fix: raise ContainerNotRunning when container dies after timeout

[GopalGB]

Summary

• Adds ContainerNotRunning exception (subclass of InterruptAgentFlow) that terminates the agent loop cleanly when the Docker container dies after container_timeout expires
• Detects container-dead markers ("No such container", "is not running") in both subprocess output and exception messages
• Prevents the agent from silently burning API calls after the container is gone

Problem

When container_timeout expires, the container's sleep process exits and Docker removes it (--rm). Subsequent docker exec calls fail, but execute() catches the exception and returns it as a normal observation. The model keeps issuing commands until step_limit or cost_limit is exhausted, wasting API calls and cost.

Solution

ContainerNotRunning extends InterruptAgentFlow, so the agent's existing run() loop handles it identically to LimitsExceeded or Submitted: it adds the exit message, saves the trajectory, and stops.

Test plan

• Unit test: ContainerNotRunning raised when subprocess returns dead-container output (non-zero returncode)
• Unit test: ContainerNotRunning raised when subprocess raises exception with dead-container message
• Integration test: ContainerNotRunning raised after real container_timeout expires (requires Docker, marked @pytest.mark.slow)
• All existing tests pass

Fixes #803

[GopalGB]

Friendly ping — this PR has been clean (mergeStateStatus: CLEAN, pre-commit.ci passed) for a month. The fix raises ContainerNotRunning instead of a misleading exception when the container dies after timeout, surfacing the real failure mode to operators. Happy to address any review feedback. cc maintainers.

[klieret]

Hi, thanks for opening this, I'm currently catching up with issues & PRs, we were in a big grind to release programbench. Will take a look soon.