Core: better integration at high resolution
Inadequate merge statistics seem to be a result of poor integration of high resolution peaks, namely due to the low quality of the shape assigned to the high resolution peaks. The shape assignment procedure assumes that strong peaks surrounding a high resolution peak are well characterised enough to give a good estimate for the shape of the peak via the mean covariance. This is often not the case because 1. there are often few neighbours of high resolution peaks; 2. the neighbours of high resolution peaks often also have poorly characterised shapes.
The solution to this could come in the form of a "bootstrapping", first using the found peaks to generate an initial guess for the shapes of the predicted peaks, but then using these intiaal guesses to iteratively refine the shapes at high resolution.