should be solved with r17829. the (already present) hack never seemed to work correctly due to the wrong default scalefactor.
musepack uses differential scf-coding in time domain. if you decode from the beginning everything is fine. if you seek with in a file, the decoder may not know the reference-scf after seek -- in the file i've tested this happened with highest subband. in this case the decoder uses a much too high scf what results in burst noise. solution (or better hack) is to use the lowest scf (equals fading out of the subband) until there was a reference scf decoded (1st non-differential coded scf).
please report, if the problem is solved now.
EDIT: The bug is reported as fixed with r17829.