A missing judge model kills the process after the agent run and loses run-meta.json entirely (SystemExit escapes both handlers)

[reacher-z]

src/clawbench/runner/run.py:578 loads the judge config after the agent run finishes, inside the judge stage:

judge_cfg = load_model_config(args.judge)

But load_model_config calls sys.exit(1) on a missing model (run_support/config.py:164-166, also :180, :191, :200). SystemExit derives from BaseException, so it escapes both the judge-stage except Exception (run.py:606) and the outer except Exception (run.py:674) — and write_run_meta (run.py:665/:722) never runs.

Net effect for a first-time user whose models.yaml lacks deepseek-v4-pro (the default judge — models.example.yaml even warns it "must exist here or the scoring stage fails"): they burn a full agent run of up to 30 minutes, the interception may well have succeeded, and then the process dies leaving no run-meta.json at all. The run is invisible to batch stats, to clawbench-rescore (which globs run-meta.json), and to HF upload.

Ask:

1. Validate the judge config at startup, before the agent run — next to the existing preflight_model_api call at run.py:352.
2. Make load_model_config raise a catchable exception rather than sys.exit, or catch SystemExit in the judge stage and degrade to a judge_setup_failed outcome that still writes run-meta.json.

Either way the invariant worth stating: a completed agent run always writes its metadata, whatever happens at scoring time.

[reacher-z]

@Perry2004 — filed from a read-only audit of main; every claim is cited to file:line and was re-verified by hand before filing. Flagging for your triage: happy to send a PR for any of these you want picked up, or to close it if you judge the behaviour intentional.