Hi,
This is actually not restricted to font locking as demonstrated by the
test case. My initial report was a bit unclear about that. It also
affects settings like:

(setq compilation-scroll-output 'first-error)

where the jump is done to the header instead of first error. My
impression from reading the code is that we would want to avoid running
(compilation--parse-region) on the header to avoid problems like this.
I am attaching a patch that superficially seems to fix these issues.
I have a hard time judging if this fix has any bad side effects or
similar though.

I am definitely missing a bit of context how the related functions are
invoked. In testing of my own package I have seen what seems to be
multiple rounds of font locking that erase previous rounds, probably
coming from buffer changes in my downstream filter-function. In that
setup, the jump to first error has been a more stable way of reproducing.

/David