[P2] rescore: truncated judge cache is permanently skipped instead of retried

[reacher-z]

Problem

The CLI's pending-selection predicate parses an existing judge file inside a broad try/except Exception: continue. A truncated JSON cache therefore skips the entire run before rescore_one gets a chance to recover it. The worker already handles invalid cached JSON, but is never scheduled in this case. Verdict files are written directly, so interruption can create this state.

Sources: pending selection, worker cache fallback and direct write.

Verified reproduction

• Create a temporary intercepted run with valid run-meta.json and data/interception.json.
• Write the incomplete text {"match": to judge.json.
• Run the real rescore.main() with --only-batch, --rubric strict, a temporary dummy model configuration and a mocked judge.
• Actual: exit 0, 0 judge calls, summary reports 1 judge error. A normal rerun repeats this behavior unless the user knows to use --force or delete the cache.

Acceptance criteria

• Treat invalid JSON, non-object verdicts and invalid verdict types as unusable caches and schedule reevaluation with a diagnostic.
• Share cache validation between the planner and worker so their recovery behavior cannot disagree.
• Write verdicts atomically so interruption cannot replace a valid cache with partial JSON.
• Test CLI-level scheduling, including both rubrics where one cache is corrupt and the other is valid.
• Define/report partial-processing errors rather than quietly treating processing as wholly successful.

Suggested priority P2: recoverability and unattended batch reliability. #299 handled match: null; this is malformed JSON rejected before that recovery path. #303 concerned batch metadata, not judge cache writes. #313 does not change this predicate.

@Perry2004 please review recovery behavior and the appropriate partial-failure exit status.