The following answer was sent by Simon Wood by e-mail:

The smoothing basis is not orthogonalized to the linear trend, so you would not expect the slope estimates to match those from lm. The point of this construction is really to enable testing of whether anything is needed beyond a linear trend, i.e. whether the smooth is needed rather than a simple linear effect.

I can't say I understand what exactly the linear slope represents in this model, but it sounds like the model is not supposed to produce a separation into a linear and a detrended nonlinear component as I had falsely assumed. To do that, Wood suggested one could

clone the smooth constructor and add the code to orthogonalize the basis