Head and neck squamous cell carcinoma (HNSC) exhibits profound intratumoral heterogeneity, driven by dynamic interactions between malignant cells and the tumour microenvironment (TME). Using consensus non-negative matrix factorisation (cNMF) on multi-site HNSC single-cell transcriptomes, we resolving conserved meta-programs define cellular ecosystems. Six major epithelial programmes emerged, including a differentiation-associated programme (Epi_Diff) correlated with SPDEF activity and favourable patient prognosis, and an invasive programme (Epi_pEMT) potentially controlled by TEAD4-mediated ECM remodelling, exhibiting partial EMT markers (VIM, TGFB1). Compartment-specific crosstalk analysis revealed Epi_pEMT cells may coordinate with mCAF1 fibroblasts and TAM(SPP1) through COL1A1-CD44 and SPP1-CD44 signalling, suggesting potential formation of a pro-invasive niche. Conversely, Epi_Diff cells may interact with NK/T cells through CEACAM5-CD8A and CCL5-ACKR2, and may contribute to inhibit immune infiltration. Multi-compartment correlation analysis revealed three ecosystem-level patterns: (1) Inverse association between Epi_Diff and Epi_pEMT (Spearman R = -0.43); (2) Negative correlation between mCAF1 abundance and cCAF frequency (R = -0.48); (3) TAM(SPP1) dominance inversely correlating with both TAM(C1Q) (R = -0.43) and NK/T infiltration (R = -0.36). These axes suggest a potential hierarchical ecology framework where lineage-specific polarisation and inter-compartment synergies may collectively govern disease progression.