feat(pdf): add layout and tables with optional local OCR

.github/workflows/ci.yml

@@ -76,6 +76,14 @@ jobs:
         run: cargo test --workspace
       - name: Run tests (all features)
         run: cargo test --workspace --all-features
+      - name: Rust OCR model integration
+        env:
+          MARKKIT_OCR_MODEL_DIR: ${{ github.workspace }}/target/ocr-models
+          RTEN_NUM_THREADS: '1'
+          RAYON_NUM_THREADS: '1'
+        run: |
+          python3 scripts/prepare_ocr_models.py
+          cargo test --release -p markkit --features pdf-ocr-ocrs --test pdf_ocr ocrs_recognizes_scanned_text_with_source_word_geometry -- --include-ignored
       - name: Run tests (office formats, no HTTP client)
         run: |
           if cargo tree -p markkit -e normal | grep -E 'reqwest|hyper'; then

.github/workflows/python.yml

@@ -35,3 +35,5 @@ jobs:
         run: |
           pip install target/wheels/*.whl
           python -c 'from importlib.metadata import metadata; import markkit; assert metadata("markkit")["Name"] == "markkit"; assert markkit.extract_html_sync("<h1>Smoke</h1>").strip() == "# Smoke"; print("Distribution and import OK")'
+      - name: Test structured documents and PDF layout
+        run: python -m unittest discover -s crates/markkit-python/tests -p 'test_*.py'

AGENTS.md

@@ -31,6 +31,7 @@ Fix root cause. Unsure: read more code; if stuck, ask w/ short options. Unrecogn
 | formats/format-support | Formats, bytes-in API, detection order, DOCX + spreadsheet mapping |
 | integrations/tool-contract | `MarkkitTool` LLM tool (`tool` feature), Everruns toolkit contract |
 | security/security-model | One-page threat/invariant summary |
+| security/native-dependencies | Hayro/ocrs provenance, model licensing, native library and host-runtime inventory |
 | security/fuzzing | cargo-fuzz targets, S17 oracle, nightly workflow, crash triage |
 | security/threat-model | Threats T1-T9, mitigations, test map, resource limits, invariants S1-S21 |
 | runtimes/python-package | PyO3 `markkit` module, maturin, PyPI |
@@ -81,6 +82,9 @@ just bench-criterion # Criterion micro-benches (crates/markkit/benches)
 just eval-public-setup   # Pinned Mira + isolated Python scorer environment
 just eval-public-prepare # Download and verify public corpora (PR slice)
 just eval-public        # Release WCXB + AgentDocBench runs and regression gate
+just eval-pdf           # Authored PDF text/layout/table expectations
+just ocr-models         # Verify optional OCR model assets and attribution
+just eval-ocr           # Tesseract/ocrs quality, geometry, timing and memory comparison
 just pre-pr        # Pre-PR checks
 ```
 

CHANGELOG.md

@@ -9,12 +9,37 @@ and this project adheres to [Semantic Versioning](https://semver.org/).
 
 ### Fixed
 
+- PDF stream guards honor escaped dictionary keys and bound compressed prefixes
+  before image filters, including when OCR is disabled.
+
 - Footnote bodies cannot become unsafe reference-link destinations in Markdown readers
   without footnote support. Bodies start after a blank line and remain valid footnotes
   in readers that support them (nightly renderer fuzz regression).
 
 ### Added
 
+- Optional Rust OCR backend (`pdf-ocr-ocrs`) using ocrs/RTen and explicit local
+  ONNX model paths. CLI `--ocr-backend ocrs`, Node `pdfOcrBackend`, and Python
+  `pdf_ocr_backend` select it; Tesseract remains the default. Pinned model setup
+  and repeated scan comparisons report text quality, geometry, latency and memory.
+  **Breaking:** exhaustive Rust `MarkkitBuilder` literals also need
+  `pdf_ocr_backend`, `pdf_ocr_detection_model` and `pdf_ocr_recognition_model`.
+
+- PDF layout, reading order and table extraction: glyph-positioned words,
+  column reconstruction, ruled tables with merged cells, conservative borderless
+  tables and wrapped cells; bold/italic, numbered lists, tagged heading roles and
+  `ActualText`; optional repeated header/footer removal. Layout JSON includes
+  source runs/words and semantic blocks/lines/cells linked to IR provenance.
+- Opt-in native PDF OCR (`pdf-ocr`, CLI `--ocr`) with Rust Hayro rendering and installed Tesseract,
+  bounded raster size/output/deadline and configurable languages. WASM now supports
+  PDF text/layout/tables; Node and Python expose layout and native OCR options.
+  Python gains structured `extract_document_sync`.
+- `just eval-pdf`: 20 authored PDF scenes across default/layout modes, including
+  column, table, wrapped-cell, tagging and list expectations.
+  **Breaking:** exhaustive Rust `MarkkitBuilder` literals must include `pdf_layout`,
+  `pdf_remove_repeated_headers`, `pdf_ocr`, and `pdf_ocr_language`; builder methods
+  and `..Default::default()` remain valid.
+
 - Public WCXB and AgentDocBench evaluations through Mira (`just eval-public`), with
   pinned/verified corpora, a release-build baseline regression gate, per-type/format
   quality and conversion timings. PR/push runs use a fixed slice; nightly runs cover