I'm in awe with how many bugs this test revealed.  All fixed.

We need to know whether the glyph exists, so we can fallback to
composing / decomposing.  Assuming that glyph==0 means "doesn't exist"
wouldn't work for applications like Pango that want to use different
"doesn't exist" glyph codes for different characters.  An explicit
return value fixes that.

Makes it clear that kerning is in visual order.

More portable.

Such that user_data and other finalizers cannot resurrect object

Test passing now.