Basically the same reason why FPGAs never caught on for general purpose embedded purposes: they are way too expensive because of their flexibility to even be considered for the kind of BOM costs you're targeting.
Just like in FPGAs, FPAAs are large seas of generic analog cells. For any given filter, you not only need to use the specific elements in these cells that you need, but you also need to factor in a whole bunch of CMOS switches to tie them together and memory to keep track of what goes where. Lastly, an FPAA by design is made to be flexible, so it will implement a set number of cells, often much more or at least a couple tens of percents more than you strictly need for a specific application. This means that there is a tremendous amount of overhead in this type of filter/control loop design.
Next, as you may know from signal theory, it's way less die-efficient to implement analog stuff than it is to do digital calculations. This means that implementing a filter inside a specialized microcontroller - e.g. a DSP - is more efficient than implementing it in an analog fashion.
All of this makes FPAAs a technology that is fundamentally less efficient and more expensive than either other, digital, flexible solutions like DSPs, or than purpose-built analog solutions.
Add to that the not-insignificant burden of requiring people to learn a non-procedural programming language that is fairly alien (FPAAs require yet something else than VHDL/Verilog/C) and the quite significant cost of developing and maintaining such an exotic toolchain, and you don't make matters much better, both on the consumer and producer side of the equation.
So slightly depressingly, it mostly boils down to money.