I decided to go with samples because it's very easy to add new wave forms. Create an array of floats for one wave period, plug it in, and you're done. AFAIK its the way sine waves were done in hardware due to how computationally expensive they are to calculate. Plus if I get around to adding normal non-looping sound files to this system it's done in the same way.
My gut feeling is that it is better to go back to implementing the core of this system as if it was on a hardware chip - using simple hard-coded algorithms, different for each tone, based on bit registers with different purpose per waveform - and then add a layer on top of that to convert instructions like "channel 1, triangle wave, 440Hz, half volume" into the right register values.
Eventually I want to play around with phase modulation, like what was used on the Yamaha chips used in Adlib sound cards or the Sega Genesis/Mega Drive, once I get my head around how the operators worked in hardware.