However, I'm sure the day is not far off when phenix.refine or the like will check if the starting R factor is too high and just "automatically" invoke a run of MR to see if something clicks.

I think the latest Phaser code actually does the reverse: if the R-factor is already relatively low, it just outputs the search model. The more problematic (and very common) situation is where the structures are slightly isomorphous and rigid-body plus restrained refinement alone could work, but MR might work better - I don't think anyone has comprehensively evaluated this. We usually just run Phaser because compared to rebuilding and refinement, it's simply not that much of a bottleneck.

-Nat