[abhishekgahlot2]

• Local results record their isolation (RunResult.isolation); unsandboxed local runs warn once.
• Closed sessions refuse further calls (SessionClosedError, 409 in the API); a failed session start is cleaned up through the closing pipeline and leak accounting.
• Harbor caches are keyed by a digest of the full URL and checked against the clone's origin; test output is kept whole.
• Console: no-verdict episodes show as unscored; episodes and runs page on the server with totals and no row cap.
• No generation caps: Z.ai sends no max_tokens or read timeout by default and reports finish_reason; truncated difficulty samples are inconclusive; make_lm_eval_task takes the model's max_gen_toks; Claude calls in the adversarial tools run at the model's maximum output.
• No trimmed output in errors, reports or QA details; verifier ids are digests.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}