Interesting - I just flipped through Ken's books and didn't find the details in between that you were looking for.

I'm used to seeing Verilog-A used for compact models (eg, BSIM4), a set of equations that end up with things like

 I(d,s) <+ id_chan;
 I(g,s) <+ ddt(qgs);
 I(g,d) <+ ddt(qgd);

In that case, the Verilog-A compiler for most simulators will try to adapt those equations to look like

 i(v(t)) + d/dt q(v(t)) + u(t) = 0

as in the transient analysis chapter (4) of Ken's Spice & Spectre book.

So, the ddt() expressions end up contributing to the "q" in that equation (current is the change in charge).  For dc, the simulator doesn't need any of the q equations (and may be able to partition the code so that it doesn't compute them).  For ac, it will linearize about the operating point, taking symbolic derivatives of the current and charge equations.