docs: add agent shutdown and timeout diagnosing guide

[dbroeglin]

Why

While investigating a DeepSWE OPA trajectory that "stopped at 95 turns" and failed, the symptoms were easy to misread: the run looked like it hung on a single command for 600s, which is the same signature you'd expect from a proxy or network stall. The real cause was the agent repeatedly running the full go test ./... suite, with the 600s waitForPendingBackgroundTasks acting as a configured shutdown ceiling rather than a measured hang. This doc captures that investigation so the next person can triage a stalled or shut-down agent quickly instead of re-deriving it.

What this adds

• docs/diagnosing-agent-shutdowns.md - a triage guide covering:
  • There is no turn limit; sessions end for other reasons.
  • The 600s waitForPendingBackgroundTasks is a shutdown ceiling, not proof of a hang.
  • A gap-analysis technique (with a Python inter-event-gap snippet) to distinguish a real hang from ordinary long-running work.
  • A proxy-vs-activity signature table to rule out egress-proxy issues.
  • The worked OPA go test ./... example plus remediation: prefer targeted tests, then the COPILOT_TASK_WAIT_TIMEOUT_SECONDS env knob, then allowlisting a host.
  • The verified env / extra_args injection points in the Pier agent.
• Links the new doc from the README docs index.
• Cross-references it from docs/collecting-run-data.md next to the aborted-sessions limitation.

Notes for reviewers

• Docs-only change; no code behavior changes. ruff check . is clean and the full pytest suite passes (188 tests).
• The branch also merges the latest origin/main to stay current, so the net PR diff is just the three documentation files (+195 / -1).