For now I'd go with option #2 - pick the lowest resolution where text is still legible and use that with "low quality" fullscreen mode.
In the C3 runtime we'll probably add a feature that lets you have low-quality and high-quality scaling on different sets of layers, but that's a long term plan at the moment. Also it can still impact performance, since each layer you render at full resolution adds another full-screen draw - so just one extra layer of high-quality text would bring you up to 2x fullscreen draws. I guess that's still better than 4x though!
When it comes to fill rate, the important thing is the size the image is drawn at. That's why low quality fullscreen mode is so effective: it draws everything at a small size, and only stretches it up at the end.