The arXiv preprint (submitted May 2026) reframes supernova classification not as a supervised learning task but as an unsupervised population mixing problem, fitting ZTF Ia and Ibc light curves to a radioactive-decay semi-analytical model then decomposing parameter distributions via Gaussian Mixture modeling to recover the Ia fraction without any spectroscopic labels. This achieves ≥90% accuracy across varying mix ratios while testing photometric, spectroscopic, or null redshifts plus limited labels. As a preprint it lacks peer review and draws from an unspecified ZTF subsample size, limiting claims about edge cases such as heavily reddened or peculiar events. Mainstream coverage overlooks the direct link to dark-energy systematics: photometric-only classification at LSST scale removes the spectroscopic bottleneck that currently caps Type Ia samples used in Hubble-diagram analyses, echoing the shift from spectroscopic redshifts to photometric redshifts in weak-lensing surveys. Related work in the Dark Energy Survey photometric-classification pipeline (arXiv:2007.14403) and the PLAsTiCC challenge results (arXiv:1810.00001) shows that label-dependent methods degrade under distribution shift; the mixing approach sidesteps this by optimizing shared population parameters directly. Limitations remain: the model assumes only two classes and a specific decay-powered template, so contamination from II or superluminous events could bias the recovered fraction. Still, the method points toward a future where cosmology constraints tighten through sheer photometric volume rather than per-object spectroscopy.