fix: recover agent state on timeout instead of losing progress

[LeeCampbell]

Summary

• agent-loop.sh: run_claude() now captures timeout exit codes instead of letting set -euo pipefail kill the script, ensuring sync_state always runs to commit and push progress
• entrypoint.sh: Timeout (exit code 124) is treated as recoverable — the loop continues to the next iteration instead of breaking, allowing the state machine to pick up where it left off

Context

When running ./scripts/run 141, the agent completed all implementation work for issue #141 but timed out right as it was about to mark tasks complete and move plan files to done/. Two compounding bugs prevented recovery:

1. set -euo pipefail caused agent-loop.sh to exit before sync_state could commit the work
2. entrypoint.sh treated exit code 124 (timeout) as fatal and broke the loop

The agent's work was lost and no PR was created despite all code changes being complete.

Test plan

• Run ./scripts/run <issue> and verify that if Claude times out mid-iteration, the work is committed and pushed before the next iteration starts
• Verify that non-timeout failures (exit codes other than 124) still break the loop as before
• Verify that after a timeout, the next iteration correctly determines state and resumes

🤖 Generated with Claude Code