings and evaluate the impact of the min_topic_size parameter on coherence (Cv), granularity, and diversity. We propose a balanced evaluation score integrating coherence and granularity. Results show higher coherence (Cv ≈ 0.78) than Latent Dirichlet Allocation (Cv ≈ 0.49), while enabling controlled topic resolution. These findings provide a practical criterion for tuning topic models in large-scale legal systems.