The automation of operational stewarding in high-speed motorsports traditionally relies on resource-intensive video tracking algorithms, which introduce severe latency, synchronization friction, and computational overhead. This paper proposes a novel Time-Series Vision-Language Model framework that bypasses spatial video processing entirely by employing a Chart-as-Image methodology. By ingesting high-frequency 20 hertz Formula 1 telemetry data via the FastF1 interface, the system deterministically isolates four-second temporal anomaly windows and renders standardized four-channel multi-variate plots encompassing velocity, throttle application, braking state, and engine revolutions per minute. These visual artifacts are subsequently evaluated by a frontier Vision-Language Model conditioned on the 2026 Federation Internationale de l Automobile Section B Sporting Regulations via an in-memory Retrieval-Augmented Generation pipeline utilizing normalized inner-product vector search. To ensure operational reliability, each generated Chain-of-Thought verdict is validated against a deterministic heuristic baseline and a three-factor hallucination guardrail prior to presentation within a human-in-the-loop adjudication interface. Empirical evaluation across a curated register of seventeen historical motorsport incidents across seven circuits demonstrates high diagnostic precision, near-zero structural hallucination, and substantial reductions in token consumption and adjudication latency compared to continuous video streams