How can I compute the bounding box of a text string without loading its glyphs before?
This is not possible in general. Reason is that hinting distorts the glyph shape for optimal rasterization, and this process sometimes creates outlines which have considerably different metrics. The TrueType format provides the (optional) ‘hdmx’ table which contains device horizontal metrics for selected pixel sizes, but even here the vertical metrics are missing. It is probably best to use both a glyph and a metrics cache to avoid recomputation.