WebNN Adds Bounded Dynamic Inputs Support

On Feb. 18, 2026, developer Tarek Ziade announced end-to-end support for bounded dynamic dimensions in WebNN, enabling autoregressive transformer inference with varying KV-cache sizes in browser backends. The change spans webnn-graph, rustnn, and pywebnn, allowing SmolLM-135M to run token-by-token without WASM fallbacks and keeping decode inside the WebNN backend to improve latency for real LLM workloads.