I will also say that in the quest for perfection, I have found a lot of imperfect so I don't mind mentioning any drawbacks to recommendations.

For instance, I've been testing a PCI-based ATI Mach32 card with some vintage OSs. It has a lot of positive qualities. In NT 3.1, it actually draws window objects faster in 64k-color modes than 256-color modes. It also doesn't handle palette swaps well in 256-color mode, so I think it runs better in 64k. At the same time, the card produces some random flickering when drawing windows in 64k-color mode.

My Matrox Impression+ (PCI, 2MB) can handle these modes better without the flickering. It isn't perfect either, but I think it's better than the Mach32 card and certainly better quality than VLB or ISA-based cards from the NT 3.1 life cycle. At the same time, the Matrox driver implementation leaves a lot of capability on the table such as the ability to set custom resolutions and refresh rates for non-standard aspect ratios.