I still maintain that we are doing a lot of extra work (LOC and
CPU-wise) just to guard against the impossible.  On the other hand there
are probably better places to look for optimization opportunities.  For
instance, I noticed that when completions arrive form the server, we
call eglot-current-column once for each candidate, although most of
those will never be used.

Also, in a few years every serious server should support the
codepoint-based counting method anyway.