Agent runs stay healthy under slow LLM responses

Background agent features (revision summaries, issue triage, domain evolution) now degrade gracefully when the configured LLM is slow instead of failing with “Finch was unable to provide a connection” and timing out agent sessions. Each agent run is now capped at 2 minutes (down from the 5-minute default) so a stalled model call frees its worker quickly, and the LLM HTTP connection pool has been doubled from 8 to 16 to absorb retry and connection-teardown churn. Real concurrency stays bounded by the agents queue, so this also adds no extra load to the model endpoint.