> > > Could you try a plain 2.0 kernel again ? Did your 2.0 work with cmdide or
> > > pciide ? At last we'll see if it's the openfirmware misconfiguring the
> > > IDE controller, or if something changed in the NetBSD kernel that cause this.
> > > There isn't much changes between 2.0 and 2.0.2 ...
> > 
> > Finally found a 2.0 kernel on a backup.
> 
> > [...]
> > DEBUG eb164_iointr() vec = 0x900 irq = 0
> > scsibus0: waiting 2 seconds for devices to settle...
> > 
> > unexpected machine check:
> 
> Did this kernel work once in this machine ?

I think so, modulo different PCI cards, and being
connected to a hard drive rather than a CD/DVD burner.

Is the CMD controller known for failing on powerup?

I'm assuming that a fresh batch of "Vector allocation failed"
wouldn't have damaged the part?