Not a question but a small contribution, since I was not finding the information anywhere.
The problem appears even with a proper utf8 locale with "older" versions versions of less.
If you cannot update to a version of less correcting the issue, I found a workaround :
export LESSUTFBINFMT=*n%C
this instructs less to process any Unicode code points that were successfully decoded and display them with normal attributes (*n) and as-they-are (%C).
Notes:
- I found the %C format by trial-and-error, I didn't find it explicitly documented.
- I found this description of the issue : redhat bug 1074489
It mentions a newer version of less solving it in 2014, but I still run into this problem with a less-458-9.el7 from 2018.