In Part IV of this series, we benchmarked local LAN and Tailscale mesh network latency for grav-ai-chatbot, proving that network transport was fast and reliable. But shortly after deploying reasoning models like qwen3.5:2b to production, users began reporting a perplexing bug: sending a sim...
