IMPACT is a scalable framework for autonomous, rapid traffic incident analysis using existing urban CCTV infrastructure. It combines a low-latency CPU-based vision module for real-time key-frame filtering with the causal reasoning capabilities of multimodal LLMs, reducing costly MLLM calls by over 92% compared to naive sparse sampling. The paper introduces TRACE10K, a dataset with three-tier textual annotations describing accident dynamics at the frame-sequence level.