Total: 1
We present a population-level spectral analysis of fast radio bursts from the second CHIME/FRB catalog using three curvature-radiation-motivated templates: point-source, one-dimensional bunch, and paired-bunch cavity models. Fits are evaluated with reduced chi-squared $χ^2_r$, AIC/BIC, and the Ljung-Box residual autocorrelation test. All three templates yield median $χ^2_r$ values close to unity for both repeating and non-repeating bursts. Repeaters show narrower $χ^2_r$ distributions than non-repeaters, with statistically significant but modest population-level differences. AIC favours the one-dimensional bunch model for the largest fraction of sources, whereas BIC increases the relative preference for the simpler point-source model. However, residual autocorrelation remains widespread across all models: only 15%-21% of sources simultaneously satisfy goodness-of-fit and residual-independence criteria, indicating persistent structured residuals beyond the tested templates. These results suggest that while curvature-radiation-motivated geometries capture the dominant spectral envelope of many FRBs, additional physical ingredients or spectral components are required to describe the fine-scale spectral structure of the data. The inferred coherence scales are $\sim$16-18 cm for the one-dimensional model and $\sim$25-28 cm for the cavity model.