This is where a design philosophy chapter in the documentation would really help,
especially since its brilliance would make for a very nice read.
The self hosted engine (SHE) is in fact extremely highly available, because it always
leaves behind a fully working 'testament' on what needs to run where e.g. in case
of a major hickup or servers (including the one running the SHE) dying.
And that includes instructions to bring up a new instance of the SHE, which will then use
this "testament" to create the next one, as workloads and systems change.
So as long as there is always a good enough testmament and an SHE running long enough to
create the next iteration, there is no need for the SHE to run at all: the VDSM daemons on
each host will faithfully do their work without stepping on each other's toes.
The principle isn't really that original to oVirt and has been used for things like
mainframe job scheduling systems for decades. But it's extremely solid in principle as
long as the "testament" or execution plan doesn't need to be to complex. You
can even run a mathematical proof on it then.
On the other hand, two servers will only create chaos, because they'd have to decide
who is right. That can take so long, the winner might die during the negotiations and then
what?