Actually, I did it the other way round, locked my MSO4000 (EXT REF) to my Rb source and connected some salvaged TTL or CMOS XTAL oscillators (the old ones in the metal cans with DIL14 footprint and only pins 1, 7, 8 and 14 populated). I supplied the oscillators with 5V from my bench supply through 1.2m long alligator leads and no local decoupling (cannot get worse than this...). Frequencies that I tried were 60MHz, 20MHz and 16MHz (no 10MHz to be found in my "collection" and I didn't want to experiment with the "odd" ones...). Those three oscillators didn't produce any jitter like seen with the unstable PLL (TB delay 100ms). Of course, there was some sweeping visible while the oscillators heated up, but this was well-defined and came slowly to a rest as the oscillators reached thermal equilibrium. If I had found a 10MHz oscillator, I would have tried it at the REF input as well.
I guess the REF oscillator inside the Rigol 4000 series isn't that bad, it's got to be the PLL that's not properly set up to lock. You get the same jitter with a good external reference if the reference is connected after configuring the oscilloscope to use the external REF input. Yet, even if one proceeds in this sequence, the external reference has some effect on the timebase (a "jittery phase shift" can be observed in the signal displayed), it just doesn't stabilize.
And yes, both DS2000 and DS1000Z perform substantially better than 0.2ppm jitter in reference to the delay time.
Cheers,
Thomas