By the sound of all of this, I would say we need to do something like X-Plane does--give the AI object its own FDM, instead of relying on the ai system to figure all this out, which seems to be impossible.

We need no more and no less than what I wrote - a numerically stable solver for orbital dynamics. The accuracy of analytical orbits seems quite enough though.

(You might not know this, but I am a theoretical physicist with more than a decade experience in numerical solutions of physics problems - so I frankly think it's unlikely that this discussion will significantly expand my insights what needs to be done - so unless someone else actually wants to start before orbiting AI objects reach the top of my to-do list, we're not in need of ideas but coded algorithms).

For those who have actually read the operations manual, we can now do a new class of interesting failure scenarios - helium pressure loss in an RCS or OMS pod, dropping fuel pressure due to blowdown dependent on the fuel fraction at the helium leak, once this fails cross feeding either RCS to RCS, OMS to OMS or OMS to RCS - with today's version, the cross-feeding valves now actually work as they should. Not sure why this would be needed, but we can also do partial cross feeding since we have separate 12 and 345 manifold valves - so you can cross-feed the 12 manifolds of the left RCS from the right RCS and feed the 345 manifolds of the left RCS from its own tank.