There are limits to what I can say here, because we don't have anyone who knows much about this algorithm (the man responsible is long gone). Further, I'm not sure that it works very well.
The modulation appears to be an artefact – basically the algorithm has to do a lot of crossfading between different bits of audio, and it may not always pick the right place. This will be material dependent, and worse with a true stereo signal. It will also be worse the higher the time change.
Because it is not possible to directly increase or decrease time (it tends to go at its own speed), the algorithm works on a buffer of audio, which it can play out at the required speed. The length of this buffer is the "delay" parameter, and the current buffer position is the "now" value. The "dtrig" meter shows when this wraps around, giving a jump in the audio. Don't know what the "ztrig" does.
The maximum length of the buffer is 16 seconds (hence your observation). I can believe that weird things will happen if you get too near the maximum.
This algorithm will work best with a mono or near mono signal, and small time changes. Your 5% should be OK. You will get some roughness with extreme pitch shifting.
If you don't need real-time performance, the "big sampler" will work better. But then, for the same reasons, so will a DAW.
