Agent runs no longer stall when the configured LLM is slow

Agent runs (revision summaries, issue triage, domain evolution, and the Slack ConversationAgent) degrade gracefully instead of failing with pool exhaustion when the configured LLM is slow. Each Condukt agent run is now capped at 2 minutes, down from the 5 minute default, and the ReqLLM HTTP pool was raised from 8 to 16 connections, so a stalled model call frees its worker instead of starving the agent pool.