With 6.9.0@master I managed to get rx5700 initialized once, but then it reverted back to the same “ring kiq_0.2.1.0 test failed” problem. I even thought it was promising that board would work with RDNA out of the box on newest kernels, but I guess it is just not a 100% chance of getting an error.
With a patch ported to 6.9.0 (one const was renamed and option should depend on 52-bit VA space) it seems to work reliably. I haven’t done any more tests yet.