post_rollout can crash + does not have completion rendered

[mikasenghaas]

2 issues with the current way post_rollout is set up:

• it does not have a try..except so can crash a run if not handled very carefully by the env designer. should handle this more holistically probably

 File "/home/ubuntu/research-environments/environments/code_env/code_env/code_env.py", line 229, in post_rollout
    completion = state["trajectory"][-1]["completion"]
                 ~~~~~~~~~~~~~~~~~~~^^^^
IndexError: list index out of range

• it does not have the completion field set in state yet because it gets called before this gets rendered -- this used to be different and is a bit unexpected given that e.g. rubric can always access the completion field

[gspeter-max]

I'm working on fixing this issue.

Root Cause Analysis:

1. post_rollout runs as a cleanup handler (via @vf.cleanup) BEFORE state['completion'] is set
2. When trajectory is empty (e.g., environment errors before any turns), accessing state['trajectory'][-1] raises IndexError

Implementation Approach:

• Add defensive error handling to base post_rollout in SandboxEnv
• Add comprehensive error handling wrapper in cleanup invocation
• Document clearly what data is available in post_rollout
• Add thorough tests for empty/missing trajectory scenarios

This follows a conservative approach without changing execution order.