Perhaps it is related to the teed line...
I was trying to figure out why that would have any effect at all, and the best answer I could come up with was multipath fading... which is usually a mobile problem.
Perhaps, having 2 copper paths of slightly different length, the signal is following both, and getting a tiny delay in one leg, enough to cause cancellation issues at certain frequencies.
Perhaps, then, the upstream frequencies are more susceptible, or perhaps there's a difference in power that helps cause it.
Have you got before/after graphs of SNR, Hlog, QLN, Bitloading etc?