Context engineering is where the real leverage is right now. Most people focus on model selection but the retrieval and memory layer around the model makes a bigger difference in practice. What's your approach to managing context window limits — chunking with overlap, or some kind of relevance scoring before injection?