In the standard curvature-perturbation scenario, PBHs form from the collapse of superhorizon curvature fluctuations after horizon re-entry. The predicted abundance is exponentially sensitive to the collapse threshold and hence to the shape of the primordial curvature profile. In this work we develop a finite-action framework to describe curvature-profile dispersion around representative peak profiles. Using a multipolar Fourier--Bessel decomposition, we separate the local peak variables of the Gaussian field from residual radial and angular deformations, normalized by their Gaussian action. We apply the formalism to spherical numerical-collapse examples in order to isolate the effect of radial shape dispersion. For finite-width spectra, and in the presence of logarithmic local non-Gaussianity, we compute the collapse threshold as a function of a coherent shape variable and combine the result with peak statistics. We find that the dominant contribution to the PBH abundance is not necessarily the mean profile, nor simply the profile with the lowest threshold. Instead, it is selected by a competition between the Gaussian cost of realizing a coherent deformation and the exponential gain associated with lowering the collapse threshold. Broad spectra and negative non-Gaussianity can make rare shape deformations dominate the abundance. In the examples studied here, the dominant branches can correspond to several-sigma coherent shape fluctuations while enhancing the integrated abundance by orders of magnitude. Equivalently, including shape dispersion can reduce the power-spectrum amplitude required to obtain a fixed PBH abundance. Our results show that residual profile dispersion is a genuine statistical ingredient in PBH formation and can be quantitatively important for accurate abundance estimates.