Allow ten hours for GenPlan containers

[joshnroy]

Summary

• increase the GenPlan Docker wall-clock timeout from one hour to ten hours
• allow $20 Claude runs to finish training and return to host-side evaluation

Why

The one-hour subprocess.run timeout can expire before a GenPlan run exhausts its model budget. The Docker client then exits while the container continues in the background, leaving training artifacts such as cost.json but preventing held-out evaluation and results.json from being written.

Verification

• python -m py_compile src/robocode/utils/docker_sandbox.py
• git diff --check
• 19 focused GenPlan tests passed; 6 rollout tests could not start their multiprocessing manager under the restricted Codex sandbox

Stack

Stacked on #178.

[joshnroy]

Folded into #177 so Claude session reuse and the container wall-clock limit can be reviewed together as one runtime reliability change.