The subsetting strategy is somewhat crazy for now: Simply zero the glyph data for all unused glyphs. While the CFF table itself doesn't shrink, the actual embedded font is compressed and greatly benefits from the repeated zeros.
This already compresses the fonts a lot (~90% for NotoSerifCJK), but they are still quite large.
Therefore, the plan of action:
- First, find more data that can be zeroed out.
- Then _maybe_ see whether we can instead properly rebuild the subsetted font.
d101612414