Could you show comparison shots to illustrate that is the case and why would it only show "sometimes"?
If one understands the compression algorithm used it is quite clear why it only shows sometimes. For example original star trail artefacts need very very specific high contrast edges to appear: bright enough star trails with a background that makes the artefacting show.
Like RAWDigger article on this says: "For a larger span, the scheme does introduce the rounding error, which may lead to posterization. In other words, if the chunk of 32 pixels spans across an area that contains a large variation in brightness, the data in the block is not exact, but is only an approximation."
Compression blocks that contain star trails have huge data span, bright trail on darkish background ==> lots of approximation/rounding errors that show very clearly if pushed.
In Lloyds example take compression looks to cause different problems, but root cause seems to be the same. Not enough different values available to draw nice patterns over a huge area. Like said, big percentage of each color channels 10M pixels are picked from more more than 30 or so luminance values. The value distributions in the histogram look bad. Not a fun case for demosaic algorithm.
I have propably 35000+ pics on the different E Mount cameras in over 4 years and I have had single incident of compressed RAW behaving really badly. But I'm a M. Sc. in Computer Science and now that I have studied the compressed algorithm enough I have no issue understanding why it would fail quite badly when certain stars align badly.