Agent runs are more resilient to slow LLM responses

Each agent run is now capped at two minutes, and the dedicated LLM HTTP pool size was doubled from 8 to 16 connections. This prevents a slow model from exhausting the agent HTTP pool and causing agent features such as revision summaries, issue triage, and domain evolution to fail with connection errors or session timeouts. The dashboard and spec writes are unaffected, because agent workers release their database connection before calling the LLM.