Waaaait a minute... How quickly could that thing switch fonts? Could the switch have occurred between two scanlines?
IIRC there was nothing to suggest that it can do any switching at all under software control - only manually. But then, that thing came in a used 5155 with no documentation whatsoever. The actual fonts in there are all different encodings for ASCII containing the Greek alphabet... perhaps because DOS code pages weren't yet standardized yet at that point in time. The thing is also hooked up to the LPT port in some way, so who knows what the point was. Could've been for a very specialized usage.
Applause for the effortpost, but I have to say that in my experience (and I have some ), the most glaring problem when "faking" APA using text mode is the attribute clash - meaning it's a bigger artistic limitation than the lack of every conceivable pixel pattern. More of those would help, but in a "diminishing returns" sort of way, as long as you're limited to two colors per group of eight pixels.
The already-mentioned CGAArt reads the the font bitmaps from an image of the CGA character ROM, so you can actually test that theory by modifying that image with the 256 different 8-bit patterns and seeing what happens even with one scanline per row.
If you're already spitballing ideas that require hardware modificiations, IMO, you might as well do a redesign with more RAM rather than ROM, and get an extra bitplane or three to decode for real 16-color APA modes!