Epithelial tumors are characterized by abundant inter- and intra-tumor heterogeneity, which complicates diagnostics and treatment. The contribution of cancer-stroma interactions to this heterogeneity is poorly understood. Here, we report a paradigm to quantify phenotypic diversity in head and neck squamous cell carcinoma (HNSCC) with single-cell resolution. By combining cell-state markers with morphological features, we identify phenotypic signatures that correlate with clinical features, including metastasis and recurrence. Integration of tumor and stromal signatures reveals that partial epithelial-mesenchymal transition (pEMT) renders disease outcome highly sensitive to stromal composition, generating a strong prognostic and predictive signature. Spatial transcriptomics and subsequent analyses of cancer spheroid dynamics identify the cancer-associated fibroblast-pEMT axis as a nexus for intercompartmental signaling that reprograms pEMT cells into an invasive phenotype. Taken together, we establish a paradigm to identify clinically relevant tumor phenotypes and discover a cell-state-dependent interplay between stromal and epithelial compartments that drives cancer aggression.