OCR pipeline and quality

OCR flows through preprocessing, backend execution, structured conversion, page acceptance, and caching. Backend outputs are not interchangeable measurements.

Backends and execution

• All backends implement OcrBackend. Tesseract is the default and is bound through the in-repo crates/xberg-tesseract C FFI crate; there is no leptess dependency.
• OcrBackendType is Tesseract | PaddleOCR | Candle | Custom. Sceptre is selected by name through Custom, not a dedicated enum variant.
• Run blocking OCR work through tokio::task::spawn_blocking, minimize runtime/FFI lock duration, and respect backend resource limits.
• Check PageOrientationHandling before assuming a backend handles rotated input.
• Validate ISO 639 language codes and required tessdata before execution. Language detection runs after extraction and does not choose traineddata automatically.

Configuration and cache

• Public types::formats::TesseractConfig and internal ocr::types::TesseractConfig have independent defaults. Change both and keep their synchronization test passing.
• The OCR cache key combines image hash, backend, config hash, and output format. The config hash includes TESSERACT_RESULT_SCHEMA_VERSION and the ordered Tesseract variable set; it contains no build or code identity.
• Bump TESSERACT_RESULT_SCHEMA_VERSION when unchanged image/config inputs can produce different output, or disable the cache for an A/B or revert check.
• Default preprocessing is 300 DPI, deskew, and Otsu binarization. Auto-rotation, denoise, contrast enhancement, and color inversion are off unless configured. Native PSM defaults to 3; WASM defaults to 6.

Quality invariants

• Query confidence_semantics() before interpreting confidence. Never threshold Uncalibrated output using a Tesseract-derived scale.
• Tesseract font size is typography from hOCR x_fsize; Sceptre/Paddle font size is a geometric detection-box proxy. Do not compare or threshold them as the same quantity.
• hocr_font_info=1 is required for Tesseract typography; without it font sizes fall back to 12 pt. Sceptre/Paddle do not provide hOCR style fractions.
• accept_or_reject_ocr_page can discard an entire page and its structured paragraphs. Compare accepted pages before word counts. A missing dictionary-invalid ratio is unknown, not zero.
• Precision is the scarce resource on the current benchmark corpus. Require an independently grounded F1 A/B for recall-oriented rewrites; use the benchmark-workflow skill.
• Measure structural output such as headings and lists in Markdown, not Plain.

Tables and geometry

• OCR table detection clusters word bounding boxes; it does not detect ruled lines or re-OCR cells.
• Detect rows, merge words into cell tokens, then detect columns on the merged tokens. Use reconstruct_table_with_columns; detecting columns on raw words creates spurious columns for multi-word cells.
• The PDF OCR raster is normalized to MediaBox-oriented user space. OCR segments carry rotation_degrees: 0.0; code reasoning about /Rotate pages must carry page rotation explicitly.
• Preserve word-level bounding boxes, confidence, and reading order through hOCR parsing and validate a reconstructed grid before emitting Markdown.