More reliable long-running agents

We have made agent runs more reliable.

Long-running agent work now survives more backend restarts, stuck statuses are handled better, and agents can be stopped when they are taking too long or heading in the wrong direction.

When something fails, UpdateMate records the error in the log and gives you a path back to the agent chat so you can fix and rerun it.

This makes scheduled and background agents safer to use for real operational work.