Because that'd be equivalent to computing the convolution explicitly. The whole approach is based on tricking Overlapp-Add to support compactly supported interpolation kernels between representative convolution kernels. Easiest would be bilinear interpolation between between the surrounding representative kernels, which mostly translated to using the bilinear weights before using the fft convolution, and deferring the summation of the interpolation to be fused with the addition done due to overlap-add.