I guess I should respond, as I'm responsible for the blog posting that you linked.
I tested the relative phase stability using amplified noise that was fed via a splitter onto the two dongles. I cross-correlated the noise and found that there was a deterministic frequency difference between the channels. I recall this was a fraction of a radian per second. I just calculated what this small deterministic frequency difference was and removed it. I found that I could use the same constant during the whole experiment and it also didn't change after power cycling the dongles. This is how I produced the flat line IQ plot that is in the blog posting.
Based on your plots, you are mostly also seeing a constant deterministic frequency difference (the linear slope on traces 3 and 4). Trace 3 has a phase jump for some reason and I don't know what is happening with trace 1.
In my tests, I only measured data for several hours at a time. I didn't see any loss of lock or weird behavior during my experiments other than what I indicated above. I have heard from several people that they have managed to reproduce the dual rtl_sdr dongle configuration and to even use it for passive radar. I have even seen people taking the proper engineering approach and feeding a clock with a fanout buffer. I haven't seen any cross-correlation plots from these devices though.
I have had issues with heating on the smaller dongles. In these cases the down converter stops working. A heat sink fixed this issue.
Keep trying, I think you are almost there. Try recabling the clock differently or try using another pair of dongles -- there is a lot of variability between the dongles. Try looking at the clock signal with a scope to see what the levels are with a working individual dongle.
I assume you are doing this for fun (that's why I did it at least). There are much easier ways to get multiple coherent channels into your computer with much better fidelity.