Ok. Let me give you a bit more specific description, simply laying down the facts of life.
Before: The RAM (KHX1600C9D3K8/32GX) is XMP DDR3-1600 rated, and XMP was working fine on my original R4E rig with SB (3960X) @ 4.6G and BIOS 4206, 24/7 (4.7-4.8G was Ok too, but I couldn't use it because of insufficient cooling).
Upgrade: New 4960X CPU, BIOS 4901 rig moved "as is" in a new case with a very powerful cooling (I only use water cooling on all my rigs). I've pushed the CPU up to 4.8G (on DDR3-1333) without much hassle. It gets slightly unstable on 4.8G but nothing serious. I'm confident there's even more headroom in this CPU sample and that it could hit 5G+ if I have had even better cooling (closing 85 C degrees @ 4.8G).
After the upgrade, the rig is 100% stable @ 4.7G on DDR3-1333, but not completely stable with XMP DDR3-1600. Nothing major, just rare Prime95 glitches within a specific FFT interval: FFTs between 32K-36K (i.e. smaller FFT are Ok and larger FFT are Ok too). This should be significant.
One of the common approaches to stabilizing the CPU-RAM collaboration is to try adjusting the VCCSA. It puzzled me that it was very low by default (0.921-0940V), and it puzzled me even more that increasing VCCSA would cause instability. Hence my original question.
Needless to say that the RAM passes all Memtest86 etc. testing (after all it's been working fine for the past 3-4 years), and so does the CPU pushed to the limits (various testing). I haven't been able to provoke errors with e.g. OCCT, AIDA64, etc. but I trust Prime95 is the ultimate stability test and as long as one of its workers crashes, I know this will happen in real life too. I'm using this rig for heavy video/music editing and a lot of image processing, so I'm not comfortable with any degree of instability (e.g. crashing a rendering job after 4 hours is not fun at all). On the other hand using DDR3-1333 instead of DDR3-1600 makes a difference for my usage of this rig, so that's why I'm so eager to find a solution.
So what do you think?