The "Apply Peak to All" feature tries to make reasonable guesses about which peaks match, but certain circumstances cause it to guess poorly. It currently uses a dot product of fragment ion abundance, along with retention time as the primary criteria; but a small weight is also given to percent peak area, which is likely the cause of failure here.
Unfortunately at the moment this algorithm has no parameters, so its results can't be changed. I've opened an issue for this to make improvements in the future.