context-compactor

// Token-based context compaction for local models (MLX, llama.cpp, Ollama) that don't report context limits.