Clamp final task scores into open interval

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

fix: remove probe_env.py from Dockerfile COPY (gitignored dev script)

Initial commit: VulnOps OpenEnv benchmark