Judge outages are scored as agent failures (match=None → exit 1 → 'failed')

[reacher-z]

When the judge provider is down, we record the model as having failed the task.

• src/clawbench/runner/judge.py:247-253 returns match=None on judge_call_failed (HTTP 402/timeout after retries).
• src/clawbench/runner/run.py:657-664 marks pass only when match is True; run.py:743-754 then exits 1.
• src/clawbench/runner/batch.py:324-327 counts exit 1 as failed.

So a judge outage is indistinguishable in batch-summary.json from genuine agent failure. This is not hypothetical — a judge account hitting HTTP 402 mid-sweep has already produced whole batches recorded as failures that needed a manual rescore pass to recover.

Ask:

1. Give match=None a distinct outcome (e.g. exit code 3 / status: "judge_inconclusive"), separate from agent failure.
2. Count it separately in batch-summary.json totals so a reader can see "N runs unjudged" instead of silently deflated reward.
3. Have --resume / clawbench-rescore target exactly those runs for re-judging.

This pairs with #243 (always report Stage-1 and Stage-2 separately): both are about not letting stage-2 problems silently corrupt the headline number.

[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.