I'm skeptical about the 'non-Markovian' claims. Most noise we see in transmons is dominated by T1 relaxation which is fairly Markovian at the scale of a single gate. Have you subtracted the pulse distortion effects?
I've been looking at the syndrome extraction cycles for the distance-3 surface code. Is there a specific reason you didn't use the rotated layout to save on qubit count?
From a hardware perspective, the wiring density for a surface code is a nightmare. How does the 0.4% threshold change if we account for the heat load of the additional microwave lines?