Replace cloud-based LLM providers (Anthropic/OpenAI) with a local Ollama
sidecar running lightweight quantized models (qwen2.5-coder:1.5b as primary,
phi3:mini as fallback). All generation stays under 60s on CPU, data never
leaves the network, and there are zero API costs.
Updated: architecture diagram, component description, technology choices,
project layout, docker-compose, env vars, and consequences sections.
