LLM application engineering: RAG, context limits, state and prompt injection

Quick conceptual questions for an applied AI engineer, each answered in a couple of minutes from real experience:

1. When would you use retrieval-augmented generation and when fine-tuning?
2. What do you do when the input does not fit in the context window?
3. How do you keep state across a long conversation or a multi-step task?
4. How do you defend an LLM application against prompt injection?