Learning between the peaks: sharp asymptotics for kernel ridge regression under power-law anisotropy