It is the RAM sticks, but it’s not the RAM sensor it turns out but the actual RAM memory chips themselves. That explains why everyone seems to find something happening around the same temperature for the sensor, but nothing directly tied to it: it’s not the sensor actually doing the throttling.
I started with your PTM shim sandwich guide (and discovered a lot of the liquid metal had been wicked under the soldered shim somehow) to confirm I was not running out of general thermal headroom in the cooling system. After that modification I couldn’t overrun the cooling system with 100% CPU usage even with multi-hour test–I highly suggest it for anyone who wants the advertised performance of the FW16.
However I was still getting performance throttling with anything making heavy use of the RAM, and stumbled upon the thread linked above (shout-out to Alex_Morozov, excellent work!) which pretty dramatically documented this issue. I tested on my own machine by adding some thermal transfer pads directly on the RAM (both top and bottom) and have been unable to induce the thermal throttling ever since in any workload that previously would trigger it. I’m not sure if it’s truly fixed for any conceivable workload, since Alex continued to have some throttling, but I used Arctic TP-3 which has a bump over most other pads for thermal conductance so it might just come down to either hitting some minimum W/mK or using something soft enough to maximize the surface area contact.