Keep inference metadata and OCR columns in sync

[davanstrien]

Summary

• Resolve the model identity and OCR text column atomically from one inference_info entry.
• For append-only list metadata, select the last entry whose declared output column exists.
• Preserve scalar metadata and heuristic fallbacks while avoiding cross-entry model/output pairing.
• Add direct and end-to-end regressions, including an incomplete final entry.

Why

OCR scripts append inference_info entries. Resolving model and text independently could pair a newer model identity with stale OCR output. Selecting one usable entry keeps them coupled.

Migration note

Existing results repositories judged by an older checkout from configs with multiple appended inference_info entries should be evaluated once with --full-rejudge. Old comparison rows do not record enough resolution provenance to prove that they used the newly selected text.

Validation

• 609 tests passed
• Ruff passed on src and tests
• ty passed on src
• git diff --check passed