Unified, provably consistent model-based clustering that jointly selects variables and handles MNAR via a data-driven penalty and explicit missingness–class modeling, validated on transcriptomics.