The C99 and C++0x standards both specify that their respective double complex types must have the same alignment and layout as an array of two doubles. This means that you can get away with passing arguments as a void * and have your routines be (relatively) easily callable from either language, and this is an approach that many libraries have taken.
The C++0x standard guarantees (§26.4) that a reinterpret_cast of std::complex<double>* to double* will do the right thing; if I remember correctly, this was not so clearly specified in earlier versions of the standard. If you are willing to target C++0x, it may be possible for you to use this to do something cleaner for your interfaces.
Given that the actual layout and alignment specifications are defined to agree, I would be tempted to just condition the type in the header file on the language; your implementation can use either language, and the data will be laid out properly in memory either way. I'm not sure how MATLAB does things internally though, so I don't know if this is compatible with MATLAB or not; if they use the standard LAPACK approach, then it will be on many but not all platforms in all circumstances; LAPACK defines its own double complex type to be a struct with two double members, which will usually be laid out the same way in memory (this isn't guaranteed), but could follow a different calling convention on some platforms.