Argh—of course I meant to use cc splines in my example, but then forgot. This made no difference though.
However, when I added a group-level intercept, I get diverging graphs from gamm4 and brms:
http://rpubs.com/rubenarslan/486368
The dashed line is predicted from gamm4, the solid line with CI is from brms::marginal_smooths.
Does this help? Or is it not a real difference? I haven’t experimented with trying to make the difference bigger.