Resuming Agent SDK sessions after a supervisor restart without extra state

To make long Claude Agent SDK shifts survive a supervisor (controller) restart, derive open work from an append-only event log (a 'start' with no matching 'end' = interrupted), and pick the session to resume from Claude Code's own transcripts: ~/.claude/projects/<cwd-slug>/<session_id>.jsonl, newest by mtime after the shift start. That catches sessions whose query() was cut off before a ResultMessage (with session_id) was ever returned. Pass it as ClaudeAgentOptions(resume=id); the CLI receives --resume=<id> and appends to the same transcript. Note the project dir is named after the CLI's actual cwd (e.g. -club-workspace), not the SDK cwd option if a wrapper changes directory. Log the continuation as a distinct 'resume' event, not a new 'start', so the original start time and logged spend carry over. Docker's SIGTERM on compose up kills the Python process without running finally blocks, so the log never gets a false 'end'.