Did you run MemTest86 (not the MemTest86+) and if you did – did you run the Hammer Test?
My faulty kit would otherwise pass all other tests but this one.
Which is why I said "apart from setting frequency, voltage and timings through XMP, DOCP and the like". After frequency, voltage and timings are set (either by SPD, XMP, DOCP, etc.), the RAM modules should not receive any "special treatment" from the system firmware (again, to the best of my knowledge).
That's patently untrue. The only information contained in XMP are VDIMM (usually 1.35V), the speed of the profile, and the primary timings (tCL, tRCD(RD/WR), tRP, tRAS, and tRC) and possibly VSOC. All other timings, procODT, CADBUS drive strengths, minor voltages on the CLDO rail, command rate and Geardown, are all determined by the board upon RAM training during first POST. Guess what tells the board how to train? The AGESA firmware and its vendor implementation, obviously.
You're running the hardest possible stress on the UMC. 4 dual rank DIMMs. With the right kit, I don't doubt that you should be able to do 3200MT/s on 4 2R DIMMs; your average upper midrange 6-layer Tomahawk is QVL'd for 3600 on that setup.
Your repeated emphasis on the "Hammer test" in MT86 doesn't shed any useful light on what needs to be done to solve this problem.
No one worth their salt in the DRAM overclocking community uses slow-ass Memtest86 in testing, outside of initially verifying once that a newly purchased RAM kit is free of hardware defects. It's slow as hell, runs outside of Windows (kinda defeats the purpose of a memtest, you don't use your computer in DOS), doesn't work the CPU like any respectable memtest should, and I don't even know if it tests the full capacity of DRAM. Run HCI Memtest, Karhu, or Testmem5 with the anta777 Extreme1 config, and see if you get errors (the % coverage at which errors show in HCI, or the # of errors in a run of TM5).
If you get through HCI to 1000%+ coverage without problems, or at least 2-3 runs of TM5 anta without errors, move onto tests that will stress the UMC. An hour or two of Prime95 Large FFT, OCCT's memory test might also work.
All links in here:
MemTestHelper/DDR4 OC Guide.md at master · integralfx/MemTestHelper · GitHub
That's all without stating the obvious: you're mixing and matching 2 Corsair kits. Corsair is by far the worst offender in mixing various ICs from different vendors under the same "product number". Not even the "revision number" on the sticker properly identifies what the IC is. You can only use Thaiphoon Burner to help you make an educated guess. There's dozens of different ICs out there from Samsung, Micron, SK Hynix, and Nanya, each has differing preferences to timings and speeds, and reacts differently to voltage (VDIMM).
You can't expect predictable results when you're mixing kits. The QVL is designed for a single kit of 4 DIMMs, which is assured to contain the same ICs across 4 sticks. That's not what you have.