These two bugs (bug 1572738 and bug 1572451) are stylo regressions.
When font-family changes, we try to recompute the font-size with a length /
percentage combinations in case the generic family changes, so the user
preferences are kept.
When calc() is involved, we clamp to non-negative too early, via
NonNegativeLength::scale_by.
I think we should generally dump this "try to track font-size across calc()"
thingie, as as various comments note it is not quite perfect, and it's not clear
how it should work in presence of min()/max().
This patch fixes the issue and simplifies code a bit, I may consider removing
this altogether in a follow-up.
Differential Revision: https://phabricator.services.mozilla.com/D41776