fix(judge): give a judge outage its own outcome, not JUDGE MISMATCH

[AmirF194]

What does this PR do?

Fixes #299, all three asks.

judge.py already returns match=None when the judge never renders a verdict (provider outage, retries exhausted, or an unparseable reply), distinct from match=False (a genuine mismatch). But downstream, run.py's final-status block treats both the same way, sharing exit 1, and batch.py's exit-code map turns exit 1 into job.status = "failed". So a judge-provider outage was indistinguishable from an agent failure in batch-summary.json.

Ask 1: give match=None a distinct outcome

run.py now exits 3 instead of 1 when judge_result.get("match") is None, keeping it separate from a genuine JUDGE MISMATCH (which still exits 1).

Ask 2: count it separately in batch-summary.json

batch.py maps exit 3 to job.status = "judge_inconclusive", its own bucket alongside passed/failed/error/skipped, carried into batch-summary.json's totals and the printed batch summary.

Ask 3: retarget those runs for re-judging

clawbench-rescore now retries a cached verdict whose match is null even without --force, since that cached file never held a scored result. --resume's own skip logic in batch.py is untouched here; it currently keys off batch-log existence rather than the recorded outcome, and reworking that is a larger, separate change.

Corpus

• v2
• v1
• both
• not applicable

_{Host-side runner change; no task data involved.}

Test plan

• tests/test_run_judge_stage.py::test_judge_failure_after_the_run_still_writes_run_meta, updated: a judge_setup_failed outcome now exits 3, not 1. New test_judge_mismatch_keeps_exit_code_1_distinct_from_inconclusive locks in that a real match=False still exits 1.
• tests/test_batch_judge_inconclusive.py, new: run_job() against a mocked subprocess maps exit 0/1/3/other to passed/failed/judge_inconclusive/error; write_summary_json() counts judge_inconclusive separately from failed in batch-summary.json.
• tests/test_rescore_retarget.py, new: rescore_one() retries a cached match: null verdict without --force, and still trusts a cached scored verdict (match: true/false) without retrying it.
• Full suite in a clean python:3.11-slim container via uv run --frozen pytest: 226 passed (218 on main, +8 from this PR). All new/changed tests fail on main and pass on this branch, confirmed both ways in the same container.
• ruff check ., ruff format --check ., pyright src/clawbench tests, uv build + twine check dist/*, and the three --help console-script checks from static-check.yml: all clean. validate-docs.yml's link checker (triggered by the CHANGELOG.md edit): clean.

Not verified: no live judge-provider outage was reproduced end-to-end; the fix is proven through judge.py's existing judge_call_failed/unsupported_api_type fallback plus the tests above, not a real API timeout.

Related issues

Fixes #299.