The issue, as I understand it, is shown in the image below:

Discussed with @mem48 (who wrote most of the current implementation) and @wangzhao0217 (who is working on a fix).

@dabreegster (thanks for flagging this), does that match the issues you've spotted? Any ideas for a quick fix welcome. My thinking: post-processing step to break linestrings at every linestring start and end point. Hope that makes sense.