Make OCR observations auditable at the parser boundary

[komaksym]

Parent

• Implement auditable UFC dataset fidelity pipeline #14 — Implement auditable UFC dataset fidelity pipeline

What to build

Treat OCR output as evidence-bearing observations rather than ground truth. Preserve enough provenance and quality information for review and later benchmarking without performing fight matching in this ticket.

Acceptance criteria

• OCR records retain image/source provenance, normalized names/date, raw judge-score values, parser version, confidence/evidence where available, and explicit quality state.
• Judge-score vectors are variable-length at the parser boundary; both fighters must have the same number of judge totals, with no silent truncation or padding.
• Missing names/date, malformed vectors, impossible scores, insufficient confidence, and parse failures remain reviewable raw observations with explicit check results rather than disappearing.
• Deterministic fixtures cover correct output, transposed digits, missing/extra scores, invalid score text, and low-confidence tokens without invoking live OCR inference.
• This ticket does not attach scorecards to fights and does not assume a fixed three-judge template.

Blocked by

• Make stable UFCStats identity authoritative end to end #15 — Make stable UFCStats identity authoritative end to end.