Action potentials in cardiac muscle and in neurons obey the same membraneequations. We treat the action potential as a physical object and ask what a singledynamical framework predicts when its parameters are swept across the rangeevolution has explored. Three legs support the claim. (1) Population of models: 852virtual cells built from twelve published excitable models by log-normalconductance perturbation at 25% variability put 97.8% of their between-cellvariance into four descriptor dimensions [family-level bootstrap CI 97.5-99.9%]; adescriptor-permutation null puts only 69.5 +/- 0.6% there, so the collapse is not anartefact of the descriptor choice, though the choice partly fixes the number. (2) Realdata: 2,336 Allen Cell Types neurons projected onto the model manifold occupy aregion adjacent to, but distinct from, the model islands. Direct measurement of thespread to 200% conductance variability (sublinear scaling) shows that conductancenoise alone cannot supply the missing variance at any feasible CV. (3) Mechanisticanchoring: an explicit sarcoplasmic-reticulum calcium loop reproduces calcium- driven alternans in a heart-failure-like regime and matches published humanrestitution kinetics and their hERG-block prolongation (ratio 2.20 model vs 1.94human). Topological charge is exactly conserved in the neural medium at everytracking setting; in the turbulent cardiac regime the ledger is dominated byboundary outflow. On open data, the drug-risk coordinate ranks the twelve CiPAtraining drugs consistently with expert-panel torsade risk (AUC 0.91, p = 0.013) - aninternal-consistency check on real pharmacology, not a blinded test. Every claim isbounded where it is made: the framework survives only in the parts its data support.