I imagine it would have difficulty with letters that differ only in scale (x, X) but for things like varying stroke order, stroke direction to achieve the same

I imagine it would have difficulty with letters that differ only in scale (x, X) but for things like varying stroke order, stroke direction to achieve the same mark - which is I think what you're getting at? - the paper suggests just recording more templates for the glyph works well.