That's an interesting plot, Jeremy, and happens to strike me at a weak moment when I'm considering whether I can find space for an old HP606B signal generator on offer locally. Your example bears close thought - interesting to see how good such a simple unit might be beyond a few kHz offset.
For my immediate MF non-linear PA application I do need to get at least to 10 Hz offset but not, as I mentioned, down to incredibly low noise floors. There are various physical mechanisms involved in the PA but, in essence, it comes down to assessing additive noise, where the original driver signal can be used as the reference. That's good, largely because it removes issues of drift and the need for any phase locking.
More generally, your comments about drift led me to wonder about how essential the phase lock is in modern practice. I guess if the drift is slow/small enough to allow a few stable spectra to be acquired, and allow good calibration of the levels measured on the baseband analyser, that may be good enough. The locked system is undoubtedly more desirable but with, for example , a reference synthesizer and a unit under test both locked to (e.g.) a typical GPSDO, it occurs to me that one might get away with no phase locking for at least some measurements. (The SML-01 has a carrier phase control which might be useful for setting the optimum operating point for the mixer).
73, Peter