Agent runs degrade gracefully under slow LLM responses

Agent LLM runs are now capped at 2 minutes, and the HTTP connection pool used by agent LLM calls has been doubled from 8 to 16 connections. This prevents a slow model from exhausting the agent HTTP pool and cascading into failed agent sessions, while leaving the dashboard and spec-write SQL paths unaffected. Agent workers release their database connection before making LLM calls, so the improvement is focused on HTTP concurrency for background agent features.