Finally cracked using pi.dev to talk to llama.cpp, using vulkan, using the igpu.
So far it's doing useful-ish work, mostly scaffolding, writing requirements markdown, boring stuff. It's slow, but it's working, and I could run it off of a solar panel.
Observations:
Prompting to generate/develop prompts is more effective than I expected it to be. It's also stubborn and frustrating. Some parts it gets completely stuck on, and doesn't respond to me wanting to rewrite or remove. The constraints of the hardware make it better to write the prompt to the filesystem early, edit manually, then go back for another round.