We found that part rotations that were close-but-not-quite-axis-aligned had about 4x worse angular error than expected due to a bug in the rotation compressor.
This post was made 3 years before mine, mine should be considered the duplicate instead.
I’d like to see whether the OP’s issue is resolved though, as it may be more general compression of orientation rather than the fixed bug due to very long parts unlike in my case (note how their example is a long straight section of track, likely many hundreds of studs, while mine was less than 50 studs)