We build a coordinate system for cellular state from 44 transcriptional modules — each defined by a sensor, transcription-factor output, cyclical negative feedback, and receptor-field specificity — scored across 154 normal cell types of the Human Protein Atlas using **single-cell RNA-seq exclusively**. The resulting nine-dimensional eigenspace recovers known biology (embryonic/germline, high-secretion, and inflammatory-myeloid cells occupy interpretable regions) before any cancer data are introduced. Projecting nineteen quality-controlled single-cell cancer datasets into this eigenspace, we find that the malignant compartment of every cancer is biased toward a small, recurrent set of non-origin normal cell types. A non-circular, recurrence-based ranking across the nineteen cancers identifies a robust **two-cell core — megakaryocytes and cytotrophoblasts** — that leads every other cell type on both convergence breadth (a top target in 74% and 68% of cancers, a clear margin to the next cell) and depth (the median per-cancer fraction of malignant cells, on which the cytotrophoblast ranks first at 10.1% and megakaryocytes second at 7.3%). No third cell type leads on both axes — the leading candidate, gastric progenitor, has moderate breadth but only middling depth (tied there with esophageal apical and pancreatic acinar) — so we report the core as a duet and the cells beside it as an enriched neighborhood rather than as fixed members. Convergence onto the duet is enriched **14.1-fold** over the unstructured baseline (18 of 19 cancers above baseline); widening the target to the majority-enriched neighborhood (10.8-fold) or to the broader contiguous eigenspace region these cells occupy (8-fold; mutual cosine 0.70 versus −0.01 for random pairs) keeps the enrichment near tenfold, so the result does not depend on where the boundary is drawn. The convergence **anchor is the placental cytotrophoblast** — the normal cell that physiologically invades, proliferates, and evades immunity, and the single highest-depth target — and the region co-localizes with embryonic and germline cells. Two controls shape the analysis. First, a low-complexity-transcriptome artifact inflates apparent convergence in poor-quality datasets; uniform per-cell quality control (n_genes ≥ 1000) removes it and we exclude datasets that do not survive it. Second, within a cancer, convergence shows **no steep progression gradient** — though the one positive trend, in hepatocellular carcinoma (Spearman ρ = +0.40, stages I–IV), is underpowered at ten patients and cannot be dismissed; colorectal cancer shows no association with grade, T-stage, or nodal status (62 patients). Convergence is largely a feature of the malignant state itself rather than a graded marker of clinical progression. Module-level decomposition resolves a two-component signature: regulatory-checkpoint suppression (the universal convergence driver) plus selective stress engagement. The result gives Abelev's embryonic-state theory of cancer a quantitative, geometric form.