> Which means that consinit code needs to know (in advance) the driver
> specific cnattach function.

Right. Only md code can know for sure what device the console
should be on. There is a number of possible sources of
information (eg passed by bootloader, compiled into kernel,
firmware/BIOS read or -callback, hardware switch) which are
inherently md.
In most cases, the driver's cnattach() has to be called anyway
in early bootstrap to do some initializations.
The cn_tab pointer has to be set early too.

> I'm not sure I got your point. To me this reads as "we already have two
> or three very evil hacks in place, just add another one."

I don't consider this scheme a hack. There are some kinks
loke the graphics card ambiguity mentioned (and USB keyboard
emulation FTM) which are more or less caused by poor hardware
design, but I don't see where #ifdefs in the friver or attach-time
flags could help here.

Or are there sparc/ofw specifics I'm missing?

best regards
Matthias