OK, cant resist jumping in.
Suggestions to simplify and and test with single link are right on. First, as suggested, setup Nstreme, polling, no framer policy. Use only one set of radio cards. Recommend even removing
second radio, no matter how painful that is.
I have similar setup with lesser hardware peaking at about 40+Mbps TCP using single cards, turbo mode, Nstreme, some framer policy, -50s signal level, very high SNR. Line of sight at 1 mile between skyscrapers, so I know the path is clean.
Absolutely use the default setting for power, but you might have to go further and go to card rates, and lower the maximum power. I recall going to 3dBm at one point. There is a point when such high signal levels and SNR are actually overloading the receiver front ends. It will naturally tell you 100/100 CCQ, but exhibit this strange low TCP performance. High power cards are not needed in some situations, but if you use them in high signal level environments, single cards only, turn down the power. The red flag you have that indicates a problem is that it wont run properly at 54Mbps. It absolutely should run max bitrate if its working properly. I like running in low 60s for RSSI so I know I wont lose the connection, but at the same time, I know it wont overload. Really good cards do what they are supposed to and will hear really well, but like to run in their comfortable range which is card dependent. Secondarily, since the second card, probably even if its disabled, still might be putting out some RF on some frequency. I have been in your situation, but was fortunate to be able to access the radios easily. I recall peaking at about 57+Mbps UDP, and I know it was easily in the 40Mbps TCP range. Also, as suggested, move that channel around and you will be surprised what you find. I have used strange methods to accomplish this. Dont look at signal levels, SNR or CCQ. I run a bandwidth test between radios, and keep changing the channel. The bandwidth test will recover and restart. You will find that one particular range of frequencies will be exhibit drastically different performance. The second method, which yields the same frequency is whatever channel runs at 54/54. In RouterOS 3.x the 'throughput' bitrate will give you the highest rate on the same channels you find get the best results, in other words, they will all give you the same result, just multiple paths to same answer.
For whatever its worth, I hope this helps, so: run constant bandwidth test, look for 54/54 bitrate, find maximum 'P throughput'. Also, lower the power as low as you dare, and maybe even turn off turbo until you find the sweet spot channel, then turn it back on and see the throughput double. I have a feeling you will be happy with the results. Its just another way to look at performance measurement, but these are the metrics that count. As far as full duplex, I dont know whats wrong with <1ms under heavy load without full duplex. I'll bet you will get great results if you can just max out the link.
Would love to hear if this helped. I have been through this. If it helps, then I will comment on a couple other issues I have experienced with certain equipment combinations later....Good Luck