Audit should fail cleanly on missing-text and empty/unjudgeable configs

[davanstrien]

Follow-up from the #49 review. ocr-bench audit can currently report OK for repositories that cannot produce a meaningful judge run:

• discovered configs with no recognized OCR text column are filtered out rather than reported;
• zero-row configs have zero rates and may pass;
• fewer than two usable OCR models cannot produce a pairwise leaderboard.

The audit should list every discovered config and return a non-zero operational/integrity status for these cases. Add tests for a mixed usable/missing-text repo, all-empty configs, and a single usable model.