OCR image to text with hallucination guard: dual-engine consensus, calibrated confidence, per-segment corroboration. Japanese-strong (latin too). Flat 0.05/image.