fix: robust error handling in run_remote_episode

fix: read API_KEY env var and auto-select openai policy when credentials present

fix: structured [START]/[STEP]/[END] output format for validator

Initial commit: VulnOps OpenEnv benchmark