refactor(vision): expose OCR API

Justified print breaks words at syllables; after paragraph lines are joined
with spaces those breaks survive as "de- fendant" — thousands of them in a
long document — and, when an emphasis span was split with the word, as
"Bap-** **tist".

Whether the hyphen belongs in the word cannot be decided locally ("de-
fendant" is one word, "Third- Party" is a hyphenated compound), so the
document is used as its own dictionary. For each break:

  1. fragments appear joined elsewhere ("defendant")      -> join plain
  2. appear hyphenated elsewhere ("six-month"), or the
     continuation is capitalized ("Hinds- Radix")         -> keep the hyphen
  3. both fragments are words the document uses and the
     continuation has 4+ letters ("commercial- type")     -> keep the hyphen
  4. no evidence                                          -> leave untouched

The policy contains zero hard-coded words: vocabulary evidence,
capitalization, and two length invariants. The 4-letter floor keeps
suspended hyphens intact in any language ("mid- and long-term", "klein- und
mittelgroß", "kuva- tai video") because conjunctions are near-universally
1-3 letters. Fragments over 40 combined characters are fused reading-order
noise and are never joined. The vocabulary is collected after scrubbing the
break pairs themselves and excludes fenced code blocks; table rows and code
blocks are never rewritten. Split emphasis spans rejoin inside their
markers. Runs under the existing fix_hyphenation option (default on).

On a 1,370-page justified legal reporter this rejoins ~8,000 broken words
(98.8% of breaks; evidence-less ones stay visibly intact); word recall
against a reference extraction rises from 97.8% to 99.1%. No "six-month" ->
"sixmonth" class errors, and no fused-column corruption by construction:
no rule joins without evidence.

Regression-checked against a ~200-document corpus with semantic scoring
against an OCR baseline: zero regressions. Three in-repo fixture snapshots
regenerated with each diff inspected. 17 unit tests cover every rule, the
vocabulary scrubbing and code-block exclusion, the length gates, chained
breaks, mismatched emphasis markers, accented and Cyrillic words, German
and Finnish suspended hyphens, and the table/code skips.

.github/workflows/ci.yml

@@ -24,6 +24,9 @@ jobs:
       - name: Cache cargo
         uses: Swatinem/rust-cache@c19371144df3bb44fab255c43d04cbc2ab54d1c4 # v2.9.1
 
+      - name: Check package version sync
+        run: python3 scripts/version.py --check
+
       - name: Run tests
         run: cargo test --verbose
 

.github/workflows/publish-crate.yml

@@ -31,6 +31,9 @@ jobs:
         with:
           fetch-depth: 2
 
+      - name: Check package version sync
+        run: python3 scripts/version.py --check
+
       - name: Check if version changed
         id: check
         run: |

.github/workflows/publish-pypi.yml

@@ -28,6 +28,9 @@ jobs:
         with:
           fetch-depth: 2
 
+      - name: Check package version sync
+        run: python3 scripts/version.py --check
+
       - name: Check if version changed
         id: check
         run: |

.github/workflows/publish-wasm.yml

@@ -28,6 +28,9 @@ jobs:
         with:
           fetch-depth: 2
 
+      - name: Check package version sync
+        run: python3 scripts/version.py --check
+
       - name: Check package version
         id: check
         run: |

.github/workflows/publish.yml

@@ -28,6 +28,9 @@ jobs:
         with:
           fetch-depth: 2
 
+      - name: Check package version sync
+        run: python3 scripts/version.py --check
+
       - name: Check if version changed
         id: check
         run: |

.gitignore

@@ -31,9 +31,8 @@ Thumbs.db
 napi/index.js
 napi/index.d.ts
 
-# Local samples and scripts
+# Local samples
 samples/
-scripts/
 
 # Test output
 test_output/

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "pdf-inspector"
-version = "0.1.7"
+version = "1.14.2"
 edition = "2021"
 autobins = false
 authors = ["Firecrawl Team"]
@@ -49,6 +49,25 @@ ttf-parser = "0.25"
 lopdf = { version = "0.42.0", features = ["rayon"] }
 rayon = "1.10"
 env_logger = "0.11"
+# Optional native page rendering for OCR pipelines. PDFium is loaded at
+# runtime, so enabling this feature does not link or download a native library.
+firecrawl-pdfium = { version = "0.1.0", optional = true }
+# Small support crates used only by the opt-in model cache. Model files remain
+# external and are never embedded in pdf-inspector artifacts.
+dirs = { version = "6.0", optional = true }
+fs2 = { version = "0.4", optional = true }
+sha2 = { version = "0.11", optional = true }
+# Optional CPU OCR backend. Models and ONNX Runtime stay external: the latter
+# is loaded dynamically from ORT_DYLIB_PATH or the platform library search path.
+image = { version = "0.25.6", default-features = false, optional = true }
+oar-ocr = { version = "0.9.1", default-features = false, features = ["simd"], optional = true }
+ort = { version = "=2.0.0-rc.13", default-features = false, features = ["load-dynamic"], optional = true }
+# HTTPS-only streaming downloader for pinned model artifacts. Kept separate
+# from model-cache so offline and package-managed deployments avoid HTTP/TLS.
+ureq = { version = "3.4", default-features = false, features = ["rustls", "platform-verifier"], optional = true }
+
+[target.'cfg(all(windows, not(target_arch = "wasm32")))'.dependencies]
+windows-sys = { version = "0.61", features = ["Win32_Storage_FileSystem"], optional = true }
 
 # Browser builds use JavaScript randomness for encrypted PDFs and embed the
 # bundled CMaps because there is no filesystem at runtime.
@@ -62,6 +81,14 @@ tempfile = "3.3"
 [features]
 default = []
 python = ["pyo3"]
+vision = []
+model-cache = ["vision", "dep:dirs", "dep:fs2", "dep:sha2", "dep:windows-sys"]
+model-download = ["model-cache", "dep:ureq"]
+ocr-oar = ["model-cache", "dep:image", "dep:oar-ocr", "dep:ort"]
+render-pdfium = ["vision", "dep:firecrawl-pdfium"]
+# Complete native OCR path. This remains opt-in so default library,
+# renderer-only, and browser consumers do not inherit inference or HTTP/TLS.
+ocr = ["render-pdfium", "ocr-oar", "model-download"]
 
 [[bin]]
 name = "pdf2md"

README.md

@@ -110,7 +110,7 @@ Or add it manually:
 
 ```toml
 [dependencies]
-pdf-inspector = "0.1"
+pdf-inspector = "1"
 ```
 
 ```rust

SECURITY.md

@@ -5,14 +5,15 @@
 If you believe you've found a security vulnerability in pdf-inspector, please
 report it privately so we can fix it before public disclosure.
 
-**Preferred:** Email **help@firecrawl.dev** with:
+**Preferred:** Submit through Firecrawl's Bugcrowd vulnerability disclosure
+program at <https://bugcrowd.com/engagements/firecrawl-vdp-ess>. Please include:
 
 - A description of the issue and its impact
 - Steps to reproduce (a minimal PDF or input that triggers the bug is ideal)
 - The version or commit hash of pdf-inspector you tested against
 
-**Alternative:** Use GitHub's private vulnerability reporting under the
-[Security tab](https://github.com/firecrawl/pdf-inspector/security/advisories/new).
+**Alternative:** If you'd rather not use Bugcrowd, email
+**help@firecrawl.dev** with the same details.
 
 We'll acknowledge your report in a timely manner and keep you updated on
 remediation progress. Please do not open a public GitHub issue for security

docs/publishing.md

@@ -1,38 +1,57 @@
 # Publishing
 
-The Rust crate is published to [crates.io](https://crates.io/crates/pdf-inspector) with trusted publishing from GitHub Actions. The first release was published manually; future releases publish from `.github/workflows/publish-crate.yml` when a `Cargo.toml` version change lands on `main`.
+Every pdf-inspector distribution uses one shared semantic version:
 
-## crates.io Trusted Publisher
+- Rust crate: `pdf-inspector`
+- Python package: `pdf-inspector`
+- Node package: `@firecrawl/pdf-inspector` and its platform packages
+- Browser package: `@firecrawl/pdf-inspector-wasm`
+- Internal NAPI and WASM Rust crates
 
-Configure the trusted publisher for the `pdf-inspector` crate with:
+`Cargo.toml` is the canonical version source. Update every manifest and lockfile
+with:
 
-- Repository: `firecrawl/pdf-inspector`
-- Workflow: `publish-crate.yml`
-- Environment: `crates-io`
+```bash
+python3 scripts/version.py <version>
+```
 
-The workflow uses `rust-lang/crates-io-auth-action@v1` to exchange GitHub's OIDC token for a short-lived crates.io token, then passes it to `cargo publish`.
+Verify that nothing has diverged with:
 
-## Release Steps
+```bash
+python3 scripts/version.py --check
+```
 
-1. Update `version` in `Cargo.toml`.
-2. Merge the version bump to `main`.
-3. The publish workflow compares the new `Cargo.toml` version with `HEAD~1`, runs `cargo publish --dry-run`, then publishes if that version is not already on crates.io.
+CI and every publishing workflow run this check before building or publishing.
 
-If `Cargo.toml` changes without a package version bump, the workflow exits without publishing.
+## Release steps
 
-## Browser WebAssembly package
+1. Choose the next shared semantic version and run `scripts/version.py`.
+2. Review the manifest and lockfile changes in the version-bump pull request.
+3. Merge the pull request to `main`.
+4. The crates.io, PyPI, Node, and WASM workflows independently build and
+   publish that version from the same commit.
+5. After all registries succeed, create one `v<version>` GitHub release that
+   links to each package and describes changes since the previous shared tag.
 
-The browser package is published as `@firecrawl/pdf-inspector-wasm`. Its version lives in `wasm/Cargo.toml`, and `.github/workflows/publish-wasm.yml` builds the `web` target with `wasm-pack` before publishing the generated package.
+The independent workflows are intentionally idempotent. A manual dispatch from
+`main` can repair a partial release, and already-published artifacts are skipped.
 
-The npm package must exist before a trusted publisher can be configured. For the first release only:
+## Trusted publishers
 
-1. Build with `wasm-pack build wasm --target web --scope firecrawl --out-dir pkg --release`.
-2. Inspect with `npm pack --dry-run ./wasm/pkg`.
-3. Publish with `npm publish ./wasm/pkg --access public` from an authorized maintainer session.
-4. In the package settings on npm, configure the GitHub Actions trusted publisher:
-   - Organization: `firecrawl`
-   - Repository: `pdf-inspector`
-   - Workflow: `publish-wasm.yml`
-   - Allowed action: `npm publish`
+The repositories use GitHub Actions OIDC instead of long-lived registry tokens.
+Configure each registry's trusted publisher for `firecrawl/pdf-inspector` and
+its corresponding workflow:
 
-After that one-time bootstrap, bumping the version in `wasm/Cargo.toml` and merging it to `main` publishes through OIDC. Until the package exists, the workflow exits cleanly without attempting an unauthenticated first publish. See npm's [trusted publishing documentation](https://docs.npmjs.com/trusted-publishers/) for the registry-side setup.
+- crates.io: `publish-crate.yml`, environment `crates-io`
+- PyPI: `publish-pypi.yml`, environment `pypi`
+- npm Node package: `publish.yml`
+- npm WASM package: `publish-wasm.yml`
+
+The WASM package must exist before npm trusted publishing can be configured. If
+it ever needs to be bootstrapped again, build and inspect it before publishing:
+
+```bash
+wasm-pack build wasm --target web --scope firecrawl --out-dir pkg --release
+npm pack --dry-run ./wasm/pkg
+npm publish ./wasm/pkg --access public
+```

docs/python.md

@@ -80,6 +80,17 @@ for page in result.pages:
 
 # Restrict to specific 0-indexed pages (preserves caller order)
 result = pdf_inspector.extract_pages_markdown("document.pdf", pages=[0, 2])
+
+# Structure-tree elements from tagged PDFs (empty list when untagged).
+# Pages are 1-indexed to match TextItem.page, so (page, mcid) joins directly
+# against extract_text_with_positions — e.g. to recover real heading levels:
+elements = pdf_inspector.extract_structure_elements("tagged.pdf")
+roles = {(e.page, e.mcid): e.role for e in elements}
+headings = [
+    item.text
+    for item in pdf_inspector.extract_text_with_positions("tagged.pdf")
+    if item.mcid is not None and roles.get((item.page, item.mcid), "").startswith("H")
+]
 ```
 
 ## API reference
@@ -100,6 +111,8 @@ result = pdf_inspector.extract_pages_markdown("document.pdf", pages=[0, 2])
 | `extract_text_in_regions_bytes(data, page_regions)` | Region extraction from bytes |
 | `extract_pages_markdown(path, pages=None)` | Per-page Markdown + layout metadata (all pages by default) |
 | `extract_pages_markdown_bytes(data, pages=None)` | Per-page Markdown from bytes |
+| `extract_structure_elements(path, pages=None)` | Structure-tree elements from tagged PDFs (page, mcid, role) |
+| `extract_structure_elements_bytes(data, pages=None)` | Structure-tree elements from bytes |
 
 ## Types
 
@@ -144,6 +157,12 @@ class TextItem:                      # extract_text_with_positions
     is_underline: bool
     is_strikeout: bool
     item_type: str
+    mcid: int | None                 # marked-content ID for tagged PDFs (None otherwise)
+
+class StructureElement:              # extract_structure_elements
+    page: int                        # 1-indexed (matches TextItem.page)
+    mcid: int
+    role: str                        # "H1".."H6", "P", "Table", ... (resolved via /RoleMap)
 
 class RegionText:                    # extract_text_in_regions
     text: str

docs/rust-api.md

@@ -1,6 +1,6 @@
 # pdf-inspector
 
-Fast PDF classification and text extraction. Detects whether a PDF is text-based or scanned, extracts text with position awareness, and converts to clean Markdown — all without OCR. Pure Rust, no ML models, no external services; the only PDF dependency is [lopdf](https://crates.io/crates/lopdf). Also available for [Python](https://pypi.org/project/pdf-inspector/) and [Node.js](https://www.npmjs.com/package/@firecrawl/pdf-inspector).
+Fast PDF classification and text extraction. Detects whether a PDF is text-based or scanned, extracts text with position awareness, and converts to clean Markdown — all without OCR. The default build is pure Rust, has no ML models or external services, and uses [lopdf](https://crates.io/crates/lopdf) for PDF parsing. Also available for [Python](https://pypi.org/project/pdf-inspector/) and [Node.js](https://www.npmjs.com/package/@firecrawl/pdf-inspector/).
 
 Built by [Firecrawl](https://firecrawl.dev) to handle text-based PDFs locally in under 200ms, skipping expensive OCR services for the ~54% of PDFs that don't need them.
 
@@ -117,6 +117,285 @@ let bytes = std::fs::read("document.pdf")?;
 let result = process_pdf_mem(&bytes)?;
 ```
 
+### Vision extension contracts
+
+The native-only `vision` feature exposes the stable seam used by OCR
+integrations without selecting or embedding an inference runtime. The
+separate `model-cache` feature adds pinned artifact management:
+
+- `PageRenderer`, `OcrEngine`, and `LayoutEngine` traits;
+- renderer-neutral owned page buffers and affine pixel↔PDF transforms;
+- `OcrOptions` and opt-in `Off`/`Auto`/`Force` routing modes;
+- positioned OCR/layout results and per-page provenance types; and
+- a versioned PP-OCRv6 Small manifest with checksum-verified, locked, atomic
+  model-cache installation and explicit offline-directory overrides.
+
+```toml
+[dependencies]
+pdf-inspector = { version = "1", features = ["vision", "model-cache"] }
+```
+
+The OCR contracts preserve existing behavior by default: OCR is `Off`, learned
+layout is disabled, and model resolution is never reached. `ModelStore` itself
+does not access the network. The optional `model-download` feature provides an
+HTTPS downloader that streams pinned artifacts into the checksum-verified
+cache only after routing has selected OCR work. Offline consumers set an
+explicit model directory and `ModelDownloadPolicy::Offline`. Renderer-only
+consumers do not enable `model-cache` or `model-download` and therefore do not
+compile their filesystem, hashing, or HTTP dependencies.
+
+```rust
+use pdf_inspector::vision::{
+    ModelDownloadPolicy, ModelStore, OcrMode, OcrOptions, PP_OCR_V6_SMALL,
+};
+
+let ocr = OcrOptions::new()
+    .mode(OcrMode::Auto)
+    .model_directory("/opt/firecrawl/models/pp-ocrv6-small")
+    .model_downloads(ModelDownloadPolicy::Offline);
+// Verifies exact sizes and SHA-256 digests before an engine opens the files.
+let models = ModelStore::from_options(&ocr)?.resolve(&PP_OCR_V6_SMALL)?;
+println!("using {} at {}", models.manifest_id(), models.revision());
+```
+
+### Optional native page rendering
+
+The `render-pdfium` feature adds a native-only page renderer backed by
+[`firecrawl-pdfium`](https://crates.io/crates/firecrawl-pdfium). It is the
+rendering boundary for OCR pipelines; enabling it does not include an OCR
+model or change the existing extraction functions. It implies `vision`,
+and `PdfiumRenderer` implements the renderer-neutral `PageRenderer` trait.
+
+```toml
+[dependencies]
+pdf-inspector = { version = "1", features = ["render-pdfium"] }
+```
+
+PDFium is loaded at runtime. Set `PDFIUM_LIB_PATH`, place its shared library
+next to the executable, or use another discovery route supported by
+`firecrawl-pdfium`.
+
+```rust
+use pdf_inspector::vision::{PdfiumRenderer, RenderOptions};
+
+let renderer = PdfiumRenderer::load()?;
+let bytes = std::fs::read("document.pdf")?;
+let pages = renderer.render_pages(
+    &bytes,
+    &[1, 3], // 1-indexed, matching pages_needing_ocr
+    None,    // optional PDF password
+    &RenderOptions::new().dpi(150.0),
+)?;
+
+for page in pages {
+    // Owned RGB pixels can leave the PDFium critical section and be sent to
+    // an OCR worker. OCR pixel boxes can be mapped back to PDF coordinates.
+    let rect = page.pixel_rect_to_pdf_rect(20.0, 30.0, 100.0, 24.0);
+    println!("page {}: {}x{}, rect={rect:?}", page.page(), page.width(), page.height());
+}
+```
+
+Browser WASM remains on the default text-only path and does not expose native
+PDFium rendering.
+
+### Optional OCR engine
+
+The native-only `ocr-oar` feature adds a CPU PP-OCRv6 Small implementation of
+`OcrEngine` backed by OAR and ONNX Runtime. It implies `model-cache`, but does
+not enable model auto-download, ONNX Runtime download, or PDF rendering. Model
+files remain external, must match the pinned manifest, and are opened only
+after `ModelStore` verifies their exact size and SHA-256 digest. Install an
+ONNX Runtime shared library separately and set `ORT_DYLIB_PATH` when it is not
+available through the platform library search path. The feature currently
+requires Rust 1.95 or newer, matching OAR 0.9.1's MSRV.
+
+```toml
+[dependencies]
+pdf-inspector = { version = "1", features = ["ocr-oar", "render-pdfium"] }
+```
+
+Direct engine invocation is intentionally separate from extraction routing and
+native/OCR fusion:
+
+```rust
+use pdf_inspector::vision::{
+    ModelDownloadPolicy, ModelStore, OarOcrEngine, OcrEngine, OcrMode,
+    OcrOptions, PdfiumRenderer, RenderOptions, PP_OCR_V6_SMALL,
+};
+
+let options = OcrOptions::new()
+    .mode(OcrMode::Force)
+    .minimum_confidence(0.45)
+    .model_directory("/opt/firecrawl/models/pp-ocrv6-small")
+    .model_downloads(ModelDownloadPolicy::Offline);
+let models = ModelStore::from_options(&options)?.resolve(&PP_OCR_V6_SMALL)?;
+let engine = OarOcrEngine::from_models(&models)?;
+
+let renderer = PdfiumRenderer::load()?;
+let bytes = std::fs::read("scan.pdf")?;
+let pages = renderer.render_pages(&bytes, &[1], None, &RenderOptions::new())?;
+let ocr_pages = engine.recognize(&pages, &options)?;
+
+for span in &ocr_pages[0].spans {
+    println!("{:.3}: {}", span.confidence, span.text);
+}
+```
+
+The engine accepts renderer-neutral RGB, RGBA, and grayscale pages, preserves
+OAR's positioned quadrilaterals in bitmap coordinates, filters spans using
+`minimum_confidence`, and records the pinned model revision in every `OcrPage`.
+`OcrMode::Off` is rejected at the engine boundary so default options cannot run
+inference accidentally.
+
+### Selective routing and lazy model acquisition
+
+`route_ocr_pages` applies the existing detector/text-quality recommendations to
+the configured mode. `Auto` processes only recommended pages, `Force` processes
+all pages (or an explicit page selection), and `Off` always returns an empty
+route. `run_ocr_pages` renders only that route, checks that both dependencies
+preserve its order, and retains each bitmap's PDF transform for fusion.
+
+```toml
+[dependencies]
+pdf-inspector = { version = "1", features = [
+  "render-pdfium",
+  "ocr-oar",
+  "model-download",
+] }
+```
+
+```rust
+use pdf_inspector::vision::{
+    route_ocr_pages, run_ocr_pages, HttpModelDownloader, ModelStore,
+    OarOcrEngine, OcrMode, OcrOptions, PdfiumRenderer, RenderOptions,
+    PP_OCR_V6_SMALL,
+};
+
+let bytes = std::fs::read("scan.pdf")?;
+let extraction = pdf_inspector::extract_pages_markdown_mem(&bytes, None)?;
+let options = OcrOptions::new().mode(OcrMode::Auto);
+let routed = route_ocr_pages(
+    options.mode,
+    extraction.pages.len() as u32,
+    &extraction.pages_needing_ocr,
+    None,
+)?;
+
+if !routed.is_empty() {
+    // No HTTP request or model initialization occurs before this point.
+    let store = ModelStore::from_options(&options)?;
+    let models = store.resolve_or_download(
+        &PP_OCR_V6_SMALL,
+        options.model_downloads,
+        &HttpModelDownloader::default(),
+    )?;
+    let run = run_ocr_pages(
+        &PdfiumRenderer::load()?,
+        &OarOcrEngine::from_models(&models)?,
+        &bytes,
+        &routed,
+        None,
+        &RenderOptions::new(),
+        &options,
+    )?;
+    println!("OCR processed {} pages", run.pages.len());
+}
+```
+
+The downloader accepts HTTPS only, checks a declared content length, caps the
+response stream to the pinned size plus one byte, and delegates final size and
+SHA-256 verification to `ModelStore`. The store serializes installation across
+processes and publishes completed artifacts atomically. Warm caches make no
+network calls; offline mode and explicit model directories never download.
+
+### OCR Markdown assembly and native fusion
+
+`fuse_ocr_pages` maps OCR polygons back into PDF coordinates and sends the
+result through pdf-inspector's existing deterministic reading-order, table,
+and Markdown pipeline. Pages whose native extraction was rejected use OCR
+output. When `Force` runs on a clean native page, normalized duplicate OCR
+blocks are removed and only additional image-backed text is retained.
+
+```rust
+use pdf_inspector::vision::{fuse_ocr_pages, OcrFusionOptions};
+
+let fused = fuse_ocr_pages(
+    &extraction.pages,
+    &run,
+    extraction.pages.len() as u32,
+    &OcrFusionOptions::new().render_dpi(150.0),
+)?;
+
+for page in &fused.pages {
+    println!("{}", page.markdown);
+    if page.provenance.hosted_recommended {
+        eprintln!("page {} needs the hosted document pipeline", page.page + 1);
+    }
+}
+```
+
+Each page carries `Native`, `Ocr`, or `Fused` provenance, the exact OCR model
+revision, accepted-page confidence, local stage timings, and non-fatal
+warnings. A page that required OCR recommends the hosted pipeline when local
+OCR is missing, empty, or below the configurable page-confidence threshold.
+This keeps the lightweight path explicit about cases it cannot finish well.
+
+### Complete OCR API
+
+The `ocr` convenience feature enables the renderer, OCR engine, verified
+model acquisition, routing, and fusion layers together. It is the intended
+downstream application integration boundary; lower-level features remain
+available for consumers that bring their own renderer, model package manager,
+or engine.
+
+```toml
+[dependencies]
+pdf-inspector = { version = "1", features = ["ocr"] }
+```
+
+```rust
+use pdf_inspector::vision::{
+    process_pdf_with_ocr, OcrMode, OcrPdfOptions,
+};
+
+let result = process_pdf_with_ocr(
+    "document.pdf",
+    OcrPdfOptions::new()
+        .mode(OcrMode::Auto)
+        .pages([1, 2, 3]),
+)?;
+
+println!("{}", result.markdown);
+println!("OCR pages: {:?}", result.pages_routed_to_ocr);
+println!(
+    "Hosted fallback pages: {:?}",
+    result.pages_recommending_hosted,
+);
+```
+
+Native extraction always runs first. In `Auto`, a clean PDF returns before
+PDFium loading, model-cache access, HTTP, or OAR initialization. Model files
+remain external and the default crate feature set remains unchanged. `Off`
+provides the same native-only behavior through the OCR result/provenance
+shape; `Force` renders every selected page. Learned layout intentionally
+returns an explicit unsupported error in this lightweight pipeline.
+
+Build the CLI with the same opt-in feature:
+
+```bash
+cargo build --release --features ocr --bin pdf2md
+pdf2md document.pdf --ocr auto --raw
+pdf2md document.pdf --ocr auto --json
+pdf2md document.pdf --ocr auto --ocr-offline --ocr-model-dir /opt/models/pp-ocrv6-small
+```
+
+CLI controls include `--ocr-dpi`, `--ocr-min-confidence`,
+`--ocr-hosted-threshold`, `--select-pages`, and the existing encrypted-PDF
+`--password` option. JSON output includes per-page Markdown, source/model
+provenance, confidence, timings, warnings, routed pages, and hosted-fallback
+recommendations. Page numbers in `OcrPdfResult` and its per-page provenance
+are 1-indexed, matching the PDF page numbers accepted by `OcrPdfOptions::pages`.
+
 Extract per-page Markdown (one string per page, plus document-wide layout
 metadata):
 
@@ -138,6 +417,34 @@ for page in &result.pages {
 println!("Complex layout? {}", result.is_complex);
 ```
 
+Extract structure-tree elements from tagged PDFs, and join them against
+`extract_text_with_positions` to attach semantic roles (heading levels,
+paragraphs, table cells) to extracted text:
+
+```rust
+use pdf_inspector::{extract_structure_elements, extract_text_with_positions};
+use std::collections::HashMap;
+
+// One entry per marked-content reference, sorted by (page, mcid); empty for
+// untagged PDFs. Pages are 1-indexed to match `TextItem::page`, so the
+// (page, mcid) pair is a direct join key.
+let elements = extract_structure_elements("tagged.pdf", None)?;
+let roles: HashMap<(u32, i64), &str> = elements
+    .iter()
+    .map(|e| ((e.page, e.mcid), e.role.as_str()))
+    .collect();
+
+for item in extract_text_with_positions("tagged.pdf")? {
+    if let Some(mcid) = item.mcid {
+        if let Some(role) = roles.get(&(item.page, mcid)) {
+            if role.starts_with('H') {
+                println!("{}: {}", role, item.text);
+            }
+        }
+    }
+}
+```
+
 ## Processing modes
 
 | Mode | What it does | Returns |
@@ -163,6 +470,8 @@ println!("Complex layout? {}", result.is_complex);
 | `to_markdown_from_items_with_rects(items, options, rects)` | Markdown with rectangle-based table detection |
 | `extract_pages_markdown(path, pages)` | Per-page Markdown + layout metadata (file) |
 | `extract_pages_markdown_mem(bytes, pages)` | Per-page Markdown from bytes |
+| `extract_structure_elements(path, pages)` | Structure-tree elements from tagged PDFs (page, mcid, role) |
+| `extract_structure_elements_mem(bytes, pages)` | Structure-tree elements from bytes |
 
 Low-level detection functions are also available via the `detector` module (`detect_pdf_type`, `detect_pdf_type_with_config`, etc.) for callers who need `PdfTypeResult` instead of `PdfProcessResult`.
 
@@ -178,7 +487,8 @@ Low-level detection functions are also available via the `detector` module (`det
 | `DetectionConfig` | Configuration for detection: scan strategy, thresholds |
 | `ScanStrategy` | `EarlyExit`, `Full`, `Sample(n)`, `Pages(vec)` |
 | `LayoutComplexity` | Layout analysis: is_complex, pages_with_tables, pages_with_columns |
-| `TextItem` | Text with position, font info, and page number |
+| `TextItem` | Text with position, font info, page number, and optional structure-tree `mcid` |
+| `StructureElement` | Tagged-PDF structure reference: page (1-indexed), mcid, role (`"H1"`..`"H6"`, `"P"`, …) |
 | `MarkdownOptions` | Configuration for Markdown formatting (page numbers, etc.) |
 | `PageMarkdown` | Per-page result: page (0-indexed), markdown, needs_ocr |
 | `PagesExtractionResult` | Per-page output + 1-indexed pages_with_tables / pages_with_columns / pages_needing_ocr, is_complex |

napi/Cargo.lock

@@ -851,7 +851,7 @@ checksum = "384b8ab6d37215f3c5301a95a4accb5d64aa607f1fcb26a11b5303878451b4fe"
 
 [[package]]
 name = "pdf-inspector"
-version = "0.1.7"
+version = "1.14.2"
 dependencies = [
  "env_logger",
  "include_dir",
@@ -867,7 +867,7 @@ dependencies = [
 
 [[package]]
 name = "pdf-inspector-napi"
-version = "0.2.2"
+version = "1.14.2"
 dependencies = [
  "napi",
  "napi-build",

napi/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "pdf-inspector-napi"
-version = "0.2.2"
+version = "1.14.2"
 edition = "2021"
 
 [lib]

napi/README.md

@@ -83,6 +83,22 @@ for (const region of result[0].regions) {
 }
 ```
 
+### Async variants
+
+`processPdf`, `classifyPdf`, and `extractPagesMarkdown` are synchronous and parse on the calling thread — in Node, that's the event loop. For a one-off call in a script that's fine, but in a server a large document can hold the loop for tens to hundreds of milliseconds.
+
+`processPdfAsync`, `classifyPdfAsync`, and `extractPagesMarkdownAsync` take the same arguments and produce the same results, but run the parse on the libuv thread pool and return a promise, keeping the event loop free. The input buffer is copied before the call returns, so it's safe to reuse or mutate immediately:
+
+```typescript
+import { classifyPdfAsync, extractPagesMarkdownAsync } from '@firecrawl/pdf-inspector'
+
+const classification = await classifyPdfAsync(pdf)
+if (classification.pdfType === 'TextBased') {
+  const { pages } = await extractPagesMarkdownAsync(pdf)
+  // ...
+}
+```
+
 ## Types
 
 ```typescript

napi/bun.lock

@@ -8,12 +8,12 @@
         "@napi-rs/cli": "^3.4.1",
       },
       "optionalDependencies": {
-        "@firecrawl/pdf-inspector-darwin-arm64": "1.12.0",
-        "@firecrawl/pdf-inspector-linux-arm64-gnu": "1.12.0",
-        "@firecrawl/pdf-inspector-linux-arm64-musl": "1.12.0",
-        "@firecrawl/pdf-inspector-linux-x64-gnu": "1.12.0",
-        "@firecrawl/pdf-inspector-linux-x64-musl": "1.12.0",
-        "@firecrawl/pdf-inspector-win32-x64-msvc": "1.12.0",
+        "@firecrawl/pdf-inspector-darwin-arm64": "1.14.2",
+        "@firecrawl/pdf-inspector-linux-arm64-gnu": "1.14.2",
+        "@firecrawl/pdf-inspector-linux-arm64-musl": "1.14.2",
+        "@firecrawl/pdf-inspector-linux-x64-gnu": "1.14.2",
+        "@firecrawl/pdf-inspector-linux-x64-musl": "1.14.2",
+        "@firecrawl/pdf-inspector-win32-x64-msvc": "1.14.2",
       },
     },
   },

napi/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@firecrawl/pdf-inspector",
-  "version": "1.12.0",
+  "version": "1.14.2",
   "description": "Fast PDF classification and text extraction. Detect text-based vs scanned PDFs, extract text by region with quality checks. Native Rust performance via napi-rs.",
   "main": "index.js",
   "types": "index.d.ts",
@@ -52,11 +52,11 @@
     "@napi-rs/cli": "^3.4.1"
   },
   "optionalDependencies": {
-    "@firecrawl/pdf-inspector-linux-x64-gnu": "1.12.0",
-    "@firecrawl/pdf-inspector-linux-x64-musl": "1.12.0",
-    "@firecrawl/pdf-inspector-linux-arm64-gnu": "1.12.0",
-    "@firecrawl/pdf-inspector-linux-arm64-musl": "1.12.0",
-    "@firecrawl/pdf-inspector-darwin-arm64": "1.12.0",
-    "@firecrawl/pdf-inspector-win32-x64-msvc": "1.12.0"
+    "@firecrawl/pdf-inspector-linux-x64-gnu": "1.14.2",
+    "@firecrawl/pdf-inspector-linux-x64-musl": "1.14.2",
+    "@firecrawl/pdf-inspector-linux-arm64-gnu": "1.14.2",
+    "@firecrawl/pdf-inspector-linux-arm64-musl": "1.14.2",
+    "@firecrawl/pdf-inspector-darwin-arm64": "1.14.2",
+    "@firecrawl/pdf-inspector-win32-x64-msvc": "1.14.2"
   }
 }

napi/src/lib.rs

@@ -89,6 +89,11 @@ pub struct TextItem {
     pub item_type: ItemType,
     /// URL for link items, `None` for other types.
     pub link_url: Option<String>,
+    /// Marked Content ID from the content stream's BDC/BMC operator, `None`
+    /// when the text is not part of marked content. Join with the
+    /// `page`/`mcid` pairs from [`extractStructureElements`] to attach
+    /// structure-tree roles (headings, paragraphs, …) in tagged PDFs.
+    pub mcid: Option<i64>,
 }
 
 /// A page's regions for text extraction: (page_index_0based, bboxes).
@@ -153,9 +158,7 @@ fn to_napi_result(r: pdf_inspector::PdfProcessResult) -> PdfResult {
     }
 }
 
-fn to_napi_page_ocr_reasons(
-    reasons: Vec<pdf_inspector::PageOcrReasons>,
-) -> Vec<PageOcrReasons> {
+fn to_napi_page_ocr_reasons(reasons: Vec<pdf_inspector::PageOcrReasons>) -> Vec<PageOcrReasons> {
     reasons
         .into_iter()
         .map(|reason| PageOcrReasons {
@@ -202,6 +205,31 @@ where
     }
 }
 
+// ---------------------------------------------------------------------------
+// Shared implementations (single body behind sync and async entry points)
+// ---------------------------------------------------------------------------
+
+fn process_pdf_impl(bytes: &[u8], pages: Option<Vec<u32>>) -> Result<PdfResult> {
+    let mut opts = pdf_inspector::PdfOptions::new();
+    if let Some(p) = pages {
+        opts = opts.pages(p);
+    }
+    let result = pdf_inspector::process_pdf_mem_with_options(bytes, opts)
+        .map_err(|e| to_napi_err(e, "process_pdf"))?;
+    Ok(to_napi_result(result))
+}
+
+fn classify_pdf_impl(bytes: &[u8]) -> Result<PdfClassification> {
+    let result =
+        pdf_inspector::classify_pdf_mem(bytes).map_err(|e| to_napi_err(e, "classify_pdf"))?;
+    Ok(PdfClassification {
+        pdf_type: convert_pdf_type(result.pdf_type),
+        page_count: result.page_count,
+        pages_needing_ocr: result.pages_needing_ocr,
+        confidence: result.confidence as f64,
+    })
+}
+
 // ---------------------------------------------------------------------------
 // Public NAPI API
 // ---------------------------------------------------------------------------
@@ -210,15 +238,7 @@ where
 #[napi]
 pub fn process_pdf(buffer: Buffer, pages: Option<Vec<u32>>) -> Result<PdfResult> {
     let bytes: Vec<u8> = buffer.to_vec();
-    catch_panic("process_pdf", move || {
-        let mut opts = pdf_inspector::PdfOptions::new();
-        if let Some(p) = pages {
-            opts = opts.pages(p);
-        }
-        let result = pdf_inspector::process_pdf_mem_with_options(&bytes, opts)
-            .map_err(|e| to_napi_err(e, "process_pdf"))?;
-        Ok(to_napi_result(result))
-    })
+    catch_panic("process_pdf", move || process_pdf_impl(&bytes, pages))
 }
 
 /// Fast detection only — no text extraction or markdown.
@@ -238,16 +258,7 @@ pub fn detect_pdf(buffer: Buffer) -> Result<PdfResult> {
 #[napi]
 pub fn classify_pdf(buffer: Buffer) -> Result<PdfClassification> {
     let bytes: Vec<u8> = buffer.to_vec();
-    catch_panic("classify_pdf", move || {
-        let result =
-            pdf_inspector::classify_pdf_mem(&bytes).map_err(|e| to_napi_err(e, "classify_pdf"))?;
-        Ok(PdfClassification {
-            pdf_type: convert_pdf_type(result.pdf_type),
-            page_count: result.page_count,
-            pages_needing_ocr: result.pages_needing_ocr,
-            confidence: result.confidence as f64,
-        })
-    })
+    catch_panic("classify_pdf", move || classify_pdf_impl(&bytes))
 }
 
 /// Extract plain text from a PDF Buffer.
@@ -300,12 +311,61 @@ pub fn extract_text_with_positions(
                     is_strikeout: item.is_strikeout,
                     item_type,
                     link_url,
+                    mcid: item.mcid,
                 }
             })
             .collect())
     })
 }
 
+/// One structure-tree element reference from a tagged PDF.
+#[napi(object)]
+pub struct StructureElementJs {
+    /// 1-indexed page number (matches `TextItem.page`).
+    pub page: u32,
+    /// Marked Content ID from the page's content stream (matches
+    /// `TextItem.mcid`).
+    pub mcid: i64,
+    /// Standard structure type name ("H1".."H6", "P", "Table", "TD", …).
+    /// Custom tags are resolved through the document's role map; tags with
+    /// no standard mapping are returned verbatim.
+    pub role: String,
+}
+
+/// Extract structure-tree element references from a tagged PDF.
+///
+/// Parses the document's structure tree (when present) and returns one
+/// entry per marked-content reference, resolved to its 1-indexed page,
+/// MCID, and structure type name. Returns an empty array when the PDF is
+/// not tagged.
+///
+/// Join `(page, mcid)` against the `page`/`mcid` fields from
+/// [`extractTextWithPositions`] to attach heading levels (H1..H6) and other
+/// semantic roles to extracted text.
+///
+/// Pass 1-indexed page numbers (matching `TextItem.page`) to restrict
+/// output; omit `pages` for the whole document. Entries are sorted by
+/// `(page, mcid)`.
+#[napi]
+pub fn extract_structure_elements(
+    buffer: Buffer,
+    pages: Option<Vec<u32>>,
+) -> Result<Vec<StructureElementJs>> {
+    let bytes: Vec<u8> = buffer.to_vec();
+    catch_panic("extract_structure_elements", move || {
+        let elements = pdf_inspector::extract_structure_elements_mem(&bytes, pages.as_deref())
+            .map_err(|e| to_napi_err(e, "extract_structure_elements"))?;
+        Ok(elements
+            .into_iter()
+            .map(|e| StructureElementJs {
+                page: e.page,
+                mcid: e.mcid,
+                role: e.role,
+            })
+            .collect())
+    })
+}
+
 /// Extract text within bounding-box regions from a PDF.
 ///
 /// For hybrid OCR: layout model detects regions in rendered images,
@@ -633,25 +693,32 @@ pub fn extract_pages_markdown(
 ) -> Result<PagesExtractionResult> {
     let bytes: Vec<u8> = buffer.to_vec();
     catch_panic("extract_pages_markdown", move || {
-        let result = pdf_inspector::extract_pages_markdown_mem(&bytes, pages.as_deref())
-            .map_err(|e| to_napi_err(e, "extract_pages_markdown"))?;
-        Ok(PagesExtractionResult {
-            pages: result
-                .pages
-                .into_iter()
-                .map(|r| PageMarkdownResult {
-                    page: r.page,
-                    markdown: r.markdown,
-                    needs_ocr: r.needs_ocr,
-                    ocr_reason: r.ocr_reason,
-                })
-                .collect(),
-            pages_with_tables: result.pages_with_tables,
-            pages_with_columns: result.pages_with_columns,
-            pages_needing_ocr: result.pages_needing_ocr,
-            ocr_reasons_by_page: to_napi_page_ocr_reasons(result.ocr_reasons_by_page),
-            is_complex: result.is_complex,
-        })
+        extract_pages_markdown_impl(&bytes, pages.as_deref())
+    })
+}
+
+fn extract_pages_markdown_impl(
+    bytes: &[u8],
+    pages: Option<&[u32]>,
+) -> Result<PagesExtractionResult> {
+    let result = pdf_inspector::extract_pages_markdown_mem(bytes, pages)
+        .map_err(|e| to_napi_err(e, "extract_pages_markdown"))?;
+    Ok(PagesExtractionResult {
+        pages: result
+            .pages
+            .into_iter()
+            .map(|r| PageMarkdownResult {
+                page: r.page,
+                markdown: r.markdown,
+                needs_ocr: r.needs_ocr,
+                ocr_reason: r.ocr_reason,
+            })
+            .collect(),
+        pages_with_tables: result.pages_with_tables,
+        pages_with_columns: result.pages_with_columns,
+        pages_needing_ocr: result.pages_needing_ocr,
+        ocr_reasons_by_page: to_napi_page_ocr_reasons(result.ocr_reasons_by_page),
+        is_complex: result.is_complex,
     })
 }
 
@@ -692,3 +759,131 @@ fn to_page_region_texts(results: Vec<pdf_inspector::PageRegionResult>) -> Vec<Pa
         })
         .collect()
 }
+
+// ---------------------------------------------------------------------------
+// Async variants (libuv thread pool via AsyncTask)
+//
+// The synchronous exports above parse on the calling thread, which in Node is
+// the event loop. These `*Async` variants run the same shared implementations
+// on the libuv thread pool and hand JavaScript a promise, so servers under
+// concurrent load keep answering requests while a document parses. The sync
+// exports keep their names, signatures, and behaviour.
+//
+// Each factory copies the input Buffer to an owned `Vec<u8>` on the calling
+// (JS) thread — deliberately. JS execution is single-threaded, so no JS code
+// can mutate the buffer while the synchronous part of the call copies it.
+// Holding the napi `Buffer` and reading it from the worker instead would be
+// zero-copy, but a caller mutating the buffer before the promise settles
+// would then race the worker's reads — undefined behavior, not a recoverable
+// error (a known napi-rs soundness hazard with cross-thread Buffer access).
+// The copy is a one-time memcpy, negligible next to the parse it unblocks.
+// ---------------------------------------------------------------------------
+
+pub struct ProcessPdfTask {
+    bytes: Vec<u8>,
+    pages: Option<Vec<u32>>,
+}
+
+impl Task for ProcessPdfTask {
+    type Output = PdfResult;
+    type JsValue = PdfResult;
+
+    fn compute(&mut self) -> Result<Self::Output> {
+        let bytes = std::mem::take(&mut self.bytes);
+        let pages = self.pages.take();
+        // AssertUnwindSafe: `bytes`/`pages` are moved into the closure and
+        // dropped on unwind — no shared state can be observed broken.
+        catch_panic(
+            "process_pdf",
+            panic::AssertUnwindSafe(move || process_pdf_impl(&bytes, pages)),
+        )
+    }
+
+    fn resolve(&mut self, _env: Env, output: Self::Output) -> Result<Self::JsValue> {
+        Ok(output)
+    }
+}
+
+/// Async variant of [`processPdf`]: same result, but the parse runs on the
+/// libuv thread pool instead of the event loop and the call returns a
+/// promise. The buffer is copied before the call returns, so it may be
+/// reused or mutated immediately.
+// ts_return_type is required: napi-rs emits `Promise<unknown>` for
+// `AsyncTask<T>` returns without it.
+#[napi(ts_return_type = "Promise<PdfResult>")]
+pub fn process_pdf_async(buffer: Buffer, pages: Option<Vec<u32>>) -> AsyncTask<ProcessPdfTask> {
+    AsyncTask::new(ProcessPdfTask {
+        bytes: buffer.to_vec(),
+        pages,
+    })
+}
+
+pub struct ClassifyPdfTask {
+    bytes: Vec<u8>,
+}
+
+impl Task for ClassifyPdfTask {
+    type Output = PdfClassification;
+    type JsValue = PdfClassification;
+
+    fn compute(&mut self) -> Result<Self::Output> {
+        let bytes = std::mem::take(&mut self.bytes);
+        catch_panic(
+            "classify_pdf",
+            panic::AssertUnwindSafe(move || classify_pdf_impl(&bytes)),
+        )
+    }
+
+    fn resolve(&mut self, _env: Env, output: Self::Output) -> Result<Self::JsValue> {
+        Ok(output)
+    }
+}
+
+/// Async variant of [`classifyPdf`]: same result, but the classification runs
+/// on the libuv thread pool instead of the event loop and the call returns a
+/// promise. The buffer is copied before the call returns, so it may be
+/// reused or mutated immediately.
+#[napi(ts_return_type = "Promise<PdfClassification>")]
+pub fn classify_pdf_async(buffer: Buffer) -> AsyncTask<ClassifyPdfTask> {
+    AsyncTask::new(ClassifyPdfTask {
+        bytes: buffer.to_vec(),
+    })
+}
+
+pub struct ExtractPagesMarkdownTask {
+    bytes: Vec<u8>,
+    pages: Option<Vec<u32>>,
+}
+
+impl Task for ExtractPagesMarkdownTask {
+    type Output = PagesExtractionResult;
+    type JsValue = PagesExtractionResult;
+
+    fn compute(&mut self) -> Result<Self::Output> {
+        let bytes = std::mem::take(&mut self.bytes);
+        let pages = self.pages.take();
+        catch_panic(
+            "extract_pages_markdown",
+            panic::AssertUnwindSafe(move || extract_pages_markdown_impl(&bytes, pages.as_deref())),
+        )
+    }
+
+    fn resolve(&mut self, _env: Env, output: Self::Output) -> Result<Self::JsValue> {
+        Ok(output)
+    }
+}
+
+/// Async variant of [`extractPagesMarkdown`]: same result, but the extraction
+/// runs on the libuv thread pool instead of the event loop and the call
+/// returns a promise. The buffer is copied before the call returns, so it
+/// may be reused or mutated immediately.
+#[napi(ts_return_type = "Promise<PagesExtractionResult>")]
+pub fn extract_pages_markdown_async(
+    buffer: Buffer,
+    pages: Option<Vec<u32>>,
+) -> AsyncTask<ExtractPagesMarkdownTask> {
+    AsyncTask::new(ExtractPagesMarkdownTask {
+        bytes: buffer.to_vec(),
+        pages,
+    })
+}

napi/test.mjs

@@ -2,16 +2,21 @@ import { readFileSync } from 'fs';
 import { strict as assert } from 'assert';
 import {
   processPdf,
+  processPdfAsync,
   detectPdf,
   classifyPdf,
+  classifyPdfAsync,
   extractText,
   extractTextWithPositions,
+  extractStructureElements,
   extractTextInRegions,
   detectVectorGridInRegion,
   extractPagesMarkdown,
+  extractPagesMarkdownAsync,
 } from './index.js';
 
 const fixture = readFileSync('../tests/fixtures/thermo-freon12.pdf');
+const taggedFixture = readFileSync('../tests/fixtures/firecrawl_docs_tagged.pdf');
 
 // --- processPdf ---
 console.log('Testing processPdf...');
@@ -79,6 +84,46 @@ assert.ok(page1Items.length > 0);
 assert.ok(page1Items.every(i => i.page === 1));
 console.log('  extractTextWithPositions with pages: OK');
 
+// mcid: undefined on untagged PDFs, numeric on tagged marked content
+assert.ok(items.every(i => i.mcid === undefined || typeof i.mcid === 'number'));
+const taggedItems = extractTextWithPositions(taggedFixture);
+assert.ok(
+  taggedItems.some(i => typeof i.mcid === 'number'),
+  'tagged PDF text items should carry Marked Content IDs',
+);
+console.log('  extractTextWithPositions mcid: OK');
+
+// --- extractStructureElements ---
+console.log('Testing extractStructureElements...');
+const structureElements = extractStructureElements(taggedFixture);
+assert.ok(structureElements.length > 0);
+assert.ok(structureElements.every(e => typeof e.page === 'number'));
+assert.ok(structureElements.every(e => typeof e.mcid === 'number'));
+assert.ok(structureElements.every(e => typeof e.role === 'string' && e.role.length > 0));
+assert.ok(
+  structureElements.some(e => e.role === 'H1'),
+  'tagged fixture should surface H1 heading roles',
+);
+
+// (page, mcid) joins against extractTextWithPositions to recover heading text
+const h1Refs = new Set(
+  structureElements.filter(e => e.role === 'H1').map(e => `${e.page}:${e.mcid}`),
+);
+const h1Text = taggedItems
+  .filter(i => typeof i.mcid === 'number' && h1Refs.has(`${i.page}:${i.mcid}`))
+  .map(i => i.text)
+  .join('');
+assert.ok(h1Text.trim().length > 0, 'H1 join should recover heading text');
+
+// pages filter is 1-indexed, matching TextItem.page
+const page1Elements = extractStructureElements(taggedFixture, [1]);
+assert.ok(page1Elements.length > 0);
+assert.ok(page1Elements.every(e => e.page === 1));
+
+// untagged PDFs yield an empty array
+assert.deepEqual(extractStructureElements(fixture), []);
+console.log('  extractStructureElements: OK');
+
 // --- extractTextInRegions ---
 console.log('Testing extractTextInRegions...');
 const regionResults = extractTextInRegions(fixture, [
@@ -124,10 +169,75 @@ assert.equal(picked.pages[0].page, 2);
 assert.equal(picked.pages[1].page, 0);
 console.log('  extractPagesMarkdown with pages: OK');
 
+// --- Async variants ---
+console.log('Testing async variants...');
+
+// processPdfAsync returns a promise and matches the sync result
+const asyncResultPromise = processPdfAsync(fixture);
+assert.ok(asyncResultPromise instanceof Promise);
+const asyncResult = await asyncResultPromise;
+assert.equal(asyncResult.pdfType, result.pdfType);
+assert.equal(asyncResult.pageCount, result.pageCount);
+assert.equal(asyncResult.markdown, result.markdown);
+console.log('  processPdfAsync: OK');
+
+// processPdfAsync with pages
+const asyncResult2 = await processPdfAsync(fixture, [1]);
+assert.equal(asyncResult2.markdown, result2.markdown);
+console.log('  processPdfAsync with pages: OK');
+
+// classifyPdfAsync matches the sync result
+const asyncClassified = await classifyPdfAsync(fixture);
+assert.equal(asyncClassified.pdfType, classified.pdfType);
+assert.equal(asyncClassified.pageCount, classified.pageCount);
+assert.equal(asyncClassified.confidence, classified.confidence);
+assert.deepEqual(asyncClassified.pagesNeedingOcr, classified.pagesNeedingOcr);
+console.log('  classifyPdfAsync: OK');
+
+// extractPagesMarkdownAsync matches the sync result
+const asyncAllPages = await extractPagesMarkdownAsync(fixture);
+assert.equal(asyncAllPages.pages.length, allPages.pages.length);
+assert.deepEqual(
+  asyncAllPages.pages.map(p => p.markdown),
+  allPages.pages.map(p => p.markdown),
+);
+assert.equal(asyncAllPages.isComplex, allPages.isComplex);
+console.log('  extractPagesMarkdownAsync: OK');
+
+// selected pages preserve caller order
+const asyncPicked = await extractPagesMarkdownAsync(fixture, [2, 0]);
+assert.equal(asyncPicked.pages.length, 2);
+assert.equal(asyncPicked.pages[0].page, 2);
+assert.equal(asyncPicked.pages[1].page, 0);
+console.log('  extractPagesMarkdownAsync with pages: OK');
+
+// input buffer is copied at call time: mutating it immediately after the
+// call must not affect the in-flight parse
+const scratch = Buffer.from(fixture);
+const inFlight = processPdfAsync(scratch);
+scratch.fill(0);
+const fromMutated = await inFlight;
+assert.equal(fromMutated.markdown, result.markdown);
+console.log('  processPdfAsync input copied at call time: OK');
+
+// concurrent async calls all settle
+const [c1, c2, c3] = await Promise.all([
+  processPdfAsync(fixture),
+  classifyPdfAsync(fixture),
+  extractPagesMarkdownAsync(fixture),
+]);
+assert.equal(c1.pdfType, 'TextBased');
+assert.equal(c2.pdfType, 'TextBased');
+assert.equal(c3.pages.length, 3);
+console.log('  concurrent async calls: OK');
+
 // --- Error handling ---
 console.log('Testing error handling...');
 assert.throws(() => processPdf(Buffer.from('not a pdf')), /process_pdf/);
 assert.throws(() => classifyPdf(Buffer.from('')), /classify_pdf/);
+await assert.rejects(processPdfAsync(Buffer.from('not a pdf')), /process_pdf/);
+await assert.rejects(classifyPdfAsync(Buffer.from('')), /classify_pdf/);
+await assert.rejects(extractPagesMarkdownAsync(Buffer.from('')), /extract_pages_markdown/);
 console.log('  error handling: OK');
 
 console.log('\nAll NAPI tests passed!');

pdf_inspector.pyi

@@ -51,6 +51,20 @@ class TextItem:
     is_underline: bool
     is_strikeout: bool
     item_type: str
+    mcid: Optional[int]
+    """Marked Content ID from the content stream's BDC/BMC operator, None when
+    the text is not part of marked content. Join with the (page, mcid) pairs
+    from extract_structure_elements to attach structure-tree roles in tagged
+    PDFs."""
+
+class StructureElement:
+    """One structure-tree element reference from a tagged PDF."""
+    page: int
+    """1-indexed page number (matches TextItem.page)."""
+    mcid: int
+    """Marked Content ID from the page's content stream (matches TextItem.mcid)."""
+    role: str
+    """Standard structure type name ("H1".."H6", "P", "Table", "TD", ...)."""
 
 class RegionText:
     """Extracted text for a single region."""
@@ -132,6 +146,27 @@ def extract_text_with_positions_bytes(data: bytes, pages: Optional[list[int]] =
     """Extract text with position information from bytes."""
     ...
 
+def extract_structure_elements(path: str, pages: Optional[list[int]] = None) -> list[StructureElement]:
+    """Extract structure-tree element references from a tagged PDF file.
+
+    Returns one entry per marked-content reference, resolved to its 1-indexed
+    page, MCID, and structure type name ("H1".."H6", "P", "Table", ...), sorted
+    by (page, mcid). Returns an empty list when the PDF is not tagged.
+
+    Args:
+        path: Path to the PDF file.
+        pages: Optional list of 1-indexed pages (matching ``TextItem.page``).
+            When ``None`` (default), the whole document is returned.
+    """
+    ...
+
+def extract_structure_elements_bytes(data: bytes, pages: Optional[list[int]] = None) -> list[StructureElement]:
+    """Extract structure-tree element references from tagged PDF bytes.
+
+    See :func:`extract_structure_elements` for details.
+    """
+    ...
+
 def extract_text_in_regions(
     path: str,
     page_regions: list[tuple[int, list[list[float]]]],

pyproject.toml

@@ -4,9 +4,9 @@ build-backend = "maturin"
 
 [project]
 name = "pdf-inspector"
-# Bump this to publish to PyPI — CI publishes automatically when the version
-# changes on main (same flow as napi/package.json for npm).
-version = "0.2.6"
+# Keep package versions in sync with `python3 scripts/version.py <version>`.
+# CI publishes automatically when the synchronized change lands on main.
+version = "1.14.2"
 description = "Fast PDF inspection, classification, and text extraction with smart scanned vs text-based detection"
 readme = "docs/python.md"
 license = { text = "MIT" }

scripts/tests/test_version.py

@@ -0,0 +1,111 @@
+import json
+import sys
+import tempfile
+import unittest
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
+
+from version import PLATFORM_PACKAGES, check_versions, set_versions
+
+
+class VersionTests(unittest.TestCase):
+    def setUp(self):
+        self.temporary = tempfile.TemporaryDirectory()
+        self.root = Path(self.temporary.name)
+        (self.root / "napi").mkdir()
+        (self.root / "site").mkdir()
+        (self.root / "wasm").mkdir()
+
+        self._write_manifest("Cargo.toml", "package", "0.1.0")
+        self._write_manifest("pyproject.toml", "project", "0.1.0")
+        self._write_manifest("napi/Cargo.toml", "package", "0.1.0")
+        self._write_manifest("wasm/Cargo.toml", "package", "0.1.0")
+
+        package = {
+            "name": "@firecrawl/pdf-inspector",
+            "version": "0.1.0",
+            "optionalDependencies": {
+                dependency: "0.1.0" for dependency in PLATFORM_PACKAGES
+            },
+        }
+        (self.root / "napi/package.json").write_text(
+            json.dumps(package), encoding="utf-8"
+        )
+        (self.root / "napi/bun.lock").write_text(
+            "\n".join(
+                f'        "{dependency}": "0.1.0",'
+                for dependency in PLATFORM_PACKAGES
+            )
+            + "\n",
+            encoding="utf-8",
+        )
+        (self.root / "site/index.html").write_text(
+            'https://cdn.jsdelivr.net/npm/@firecrawl/pdf-inspector-wasm@0.1.0/'
+            'pdf_inspector_wasm.js\n',
+            encoding="utf-8",
+        )
+        self._write_lock(
+            "napi/Cargo.lock", ("pdf-inspector", "pdf-inspector-napi")
+        )
+        self._write_lock(
+            "wasm/Cargo.lock", ("pdf-inspector", "pdf-inspector-wasm")
+        )
+
+    def tearDown(self):
+        self.temporary.cleanup()
+
+    def _write_manifest(self, relative, section, version):
+        (self.root / relative).write_text(
+            f'[{section}]\nname = "fixture"\nversion = "{version}"\n',
+            encoding="utf-8",
+        )
+
+    def _write_lock(self, relative, packages):
+        content = "\n".join(
+            f'[[package]]\nname = "{package}"\nversion = "0.1.0"\n'
+            for package in packages
+        )
+        (self.root / relative).write_text(content, encoding="utf-8")
+
+    def test_updates_every_version_location(self):
+        set_versions("1.14.0", self.root)
+
+        self.assertEqual(check_versions(self.root), "1.14.0")
+
+    def test_reports_a_divergent_package(self):
+        self._write_manifest("wasm/Cargo.toml", "package", "0.2.0")
+
+        with self.assertRaisesRegex(ValueError, "WASM package: 0.2.0"):
+            check_versions(self.root)
+
+    def test_rejects_an_invalid_version(self):
+        with self.assertRaisesRegex(ValueError, "Invalid semantic version"):
+            set_versions("next", self.root)
+
+    def test_rejects_numeric_prerelease_with_leading_zero(self):
+        before = (self.root / "Cargo.toml").read_text(encoding="utf-8")
+
+        with self.assertRaisesRegex(ValueError, "Invalid semantic version"):
+            set_versions("1.2.3-01", self.root)
+
+        self.assertEqual(
+            (self.root / "Cargo.toml").read_text(encoding="utf-8"), before
+        )
+
+    def test_preflight_failure_does_not_partially_update(self):
+        before = (self.root / "Cargo.toml").read_text(encoding="utf-8")
+        (self.root / "site/index.html").write_text(
+            "missing module URL\n", encoding="utf-8"
+        )
+
+        with self.assertRaisesRegex(ValueError, "Missing pinned WASM package URL"):
+            set_versions("1.14.0", self.root)
+
+        self.assertEqual(
+            (self.root / "Cargo.toml").read_text(encoding="utf-8"), before
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()

scripts/version.py

@@ -0,0 +1,262 @@
+#!/usr/bin/env python3
+"""Keep every pdf-inspector package on one release version."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import re
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parents[1]
+PRERELEASE_IDENTIFIER = (
+    r"(?:0|[1-9]\d*|[0-9A-Za-z-]*[A-Za-z-][0-9A-Za-z-]*)"
+)
+SEMVER = re.compile(
+    r"^(0|[1-9]\d*)\.(0|[1-9]\d*)\.(0|[1-9]\d*)"
+    rf"(?:-{PRERELEASE_IDENTIFIER}(?:\.{PRERELEASE_IDENTIFIER})*)?"
+    r"(?:\+[0-9A-Za-z-]+(?:\.[0-9A-Za-z-]+)*)?$"
+)
+VERSION_LINE = re.compile(r'^(\s*version\s*=\s*")[^"]+(".*)$')
+SECTION_LINE = re.compile(r"^\s*\[([^]]+)]\s*$")
+PLATFORM_PACKAGES = (
+    "@firecrawl/pdf-inspector-linux-x64-gnu",
+    "@firecrawl/pdf-inspector-linux-x64-musl",
+    "@firecrawl/pdf-inspector-linux-arm64-gnu",
+    "@firecrawl/pdf-inspector-linux-arm64-musl",
+    "@firecrawl/pdf-inspector-darwin-arm64",
+    "@firecrawl/pdf-inspector-win32-x64-msvc",
+)
+TOML_VERSIONS = (
+    ("Rust crate", Path("Cargo.toml"), "package"),
+    ("Python package", Path("pyproject.toml"), "project"),
+    ("NAPI crate", Path("napi/Cargo.toml"), "package"),
+    ("WASM package", Path("wasm/Cargo.toml"), "package"),
+)
+LOCK_VERSIONS = (
+    ("NAPI lock: core", Path("napi/Cargo.lock"), "pdf-inspector"),
+    ("NAPI lock: binding", Path("napi/Cargo.lock"), "pdf-inspector-napi"),
+    ("WASM lock: core", Path("wasm/Cargo.lock"), "pdf-inspector"),
+    ("WASM lock: binding", Path("wasm/Cargo.lock"), "pdf-inspector-wasm"),
+)
+SITE_WASM_VERSION = re.compile(
+    r"(@firecrawl/pdf-inspector-wasm@)([^/\"]+)(/pdf_inspector_wasm\.js)"
+)
+
+
+def _read_section_version(path: Path, section: str) -> str:
+    active = False
+    for line in path.read_text(encoding="utf-8").splitlines():
+        section_match = SECTION_LINE.match(line)
+        if section_match:
+            active = section_match.group(1) == section
+        elif active:
+            version_match = VERSION_LINE.match(line)
+            if version_match:
+                return line.split('"', 2)[1]
+    raise ValueError(f"No version found in [{section}] of {path}")
+
+
+def _write_section_version(path: Path, section: str, version: str) -> None:
+    lines = path.read_text(encoding="utf-8").splitlines(keepends=True)
+    active = False
+    for index, line in enumerate(lines):
+        section_match = SECTION_LINE.match(line)
+        if section_match:
+            active = section_match.group(1) == section
+        elif active:
+            version_match = VERSION_LINE.match(line)
+            if version_match:
+                newline = "\n" if line.endswith("\n") else ""
+                replacement = (
+                    f"{version_match.group(1)}{version}"
+                    f"{version_match.group(2).rstrip()}"
+                )
+                lines[index] = (
+                    f"{replacement}{newline}"
+                )
+                path.write_text("".join(lines), encoding="utf-8")
+                return
+    raise ValueError(f"No version found in [{section}] of {path}")
+
+
+def _package_block(lines: list[str], package: str) -> tuple[int, int]:
+    for start, line in enumerate(lines):
+        if line.strip() != "[[package]]":
+            continue
+        end = next(
+            (
+                index
+                for index in range(start + 1, len(lines))
+                if lines[index].strip() == "[[package]]"
+            ),
+            len(lines),
+        )
+        if any(line.strip() == f'name = "{package}"' for line in lines[start:end]):
+            return start, end
+    raise ValueError(f"No lockfile entry found for {package}")
+
+
+def _read_lock_version(path: Path, package: str) -> str:
+    lines = path.read_text(encoding="utf-8").splitlines()
+    start, end = _package_block(lines, package)
+    for line in lines[start:end]:
+        version_match = VERSION_LINE.match(line)
+        if version_match:
+            return line.split('"', 2)[1]
+    raise ValueError(f"No version found for {package} in {path}")
+
+
+def _write_lock_version(path: Path, package: str, version: str) -> None:
+    lines = path.read_text(encoding="utf-8").splitlines(keepends=True)
+    start, end = _package_block(lines, package)
+    for index in range(start, end):
+        version_match = VERSION_LINE.match(lines[index])
+        if version_match:
+            newline = "\n" if lines[index].endswith("\n") else ""
+            lines[index] = (
+                f'{version_match.group(1)}{version}{version_match.group(2).rstrip()}'
+                f"{newline}"
+            )
+            path.write_text("".join(lines), encoding="utf-8")
+            return
+    raise ValueError(f"No version found for {package} in {path}")
+
+
+def _node_versions(root: Path) -> dict[str, str]:
+    package = json.loads((root / "napi/package.json").read_text(encoding="utf-8"))
+    versions = {"Node package": package["version"]}
+    optional = package.get("optionalDependencies", {})
+    for dependency in PLATFORM_PACKAGES:
+        if dependency not in optional:
+            raise ValueError(f"Missing Node optional dependency: {dependency}")
+        versions[f"Node optional dependency: {dependency}"] = optional[dependency]
+    return versions
+
+
+def _bun_versions(root: Path) -> dict[str, str]:
+    text = (root / "napi/bun.lock").read_text(encoding="utf-8")
+    versions = {}
+    for dependency in PLATFORM_PACKAGES:
+        match = re.search(
+            rf'"{re.escape(dependency)}": "([^"]+)"[,]', text
+        )
+        if not match:
+            raise ValueError(f"Missing Bun lock dependency: {dependency}")
+        versions[f"Bun lock: {dependency}"] = match.group(1)
+    return versions
+
+
+def _site_wasm_version(root: Path) -> str:
+    text = (root / "site/index.html").read_text(encoding="utf-8")
+    match = SITE_WASM_VERSION.search(text)
+    if not match:
+        raise ValueError("Missing pinned WASM package URL in site/index.html")
+    return match.group(2)
+
+
+def package_versions(root: Path = ROOT) -> dict[str, str]:
+    versions = {
+        label: _read_section_version(root / relative, section)
+        for label, relative, section in TOML_VERSIONS
+    }
+    versions.update(_node_versions(root))
+    versions.update(_bun_versions(root))
+    versions["Website WASM module"] = _site_wasm_version(root)
+    versions.update(
+        {
+            label: _read_lock_version(root / relative, package)
+            for label, relative, package in LOCK_VERSIONS
+        }
+    )
+    return versions
+
+
+def check_versions(root: Path = ROOT) -> str:
+    versions = package_versions(root)
+    expected = versions["Rust crate"]
+    if not SEMVER.fullmatch(expected):
+        raise ValueError(f"Rust crate has an invalid semantic version: {expected}")
+    mismatches = {
+        label: version for label, version in versions.items() if version != expected
+    }
+    if mismatches:
+        details = "\n".join(
+            f"  - {label}: {version}" for label, version in mismatches.items()
+        )
+        raise ValueError(f"Expected every package to use {expected}:\n{details}")
+    return expected
+
+
+def set_versions(version: str, root: Path = ROOT) -> None:
+    if not SEMVER.fullmatch(version):
+        raise ValueError(f"Invalid semantic version: {version}")
+
+    # Validate every expected location before writing the first file. This
+    # prevents a stale manifest or generated file from leaving a partial bump.
+    package_versions(root)
+
+    for _, relative, section in TOML_VERSIONS:
+        _write_section_version(root / relative, section, version)
+
+    package_path = root / "napi/package.json"
+    package = json.loads(package_path.read_text(encoding="utf-8"))
+    package["version"] = version
+    optional = package.get("optionalDependencies", {})
+    for dependency in PLATFORM_PACKAGES:
+        if dependency not in optional:
+            raise ValueError(f"Missing Node optional dependency: {dependency}")
+        optional[dependency] = version
+    package_path.write_text(json.dumps(package, indent=2) + "\n", encoding="utf-8")
+
+    bun_path = root / "napi/bun.lock"
+    bun_text = bun_path.read_text(encoding="utf-8")
+    for dependency in PLATFORM_PACKAGES:
+        pattern = rf'("{re.escape(dependency)}": ")[^"]+("[,])'
+        bun_text, count = re.subn(
+            pattern, rf"\g<1>{version}\g<2>", bun_text, count=1
+        )
+        if count != 1:
+            raise ValueError(f"Missing Bun lock dependency: {dependency}")
+    bun_path.write_text(bun_text, encoding="utf-8")
+
+    site_path = root / "site/index.html"
+    site_text = site_path.read_text(encoding="utf-8")
+    site_text, count = SITE_WASM_VERSION.subn(
+        rf"\g<1>{version}\g<3>", site_text, count=1
+    )
+    if count != 1:
+        raise ValueError("Missing pinned WASM package URL in site/index.html")
+    site_path.write_text(site_text, encoding="utf-8")
+
+    for _, relative, package_name in LOCK_VERSIONS:
+        _write_lock_version(root / relative, package_name, version)
+
+    check_versions(root)
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("version", nargs="?", help="new shared semantic version")
+    parser.add_argument(
+        "--check", action="store_true", help="fail if package versions have diverged"
+    )
+    arguments = parser.parse_args()
+    if arguments.check == bool(arguments.version):
+        parser.error("provide either a version or --check")
+
+    try:
+        if arguments.check:
+            version = check_versions()
+            print(f"All packages use {version}")
+        else:
+            set_versions(arguments.version)
+            print(f"Updated all packages to {arguments.version}")
+    except ValueError as error:
+        parser.exit(1, f"{error}\n")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

site/index.html

@@ -975,7 +975,7 @@ result = pdf_inspector.<span class="fn">process_pdf</span>(<span class="str">"do
 <script>
   (() => {
     const MAX_FILE_SIZE = 25 * 1024 * 1024;
-    const WASM_MODULE_URL = "https://cdn.jsdelivr.net/npm/@firecrawl/pdf-inspector-wasm@0.1.1/pdf_inspector_wasm.js";
+    const WASM_MODULE_URL = "https://cdn.jsdelivr.net/npm/@firecrawl/pdf-inspector-wasm@1.14.2/pdf_inspector_wasm.js";
     const input = document.querySelector("#pdf-input");
     const dropZone = document.querySelector("#drop-zone");
     const filePanel = document.querySelector("#demo-file");

src/bin/pdf2md.rs

@@ -1,6 +1,11 @@
 //! CLI tool for PDF to Markdown conversion
 
 use pdf_inspector::extractor::ItemType;
+#[cfg(all(feature = "ocr", not(target_arch = "wasm32")))]
+use pdf_inspector::vision::{
+    process_pdf_with_ocr, ModelDownloadPolicy, OcrMode, OcrOptions, OcrPdfOptions, OcrPdfResult,
+    PageContentSource, RenderOptions,
+};
 use pdf_inspector::{
     extract_text_with_positions_pages_with_password, process_pdf_with_options, LayoutComplexity,
     PdfOptions, PdfType, ProcessMode, TextItem,
@@ -103,6 +108,134 @@ fn format_items_json(items: &[TextItem]) -> String {
     )
 }
 
+#[cfg(all(feature = "ocr", not(target_arch = "wasm32")))]
+fn optional_json_number(value: Option<f32>) -> String {
+    value
+        .filter(|value| value.is_finite())
+        .map(|value| format!("{value:.4}"))
+        .unwrap_or_else(|| "null".to_string())
+}
+
+#[cfg(all(feature = "ocr", not(target_arch = "wasm32")))]
+fn format_ocr_json(result: &OcrPdfResult) -> String {
+    let routed = result
+        .pages_routed_to_ocr
+        .iter()
+        .map(u32::to_string)
+        .collect::<Vec<_>>()
+        .join(",");
+    let recommended = result
+        .pages_recommended_for_ocr
+        .iter()
+        .map(u32::to_string)
+        .collect::<Vec<_>>()
+        .join(",");
+    let hosted = result
+        .pages_recommending_hosted
+        .iter()
+        .map(u32::to_string)
+        .collect::<Vec<_>>()
+        .join(",");
+    let pages = result
+        .pages
+        .iter()
+        .map(|page| {
+            let provenance = &page.provenance;
+            let source = match provenance.source {
+                PageContentSource::Native => "native",
+                PageContentSource::Ocr => "ocr",
+                PageContentSource::Fused => "fused",
+                _ => "unknown",
+            };
+            let model = provenance
+                .ocr_model
+                .as_ref()
+                .map(|model| {
+                    format!(
+                        r#"{{"name":"{}","revision":"{}"}}"#,
+                        json_escape(&model.name),
+                        json_escape(&model.revision)
+                    )
+                })
+                .unwrap_or_else(|| "null".to_string());
+            let warnings = provenance
+                .warnings
+                .iter()
+                .map(|warning| format!(r#""{}""#, json_escape(warning)))
+                .collect::<Vec<_>>()
+                .join(",");
+            format!(
+                r#"{{"page":{},"source":"{}","markdown":"{}","ocr_model":{},"render_dpi":{},"ocr_confidence":{},"hosted_recommended":{},"timings":{{"render_ms":{},"ocr_ms":{},"layout_ms":{},"assembly_ms":{}}},"warnings":[{}]}}"#,
+                provenance.page,
+                source,
+                json_escape(&page.markdown),
+                model,
+                optional_json_number(provenance.render_dpi),
+                optional_json_number(provenance.ocr_confidence),
+                provenance.hosted_recommended,
+                provenance.timings.render_ms,
+                provenance.timings.ocr_ms,
+                provenance.timings.layout_ms,
+                provenance.timings.assembly_ms,
+                warnings,
+            )
+        })
+        .collect::<Vec<_>>()
+        .join(",");
+    let table_pages = result
+        .pages_with_tables
+        .iter()
+        .map(u32::to_string)
+        .collect::<Vec<_>>()
+        .join(",");
+    let column_pages = result
+        .pages_with_columns
+        .iter()
+        .map(u32::to_string)
+        .collect::<Vec<_>>()
+        .join(",");
+    let ocr_reasons = format_ocr_reasons_by_page(&result.ocr_reasons_by_page);
+    format!(
+        r#"{{"page_count":{},"processing_time_ms":{},"render_time_ms":{},"ocr_time_ms":{},"pages_recommended_for_ocr":[{}],"pages_routed_to_ocr":[{}],"pages_recommending_hosted":[{}],"ocr_reasons_by_page":[{}],"is_complex":{},"pages_with_tables":[{}],"pages_with_columns":[{}],"pages":[{}],"markdown":"{}"}}"#,
+        result.page_count,
+        result.processing_time_ms,
+        result.render_time_ms,
+        result.ocr_time_ms,
+        recommended,
+        routed,
+        hosted,
+        ocr_reasons,
+        result.is_complex,
+        table_pages,
+        column_pages,
+        pages,
+        json_escape(&result.markdown),
+    )
+}
+
+fn argument_value<'a>(args: &'a [String], name: &str) -> Result<Option<&'a str>, String> {
+    args.iter()
+        .position(|argument| argument == name)
+        .map(|index| {
+            args.get(index + 1)
+                .map(String::as_str)
+                .ok_or_else(|| format!("{name} requires a value"))
+        })
+        .transpose()
+}
+
+#[cfg(all(feature = "ocr", not(target_arch = "wasm32")))]
+fn float_argument(args: &[String], name: &str, default: f32) -> Result<f32, String> {
+    argument_value(args, name)?
+        .map(|value| {
+            value
+                .parse::<f32>()
+                .map_err(|_| format!("{name} requires a number, got {value:?}"))
+        })
+        .transpose()
+        .map(|value| value.unwrap_or(default))
+}
+
 fn extract_items_json(
     pdf_path: &str,
     page_filter: Option<&HashSet<u32>>,
@@ -242,6 +375,12 @@ fn main() {
         eprintln!("  --password PW       Password for an encrypted PDF");
         eprintln!("  --detect-only       Only detect PDF type (no extraction)");
         eprintln!("  --analyze           Detect + extract + layout analysis (no markdown)");
+        eprintln!("  --ocr MODE          OCR mode: off, auto, or force (requires feature `ocr`)");
+        eprintln!("  --ocr-dpi N         OCR render resolution (default: 150)");
+        eprintln!("  --ocr-min-confidence N  Drop OCR spans below N (default: 0)");
+        eprintln!("  --ocr-hosted-threshold N  Recommend hosted parsing below N (default: 0.5)");
+        eprintln!("  --ocr-model-dir DIR Use a package-managed local model directory");
+        eprintln!("  --ocr-offline       Never download missing OCR models");
         process::exit(1);
     }
 
@@ -253,6 +392,10 @@ fn main() {
     let page_numbers = args.iter().any(|a| a == "--pages");
     let detect_only = args.iter().any(|a| a == "--detect-only");
     let analyze = args.iter().any(|a| a == "--analyze");
+    let ocr_mode_argument = argument_value(&args, "--ocr").unwrap_or_else(|error| {
+        eprintln!("Error: {error}");
+        process::exit(1);
+    });
 
     // Parse --password value
     let password = args.iter().position(|a| a == "--password").map(|i| {
@@ -283,6 +426,141 @@ fn main() {
             })
         });
 
+    let output_file = args
+        .get(2)
+        .filter(|a| !a.starts_with("--"))
+        .map(|s| s.as_str());
+
+    let has_ocr_only_option = [
+        "--ocr-dpi",
+        "--ocr-min-confidence",
+        "--ocr-hosted-threshold",
+        "--ocr-model-dir",
+        "--ocr-offline",
+    ]
+    .iter()
+    .any(|option| args.iter().any(|argument| argument == option));
+    if ocr_mode_argument.is_none() && has_ocr_only_option {
+        eprintln!("Error: OCR options require --ocr off, --ocr auto, or --ocr force");
+        process::exit(1);
+    }
+
+    if let Some(mode) = ocr_mode_argument {
+        if items_json_output || detect_only || analyze {
+            eprintln!(
+                "Error: --ocr cannot be combined with --items-json, --detect-only, or --analyze"
+            );
+            process::exit(1);
+        }
+
+        #[cfg(not(all(feature = "ocr", not(target_arch = "wasm32"))))]
+        {
+            let _ = mode;
+            eprintln!("Error: this pdf2md build does not include OCR; rebuild with --features ocr");
+            process::exit(1);
+        }
+
+        #[cfg(all(feature = "ocr", not(target_arch = "wasm32")))]
+        {
+            let mode = match mode {
+                "off" => OcrMode::Off,
+                "auto" => OcrMode::Auto,
+                "force" => OcrMode::Force,
+                value => {
+                    eprintln!("Error: invalid --ocr mode {value:?}; expected off, auto, or force");
+                    process::exit(1);
+                }
+            };
+            let dpi = float_argument(&args, "--ocr-dpi", 150.0).unwrap_or_else(|error| {
+                eprintln!("Error: {error}");
+                process::exit(1);
+            });
+            let minimum_confidence = float_argument(&args, "--ocr-min-confidence", 0.0)
+                .unwrap_or_else(|error| {
+                    eprintln!("Error: {error}");
+                    process::exit(1);
+                });
+            let hosted_threshold = float_argument(&args, "--ocr-hosted-threshold", 0.5)
+                .unwrap_or_else(|error| {
+                    eprintln!("Error: {error}");
+                    process::exit(1);
+                });
+            let model_directory =
+                argument_value(&args, "--ocr-model-dir").unwrap_or_else(|error| {
+                    eprintln!("Error: {error}");
+                    process::exit(1);
+                });
+
+            let mut ocr = OcrOptions::new()
+                .mode(mode)
+                .minimum_confidence(minimum_confidence);
+            if let Some(directory) = model_directory {
+                ocr = ocr.model_directory(directory);
+            }
+            if args.iter().any(|argument| argument == "--ocr-offline") {
+                ocr = ocr.model_downloads(ModelDownloadPolicy::Offline);
+            }
+            let mut markdown = pdf_inspector::MarkdownOptions::default();
+            if compact_output {
+                markdown.profile = pdf_inspector::MarkdownProfile::Compact;
+            }
+            markdown.include_page_numbers = page_numbers;
+            let mut pdf_options = OcrPdfOptions::new()
+                .render(RenderOptions::new().dpi(dpi))
+                .ocr(ocr)
+                .markdown(markdown)
+                .hosted_recommendation_confidence(hosted_threshold);
+            if let Some(pages) = page_filter.clone() {
+                pdf_options = pdf_options.pages(pages);
+            }
+            if let Some(password) = password.clone() {
+                pdf_options = pdf_options.password(password);
+            }
+
+            match process_pdf_with_ocr(pdf_path, pdf_options) {
+                Ok(result) => {
+                    if json_output {
+                        println!("{}", format_ocr_json(&result));
+                    } else if raw_output {
+                        print!("{}", result.markdown);
+                    } else {
+                        eprintln!("PDF to Markdown Conversion (OCR)");
+                        eprintln!("======================================");
+                        eprintln!("File: {pdf_path}");
+                        eprintln!("Pages: {}", result.page_count);
+                        eprintln!("Pages routed to OCR: {:?}", result.pages_routed_to_ocr);
+                        if !result.pages_recommending_hosted.is_empty() {
+                            eprintln!(
+                                "Hosted parsing recommended for pages: {:?}",
+                                result.pages_recommending_hosted
+                            );
+                        }
+                        eprintln!("Processing time: {}ms", result.processing_time_ms);
+                        if let Some(output) = output_file {
+                            fs::write(output, &result.markdown)
+                                .expect("Failed to write output file");
+                            eprintln!("Markdown written to: {output}");
+                        } else {
+                            eprintln!();
+                            eprintln!("--- Markdown Output ---");
+                            eprintln!();
+                            print!("{}", result.markdown);
+                        }
+                    }
+                }
+                Err(error) => {
+                    if json_output {
+                        println!(r#"{{"error":"{}"}}"#, json_escape(&error.to_string()));
+                    } else {
+                        eprintln!("Error: {error}");
+                    }
+                    process::exit(1);
+                }
+            }
+            return;
+        }
+    }
+
     if items_json_output {
         match extract_items_json(pdf_path, page_filter.as_ref(), password.as_deref()) {
             Ok(json) => println!("{}", json),
@@ -294,11 +572,6 @@ fn main() {
         return;
     }
 
-    let output_file = args
-        .get(2)
-        .filter(|a| !a.starts_with("--"))
-        .map(|s| s.as_str());
-
     let process_mode = if detect_only {
         ProcessMode::DetectOnly
     } else if analyze {

src/detector.rs

@@ -1382,7 +1382,13 @@ fn scan_content_for_text_operators(
     let is_word_end =
         |pos: usize| -> bool { pos + 1 >= content.len() || content[pos + 1].is_ascii_whitespace() };
 
-    // Simple state machine to find operators
+    // Simple state machine to find operators.
+    // Each Tj/TJ/Tf lookback stops at the previous text/font operator so a
+    // malformed `] TJ` (no `[`) cannot rescan the entire prefix — that was
+    // quadratic in the number of operators.
+    // `Tj`/`TJ` are only counted when the preceding token closes a string or
+    // array (')', '>', ']'), so `Tj` inside `(Hello Tj World)` cannot pin the floor.
+    let mut operand_floor = 0usize;
     let mut i = 0;
     while i < content.len() {
         let b = content[i];
@@ -1392,14 +1398,15 @@ fn scan_content_for_text_operators(
             let next = content[i + 1];
             if next == b'j' || next == b'J' {
                 // Verify it's an operator (followed by whitespace or newline)
-                if i + 2 >= content.len()
+                if (i + 2 >= content.len()
                     || content[i + 2].is_ascii_whitespace()
                     || content[i + 2] == b'\n'
-                    || content[i + 2] == b'\r'
+                    || content[i + 2] == b'\r')
+                    && preceding_operand_closer(content, i, operand_floor)
                 {
                     text_ops += 1;
-                    // Scan backward for text string operand to collect unique chars
-                    collect_text_chars_before(content, i, unique_chars);
+                    collect_text_chars_before(content, i, unique_chars, operand_floor);
+                    operand_floor = i;
                 }
             } else if next == b'f' {
                 // Tf = set font operator
@@ -1415,12 +1422,10 @@ fn scan_content_for_text_operators(
                     || content[i + 2] == b'<'
                     || content[i + 2] == b'/'
                 {
-                    font_changes += 1;
-                    // Extract the font name operand preceding the size + Tf.
-                    // Pattern: /FontName <size> Tf
-                    // Scan backward past the size number and whitespace to find /Name.
-                    if let Some(name) = extract_font_name_before_tf(content, i) {
+                    if let Some(name) = extract_font_name_before_tf(content, i, operand_floor) {
                         used_font_names.insert(name);
+                        font_changes += 1;
+                        operand_floor = i;
                     }
                 }
             }
@@ -1466,6 +1471,20 @@ fn scan_content_for_text_operators(
     (text_ops, image_count, path_ops, font_changes)
 }
 
+/// True when the token before `op_pos` (skipping whitespace, not crossing
+/// `floor`) is a string/array closer. Used so `Tj` inside `(Hello Tj World)`
+/// is not treated as an operator.
+fn preceding_operand_closer(content: &[u8], op_pos: usize, floor: usize) -> bool {
+    let mut j = op_pos;
+    while j > floor {
+        j -= 1;
+        if !content[j].is_ascii_whitespace() {
+            return matches!(content[j], b')' | b'>' | b']');
+        }
+    }
+    false
+}
+
 /// Extract the font name operand from content stream bytes preceding a Tf operator.
 ///
 /// The Tf operator syntax is: `/FontName size Tf`
@@ -1473,25 +1492,27 @@ fn scan_content_for_text_operators(
 /// whitespace to find the `/Name` token.
 ///
 /// Returns the font name bytes (without the leading `/`), e.g. `b"F1"` for `/F1`.
-fn extract_font_name_before_tf(content: &[u8], tf_pos: usize) -> Option<Vec<u8>> {
+/// `floor` is the start of the previous text/font operator (or 0); lookback
+/// must not cross it.
+fn extract_font_name_before_tf(content: &[u8], tf_pos: usize, floor: usize) -> Option<Vec<u8>> {
     // Scan backward past whitespace before "Tf"
     let mut j = tf_pos;
-    while j > 0 && content[j - 1].is_ascii_whitespace() {
+    while j > floor && content[j - 1].is_ascii_whitespace() {
         j -= 1;
     }
     // Scan backward past the size number (digits, '.', '-')
-    while j > 0
+    while j > floor
         && (content[j - 1].is_ascii_digit() || content[j - 1] == b'.' || content[j - 1] == b'-')
     {
         j -= 1;
     }
     // Scan backward past whitespace between font name and size
-    while j > 0 && content[j - 1].is_ascii_whitespace() {
+    while j > floor && content[j - 1].is_ascii_whitespace() {
         j -= 1;
     }
     // Now j should point just after the font name. Scan backward to find '/'.
     let name_end = j;
-    while j > 0 && content[j - 1] != b'/' {
+    while j > floor && content[j - 1] != b'/' {
         // Font names consist of regular characters (not whitespace, not delimiters)
         if content[j - 1].is_ascii_whitespace() || content[j - 1] == b'(' || content[j - 1] == b')'
         {
@@ -1499,7 +1520,7 @@ fn extract_font_name_before_tf(content: &[u8], tf_pos: usize) -> Option<Vec<u8>>
         }
         j -= 1;
     }
-    if j == 0 || content[j - 1] != b'/' {
+    if j <= floor || content[j - 1] != b'/' {
         return None;
     }
     // j-1 is the '/', font name is content[j..name_end]
@@ -1514,16 +1535,24 @@ fn extract_font_name_before_tf(content: &[u8], tf_pos: usize) -> Option<Vec<u8>>
 /// and collect unique non-whitespace bytes from it.
 ///
 /// Handles both literal strings `(...)` and hex strings `<...>`.
-fn collect_text_chars_before(content: &[u8], op_pos: usize, unique_chars: &mut HashSet<u8>) {
+/// `floor` is the start of the previous text/font operator (or 0); lookback
+/// must not cross it, or a missing `[` before `TJ` rescans the whole prefix.
+fn collect_text_chars_before(
+    content: &[u8],
+    op_pos: usize,
+    unique_chars: &mut HashSet<u8>,
+    floor: usize,
+) {
     // Walk backward past whitespace to find the closing delimiter
     let mut j = op_pos;
-    while j > 0 {
+    while j > floor {
         j -= 1;
         if !content[j].is_ascii_whitespace() {
             break;
         }
     }
-    if j == 0 {
+    // All whitespace, or we landed on the previous operator token.
+    if j == floor {
         return;
     }
 
@@ -1533,7 +1562,7 @@ fn collect_text_chars_before(content: &[u8], op_pos: usize, unique_chars: &mut H
         // Literal string: scan backward for matching '('
         let mut depth = 1i32;
         let mut k = j;
-        while k > 0 && depth > 0 {
+        while k > floor && depth > 0 {
             k -= 1;
             match content[k] {
                 b')' if k == 0 || content[k - 1] != b'\\' => depth += 1,
@@ -1552,7 +1581,7 @@ fn collect_text_chars_before(content: &[u8], op_pos: usize, unique_chars: &mut H
     } else if closing == b'>' {
         // Hex string: scan backward for '<'
         let mut k = j;
-        while k > 0 {
+        while k > floor {
             k -= 1;
             if content[k] == b'<' {
                 break;
@@ -1582,7 +1611,7 @@ fn collect_text_chars_before(content: &[u8], op_pos: usize, unique_chars: &mut H
     } else if closing == b']' {
         // TJ array: scan backward for '[' and collect from all strings inside
         let mut k = j;
-        while k > 0 {
+        while k > floor {
             k -= 1;
             if content[k] == b'[' {
                 break;
@@ -2014,6 +2043,51 @@ mod tests {
         assert_eq!(imgs3, 0);
     }
 
+    #[test]
+    fn test_scan_content_successive_tj_collects_each_operand() {
+        // Lookback is floored at the previous Tj/TJ/Tf so later operators must
+        // still see their own operands.
+        let content = b"[(Hello)] TJ [(World)] TJ (More) Tj";
+        let mut uchars = HashSet::new();
+        let (ops, _, _, _) =
+            scan_content_for_text_operators(content, &mut uchars, &mut HashSet::new());
+        assert_eq!(ops, 3);
+        for &ch in b"HeloWrdM" {
+            assert!(uchars.contains(&ch), "missing char {}", ch as char);
+        }
+    }
+
+    #[test]
+    fn test_scan_content_tj_inside_literal_is_not_an_operator() {
+        // `Tj` followed by space inside a literal must not count as an operator
+        // or pin the lookback floor; the real `Tj` still collects the string.
+        let content = b"BT (Hello Tj World) Tj ET";
+        let mut uchars = HashSet::new();
+        let (ops, _, _, _) =
+            scan_content_for_text_operators(content, &mut uchars, &mut HashSet::new());
+        assert_eq!(ops, 1);
+        for &ch in b"HeloTjWrd" {
+            assert!(uchars.contains(&ch), "missing char {}", ch as char);
+        }
+    }
+
+    #[test]
+    fn test_scan_content_malformed_tj_lookback_stays_linear() {
+        // `] TJ` with no `[` used to walk the entire prefix for every operator
+        // (quadratic). 30k repeats is enough that a prefix rescan would dominate
+        // the test runtime; with the floor it is a single linear pass.
+        let n = 30_000usize;
+        let mut content = Vec::with_capacity(n * 5);
+        for _ in 0..n {
+            content.extend_from_slice(b"] TJ\n");
+        }
+        let mut uchars = HashSet::new();
+        let (ops, _, _, _) =
+            scan_content_for_text_operators(&content, &mut uchars, &mut HashSet::new());
+        assert_eq!(ops, n as u32);
+        assert!(uchars.is_empty());
+    }
+
     #[test]
     fn test_image_dominated_detection() {
         // Do operators are no longer counted as images by scan_content_for_text_operators.
@@ -2772,14 +2846,14 @@ mod tests {
     fn test_extract_font_name_basic() {
         // Standard pattern: /F1 12 Tf
         let content = b"/F1 12 Tf";
-        let name = extract_font_name_before_tf(content, 6); // 'T' is at index 6
+        let name = extract_font_name_before_tf(content, 6, 0); // 'T' is at index 6
         assert_eq!(name, Some(b"F1".to_vec()));
     }
 
     #[test]
     fn test_extract_font_name_long_name() {
         let content = b"/ArialMT-Bold 9.5 Tf";
-        let name = extract_font_name_before_tf(content, 18);
+        let name = extract_font_name_before_tf(content, 18, 0);
         assert_eq!(name, Some(b"ArialMT-Bold".to_vec()));
     }
 

src/extractor/content_decode.rs

@@ -0,0 +1,328 @@
+//! Bounded content-stream decoding.
+//!
+//! `lopdf::content::Content::decode` materializes every operator before any
+//! caller can apply a limit. A compact page of `q Q` pairs can therefore
+//! allocate hundreds of megabytes and abort. Count operators first (without
+//! allocating `Operation` objects) and skip decode when the cap is exceeded.
+
+use crate::PdfError;
+use lopdf::content::Content;
+
+/// Maximum content-stream operators decoded for a page or a single Form
+/// XObject. Matches the previous post-decode skip threshold.
+pub(crate) const MAX_PAGE_OPERATIONS: usize = 1_000_000;
+
+/// Decode `data` unless it contains more than `max_operations` operators.
+///
+/// Returns `Ok(None)` when the stream exceeds the cap, so callers can skip
+/// extraction without first allocating the operation vector.
+pub(crate) fn decode_content_bounded(
+    data: &[u8],
+    max_operations: usize,
+) -> Result<Option<Content>, PdfError> {
+    if content_exceeds_operation_limit(data, max_operations) {
+        return Ok(None);
+    }
+    Content::decode(data)
+        .map(Some)
+        .map_err(|e| PdfError::Parse(e.to_string()))
+}
+
+fn content_exceeds_operation_limit(data: &[u8], max_operations: usize) -> bool {
+    count_content_operators(data, max_operations.saturating_add(1)) > max_operations
+}
+
+/// Count operators using the same token rules as lopdf's content parser,
+/// stopping at `limit`. Does not allocate `Operation` / `Object` values.
+fn count_content_operators(data: &[u8], limit: usize) -> usize {
+    let mut i = 0;
+    let mut count = 0;
+    while i < data.len() && count < limit {
+        skip_content_space(data, &mut i);
+        if i >= data.len() {
+            break;
+        }
+        if data[i] == b'%' {
+            skip_comment(data, &mut i);
+            continue;
+        }
+        match data[i] {
+            b'(' => i = skip_literal_string(data, i),
+            b'<' => {
+                if data.get(i + 1) == Some(&b'<') {
+                    i += 2;
+                } else {
+                    i = skip_hex_string(data, i);
+                }
+            }
+            b'>' => {
+                i += 1;
+                if data.get(i) == Some(&b'>') {
+                    i += 1;
+                }
+            }
+            b'[' | b']' => i += 1,
+            b'/' => skip_name(data, &mut i),
+            b'+' | b'-' | b'.' => skip_number(data, &mut i),
+            b if b.is_ascii_digit() => skip_number(data, &mut i),
+            b if is_operator_byte(b) => {
+                let start = i;
+                i += 1;
+                while i < data.len() && is_operator_byte(data[i]) {
+                    i += 1;
+                }
+                let token = &data[start..i];
+                if token == b"true" || token == b"false" || token == b"null" {
+                    continue;
+                }
+                count += 1;
+                if token == b"BI" && (i >= data.len() || is_content_space(data[i])) {
+                    i = skip_inline_image_after_bi(data, i);
+                }
+            }
+            _ => i += 1,
+        }
+    }
+    count
+}
+
+fn is_content_space(b: u8) -> bool {
+    // PDF whitespace (ISO 32000): NUL, tab, LF, FF, CR, space. Names must
+    // stop on these so a following operator is not absorbed into `/Name`.
+    matches!(b, b'\0' | b'\t' | b'\n' | b'\x0c' | b'\r' | b' ')
+}
+
+fn is_operator_byte(b: u8) -> bool {
+    b.is_ascii_alphabetic() || matches!(b, b'*' | b'\'' | b'"')
+}
+
+fn is_delimiter(b: u8) -> bool {
+    matches!(
+        b,
+        b'(' | b')' | b'<' | b'>' | b'[' | b']' | b'{' | b'}' | b'/' | b'%'
+    )
+}
+
+fn skip_content_space(data: &[u8], i: &mut usize) {
+    while *i < data.len() && is_content_space(data[*i]) {
+        *i += 1;
+    }
+}
+
+fn skip_comment(data: &[u8], i: &mut usize) {
+    while *i < data.len() && data[*i] != b'\n' && data[*i] != b'\r' {
+        *i += 1;
+    }
+}
+
+fn skip_literal_string(data: &[u8], mut i: usize) -> usize {
+    let mut depth = 1i32;
+    i += 1;
+    while i < data.len() && depth > 0 {
+        match data[i] {
+            b'\\' => {
+                i += 1;
+                if i < data.len() {
+                    i += 1;
+                }
+            }
+            b'(' => {
+                depth += 1;
+                i += 1;
+            }
+            b')' => {
+                depth -= 1;
+                i += 1;
+            }
+            _ => i += 1,
+        }
+    }
+    i
+}
+
+fn skip_hex_string(data: &[u8], mut i: usize) -> usize {
+    i += 1;
+    while i < data.len() && data[i] != b'>' {
+        i += 1;
+    }
+    if i < data.len() {
+        i += 1;
+    }
+    i
+}
+
+fn skip_name(data: &[u8], i: &mut usize) {
+    *i += 1;
+    while *i < data.len() && !is_content_space(data[*i]) && !is_delimiter(data[*i]) {
+        *i += 1;
+    }
+}
+
+fn skip_number(data: &[u8], i: &mut usize) {
+    if *i < data.len() && matches!(data[*i], b'+' | b'-') {
+        *i += 1;
+    }
+    while *i < data.len() && data[*i].is_ascii_digit() {
+        *i += 1;
+    }
+    if *i < data.len() && data[*i] == b'.' {
+        *i += 1;
+        while *i < data.len() && data[*i].is_ascii_digit() {
+            *i += 1;
+        }
+    }
+}
+
+/// After a `BI` operator, skip inline-image data through `EI`.
+/// Uses the same PDF whitespace set as `is_content_space`. If `EI` is not
+/// found, leave the cursor in place so later operators are still counted
+/// (undercounting would let decode allocate the full vector).
+fn skip_inline_image_after_bi(data: &[u8], mut i: usize) -> usize {
+    skip_content_space(data, &mut i);
+    let rest = &data[i..];
+    if let Some(pos) = rest.windows(4).position(|w| {
+        is_content_space(w[0]) && w[1] == b'E' && w[2] == b'I' && is_content_space(w[3])
+    }) {
+        return i + pos + 3;
+    }
+    i
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn lopdf_op_count(data: &[u8]) -> usize {
+        Content::decode(data)
+            .map(|c| c.operations.len())
+            .unwrap_or(0)
+    }
+
+    /// DoS safety: never report fewer operators than lopdf would allocate.
+    /// Overcount is acceptable (skip a page); undercount would re-open decode.
+    fn assert_count_does_not_undercount(data: &[u8]) {
+        let ours = count_content_operators(data, usize::MAX);
+        match Content::decode(data) {
+            Ok(content) => assert!(
+                ours >= content.operations.len(),
+                "undercount: ours={ours} lopdf={} for {:?}",
+                content.operations.len(),
+                String::from_utf8_lossy(data)
+            ),
+            Err(_) => {}
+        }
+    }
+
+    #[test]
+    fn operator_count_matches_lopdf_for_typical_streams() {
+        let samples: &[&[u8]] = &[
+            b"q 1 0 0 1 0 0 cm BT /F1 12 Tf 72 720 Td (Hello) Tj ET Q",
+            b"q Q q Q",
+            b"BT /F1 12 Tf 12 TL 1 0 0 1 100 512 Tm (first) Tj (struck) ' ET",
+            b"1 0 0 rg 0 0 10 10 re f",
+            b"true false null q",
+            b"% comment\nq Q\n",
+            b"[ (a) 1 (b) ] TJ",
+            b"1 0 0 1 0 0 cm /Im0 Do",
+        ];
+        for data in samples {
+            assert_eq!(
+                count_content_operators(data, usize::MAX),
+                lopdf_op_count(data),
+                "count mismatch for {}",
+                String::from_utf8_lossy(data)
+            );
+        }
+    }
+
+    #[test]
+    fn strings_and_comments_are_not_operators() {
+        let data = b"(q Q Tj) Tj % q Q\nET";
+        assert_eq!(
+            count_content_operators(data, usize::MAX),
+            lopdf_op_count(data)
+        );
+        assert_eq!(count_content_operators(data, usize::MAX), 2); // Tj, ET
+    }
+
+    #[test]
+    fn inline_image_counts_as_one_operator() {
+        let data = b"BI /W 2 /H 2 /CS /RGB /BPC 8 ID \x00\x01\x02\x03 EI q";
+        assert_eq!(
+            count_content_operators(data, usize::MAX),
+            lopdf_op_count(data)
+        );
+        assert_eq!(count_content_operators(data, usize::MAX), 2); // BI, q
+    }
+
+    #[test]
+    fn inline_image_ei_accepts_pdf_whitespace() {
+        let tab = b"BI /W 1 /H 1 ID \xff\tEI\t q Q";
+        let nul = b"BI /W 1 /H 1 ID \xff\x00EI\x00 q Q";
+        let ff = b"BI /W 1 /H 1 ID \xff\x0cEI\x0c q Q";
+        for data in [tab.as_slice(), nul.as_slice(), ff.as_slice()] {
+            assert_count_does_not_undercount(data);
+            assert!(
+                count_content_operators(data, usize::MAX) >= 3,
+                "BI plus following q Q must remain visible after EI, got {} for {:?}",
+                count_content_operators(data, usize::MAX),
+                String::from_utf8_lossy(data)
+            );
+        }
+    }
+
+    #[test]
+    fn decode_is_skipped_when_operator_cap_is_exceeded() {
+        let mut data = Vec::new();
+        for _ in 0..20 {
+            data.extend_from_slice(b"q Q\n");
+        }
+        assert!(decode_content_bounded(&data, 10).unwrap().is_none());
+        let decoded = decode_content_bounded(&data, 50).unwrap().unwrap();
+        assert_eq!(decoded.operations.len(), 40);
+    }
+
+    #[test]
+    fn name_whitespace_does_not_swallow_following_operator() {
+        // NUL / form-feed end a name (PDF whitespace). Absorbing `q` into
+        // `/x` would undercount and let decode allocate the operator vector.
+        let mut nul_sep = Vec::new();
+        let mut ff_sep = Vec::new();
+        for _ in 0..8_000 {
+            nul_sep.extend_from_slice(b"/x\x00q");
+            ff_sep.extend_from_slice(b"/x\x0cq");
+        }
+        assert_count_does_not_undercount(&nul_sep);
+        assert_count_does_not_undercount(&ff_sep);
+        assert!(count_content_operators(&ff_sep, usize::MAX) >= 8_000);
+    }
+
+    #[test]
+    fn edge_streams_do_not_undercount_vs_lopdf() {
+        let samples: &[&[u8]] = &[
+            b".5 0 0 .5 0 0 cm",
+            b"+1 -2 3.0 rg",
+            b"<0041> Tj",
+            b"(unbalanced",
+            b"BI /W 1 /H 1 ID \xff\xff no EI here q Q q Q",
+            b"q\x00Q\x00q\x00Q",
+            b"/F1\x0c12 Tf (Hi) Tj",
+            b"{ 1 2 add } cvx",
+        ];
+        for data in samples {
+            assert_count_does_not_undercount(data);
+        }
+    }
+
+    #[test]
+    fn million_q_pairs_are_rejected_without_decode() {
+        let mut data = Vec::with_capacity((MAX_PAGE_OPERATIONS + 1) * 2);
+        for _ in 0..=MAX_PAGE_OPERATIONS {
+            data.extend_from_slice(b"q\n");
+        }
+        assert!(content_exceeds_operation_limit(&data, MAX_PAGE_OPERATIONS));
+        assert!(decode_content_bounded(&data, MAX_PAGE_OPERATIONS)
+            .unwrap()
+            .is_none());
+    }
+}

src/extractor/content_stream.rs

@@ -19,7 +19,7 @@ use super::fonts::{
     CMapDecisionCache, FontStyleCache,
 };
 use super::underline::UnderlineLine;
-use super::xobjects::{extract_form_xobject_text, get_page_xobjects, XObjectType};
+use super::xobjects::{extract_form_xobject_text, get_page_xobjects, FormWalkBudget, XObjectType};
 use super::{get_number, image_bbox_from_ctm, multiply_matrices};
 
 /// Strip PDF comments (% to end of line) from content stream bytes.
@@ -137,8 +137,11 @@ fn rise_adjusted(tm: &[f32; 6], rise: f32) -> [f32; 6] {
     ]
 }
 
-/// Returns `(page_extraction, has_gid_fonts)` where `has_gid_fonts` indicates
-/// the page uses fonts with unresolvable gid-encoded glyphs.
+/// Returns `(page_extraction, has_gid_fonts, coords_rotated, skipped_invisible)`
+/// where `has_gid_fonts` indicates the page uses fonts with unresolvable
+/// gid-encoded glyphs and `skipped_invisible` reports that invisible (Tr 3)
+/// text was present but suppressed — callers can use it to decide whether an
+/// `include_invisible` retry could recover anything at all.
 pub(crate) fn extract_page_text_items(
     doc: &Document,
     page_id: ObjectId,
@@ -146,9 +149,8 @@ pub(crate) fn extract_page_text_items(
     font_cmaps: &FontCMaps,
     include_invisible: bool,
     style_cache: &mut FontStyleCache,
-) -> Result<(PageExtraction, bool, bool), PdfError> {
-    use lopdf::content::Content;
-
+    form_budget: &mut FormWalkBudget,
+) -> Result<(PageExtraction, bool, bool, bool), PdfError> {
     let mut items = Vec::new();
     let mut rects: Vec<PdfRect> = Vec::new();
     let mut clip_rects: Vec<PdfRect> = Vec::new();
@@ -252,22 +254,27 @@ pub(crate) fn extract_page_text_items(
     // Content::decode parser, causing it to skip operators like ET and Q.
     let content_data = strip_pdf_comments(&content_data);
 
-    let content = Content::decode(&content_data).map_err(|e| PdfError::Parse(e.to_string()))?;
-
-    const MAX_OPERATIONS: usize = 1_000_000;
-    if content.operations.len() > MAX_OPERATIONS {
-        log::warn!(
-            "page {}: skipping extraction — {} operations exceeds limit ({})",
-            page_num,
-            content.operations.len(),
-            MAX_OPERATIONS
-        );
-        return Ok(((Vec::new(), Vec::new(), Vec::new()), false, false));
-    }
+    let content = match super::content_decode::decode_content_bounded(
+        &content_data,
+        super::content_decode::MAX_PAGE_OPERATIONS,
+    )? {
+        Some(content) => content,
+        None => {
+            log::warn!(
+                "page {}: skipping extraction — content stream exceeds {} operations",
+                page_num,
+                super::content_decode::MAX_PAGE_OPERATIONS
+            );
+            return Ok(((Vec::new(), Vec::new(), Vec::new()), false, false, false));
+        }
+    };
 
     // Graphics state tracking
     let mut ctm = [1.0f32, 0.0, 0.0, 1.0, 0.0, 0.0]; // Current Transformation Matrix
     let mut text_rendering_mode: i32 = 0; // 0=fill, 1=stroke, 2=fill+stroke, 3=invisible
+                                          // Invisible (Tr 3) text was present but suppressed — reported to callers
+                                          // so an include_invisible retry is attempted only when it can recover.
+    let mut skipped_invisible = false;
     let mut line_width: f32 = 1.0;
     #[derive(Clone)]
     struct SavedGraphicsState {
@@ -494,6 +501,14 @@ pub(crate) fn extract_page_text_items(
                     // For Mixed/template PDFs, include_invisible=true extracts
                     // the OCR text layer that sits behind scanned images.
                     if text_rendering_mode == 3 && !include_invisible {
+                        if op
+                            .operands
+                            .first()
+                            .and_then(get_operand_bytes)
+                            .is_some_and(|raw| !raw.is_empty())
+                        {
+                            skipped_invisible = true;
+                        }
                         if let Some(w_ts) = w_ts_opt {
                             text_matrix[4] += w_ts * text_matrix[0];
                             text_matrix[5] += w_ts * text_matrix[1];
@@ -565,6 +580,16 @@ pub(crate) fn extract_page_text_items(
                 if in_text_block && !op.operands.is_empty() {
                     if let Ok(array) = op.operands[0].as_array() {
                         let font_info = font_widths.get(&current_font);
+                        // Numeric-only TJ arrays (pure kerning) show no
+                        // text — they must not trigger the invisible retry.
+                        if text_rendering_mode == 3
+                            && !include_invisible
+                            && array
+                                .iter()
+                                .any(|el| get_operand_bytes(el).is_some_and(|raw| !raw.is_empty()))
+                        {
+                            skipped_invisible = true;
+                        }
                         let is_invisible = (text_rendering_mode == 3 && !include_invisible)
                             || suppress_glyph_extraction;
                         // Capture first-glyph position for ActualText
@@ -770,6 +795,16 @@ pub(crate) fn extract_page_text_items(
                         )
                     })
                 });
+                if text_rendering_mode == 3
+                    && !include_invisible
+                    && op
+                        .operands
+                        .first()
+                        .and_then(get_operand_bytes)
+                        .is_some_and(|raw| !raw.is_empty())
+                {
+                    skipped_invisible = true;
+                }
                 if !((text_rendering_mode == 3 && !include_invisible)
                     || suppress_glyph_extraction
                     || op.operands.is_empty())
@@ -882,6 +917,7 @@ pub(crate) fn extract_page_text_items(
                                         &ctm,
                                         &mut cmap_decisions,
                                         style_cache,
+                                        form_budget,
                                     );
                                     items.extend(form_items);
                                 }
@@ -1251,6 +1287,12 @@ pub(crate) fn extract_page_text_items(
         }
     }
 
+    if form_budget.was_truncated() {
+        log::warn!(
+            "page {page_num}: Form XObject expansion truncated (invocation or operation budget reached); nested form text may be incomplete"
+        );
+    }
+
     // Underline detection reads only painted ink: `re` rects confirmed by
     // a paint operator plus filled-subpath rects — never clip-only rects,
     // which draw nothing.
@@ -1300,7 +1342,12 @@ pub(crate) fn extract_page_text_items(
 
     let items = super::merge_text_items(items);
     let items = super::merge_subscript_items(items);
-    Ok(((items, rects, lines), has_gid_fonts, coords_rotated))
+    Ok((
+        (items, rects, lines),
+        has_gid_fonts,
+        coords_rotated,
+        skipped_invisible,
+    ))
 }
 
 /// Counts of text operators with horizontal vs rotated combined matrices.
@@ -1498,13 +1545,14 @@ mod tests {
 
         let (doc, page_id) = simple_doc_with_content(content);
         let font_cmaps = FontCMaps::from_doc(&doc);
-        let ((items, _, _), _, _) = extract_page_text_items(
+        let ((items, _, _), _, _, _) = extract_page_text_items(
             &doc,
             page_id,
             1,
             &font_cmaps,
             false,
             &mut FontStyleCache::new(),
+            &mut FormWalkBudget::new(),
         )
         .unwrap();
         items
@@ -1734,9 +1782,10 @@ BT /F1 12 Tf 0 1 -1 0 240 100 Tm (WORLD) Tj ET
             &font_cmaps,
             false,
             &mut FontStyleCache::new(),
+            &mut FormWalkBudget::new(),
         )
         .unwrap();
-        let ((items, rects, lines), _has_gid, _coords_rotated) = result;
+        let ((items, rects, lines), _has_gid, _coords_rotated, _skipped_invisible) = result;
         assert!(items.is_empty());
         assert!(rects.is_empty());
         assert!(lines.is_empty());
@@ -1817,13 +1866,14 @@ BT 30 700 Tm <41> Tj ET";
         doc.trailer.set("Root", Object::Reference(catalog_id));
 
         let font_cmaps = FontCMaps::from_doc(&doc);
-        let ((items, _, _), _, _) = extract_page_text_items(
+        let ((items, _, _), _, _, _) = extract_page_text_items(
             &doc,
             page_id,
             1,
             &font_cmaps,
             false,
             &mut FontStyleCache::new(),
+            &mut FormWalkBudget::new(),
         )
         .unwrap();
         let text = items
@@ -1887,4 +1937,19 @@ BT 30 700 Tm <41> Tj ET";
         let output = strip_pdf_comments(input);
         assert_eq!(output, b"(x\\\\) Tj  \nET\n");
     }
+
+    #[test]
+    fn oversized_content_stream_skips_extraction() {
+        let mut content =
+            Vec::with_capacity((super::super::content_decode::MAX_PAGE_OPERATIONS + 1) * 2);
+        for _ in 0..=super::super::content_decode::MAX_PAGE_OPERATIONS {
+            content.extend_from_slice(b"q\n");
+        }
+        content.extend_from_slice(b"BT /F1 12 Tf 72 720 Td (Hello) Tj ET\n");
+        let items = extract_simple_items(&content);
+        assert!(
+            items.is_empty(),
+            "pages over the operator cap must not be decoded"
+        );
+    }
 }

src/extractor/fonts.rs

@@ -482,7 +482,11 @@ pub(crate) fn parse_cid_w_array(
     widths: &mut HashMap<u16, u16>,
 ) {
     let mut i = 0;
+    let mut assigned = 0usize;
     while i < w_array.len() {
+        if assigned >= crate::tounicode::MAX_CID_W_EXPANSION {
+            return;
+        }
         let start_cid = match &w_array[i] {
             Object::Integer(n) => *n as u16,
             Object::Real(n) => *n as u16,
@@ -501,12 +505,14 @@ pub(crate) fn parse_cid_w_array(
             Object::Array(arr) => {
                 // [c [w1 w2 ...]] — consecutive widths starting at c
                 for (j, w_obj) in arr.iter().enumerate() {
-                    let w = match w_obj {
-                        Object::Integer(n) => *n as u16,
-                        Object::Real(n) => *n as u16,
-                        _ => continue,
-                    };
-                    widths.insert(start_cid + j as u16, w);
+                    if !assign_cid_width(
+                        widths,
+                        start_cid.wrapping_add(j as u16),
+                        w_obj,
+                        &mut assigned,
+                    ) {
+                        return;
+                    }
                 }
                 i += 1;
             }
@@ -514,12 +520,14 @@ pub(crate) fn parse_cid_w_array(
                 // Could be a reference to an array
                 if let Ok(Object::Array(arr)) = doc.get_object(*r) {
                     for (j, w_obj) in arr.iter().enumerate() {
-                        let w = match w_obj {
-                            Object::Integer(n) => *n as u16,
-                            Object::Real(n) => *n as u16,
-                            _ => continue,
-                        };
-                        widths.insert(start_cid + j as u16, w);
+                        if !assign_cid_width(
+                            widths,
+                            start_cid.wrapping_add(j as u16),
+                            w_obj,
+                            &mut assigned,
+                        ) {
+                            return;
+                        }
                     }
                     i += 1;
                 } else {
@@ -542,8 +550,8 @@ pub(crate) fn parse_cid_w_array(
                         continue;
                     }
                 };
-                for cid in start_cid..=end {
-                    widths.insert(cid, w);
+                if !assign_cid_width_range(widths, start_cid, end, w, &mut assigned) {
+                    return;
                 }
                 i += 1;
             }
@@ -561,8 +569,8 @@ pub(crate) fn parse_cid_w_array(
                         continue;
                     }
                 };
-                for cid in start_cid..=end {
-                    widths.insert(cid, w);
+                if !assign_cid_width_range(widths, start_cid, end, w, &mut assigned) {
+                    return;
                 }
                 i += 1;
             }
@@ -573,6 +581,45 @@ pub(crate) fn parse_cid_w_array(
     }
 }
 
+fn assign_cid_width(
+    widths: &mut HashMap<u16, u16>,
+    cid: u16,
+    w_obj: &Object,
+    assigned: &mut usize,
+) -> bool {
+    let w = match w_obj {
+        Object::Integer(n) => *n as u16,
+        Object::Real(n) => *n as u16,
+        _ => return true,
+    };
+    if *assigned >= crate::tounicode::MAX_CID_W_EXPANSION {
+        return false;
+    }
+    widths.insert(cid, w);
+    *assigned += 1;
+    true
+}
+
+fn assign_cid_width_range(
+    widths: &mut HashMap<u16, u16>,
+    start: u16,
+    end: u16,
+    w: u16,
+    assigned: &mut usize,
+) -> bool {
+    if start > end {
+        return true;
+    }
+    for cid in start..=end {
+        if *assigned >= crate::tounicode::MAX_CID_W_EXPANSION {
+            return false;
+        }
+        widths.insert(cid, w);
+        *assigned += 1;
+    }
+    true
+}
+
 /// Compute the width of a string in text space units,
 /// given raw bytes and font width info.
 /// Returns width in text space units (font_units * units_scale * font_size).
@@ -2313,4 +2360,48 @@ end",
         // invalid CMap result — so it must not clear the gid flag.
         assert!(gid_flagged(Some("<01> <FFFD>\n<02> <FFFD>")));
     }
+
+    #[test]
+    fn parse_cid_w_array_range_and_consecutive() {
+        use super::parse_cid_w_array;
+        use lopdf::{Document, Object};
+        use std::collections::HashMap;
+
+        let doc = Document::new();
+        let mut widths = HashMap::new();
+        let w = vec![
+            Object::Integer(10),
+            Object::Integer(12),
+            Object::Integer(500),
+            Object::Integer(20),
+            Object::Array(vec![Object::Integer(100), Object::Integer(200)]),
+        ];
+        parse_cid_w_array(&doc, &w, &mut widths);
+        assert_eq!(widths.get(&10), Some(&500));
+        assert_eq!(widths.get(&11), Some(&500));
+        assert_eq!(widths.get(&12), Some(&500));
+        assert_eq!(widths.get(&20), Some(&100));
+        assert_eq!(widths.get(&21), Some(&200));
+    }
+
+    #[test]
+    fn parse_cid_w_array_repeated_full_ranges_stay_bounded() {
+        use super::parse_cid_w_array;
+        use crate::tounicode::MAX_CID_W_EXPANSION;
+        use lopdf::{Document, Object};
+        use std::collections::HashMap;
+
+        let doc = Document::new();
+        let mut widths = HashMap::new();
+        let mut w = Vec::new();
+        for _ in 0..5_000 {
+            w.push(Object::Integer(0));
+            w.push(Object::Integer(65535));
+            w.push(Object::Integer(500));
+        }
+        parse_cid_w_array(&doc, &w, &mut widths);
+        assert!(widths.len() <= MAX_CID_W_EXPANSION);
+        assert_eq!(widths.get(&0), Some(&500));
+        assert_eq!(widths.get(&65535), Some(&500));
+    }
 }

src/extractor/layout.rs

@@ -41,15 +41,110 @@ pub(crate) fn detect_columns(
     }
     debug!("page {}: detect_columns: {} items", page, page_items.len());
 
-    // Find page bounds
-    let x_min = page_items.iter().map(|i| i.x).fold(f32::INFINITY, f32::min);
-    let x_max = page_items
-        .iter()
-        .map(|i| i.x + effective_width(i))
-        .fold(f32::NEG_INFINITY, f32::max);
+    // The width of one ordinary page, used three ways below: as the largest
+    // credible width for a single text run, as the size of empty gap that marks
+    // content as detached, and as the span past which those checks run at all.
+    // This is a heuristic, not a format rule: PDF 2.0 sets no page-size limit,
+    // and since PDF 1.6 `UserUnit` scales a page's physical size independently
+    // of its coordinates. 14_400 units (200in at the default 1/72in unit) is
+    // the traditional Acrobat architectural limit, which makes it a reasonable
+    // "wider than any ordinary page" mark in coordinate space.
+    const MAX_PAGE_EXTENT: f32 = 14_400.0;
+    // A detached cluster is only dropped if it also holds a small minority of
+    // the items, so a genuine two-part layout keeps its full bounds even when
+    // the halves are far apart.
+    const MAX_TRIM_FRACTION: f32 = 0.10;
+
+    // Position and width of each item, skipping only non-finite geometry.
+    let finite_span = |i: &&TextItem| -> Option<(f32, f32)> {
+        let (left, width) = (i.x, effective_width(i));
+        (left.is_finite() && (left + width).is_finite()).then_some((left, width))
+    };
+
+    let (min_left, max_right, total) = page_items.iter().filter_map(finite_span).fold(
+        (f32::INFINITY, f32::NEG_INFINITY, 0usize),
+        |(lo, hi, n), (left, width)| (lo.min(left), hi.max(left + width), n + 1),
+    );
+
+    // No item had usable geometry, so there is no layout to report.
+    if total == 0 {
+        return vec![];
+    }
+
+    // Every threshold below (gutter margins, spanning-item width, the XY-cut
+    // margin) is a fraction of the page width, so a far item can set the scale
+    // for the whole page and shrink the effective detection window to a
+    // rounding error — real gutters then fall inside the margin band and a
+    // genuine multi-column page collapses to one region.
+    //
+    // Anything inside one page extent is ordinary, so the common case keeps the
+    // plain bounds and skips the work below entirely.
+    let (x_min, x_max) = if max_right - min_left <= MAX_PAGE_EXTENT {
+        (min_left, max_right)
+    } else {
+        // Discarding content needs positive evidence that it is not part of the
+        // layout, because a count-based rule alone cannot tell a stray from a
+        // sparse far sidebar. The evidence is geometric: positions are grouped
+        // into clusters separated by more than a whole page of continuous
+        // emptiness. Real content, however sparse, does not leave a void that
+        // large; a malformed coordinate sits alone beyond one.
+        let mut spans: Vec<(f32, f32)> = page_items.iter().filter_map(finite_span).collect();
+        spans.sort_by(|a, b| a.0.total_cmp(&b.0));
+
+        let mut core: Option<std::ops::Range<usize>> = None;
+        let mut start = 0usize;
+        for i in 1..=spans.len() {
+            if i < spans.len() && spans[i].0 - spans[i - 1].0 <= MAX_PAGE_EXTENT {
+                continue;
+            }
+            if core.as_ref().is_none_or(|best| i - start > best.len()) {
+                core = Some(start..i);
+            }
+            start = i;
+        }
+        let mut core = core.unwrap_or(0..spans.len());
+
+        // Only drop the detached clusters when they are a small minority, so a
+        // genuine two-part layout keeps its full bounds.
+        let dropped = spans.len() - core.len();
+        if dropped as f32 > spans.len() as f32 * MAX_TRIM_FRACTION {
+            core = 0..spans.len();
+        }
+        let core = &spans[core];
+
+        // Positions cannot be inflated by a bogus width, so the spread of the
+        // content is a sound scale for judging one. A run much wider than the
+        // page's own content is a malformed width — the test is relative, so a
+        // genuinely large page keeps its genuinely long runs.
+        let (lo, widest_left) = (core[0].0, core[core.len() - 1].0);
+        let max_run_width = (widest_left - lo) + MAX_PAGE_EXTENT;
+        let hi = core
+            .iter()
+            .filter(|&&(_, width)| width <= max_run_width)
+            .map(|&(left, width)| left + width)
+            .fold(widest_left, f32::max);
+
+        if lo != min_left || hi != max_right {
+            debug!(
+                "page {page}: bounds {min_left}..{max_right} exceed one page; \
+                 dropped {dropped}/{} detached item(s), using {lo}..{hi}",
+                spans.len()
+            );
+        }
+        (lo, hi)
+    };
+
+    // Hard ceiling on the histogram size, independent of the trimming above:
+    // the bounds are attacker-influenced, so an unclamped
+    // `page_width / BIN_WIDTH` lets a crafted PDF force an arbitrarily large
+    // `vec![0u32; num_bins]` allocation. 65_536 bins covers ~128k points at
+    // BIN_WIDTH 2.0 — roughly 9x the largest legal page — so this never binds
+    // on a real layout. Kept as a bound that does not depend on the outlier
+    // heuristic staying correct.
+    const MAX_BINS: usize = 65_536;
 
     let page_width = x_max - x_min;
-    if page_width < 200.0 {
+    if !page_width.is_finite() || page_width < 200.0 {
         return vec![ColumnRegion { x_min, x_max }];
     }
 
@@ -57,13 +152,20 @@ pub(crate) fn detect_columns(
         return vec![ColumnRegion { x_min, x_max }];
     }
 
+    // Widen the bins rather than dropping the tail of the page. Clamping the
+    // count alone would leave anything past MAX_BINS * BIN_WIDTH outside the
+    // histogram, folded into the last bin, which places gutters at the wrong
+    // coordinates. Scaling keeps full coverage under the same allocation
+    // ceiling; only the resolution degrades, and only beyond ~131k points.
+    let bin_width = BIN_WIDTH.max(page_width / MAX_BINS as f32);
+
     // Build occupancy histogram.
     // Exclude items wider than 60% of page width — these are spanning items
     // (titles, full-width paragraphs) that would fill the gutter and prevent
     // detection of partial-page column layouts (e.g. two-column abstracts on
     // a page that also has single-column introduction text).
     let wide_threshold = page_width * 0.6;
-    let num_bins = ((page_width / BIN_WIDTH).ceil() as usize).max(1);
+    let num_bins = ((page_width / bin_width).ceil() as usize).clamp(1, MAX_BINS);
     let mut histogram = vec![0u32; num_bins];
 
     for item in &page_items {
@@ -71,8 +173,8 @@ pub(crate) fn detect_columns(
         if w > wide_threshold {
             continue;
         }
-        let left = ((item.x - x_min) / BIN_WIDTH).floor() as usize;
-        let right = (((item.x + w) - x_min) / BIN_WIDTH).ceil() as usize;
+        let left = ((item.x - x_min) / bin_width).floor() as usize;
+        let right = (((item.x + w) - x_min) / bin_width).ceil() as usize;
         let left = left.min(num_bins);
         let right = right.min(num_bins);
         for count in histogram.iter_mut().take(right).skip(left) {
@@ -109,12 +211,12 @@ pub(crate) fn detect_columns(
     let valleys: Vec<(usize, usize)> = valleys
         .into_iter()
         .filter(|&(start, end)| {
-            let width_pts = (end - start) as f32 * BIN_WIDTH;
+            let width_pts = (end - start) as f32 * bin_width;
             if width_pts < MIN_GUTTER_WIDTH {
                 return false;
             }
             // Valley center must not be within 5% of page edges
-            let center_pts = ((start + end) as f32 / 2.0) * BIN_WIDTH;
+            let center_pts = ((start + end) as f32 / 2.0) * bin_width;
             center_pts > margin_threshold && center_pts < (page_width - margin_threshold)
         })
         .collect();
@@ -132,7 +234,7 @@ pub(crate) fn detect_columns(
             &histogram,
             num_bins,
             x_min,
-            BIN_WIDTH,
+            bin_width,
             page_width,
             margin_threshold,
         );
@@ -141,7 +243,7 @@ pub(crate) fn detect_columns(
                 &rel_valleys,
                 &page_items,
                 x_min,
-                BIN_WIDTH,
+                bin_width,
                 x_max,
                 MIN_ITEMS_PER_COLUMN,
                 MIN_VERTICAL_SPAN_RATIO,
@@ -182,7 +284,7 @@ pub(crate) fn detect_columns(
         &valleys,
         &page_items,
         x_min,
-        BIN_WIDTH,
+        bin_width,
         x_max,
         MIN_ITEMS_PER_COLUMN,
         MIN_VERTICAL_SPAN_RATIO,
@@ -196,7 +298,7 @@ pub(crate) fn detect_columns(
         &valleys,
         &page_items,
         x_min,
-        BIN_WIDTH,
+        bin_width,
         x_max,
         MIN_ITEMS_PER_COLUMN,
         MIN_VERTICAL_SPAN_RATIO,
@@ -1827,7 +1929,7 @@ fn split_column_stragglers(lines: Vec<TextLine>) -> (Vec<TextLine>, Vec<TextLine
         .unwrap();
 
     let (cs, ce) = segments[core_seg];
-    let mut core = Vec::with_capacity(ce - cs);
+    let mut core = Vec::with_capacity(ce.saturating_sub(cs));
     let mut stragglers = Vec::new();
     for (i, line) in lines.into_iter().enumerate() {
         if i >= cs && i < ce {
@@ -2533,6 +2635,214 @@ mod tests {
         );
     }
 
+    #[test]
+    fn extreme_far_coordinate_does_not_allocate_unboundedly() {
+        // A crafted PDF can place a text run at an arbitrary coordinate via the
+        // text matrix. The derived page width must not drive an unbounded
+        // histogram allocation (previously `page_width / BIN_WIDTH` bins with no
+        // upper bound would try to reserve terabytes and abort the process).
+        let mut items = Vec::new();
+        for i in 0..24 {
+            items.push(make_item(1, i as f32 * 10.0, 700.0 - i as f32 * 5.0, "A"));
+        }
+        // Item placed 1e12 points away — 5e11 bins if left unclamped.
+        items.push(make_item(1, 1e12, 700.0, "Z"));
+
+        // Must return without aborting; content is preserved as a single region.
+        let cols = detect_columns(&items, 1, false);
+        assert!(!cols.is_empty());
+    }
+
+    #[test]
+    fn non_finite_coordinates_never_leak_into_region_bounds() {
+        // An inf/NaN coordinate must not escape as a column boundary: callers
+        // treat these as page/column edges.
+        for bad_x in [f32::INFINITY, f32::NEG_INFINITY, f32::NAN] {
+            let mut items = Vec::new();
+            for i in 0..24 {
+                items.push(make_item(1, i as f32 * 10.0, 700.0 - i as f32 * 5.0, "A"));
+            }
+            items.push(make_item(1, bad_x, 700.0, "Z"));
+
+            for col in detect_columns(&items, 1, false) {
+                assert!(
+                    col.x_min.is_finite() && col.x_max.is_finite(),
+                    "bad_x {bad_x} leaked bounds {}..{}",
+                    col.x_min,
+                    col.x_max
+                );
+            }
+        }
+    }
+
+    #[test]
+    fn all_non_finite_coordinates_yield_no_columns() {
+        let items: Vec<TextItem> = (0..24)
+            .map(|i| make_item(1, f32::NAN, 700.0 - i as f32 * 5.0, "A"))
+            .collect();
+
+        assert!(detect_columns(&items, 1, false).is_empty());
+    }
+
+    #[test]
+    fn one_bad_item_does_not_disable_column_detection() {
+        // A single stray item should not collapse a clean two-column page to
+        // one region. Every gutter threshold is a fraction of the page width,
+        // so an untrimmed outlier pushes real gutters inside the rejected
+        // margin band. A malformed *width* at an ordinary position poisons the
+        // bounds just as a malformed position does.
+        for (label, bad_x, bad_width) in [
+            ("nan position", f32::NAN, 0.0),
+            ("inf position", f32::INFINITY, 0.0),
+            ("far position", 50_000.0, 0.0),
+            ("very far position", 1e12, 0.0),
+            ("huge width", 100.0, 1e12),
+            ("inf width", 100.0, f32::INFINITY),
+        ] {
+            let mut items = Vec::new();
+            items.extend(fill_zone(1, 30.0, 280.0, 750.0, 50.0));
+            items.extend(fill_zone(1, 320.0, 570.0, 750.0, 50.0));
+            let mut bad = make_item(1, bad_x, 400.0, "Z");
+            bad.width = bad_width;
+            items.push(bad);
+
+            let cols = detect_columns(&items, 1, false);
+            assert_eq!(
+                cols.len(),
+                2,
+                "{label}: expected 2 columns, got {}",
+                cols.len()
+            );
+            for col in &cols {
+                assert!(
+                    col.x_max - col.x_min <= MAX_PAGE_EXTENT_FOR_TEST,
+                    "{label}: region {}..{} exceeds one page",
+                    col.x_min,
+                    col.x_max
+                );
+            }
+        }
+    }
+
+    /// Mirrors `MAX_PAGE_EXTENT` in `detect_columns`.
+    const MAX_PAGE_EXTENT_FOR_TEST: f32 = 14_400.0;
+
+    #[test]
+    fn very_wide_page_keeps_full_histogram_coverage() {
+        // Beyond MAX_BINS * BIN_WIDTH (~131k points) the bins must widen rather
+        // than stop covering the page. Three zones: the first gutter is inside
+        // the old coverage limit, the second is past it. Because the first
+        // gutter is found, the XY-cut fallback never runs, so a truncated
+        // histogram silently reports two columns instead of three.
+        let mut items = Vec::new();
+        items.extend(fill_zone(1, 0.0, 60_000.0, 750.0, 700.0));
+        items.extend(fill_zone(1, 70_000.0, 140_000.0, 750.0, 700.0));
+        items.extend(fill_zone(1, 160_000.0, 200_000.0, 750.0, 700.0));
+
+        let cols = detect_columns(&items, 1, false);
+        assert_eq!(
+            cols.len(),
+            3,
+            "Expected 3 columns across a 200k-wide page, got {}",
+            cols.len()
+        );
+        assert!(
+            (140_000.0..=160_000.0).contains(&cols[1].x_max),
+            "second gutter at {}, expected inside the real 140k..160k gap",
+            cols[1].x_max
+        );
+    }
+
+    #[test]
+    fn large_page_with_legitimately_long_runs_is_kept() {
+        // On a very large page, individual runs can exceed one ordinary page's
+        // width. They are real content, so they must not be judged malformed:
+        // the page keeps its columns and its full right edge.
+        let mut items = Vec::new();
+        for row in 0..30 {
+            let y = 750.0 - row as f32 * 14.0;
+            let mut left = make_item(1, 0.0, y, "Left run");
+            left.width = 20_000.0;
+            let mut right = make_item(1, 25_000.0, y, "Right run");
+            right.width = 20_000.0;
+            items.extend([left, right]);
+        }
+
+        let cols = detect_columns(&items, 1, false);
+        assert!(
+            !cols.is_empty(),
+            "a page of long-but-valid runs must still report a layout"
+        );
+        let right_edge = cols
+            .iter()
+            .map(|c| c.x_max)
+            .fold(f32::NEG_INFINITY, f32::max);
+        assert!(
+            right_edge > 44_000.0,
+            "long runs were treated as malformed: right edge {right_edge}, expected ~45_000"
+        );
+    }
+
+    #[test]
+    fn sparse_far_sidebar_on_a_large_page_is_kept() {
+        // A large-format page with a thin, sparsely-populated sidebar far from
+        // the main block. The sidebar is a small minority of the items, so an
+        // item-count rule alone would discard it — but nothing about its
+        // geometry says it is invalid, so its bounds must survive.
+        let mut items = Vec::new();
+        items.extend(fill_zone(1, 0.0, 12_000.0, 750.0, 500.0));
+        for i in 0..12 {
+            items.push(make_item(1, 24_000.0, 750.0 - i as f32 * 14.0, "Sidebar"));
+        }
+
+        let cols = detect_columns(&items, 1, false);
+        let right_edge = cols
+            .iter()
+            .map(|c| c.x_max)
+            .fold(f32::NEG_INFINITY, f32::max);
+        assert!(
+            right_edge > 24_000.0,
+            "sidebar was trimmed away: right edge {right_edge}, expected >24_000"
+        );
+    }
+
+    #[test]
+    fn genuinely_wide_layout_keeps_its_true_bounds() {
+        // A large-format page whose content really is spread beyond one
+        // ordinary page must not be trimmed to the median cluster: its far
+        // items are the majority, not strays.
+        let mut items = Vec::new();
+        items.extend(fill_zone(1, 100.0, 20_000.0, 750.0, 600.0));
+        items.extend(fill_zone(1, 22_000.0, 40_000.0, 750.0, 600.0));
+
+        let cols = detect_columns(&items, 1, false);
+        let widest = cols
+            .iter()
+            .map(|c| c.x_max)
+            .fold(f32::NEG_INFINITY, f32::max);
+        assert!(
+            widest > 35_000.0,
+            "wide layout was trimmed: right edge {widest}, expected ~40_000"
+        );
+    }
+
+    #[test]
+    fn oversized_but_legal_page_is_not_trimmed() {
+        // A wide-format page well inside the 14_400pt spec limit must keep its
+        // real bounds — outlier trimming is only for spans beyond a legal page.
+        let mut items = Vec::new();
+        items.extend(fill_zone(1, 100.0, 4_000.0, 750.0, 400.0));
+        items.extend(fill_zone(1, 4_400.0, 8_000.0, 750.0, 400.0));
+
+        let cols = detect_columns(&items, 1, false);
+        assert_eq!(cols.len(), 2, "Expected 2 columns, got {}", cols.len());
+        assert!(
+            cols[1].x_max > 7_000.0,
+            "right column should keep its true extent, got {}",
+            cols[1].x_max
+        );
+    }
+
     #[test]
     fn two_column_regression_guard() {
         // Standard 2-column layout with clear gutter at center

src/extractor/links.rs

@@ -2,11 +2,51 @@
 
 use crate::types::{ItemType, TextItem};
 use lopdf::{Document, Object, ObjectId};
-use std::collections::HashMap;
+use std::collections::{HashMap, HashSet};
 
 use super::fonts::{resolve_array, resolve_dict};
 use super::get_number;
 
+/// Upper bound on the number of form-field nodes visited during a single
+/// `extract_form_fields` pass. A crafted PDF can chain thousands of distinct
+/// `/Kids` fields to blow the stack even without an outright reference cycle,
+/// so we cap total traversal work in addition to detecting cycles.
+const MAX_FORM_FIELD_NODES: usize = 100_000;
+
+/// Upper bound on `/Kids` recursion depth. Real AcroForm hierarchies are only
+/// a few levels deep (fields → child fields → widgets); a crafted PDF can chain
+/// tens of thousands of distinct fields into a linear `/Kids` list that would
+/// overflow the stack via depth-first recursion long before the node budget is
+/// reached. This depth cap bounds the stack independently of total node count.
+const MAX_FORM_FIELD_DEPTH: usize = 100;
+
+/// Traversal budget for the AcroForm field walk. Bounds both the number of
+/// distinct nodes visited *and* the total number of `/Fields`/`/Kids` entries
+/// examined.
+///
+/// Counting `visited` alone is not enough: invalid entries (non-references) and
+/// duplicate references never grow `visited`, so an oversized array full of them
+/// would iterate to completion no matter how large. Charging every examined
+/// entry against the same budget makes it a real cap on traversal work.
+pub(crate) struct FieldWalkBudget {
+    visited: HashSet<ObjectId>,
+    examined: usize,
+}
+
+impl FieldWalkBudget {
+    fn new() -> Self {
+        Self {
+            visited: HashSet::new(),
+            examined: 0,
+        }
+    }
+
+    /// True once the budget is spent; callers must stop iterating and recursing.
+    fn exhausted(&self) -> bool {
+        self.visited.len() >= MAX_FORM_FIELD_NODES || self.examined >= MAX_FORM_FIELD_NODES
+    }
+}
+
 pub fn extract_page_links(doc: &Document, page_id: ObjectId, page_num: u32) -> Vec<TextItem> {
     let mut links = Vec::new();
 
@@ -146,9 +186,12 @@ pub(crate) fn extract_form_fields(
         Err(_) => return items,
     };
 
+    // Borrow the array rather than cloning it: a crafted `/Fields` can be huge,
+    // and cloning would pay an O(n) allocation/copy before the budget check
+    // below can stop the work.
     let fields = match acroform.get(b"Fields") {
         Ok(obj) => match resolve_array(doc, obj) {
-            Some(arr) => arr.clone(),
+            Some(arr) => arr,
             None => return items,
         },
         Err(_) => return items,
@@ -158,7 +201,19 @@ pub(crate) fn extract_form_fields(
     }
     let annotation_pages = annotation_page_map(doc, page_map);
 
-    for field_obj in &fields {
+    // Bound the walk so a crafted PDF cannot send us into unbounded recursion
+    // via a `/Kids` cycle, a deep chain, or an oversized array of invalid or
+    // duplicate entries.
+    let mut budget = FieldWalkBudget::new();
+
+    for field_obj in fields {
+        // Stop once the budget is spent so a `/Fields` array wider than the
+        // budget can't burn CPU iterating entries whose walk would no-op. Charge
+        // every entry (including invalid ones) against the budget.
+        if budget.exhausted() {
+            break;
+        }
+        budget.examined += 1;
         if let Ok(field_ref) = field_obj.as_reference() {
             walk_form_fields(
                 doc,
@@ -168,6 +223,8 @@ pub(crate) fn extract_form_fields(
                 page_map,
                 &annotation_pages,
                 &mut items,
+                &mut budget,
+                0,
             );
         }
     }
@@ -202,6 +259,7 @@ fn annotation_page_map(
 }
 
 /// Recursively walk the form field tree, extracting leaf field values.
+#[allow(clippy::too_many_arguments)]
 pub(crate) fn walk_form_fields(
     doc: &Document,
     field_id: ObjectId,
@@ -210,7 +268,22 @@ pub(crate) fn walk_form_fields(
     page_map: &HashMap<ObjectId, u32>,
     annotation_pages: &HashMap<ObjectId, u32>,
     items: &mut Vec<TextItem>,
+    budget: &mut FieldWalkBudget,
+    depth: usize,
 ) {
+    // Guard against `/Kids` cycles and pathologically large field trees.
+    // Exceeding the depth cap means the chain is too deep to be a legitimate
+    // form (and would overflow the stack); an exhausted budget means the tree is
+    // too large. Both checks run *before* inserting so the visited set can never
+    // grow past the budget.
+    if depth > MAX_FORM_FIELD_DEPTH || budget.exhausted() {
+        return;
+    }
+    // Revisiting an object ID means we hit a `/Kids` cycle.
+    if !budget.visited.insert(field_id) {
+        return;
+    }
+
     let field_dict = match doc.get_dictionary(field_id) {
         Ok(d) => d,
         Err(_) => return,
@@ -241,9 +314,19 @@ pub(crate) fn walk_form_fields(
 
     // Check for /Kids — if present, recurse into children
     if let Ok(kids_obj) = field_dict.get(b"Kids") {
+        // Iterate the borrowed array directly — cloning a crafted, oversized
+        // `/Kids` would allocate and copy every entry before the budget check
+        // below could stop the work.
         if let Some(kids) = resolve_array(doc, kids_obj) {
-            let kids = kids.clone();
-            for kid in &kids {
+            for kid in kids {
+                // Stop once the budget is spent so a `/Kids` array wider than the
+                // budget can't burn CPU iterating entries whose walk would no-op.
+                // Charge every entry (including invalid/duplicate ones) against
+                // the budget so this is a true traversal-work cap.
+                if budget.exhausted() {
+                    break;
+                }
+                budget.examined += 1;
                 if let Ok(kid_ref) = kid.as_reference() {
                     walk_form_fields(
                         doc,
@@ -253,6 +336,8 @@ pub(crate) fn walk_form_fields(
                         page_map,
                         annotation_pages,
                         items,
+                        budget,
+                        depth + 1,
                     );
                 }
             }
@@ -411,4 +496,225 @@ mod tests {
         assert_eq!(items[0].page, 2);
         assert_eq!(items[0].text, "customer: Alice");
     }
+
+    #[test]
+    fn kids_self_cycle_does_not_overflow_stack() {
+        // A crafted AcroForm field that lists itself in `/Kids` must not send
+        // the traversal into unbounded recursion.
+        let mut doc = Document::new();
+        let field_id = doc.new_object_id();
+        doc.set_object(
+            field_id,
+            dictionary! {
+                "FT" => "Tx",
+                "T" => Object::string_literal("loop"),
+                "Kids" => vec![Object::Reference(field_id)],
+            },
+        );
+        let catalog_id = doc.add_object(dictionary! {
+            "Type" => "Catalog",
+            "AcroForm" => dictionary! {
+                "Fields" => vec![Object::Reference(field_id)],
+            },
+        });
+        doc.trailer.set("Root", Object::Reference(catalog_id));
+
+        let page_map = HashMap::new();
+        // Completes (rather than overflowing the stack) and yields no items.
+        let items = extract_form_fields(&doc, &page_map);
+        assert!(items.is_empty());
+    }
+
+    #[test]
+    fn kids_mutual_cycle_terminates() {
+        // Two fields that reference each other via `/Kids` form a cycle that
+        // must also terminate.
+        let mut doc = Document::new();
+        let field_a = doc.new_object_id();
+        let field_b = doc.new_object_id();
+        doc.set_object(
+            field_a,
+            dictionary! {
+                "T" => Object::string_literal("a"),
+                "Kids" => vec![Object::Reference(field_b)],
+            },
+        );
+        doc.set_object(
+            field_b,
+            dictionary! {
+                "T" => Object::string_literal("b"),
+                "Kids" => vec![Object::Reference(field_a)],
+            },
+        );
+        let catalog_id = doc.add_object(dictionary! {
+            "Type" => "Catalog",
+            "AcroForm" => dictionary! {
+                "Fields" => vec![Object::Reference(field_a)],
+            },
+        });
+        doc.trailer.set("Root", Object::Reference(catalog_id));
+
+        let page_map = HashMap::new();
+        let items = extract_form_fields(&doc, &page_map);
+        assert!(items.is_empty());
+    }
+
+    #[test]
+    fn deep_acyclic_kids_chain_does_not_overflow_stack() {
+        // A long chain of *distinct* fields (no cycle) must also terminate:
+        // the visited set alone would still recurse to the chain length, so
+        // the depth cap is what prevents a stack overflow here.
+        let mut doc = Document::new();
+        let n = MAX_FORM_FIELD_DEPTH * 500;
+        let ids: Vec<ObjectId> = (0..=n).map(|_| doc.new_object_id()).collect();
+        for i in 0..n {
+            doc.set_object(
+                ids[i],
+                dictionary! {
+                    "FT" => "Tx",
+                    "Kids" => vec![Object::Reference(ids[i + 1])],
+                },
+            );
+        }
+        // Leaf carries a value; it sits far below the depth cap so it is never
+        // reached, proving traversal stops early rather than crashing.
+        doc.set_object(
+            ids[n],
+            dictionary! {
+                "FT" => "Tx",
+                "T" => Object::string_literal("leaf"),
+                "V" => Object::string_literal("x"),
+                "Rect" => vec![10.into(), 20.into(), 110.into(), 40.into()],
+            },
+        );
+        let catalog_id = doc.add_object(dictionary! {
+            "Type" => "Catalog",
+            "AcroForm" => dictionary! {
+                "Fields" => vec![Object::Reference(ids[0])],
+            },
+        });
+        doc.trailer.set("Root", Object::Reference(catalog_id));
+
+        let page_map = HashMap::new();
+        let items = extract_form_fields(&doc, &page_map);
+        assert!(items.is_empty());
+    }
+
+    #[test]
+    fn wide_tree_traversal_stops_at_node_budget() {
+        // A single field with a `/Kids` array wider than the node budget must
+        // stop traversal at the cap rather than growing `visited` (and the work)
+        // without bound. Each processed leaf emits one item, so the item count
+        // is bounded by the budget and reaches right up to it (a couple of
+        // slots go to the root and the boundary node charged against the cap).
+        let mut doc = Document::new();
+        let fanout = MAX_FORM_FIELD_NODES + 50;
+        let leaf_ids: Vec<ObjectId> = (0..fanout).map(|_| doc.new_object_id()).collect();
+        for &leaf in &leaf_ids {
+            doc.set_object(
+                leaf,
+                dictionary! {
+                    "FT" => "Tx",
+                    "V" => Object::string_literal("v"),
+                    "Rect" => vec![10.into(), 20.into(), 110.into(), 40.into()],
+                },
+            );
+        }
+        let kids: Vec<Object> = leaf_ids.iter().map(|&id| Object::Reference(id)).collect();
+        let root_id = doc.add_object(dictionary! {
+            "T" => Object::string_literal("root"),
+            "Kids" => kids,
+        });
+        let catalog_id = doc.add_object(dictionary! {
+            "Type" => "Catalog",
+            "AcroForm" => dictionary! {
+                "Fields" => vec![Object::Reference(root_id)],
+            },
+        });
+        doc.trailer.set("Root", Object::Reference(catalog_id));
+
+        let page_map = HashMap::new();
+        let items = extract_form_fields(&doc, &page_map);
+        // Extraction stops at the budget: bounded above by the cap, and it gets
+        // right up to it (allowing a small delta for the root/boundary nodes
+        // charged against the budget).
+        assert!(items.len() <= MAX_FORM_FIELD_NODES);
+        assert!(items.len() >= MAX_FORM_FIELD_NODES - 3);
+    }
+
+    #[test]
+    fn wide_top_level_fields_stop_at_node_budget() {
+        // A top-level `/Fields` array wider than the budget must also stop at
+        // the cap: the item count is bounded by the budget and reaches right up
+        // to it.
+        let mut doc = Document::new();
+        let fanout = MAX_FORM_FIELD_NODES + 50;
+        let leaf_ids: Vec<ObjectId> = (0..fanout).map(|_| doc.new_object_id()).collect();
+        for &leaf in &leaf_ids {
+            doc.set_object(
+                leaf,
+                dictionary! {
+                    "FT" => "Tx",
+                    "V" => Object::string_literal("v"),
+                    "Rect" => vec![10.into(), 20.into(), 110.into(), 40.into()],
+                },
+            );
+        }
+        let fields: Vec<Object> = leaf_ids.iter().map(|&id| Object::Reference(id)).collect();
+        let catalog_id = doc.add_object(dictionary! {
+            "Type" => "Catalog",
+            "AcroForm" => dictionary! {
+                "Fields" => fields,
+            },
+        });
+        doc.trailer.set("Root", Object::Reference(catalog_id));
+
+        let page_map = HashMap::new();
+        let items = extract_form_fields(&doc, &page_map);
+        assert!(items.len() <= MAX_FORM_FIELD_NODES);
+        assert!(items.len() >= MAX_FORM_FIELD_NODES - 3);
+    }
+
+    #[test]
+    fn duplicate_and_invalid_kids_entries_stop_at_budget() {
+        // Duplicate references and non-reference junk never grow `visited`, so
+        // without charging examined entries against the budget an oversized
+        // array of them would iterate to completion. The walk must still
+        // terminate and extract the single real leaf exactly once.
+        let mut doc = Document::new();
+        let leaf_id = doc.new_object_id();
+        doc.set_object(
+            leaf_id,
+            dictionary! {
+                "FT" => "Tx",
+                "V" => Object::string_literal("v"),
+                "Rect" => vec![10.into(), 20.into(), 110.into(), 40.into()],
+            },
+        );
+        // A `/Kids` array far wider than the budget: half duplicate references
+        // to the same leaf, half invalid (null) entries.
+        let mut kids: Vec<Object> = Vec::new();
+        for i in 0..(MAX_FORM_FIELD_NODES * 2) {
+            if i % 2 == 0 {
+                kids.push(Object::Reference(leaf_id));
+            } else {
+                kids.push(Object::Null);
+            }
+        }
+        let root_id = doc.add_object(dictionary! {
+            "T" => Object::string_literal("root"),
+            "Kids" => kids,
+        });
+        let catalog_id = doc.add_object(dictionary! {
+            "Type" => "Catalog",
+            "AcroForm" => dictionary! {
+                "Fields" => vec![Object::Reference(root_id)],
+            },
+        });
+        doc.trailer.set("Root", Object::Reference(catalog_id));
+
+        let page_map = HashMap::new();
+        let items = extract_form_fields(&doc, &page_map);
+        assert_eq!(items.len(), 1);
+    }
 }

src/extractor/mod.rs

@@ -3,6 +3,7 @@
 //! This module extracts text with position information for structure detection.
 
 mod base14;
+mod content_decode;
 pub(crate) mod content_stream;
 mod fonts;
 mod layout;
@@ -37,6 +38,7 @@ pub(crate) use layout::group_prefiltered_items_into_lines_with_thresholds_and_re
 pub(crate) use layout::is_newspaper_layout;
 pub(crate) use layout::ColumnRegion;
 pub use layout::{group_into_lines, group_into_lines_preserving_all_text};
+pub(crate) use xobjects::FormWalkBudget;
 
 // ---------------------------------------------------------------------------
 // Public API
@@ -282,18 +284,22 @@ fn extract_positioned_text_impl(
             font_cmaps,
             include_invisible,
             &mut style_cache,
+            &mut FormWalkBudget::new(),
         );
-        let ((mut items, mut rects, mut lines), has_gid_fonts, coords_rotated) = match page_result {
-            Ok(extraction) => extraction,
-            Err(error) if required_pages.is_some_and(|required| !required.contains(page_num)) => {
-                debug!(
-                    "page {}: skipping context-only extraction error: {}",
-                    page_num, error
-                );
-                continue;
-            }
-            Err(error) => return Err(error),
-        };
+        let ((mut items, mut rects, mut lines), has_gid_fonts, coords_rotated, _skipped_invisible) =
+            match page_result {
+                Ok(extraction) => extraction,
+                Err(error)
+                    if required_pages.is_some_and(|required| !required.contains(page_num)) =>
+                {
+                    debug!(
+                        "page {}: skipping context-only extraction error: {}",
+                        page_num, error
+                    );
+                    continue;
+                }
+                Err(error) => return Err(error),
+            };
         // Clip to the visible page box: single-page extracts and imposed
         // spreads keep neighboring pages' content in the stream, positioned
         // outside the CropBox. Extracting it interleaves invisible text into
@@ -884,6 +890,92 @@ fn tracked_run_space_floor(group: &[&TextItem], start: usize) -> Option<(usize,
     Some((end, floor * fs))
 }
 
+/// Fractional font-size band within which `merge_text_items` treats two runs as
+/// the same size. Shared with `is_small_caps_continuation`, which exists only to
+/// rescue junctions this band would otherwise break.
+const MERGE_FONT_SIZE_BAND: f32 = 0.20;
+
+/// Detect a small-caps continuation: typesetters render small caps as a
+/// full-size capital immediately followed by shrunken capitals in the same
+/// font (`(R) Tj` at 9.98pt, then `(OLANDO) Tj` at 6.74pt). Those runs are one
+/// word, but the font-size band in `merge_text_items` would split them,
+/// leaving "R" and "OLANDO" as separate items — which then read as separate
+/// table columns, since column boundaries cluster on item start positions.
+///
+/// Gated tightly so it cannot absorb the other reasons a smaller run follows a
+/// larger one:
+///   - runs the size band already accepts — excluded by requiring the junction
+///     to *cross* the band, so within-band pairs keep the normal word-spacing
+///     logic instead of having their space suppressed
+///   - superscripts / footnote markers — excluded by requiring an uppercase
+///     *letter* on both sides, so digits never qualify
+///   - drop caps — excluded because the body text that follows is mixed case
+///   - adjacent table cells or separate words — excluded by requiring the runs
+///     to be visually contiguous (essentially no gap)
+fn is_small_caps_continuation(
+    text_so_far: &str,
+    first: &TextItem,
+    next: &TextItem,
+    gap: f32,
+) -> bool {
+    // Must shrink. Real small caps sit near 0.7-0.8 of the full cap height;
+    // anything smaller is a superscript or a different run entirely.
+    if first.font_size <= 0.0 || next.font_size >= first.font_size {
+        return false;
+    }
+    // Only rescue junctions the size band would have broken. Within-band pairs
+    // merge on their own, and suppressing their space would swallow real word
+    // gaps between two similarly-sized uppercase words.
+    if (next.font_size - first.font_size).abs() <= first.font_size * MERGE_FONT_SIZE_BAND {
+        return false;
+    }
+    if next.font_size / first.font_size < 0.55 {
+        return false;
+    }
+    // Visually contiguous: the capital and its small caps touch. A real word
+    // space or a column gap disqualifies.
+    if !(-first.font_size * 0.2..=first.font_size * 0.15).contains(&gap) {
+        return false;
+    }
+    // The continuation must be all-uppercase letters (digits and lowercase
+    // both disqualify), and must contain at least one letter.
+    let mut saw_letter = false;
+    for ch in next.text.chars() {
+        if ch.is_alphabetic() {
+            saw_letter = true;
+            if !ch.is_uppercase() {
+                return false;
+            }
+        } else if ch.is_numeric() {
+            return false;
+        }
+    }
+    if !saw_letter {
+        return false;
+    }
+    // What we are continuing must itself end in a capital. Check the actual
+    // trailing character rather than skipping back to the nearest letter: after
+    // "ANGELA M. MAZZARELLI1" the run to continue is the footnote marker, not
+    // the "I" before it.
+    let trimmed = text_so_far.trim_end();
+    if trimmed.chars().last().is_some_and(|c| c.is_numeric()) {
+        // One legitimate exception: an ordinal suffix set as a smaller run,
+        // e.g. "JULY 4" + "TH". Only the four English suffixes qualify —
+        // anything else after a digit is a footnote marker or numeric suffix.
+        return matches!(trimmed_suffix(next), "TH" | "ST" | "ND" | "RD");
+    }
+    trimmed
+        .chars()
+        .rev()
+        .find(|c| c.is_alphabetic())
+        .is_some_and(|c| c.is_uppercase())
+}
+
+/// The continuation run's text, trimmed — used to spot ordinal suffixes.
+fn trimmed_suffix(next: &TextItem) -> &str {
+    next.text.trim()
+}
+
 pub(crate) fn merge_text_items(items: Vec<TextItem>) -> Vec<TextItem> {
     if items.is_empty() {
         return items;
@@ -942,8 +1034,16 @@ pub(crate) fn merge_text_items(items: Vec<TextItem>) -> Vec<TextItem> {
             let mut j = i + 1;
             while j < group.len() {
                 let next = group[j];
-                // Must be similar font size (within 20%)
-                if (next.font_size - first.font_size).abs() > first.font_size * 0.20 {
+                // A small-caps junction is mid-word: it both survives the
+                // font-size band below and must never take a space.
+                let small_caps_join =
+                    is_small_caps_continuation(&text, first, next, next.x - end_x);
+                // Must be similar font size, except for genuine small-caps
+                // runs, where the shrunken capitals are the same word as the
+                // full-size initial (see helper).
+                if (next.font_size - first.font_size).abs() > first.font_size * MERGE_FONT_SIZE_BAND
+                    && !small_caps_join
+                {
                     break;
                 }
                 // Never merge across style boundaries: the merged item
@@ -997,7 +1097,7 @@ pub(crate) fn merge_text_items(items: Vec<TextItem>) -> Vec<TextItem> {
                     Some((run_end, floor)) if j <= run_end => floor,
                     _ => threshold,
                 };
-                if needs_bullet_space || gap > effective_threshold {
+                if !small_caps_join && (needs_bullet_space || gap > effective_threshold) {
                     text.push(' ');
                 }
                 text.push_str(&next.text);
@@ -3001,6 +3101,145 @@ mod tests {
         }
     }
 
+    /// Small caps as typesetters emit them: a full-size capital at 9.98pt
+    /// immediately followed by shrunken capitals at 6.74pt, touching.
+    /// Modelled on `199AD3d.pdf` p.5 ("ROLANDO T. ACOSTA, P.J.").
+    #[test]
+    fn small_caps_run_merges_into_one_word() {
+        let items = vec![
+            make_item_fs("R", 144.36, 581.84, 7.20, 9.98),
+            make_item_fs("OLANDO", 151.56, 581.84, 30.56, 6.74),
+            make_item_fs("T. A", 185.45, 581.84, 17.58, 9.98),
+            make_item_fs("COSTA", 203.94, 581.84, 23.15, 6.74),
+            make_item_fs(", P.J.", 227.09, 581.84, 22.56, 9.98),
+        ];
+        let merged = merge_text_items(items);
+        assert_eq!(merged.len(), 1, "got {:?}", merged);
+        assert_eq!(merged[0].text, "ROLANDO T. ACOSTA, P.J.");
+    }
+
+    /// The full two-column row: both names must merge independently and the
+    /// 72pt column gap between them must survive as an item boundary.
+    #[test]
+    fn small_caps_merge_does_not_swallow_a_second_column() {
+        let items = vec![
+            // Column 1: "ROLANDO T. ACOSTA, P.J." ending at x=249.65
+            make_item_fs("R", 144.36, 581.84, 7.20, 9.98),
+            make_item_fs("OLANDO", 151.56, 581.84, 30.56, 6.74),
+            make_item_fs("T. A", 185.45, 581.84, 17.58, 9.98),
+            make_item_fs("COSTA", 203.94, 581.84, 23.15, 6.74),
+            make_item_fs(", P.J.", 227.09, 581.84, 22.56, 9.98),
+            // Column 2 starts at x=321.96 — a 72pt gap.
+            make_item_fs("A", 321.96, 581.84, 7.20, 9.98),
+            make_item_fs("NIL", 329.17, 581.84, 12.72, 6.74),
+            make_item_fs("C. S", 345.04, 581.84, 19.59, 9.98),
+            make_item_fs("INGH", 364.62, 581.84, 19.08, 6.74),
+        ];
+        let merged = merge_text_items(items);
+        let texts: Vec<&str> = merged.iter().map(|i| i.text.as_str()).collect();
+        assert_eq!(
+            texts,
+            vec!["ROLANDO T. ACOSTA, P.J.", "ANIL C. SINGH"],
+            "column gap should keep the two names apart"
+        );
+    }
+
+    #[test]
+    fn small_caps_merge_keeps_word_space_between_same_size_capitals() {
+        // Two uppercase words at sizes the merge band already accepts (9.98 and
+        // 9.0, a 10% drop) separated by a real word gap. The small-caps path
+        // must not claim this junction and swallow the space.
+        let items = vec![
+            make_item_fs("SEE", 100.0, 500.0, 18.0, 9.98),
+            make_item_fs("ALSO", 119.2, 500.0, 24.0, 9.0),
+        ];
+        let merged = merge_text_items(items);
+        assert_eq!(merged.len(), 1, "got {:?}", merged);
+        assert_eq!(merged[0].text, "SEE ALSO");
+    }
+
+    #[test]
+    fn trailing_digit_is_not_a_capital_awaiting_small_caps() {
+        // "...MAZZARELLI1" ends in a footnote marker; the backward search for an
+        // uppercase letter must not skip the digit and glue the next run.
+        assert!(!is_small_caps_continuation(
+            "ANGELA M. MAZZARELLI1",
+            &make_item_fs("ANGELA", 100.0, 500.0, 40.0, 9.98),
+            &make_item_fs("SHULMAN", 140.0, 500.0, 30.0, 6.74),
+            0.0,
+        ));
+    }
+
+    #[test]
+    fn ordinal_suffix_after_a_digit_still_merges() {
+        // "TUESDAY, JULY 4" + "TH" is one word in the source; the digit guard
+        // must not block the four English ordinal suffixes.
+        for suffix in ["TH", "ST", "ND", "RD"] {
+            assert!(
+                is_small_caps_continuation(
+                    "TUESDAY, JULY 4",
+                    &make_item_fs("JULY", 100.0, 500.0, 30.0, 12.0),
+                    &make_item_fs(suffix, 130.0, 500.0, 8.0, 8.0),
+                    0.0,
+                ),
+                "{suffix} should merge after a digit"
+            );
+        }
+    }
+
+    #[test]
+    fn superscript_footnote_marker_is_not_a_small_caps_continuation() {
+        // A digit must never qualify — otherwise footnote markers get glued on
+        // without the superscript handling.
+        assert!(!is_small_caps_continuation(
+            "MAZZARELLI",
+            &make_item_fs("MAZZARELLI", 100.0, 500.0, 50.0, 9.98),
+            &make_item_fs("1", 150.0, 503.0, 3.0, 6.74),
+            0.0,
+        ));
+    }
+
+    #[test]
+    fn drop_cap_is_not_a_small_caps_continuation() {
+        // Mixed-case body text after a large initial is a drop cap, not small
+        // caps.
+        assert!(!is_small_caps_continuation(
+            "T",
+            &make_item_fs("T", 100.0, 500.0, 20.0, 30.0),
+            &make_item_fs("he court held", 120.0, 500.0, 60.0, 10.0),
+            0.0,
+        ));
+    }
+
+    #[test]
+    fn separate_word_is_not_a_small_caps_continuation() {
+        // A real word space disqualifies even when both runs are uppercase.
+        let first = make_item_fs("SEE", 100.0, 500.0, 20.0, 9.98);
+        let next = make_item_fs("ALSO", 128.0, 500.0, 25.0, 6.74);
+        assert!(!is_small_caps_continuation("SEE", &first, &next, 8.0));
+    }
+
+    #[test]
+    fn lowercase_continuation_is_not_small_caps() {
+        assert!(!is_small_caps_continuation(
+            "SMALL",
+            &make_item_fs("SMALL", 100.0, 500.0, 30.0, 9.98),
+            &make_item_fs("caps", 130.0, 500.0, 20.0, 6.74),
+            0.0,
+        ));
+    }
+
+    #[test]
+    fn too_small_a_ratio_is_not_small_caps() {
+        // 0.4 ratio is a superscript/sub-run, outside the small-caps band.
+        assert!(!is_small_caps_continuation(
+            "A",
+            &make_item_fs("A", 100.0, 500.0, 7.0, 10.0),
+            &make_item_fs("BC", 107.0, 500.0, 8.0, 4.0),
+            0.0,
+        ));
+    }
+
     #[test]
     fn test_merge_subscript_items_chemical_formula() {
         // NH₃: "NH" at fs=8 followed by subscript "3" at fs=4.7

src/extractor/xobjects.rs

@@ -16,6 +16,76 @@ use super::{get_number, image_bbox_from_ctm, multiply_matrices};
 
 const MAX_FORM_XOBJECT_DEPTH: u8 = 5;
 
+/// Upper bound on Form XObject invocations during a single page extraction.
+/// Depth alone is not enough: an acyclic DAG where each form invokes the next
+/// N times expands to N^depth work before the depth cap is reached.
+const MAX_FORM_XOBJECT_INVOCATIONS: usize = 10_000;
+
+/// Upper bound on content-stream operations walked across all Form XObject
+/// expansions for a page. Nested forms are decoded independently of the
+/// page-level operation cap, so this keeps total form work in the same
+/// ballpark as that page cap.
+const MAX_FORM_XOBJECT_OPERATIONS: usize = 1_000_000;
+
+/// Shared budget for Form XObject expansion on a page. Bounds both nested DAG
+/// expansion and repeated sibling `/Do` invocations of the same form.
+pub(crate) struct FormWalkBudget {
+    invocations: usize,
+    operations: usize,
+    max_invocations: usize,
+    max_operations: usize,
+    truncated: bool,
+}
+
+impl FormWalkBudget {
+    pub(crate) fn new() -> Self {
+        Self::with_limits(MAX_FORM_XOBJECT_INVOCATIONS, MAX_FORM_XOBJECT_OPERATIONS)
+    }
+
+    fn with_limits(max_invocations: usize, max_operations: usize) -> Self {
+        Self {
+            invocations: 0,
+            operations: 0,
+            max_invocations,
+            max_operations,
+            truncated: false,
+        }
+    }
+
+    fn exhausted(&mut self) -> bool {
+        if self.invocations >= self.max_invocations || self.operations >= self.max_operations {
+            self.truncated = true;
+            true
+        } else {
+            false
+        }
+    }
+
+    fn charge_invocation(&mut self) -> bool {
+        if self.exhausted() {
+            return false;
+        }
+        self.invocations += 1;
+        true
+    }
+
+    /// Charge one walked content-stream operator. Independent of the
+    /// invocation cap so a form that was already admitted can finish its
+    /// stream (up to the operation cap).
+    fn charge_operation(&mut self) -> bool {
+        if self.operations >= self.max_operations {
+            self.truncated = true;
+            return false;
+        }
+        self.operations += 1;
+        true
+    }
+
+    pub(crate) fn was_truncated(&self) -> bool {
+        self.truncated
+    }
+}
+
 pub(crate) enum XObjectType {
     Image,
     Form(ObjectId),
@@ -109,6 +179,7 @@ fn collect_xobjects_from_dict(
 }
 
 /// Extract text items from a Form XObject
+#[allow(clippy::too_many_arguments)]
 pub(crate) fn extract_form_xobject_text(
     doc: &Document,
     form_id: ObjectId,
@@ -117,6 +188,7 @@ pub(crate) fn extract_form_xobject_text(
     parent_ctm: &[f32; 6],
     cmap_decisions: &mut CMapDecisionCache,
     style_cache: &mut FontStyleCache,
+    budget: &mut FormWalkBudget,
 ) -> Vec<TextItem> {
     extract_form_xobject_text_inner(
         doc,
@@ -127,6 +199,7 @@ pub(crate) fn extract_form_xobject_text(
         cmap_decisions,
         style_cache,
         0,
+        budget,
     )
 }
 
@@ -140,11 +213,14 @@ fn extract_form_xobject_text_inner(
     cmap_decisions: &mut CMapDecisionCache,
     style_cache: &mut FontStyleCache,
     depth: u8,
+    budget: &mut FormWalkBudget,
 ) -> Vec<TextItem> {
-    use lopdf::content::Content;
-
     let mut items = Vec::new();
 
+    if !budget.charge_invocation() {
+        return items;
+    }
+
     // Get the Form XObject stream
     let Ok(Object::Stream(stream)) = doc.get_object(form_id) else {
         return items;
@@ -156,8 +232,12 @@ fn extract_form_xobject_text_inner(
         Err(_) => stream.content.clone(),
     };
 
-    // Decode the content stream
-    let Ok(content) = Content::decode(&content_data) else {
+    // Decode the content stream. Cap before lopdf materializes the operator
+    // vector — the walk budget cannot help if decode itself allocates first.
+    let Ok(Some(content)) = super::content_decode::decode_content_bounded(
+        &content_data,
+        super::content_decode::MAX_PAGE_OPERATIONS,
+    ) else {
         return items;
     };
 
@@ -246,19 +326,55 @@ fn extract_form_xobject_text_inner(
     let mut current_font = String::new();
     let mut current_font_size: f32 = 12.0;
     let mut text_matrix = [1.0f32, 0.0, 0.0, 1.0, 0.0, 0.0];
+    // Text line matrix (TLM) — Td/TD/T* move relative to the start of the
+    // current line, not to the position left by the last show operator.
+    let mut line_matrix = [1.0f32, 0.0, 0.0, 1.0, 0.0, 0.0];
+    let mut text_leading: f32 = 0.0; // TL parameter (text-space units)
+    let mut char_spacing: f32 = 0.0; // Tc parameter
+    let mut word_spacing: f32 = 0.0; // Tw parameter
     let mut in_text_block = false;
     let mut fill_is_white = false;
     let mut ctm = base_ctm;
-    let mut ctm_stack: Vec<[f32; 6]> = Vec::new();
+
+    // Text state (Tc/Tw/TL/Tf) and the fill colour are part of the graphics
+    // state and must be saved/restored by q/Q alongside the CTM.
+    #[derive(Clone)]
+    struct GraphicsState {
+        ctm: [f32; 6],
+        char_spacing: f32,
+        word_spacing: f32,
+        text_leading: f32,
+        current_font: String,
+        current_font_size: f32,
+        fill_is_white: bool,
+    }
+    let mut ctm_stack: Vec<GraphicsState> = Vec::new();
 
     for op in &content.operations {
+        if !budget.charge_operation() {
+            break;
+        }
         match op.operator.as_str() {
             "q" => {
-                ctm_stack.push(ctm);
+                ctm_stack.push(GraphicsState {
+                    ctm,
+                    char_spacing,
+                    word_spacing,
+                    text_leading,
+                    current_font: current_font.clone(),
+                    current_font_size,
+                    fill_is_white,
+                });
             }
             "Q" => {
                 if let Some(saved) = ctm_stack.pop() {
-                    ctm = saved;
+                    ctm = saved.ctm;
+                    char_spacing = saved.char_spacing;
+                    word_spacing = saved.word_spacing;
+                    text_leading = saved.text_leading;
+                    current_font = saved.current_font;
+                    current_font_size = saved.current_font_size;
+                    fill_is_white = saved.fill_is_white;
                 }
             }
             "cm" => {
@@ -276,7 +392,7 @@ fn extract_form_xobject_text_inner(
                         let xobj_name = String::from_utf8_lossy(name).to_string();
                         match form_xobjects.get(&xobj_name) {
                             Some(XObjectType::Form(nested_id)) => {
-                                if depth < MAX_FORM_XOBJECT_DEPTH {
+                                if depth < MAX_FORM_XOBJECT_DEPTH && !budget.exhausted() {
                                     let nested_items = extract_form_xobject_text_inner(
                                         doc,
                                         *nested_id,
@@ -286,6 +402,7 @@ fn extract_form_xobject_text_inner(
                                         cmap_decisions,
                                         style_cache,
                                         depth + 1,
+                                        budget,
                                     );
                                     items.extend(nested_items);
                                 }
@@ -321,6 +438,7 @@ fn extract_form_xobject_text_inner(
             "BT" => {
                 in_text_block = true;
                 text_matrix = [1.0, 0.0, 0.0, 1.0, 0.0, 0.0];
+                line_matrix = text_matrix;
             }
             "ET" => {
                 in_text_block = false;
@@ -333,12 +451,33 @@ fn extract_form_xobject_text_inner(
                     current_font_size = get_number(&op.operands[1]).unwrap_or(12.0);
                 }
             }
+            "TL" => {
+                // Set text leading (used by T*, ', and ")
+                if let Some(tl) = op.operands.first().and_then(get_number) {
+                    text_leading = tl;
+                }
+            }
+            "Tc" => {
+                if let Some(tc) = op.operands.first().and_then(get_number) {
+                    char_spacing = tc;
+                }
+            }
+            "Tw" => {
+                if let Some(tw) = op.operands.first().and_then(get_number) {
+                    word_spacing = tw;
+                }
+            }
             "Td" | "TD" => {
+                // Move text position: TLM = T(tx,ty) x TLM; Tm = TLM
                 if op.operands.len() >= 2 {
                     let tx = get_number(&op.operands[0]).unwrap_or(0.0);
                     let ty = get_number(&op.operands[1]).unwrap_or(0.0);
-                    text_matrix[4] += tx * text_matrix[0] + ty * text_matrix[2];
-                    text_matrix[5] += tx * text_matrix[1] + ty * text_matrix[3];
+                    line_matrix[4] += tx * line_matrix[0] + ty * line_matrix[2];
+                    line_matrix[5] += tx * line_matrix[1] + ty * line_matrix[3];
+                    text_matrix = line_matrix;
+                    if op.operator == "TD" {
+                        text_leading = -ty;
+                    }
                 }
             }
             "Tm" => {
@@ -347,8 +486,20 @@ fn extract_form_xobject_text_inner(
                         text_matrix[i] =
                             get_number(operand).unwrap_or(if i == 0 || i == 3 { 1.0 } else { 0.0 });
                     }
+                    line_matrix = text_matrix;
                 }
             }
+            "T*" => {
+                // Move to start of next line: equivalent to `0 -TL Td`
+                let tl = if text_leading != 0.0 {
+                    text_leading
+                } else {
+                    current_font_size * 1.2
+                };
+                line_matrix[4] += (-tl) * line_matrix[2];
+                line_matrix[5] += (-tl) * line_matrix[3];
+                text_matrix = line_matrix;
+            }
             "g" => {
                 if let Some(gray) = op.operands.first().and_then(get_number) {
                     fill_is_white = gray > 0.95;
@@ -384,17 +535,33 @@ fn extract_form_xobject_text_inner(
                     _ => fill_is_white = false,
                 }
             }
-            "Tj" => {
-                if in_text_block && !op.operands.is_empty() {
+            "Tj" | "'" | "\"" => {
+                // `'` = move to next line then show; `"` = set word/char spacing,
+                // move to next line, then show (string is the last operand).
+                if op.operator != "Tj" {
+                    if op.operator == "\"" && op.operands.len() >= 3 {
+                        word_spacing = get_number(&op.operands[0]).unwrap_or(word_spacing);
+                        char_spacing = get_number(&op.operands[1]).unwrap_or(char_spacing);
+                    }
+                    let tl = if text_leading != 0.0 {
+                        text_leading
+                    } else {
+                        current_font_size * 1.2
+                    };
+                    line_matrix[4] += (-tl) * line_matrix[2];
+                    line_matrix[5] += (-tl) * line_matrix[3];
+                    text_matrix = line_matrix;
+                }
+                if let (true, Some(show_operand)) = (in_text_block, op.operands.last()) {
                     if fill_is_white {
                         if let Some(font_info) = font_widths.get(&current_font) {
-                            if let Some(raw_bytes) = get_operand_bytes(&op.operands[0]) {
+                            if let Some(raw_bytes) = get_operand_bytes(show_operand) {
                                 let w_ts = compute_string_width_ts(
                                     raw_bytes,
                                     font_info,
                                     current_font_size,
-                                    0.0,
-                                    0.0,
+                                    char_spacing,
+                                    word_spacing,
                                 );
                                 text_matrix[4] += w_ts * text_matrix[0];
                                 text_matrix[5] += w_ts * text_matrix[1];
@@ -403,7 +570,7 @@ fn extract_form_xobject_text_inner(
                         continue;
                     }
                     if let Some(text) = extract_text_from_operand(
-                        &op.operands[0],
+                        show_operand,
                         &current_font,
                         font_base_names.get(&current_font).map(|s| s.as_str()),
                         font_cmaps,
@@ -419,13 +586,13 @@ fn extract_form_xobject_text_inner(
                             * type3_scales.get(&current_font).copied().unwrap_or(1.0);
                         let (x, y) = (combined[4], combined[5]);
                         let width = if let Some(font_info) = font_widths.get(&current_font) {
-                            if let Some(raw_bytes) = get_operand_bytes(&op.operands[0]) {
+                            if let Some(raw_bytes) = get_operand_bytes(show_operand) {
                                 let w_ts = compute_string_width_ts(
                                     raw_bytes,
                                     font_info,
                                     current_font_size,
-                                    0.0,
-                                    0.0,
+                                    char_spacing,
+                                    word_spacing,
                                 );
                                 text_matrix[4] += w_ts * text_matrix[0];
                                 text_matrix[5] += w_ts * text_matrix[1];
@@ -548,8 +715,8 @@ fn extract_form_xobject_text_inner(
                                         raw_bytes,
                                         fi,
                                         current_font_size,
-                                        0.0,
-                                        0.0,
+                                        char_spacing,
+                                        word_spacing,
                                     );
                                 }
                             }
@@ -683,3 +850,415 @@ pub(crate) fn get_form_fonts<'a>(
 
     fonts
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::extractor::content_stream::extract_page_text_items;
+    use lopdf::{dictionary, Dictionary, Stream};
+
+    /// Build an acyclic Form XObject DAG: `levels` form objects, each non-leaf
+    /// invoking the next form `branches` times. The leaf draws a single `(X)`.
+    /// Returns `(doc, root_form_id)`.
+    fn form_dag(branches: usize, levels: usize) -> (Document, ObjectId) {
+        assert!(levels >= 2);
+        let mut doc = Document::new();
+        let font_id = doc.add_object(dictionary! {
+            "Type" => "Font",
+            "Subtype" => "Type1",
+            "BaseFont" => "Helvetica",
+        });
+        let ids: Vec<ObjectId> = (0..levels).map(|_| doc.new_object_id()).collect();
+        for level in 0..levels {
+            let stream = if level + 1 == levels {
+                Stream::new(
+                    dictionary! {
+                        "Type" => "XObject",
+                        "Subtype" => "Form",
+                        "BBox" => vec![0.into(), 0.into(), 100.into(), 100.into()],
+                        "Resources" => dictionary! {
+                            "Font" => dictionary! {
+                                "F1" => Object::Reference(font_id),
+                            },
+                        },
+                    },
+                    b"BT /F1 10 Tf 10 10 Td (X) Tj ET\n".to_vec(),
+                )
+            } else {
+                let next_name = format!("Fm{}", level + 1);
+                let content = format!("/{next_name} Do\n").repeat(branches);
+                let mut xobjects = Dictionary::new();
+                xobjects.set(next_name, Object::Reference(ids[level + 1]));
+                let mut resources = Dictionary::new();
+                resources.set("XObject", Object::Dictionary(xobjects));
+                let mut dict = dictionary! {
+                    "Type" => "XObject",
+                    "Subtype" => "Form",
+                    "BBox" => vec![0.into(), 0.into(), 100.into(), 100.into()],
+                };
+                dict.set("Resources", Object::Dictionary(resources));
+                Stream::new(dict, content.into_bytes())
+            };
+            doc.set_object(ids[level], Object::Stream(stream));
+        }
+        (doc, ids[0])
+    }
+
+    fn page_invoking_form(mut doc: Document, form_id: ObjectId) -> (Document, ObjectId) {
+        let content_id = doc.add_object(Object::Stream(Stream::new(
+            dictionary! {},
+            b"/Fm0 Do\n".to_vec(),
+        )));
+        let page_id = doc.add_object(dictionary! {
+            "Type" => "Page",
+            "Contents" => Object::Reference(content_id),
+            "Resources" => dictionary! {
+                "XObject" => dictionary! {
+                    "Fm0" => Object::Reference(form_id),
+                },
+            },
+            "MediaBox" => vec![0.into(), 0.into(), 612.into(), 792.into()],
+        });
+        let pages_id = doc.add_object(dictionary! {
+            "Type" => "Pages",
+            "Count" => Object::Integer(1),
+            "Kids" => vec![Object::Reference(page_id)],
+        });
+        let catalog_id = doc.add_object(dictionary! {
+            "Type" => "Catalog",
+            "Pages" => Object::Reference(pages_id),
+        });
+        doc.trailer.set("Root", Object::Reference(catalog_id));
+        (doc, page_id)
+    }
+
+    fn extract_form(
+        doc: &Document,
+        form_id: ObjectId,
+        budget: &mut FormWalkBudget,
+    ) -> Vec<TextItem> {
+        extract_form_xobject_text(
+            doc,
+            form_id,
+            1,
+            &FontCMaps::from_doc(doc),
+            &[1.0, 0.0, 0.0, 1.0, 0.0, 0.0],
+            &mut CMapDecisionCache::new(),
+            &mut FontStyleCache::new(),
+            budget,
+        )
+    }
+
+    #[test]
+    fn nested_form_still_extracts_leaf_text() {
+        let (doc, root) = form_dag(1, 3);
+        let items = extract_form(&doc, root, &mut FormWalkBudget::new());
+        assert_eq!(items.len(), 1);
+        assert_eq!(items[0].text, "X");
+    }
+
+    #[test]
+    fn acyclic_form_dag_within_budget_keeps_all_leaves() {
+        // 4 sibling invocations across 4 nested levels → 4^4 leaf drawings.
+        // Default budgets are far above 256, so legitimate nesting is intact.
+        let (doc, root) = form_dag(4, 5);
+        let items = extract_form(&doc, root, &mut FormWalkBudget::new());
+        assert_eq!(items.len(), 4usize.pow(4));
+        assert!(items.iter().all(|item| item.text == "X"));
+    }
+
+    #[test]
+    fn acyclic_form_dag_stops_at_invocation_budget() {
+        // Same DAG as above would draw 256 leaves; a tiny invocation cap must
+        // stop expansion rather than walking the full tree.
+        let (doc, root) = form_dag(4, 5);
+        let mut budget = FormWalkBudget::with_limits(20, MAX_FORM_XOBJECT_OPERATIONS);
+        let items = extract_form(&doc, root, &mut budget);
+        assert!(
+            items.len() < 4usize.pow(4),
+            "invocation budget must truncate DAG expansion; got {} items",
+            items.len()
+        );
+        assert!(budget.was_truncated());
+    }
+
+    #[test]
+    fn form_operations_stop_at_budget() {
+        let mut doc = Document::new();
+        let font_id = doc.add_object(dictionary! {
+            "Type" => "Font",
+            "Subtype" => "Type1",
+            "BaseFont" => "Helvetica",
+        });
+        let mut content = b"q Q\n".repeat(50);
+        content.extend_from_slice(b"BT /F1 10 Tf 10 10 Td (X) Tj ET\n");
+        let form_id = doc.add_object(Object::Stream(Stream::new(
+            dictionary! {
+                "Type" => "XObject",
+                "Subtype" => "Form",
+                "BBox" => vec![0.into(), 0.into(), 100.into(), 100.into()],
+                "Resources" => dictionary! {
+                    "Font" => dictionary! {
+                        "F1" => Object::Reference(font_id),
+                    },
+                },
+            },
+            content,
+        )));
+
+        let mut budget = FormWalkBudget::with_limits(MAX_FORM_XOBJECT_INVOCATIONS, 10);
+        let items = extract_form(&doc, form_id, &mut budget);
+        assert!(
+            items.is_empty(),
+            "operation budget must stop before the trailing text show"
+        );
+        assert!(budget.was_truncated());
+    }
+
+    #[test]
+    fn page_level_form_dag_stays_within_production_budget() {
+        // A page-level `/Do` of an 8-wide, 6-level Form DAG would expand to
+        // 8^5 = 32_768 leaf drawings without a budget. The production
+        // invocation cap must keep extraction bounded.
+        let (doc, root) = form_dag(8, 6);
+        let (doc, page_id) = page_invoking_form(doc, root);
+
+        let font_cmaps = FontCMaps::from_doc(&doc);
+        let ((items, _, _), _, _, _) = extract_page_text_items(
+            &doc,
+            page_id,
+            1,
+            &font_cmaps,
+            false,
+            &mut FontStyleCache::new(),
+            &mut FormWalkBudget::new(),
+        )
+        .unwrap();
+        assert!(
+            items.len() <= MAX_FORM_XOBJECT_INVOCATIONS,
+            "page-level Form expansion must stay within the invocation cap; got {}",
+            items.len()
+        );
+        assert!(
+            !items.is_empty(),
+            "budget must still allow some nested form text through"
+        );
+    }
+
+    #[test]
+    fn shared_form_budget_spans_two_extraction_passes() {
+        // The invisible-layer retry calls extract_page_text_items twice for
+        // the same page; both passes must share one budget.
+        let (doc, root) = form_dag(1, 2);
+        let (doc, page_id) = page_invoking_form(doc, root);
+        let font_cmaps = FontCMaps::from_doc(&doc);
+        // Root + leaf = 2 invocations on the first pass.
+        let mut budget = FormWalkBudget::with_limits(2, MAX_FORM_XOBJECT_OPERATIONS);
+        let ((first, _, _), _, _, _) = extract_page_text_items(
+            &doc,
+            page_id,
+            1,
+            &font_cmaps,
+            false,
+            &mut FontStyleCache::new(),
+            &mut budget,
+        )
+        .unwrap();
+        assert_eq!(first.iter().filter(|item| item.text == "X").count(), 1);
+        assert!(!budget.was_truncated());
+
+        let ((second, _, _), _, _, _) = extract_page_text_items(
+            &doc,
+            page_id,
+            1,
+            &font_cmaps,
+            true,
+            &mut FontStyleCache::new(),
+            &mut budget,
+        )
+        .unwrap();
+        assert!(
+            second.iter().all(|item| item.text != "X"),
+            "second pass must not get a fresh invocation budget"
+        );
+        assert!(budget.was_truncated());
+    }
+
+    /// Build a document whose page draws *all* of its content through a single
+    /// Form XObject — the shape emitted by print-to-PDF producers like PDFlib,
+    /// where the page stream itself is only `q /X1 Do Q`.
+    fn doc_with_form_content(form_content: &[u8]) -> (Document, ObjectId) {
+        let mut doc = Document::new();
+        let widths: Vec<Object> = (0..=255).map(|_| 600.into()).collect();
+        let font_id = doc.add_object(dictionary! {
+            "Type" => "Font",
+            "Subtype" => "Type1",
+            "BaseFont" => "Helvetica",
+            "FirstChar" => 0,
+            "LastChar" => 255,
+            "Widths" => Object::Array(widths),
+        });
+        let form_id = doc.add_object(Object::Stream(Stream::new(
+            dictionary! {
+                "Type" => "XObject",
+                "Subtype" => "Form",
+                "BBox" => vec![0.into(), 0.into(), 612.into(), 792.into()],
+                "Resources" => dictionary! {
+                    "Font" => dictionary! { "F1" => Object::Reference(font_id) },
+                },
+            },
+            form_content.to_vec(),
+        )));
+        let content_id = doc.add_object(Object::Stream(Stream::new(
+            dictionary! {},
+            b"q /X1 Do Q".to_vec(),
+        )));
+        let page_id = doc.add_object(dictionary! {
+            "Type" => "Page",
+            "Contents" => Object::Reference(content_id),
+            "Resources" => dictionary! {
+                "XObject" => dictionary! { "X1" => Object::Reference(form_id) },
+            },
+            "MediaBox" => vec![0.into(), 0.into(), 612.into(), 792.into()],
+        });
+        let pages_id = doc.add_object(dictionary! {
+            "Type" => "Pages",
+            "Count" => Object::Integer(1),
+            "Kids" => vec![Object::Reference(page_id)],
+        });
+        let catalog_id = doc.add_object(dictionary! {
+            "Type" => "Catalog",
+            "Pages" => Object::Reference(pages_id),
+        });
+        doc.trailer.set("Root", Object::Reference(catalog_id));
+        (doc, page_id)
+    }
+
+    fn form_items(form_content: &[u8]) -> Vec<TextItem> {
+        let (doc, page_id) = doc_with_form_content(form_content);
+        let font_cmaps = FontCMaps::from_doc(&doc);
+        let ((items, _, _), _, _, _) = extract_page_text_items(
+            &doc,
+            page_id,
+            1,
+            &font_cmaps,
+            false,
+            &mut FontStyleCache::new(),
+            &mut FormWalkBudget::new(),
+        )
+        .unwrap();
+        items
+    }
+
+    fn find<'a>(items: &'a [TextItem], text: &str) -> &'a TextItem {
+        items
+            .iter()
+            .find(|item| item.text == text)
+            .unwrap_or_else(|| {
+                let found: Vec<&String> = items.iter().map(|i| &i.text).collect();
+                panic!("no item {text:?} in {found:?}")
+            })
+    }
+
+    #[test]
+    fn t_star_inside_form_moves_to_next_line() {
+        // T* was previously unhandled inside Form XObjects, so every line after
+        // the first piled onto the preceding baseline and drifted right.
+        let items =
+            form_items(b"BT /F1 12 Tf 12 TL 1 0 0 1 100 700 Tm (first) Tj T* (second) Tj ET");
+
+        let first = find(&items, "first");
+        let second = find(&items, "second");
+        assert!((first.y - 700.0).abs() < 0.1, "first y = {}", first.y);
+        assert!((second.y - 688.0).abs() < 0.1, "second y = {}", second.y);
+        assert!((second.x - 100.0).abs() < 0.1, "second x = {}", second.x);
+    }
+
+    #[test]
+    fn td_inside_form_is_relative_to_line_start_not_shown_text() {
+        // Td moves relative to the text *line* matrix. Applying it to the
+        // matrix already advanced by Tj marched each line off the right edge.
+        let items = form_items(b"BT /F1 12 Tf 1 0 0 1 100 700 Tm (AAAAA) Tj 0 -12 Td (B) Tj ET");
+
+        let b = find(&items, "B");
+        assert!((b.x - 100.0).abs() < 0.1, "B x = {} (expected 100)", b.x);
+        assert!((b.y - 688.0).abs() < 0.1, "B y = {}", b.y);
+    }
+
+    #[test]
+    fn td_inside_form_sets_leading_for_later_t_star() {
+        // `TD` sets the leading to -ty as a side effect; a following T* must
+        // reuse it.
+        let items = form_items(
+            b"BT /F1 12 Tf 1 0 0 1 100 700 Tm (one) Tj 0 -15 TD (two) Tj T* (three) Tj ET",
+        );
+
+        assert!((find(&items, "two").y - 685.0).abs() < 0.1);
+        let three = find(&items, "three");
+        assert!((three.y - 670.0).abs() < 0.1, "three y = {}", three.y);
+        assert!((three.x - 100.0).abs() < 0.1, "three x = {}", three.x);
+    }
+
+    #[test]
+    fn quote_operator_inside_form_moves_to_next_line() {
+        let items = form_items(b"BT /F1 12 Tf 12 TL 1 0 0 1 100 700 Tm (first) Tj (second) ' ET");
+
+        let second = find(&items, "second");
+        assert!((second.y - 688.0).abs() < 0.1, "second y = {}", second.y);
+        assert!((second.x - 100.0).abs() < 0.1, "second x = {}", second.x);
+    }
+
+    #[test]
+    fn double_quote_operator_inside_form_sets_spacing_and_moves() {
+        // `aw ac (string) "` — set word spacing and char spacing, then T* and show.
+        let items =
+            form_items(b"BT /F1 12 Tf 12 TL 1 0 0 1 100 700 Tm (first) Tj 0 0 (second) \" ET");
+
+        let second = find(&items, "second");
+        assert!((second.y - 688.0).abs() < 0.1, "second y = {}", second.y);
+        assert!((second.x - 100.0).abs() < 0.1, "second x = {}", second.x);
+    }
+
+    #[test]
+    fn char_spacing_inside_form_widens_advance() {
+        // Tc was hardcoded to 0 in the form parser, so advance widths drifted.
+        // 2 glyphs x 600/1000 x 12pt = 14.4, plus 2 x Tc(2.0) = 18.4.
+        let items = form_items(b"BT /F1 12 Tf 1 0 0 1 100 700 Tm 2 Tc (AB) Tj ET");
+
+        let ab = find(&items, "AB");
+        assert!((ab.width - 18.4).abs() < 0.1, "AB width = {}", ab.width);
+    }
+
+    #[test]
+    fn q_restores_fill_colour_inside_form() {
+        // A white fill set inside q/Q must not leak past the Q — otherwise the
+        // following black text is treated as invisible and dropped entirely.
+        let items = form_items(
+            b"BT /F1 12 Tf 12 TL 1 0 0 1 100 700 Tm q 1 g (hidden) Tj Q T* (visible) Tj ET",
+        );
+
+        assert!(
+            items.iter().any(|item| item.text == "visible"),
+            "text after Q was dropped: {:?}",
+            items.iter().map(|i| &i.text).collect::<Vec<_>>()
+        );
+        assert!(
+            !items.iter().any(|item| item.text == "hidden"),
+            "white-filled text should still be suppressed"
+        );
+    }
+
+    #[test]
+    fn q_restores_text_state_inside_form() {
+        // Tc/TL live in the graphics state; `Q` must roll them back.
+        let items =
+            form_items(b"BT /F1 12 Tf 12 TL 1 0 0 1 100 700 Tm q 30 TL (a) Tj Q T* (b) Tj ET");
+
+        let b = find(&items, "b");
+        assert!(
+            (b.y - 688.0).abs() < 0.1,
+            "b y = {} (leading should restore to 12)",
+            b.y
+        );
+    }
+}

src/glyph_names.rs

@@ -4566,9 +4566,13 @@ pub fn glyph_to_char(name: &str) -> Option<char> {
         }
     }
 
-    // Try to parse uniXXXX format
-    if name.starts_with("uni") && name.len() >= 7 {
-        if let Ok(code) = u32::from_str_radix(&name[3..7], 16) {
+    // Try to parse uniXXXX format.
+    // Use `get` rather than a byte-length check + slice: `name` can contain
+    // non-ASCII bytes (e.g. U+FFFD from lossy UTF-8 decoding of an attacker
+    // controlled /Differences name), so byte index 7 may not be a char
+    // boundary and `&name[3..7]` would panic.
+    if let Some(hex) = name.strip_prefix("uni").and_then(|rest| rest.get(..4)) {
+        if let Ok(code) = u32::from_str_radix(hex, 16) {
             // Strip PUA F000 offset: uniF0XX → U+00XX (Windows Symbol encoding convention)
             let code = if (0xF000..=0xF0FF).contains(&code) {
                 code - 0xF000
@@ -4588,3 +4592,36 @@ pub fn glyph_to_char(name: &str) -> Option<char> {
 
     None
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn uni_hex_parsing() {
+        assert_eq!(glyph_to_char("uni0041"), Some('A'));
+        assert_eq!(glyph_to_char("uni00e9"), Some('\u{00e9}'));
+        // PUA F0xx symbol-encoding offset is stripped.
+        assert_eq!(glyph_to_char("uniF041"), Some('A'));
+    }
+
+    #[test]
+    fn u_hex_parsing() {
+        assert_eq!(glyph_to_char("u0041"), Some('A'));
+        assert_eq!(glyph_to_char("u1F600"), Some('\u{1F600}'));
+    }
+
+    #[test]
+    fn non_ascii_uni_name_does_not_panic() {
+        // A crafted /Differences name like `/uni#80#80#80#80` decodes via
+        // from_utf8_lossy into "uni" followed by four U+FFFD replacements.
+        // Byte index 7 lands mid-character, so a naive `&name[3..7]` slice
+        // would panic. It must be handled gracefully instead.
+        let crafted = format!("uni{0}{0}{0}{0}", '\u{FFFD}');
+        assert_eq!(glyph_to_char(&crafted), None);
+
+        // Assorted non-ASCII bytes right after the "uni" prefix.
+        assert_eq!(glyph_to_char("uni\u{FFFD}bc"), None);
+        assert_eq!(glyph_to_char("uni\u{00e9}00"), None);
+    }
+}

src/lib.rs

@@ -43,6 +43,7 @@ mod text_quality;
 pub mod text_utils;
 pub mod tounicode;
 pub mod types;
+pub mod vision;
 
 pub use detector::{
     detect_pdf_type, detect_pdf_type_mem, detect_pdf_type_mem_with_config,
@@ -458,8 +459,35 @@ pub fn extract_pages_markdown_mem(
     buffer: &[u8],
     pages: Option<&[u32]>,
 ) -> Result<PagesExtractionResult, PdfError> {
+    extract_pages_markdown_mem_impl(buffer, pages, None, &MarkdownOptions::default(), false)
+        .map(|(result, _)| result)
+}
+
+#[cfg(all(feature = "ocr", not(target_arch = "wasm32")))]
+pub(crate) fn extract_pages_markdown_mem_for_ocr(
+    buffer: &[u8],
+    pages: Option<&[u32]>,
+    password: Option<&str>,
+    markdown_options: &MarkdownOptions,
+) -> Result<(PagesExtractionResult, u32), PdfError> {
+    extract_pages_markdown_mem_impl(
+        buffer,
+        pages,
+        password,
+        markdown_options,
+        markdown_options.strip_headers_footers,
+    )
+}
+
+fn extract_pages_markdown_mem_impl(
+    buffer: &[u8],
+    pages: Option<&[u32]>,
+    password: Option<&str>,
+    markdown_options: &MarkdownOptions,
+    strip_repeated_headers_footers: bool,
+) -> Result<(PagesExtractionResult, u32), PdfError> {
     validate_pdf_bytes(buffer)?;
-    let (doc, page_count) = load_document_from_mem(buffer)?;
+    let (doc, page_count) = load_document_from_mem_with_password(buffer, password)?;
     let font_cmaps = FontCMaps::from_doc(&doc);
 
     // Extract ALL pages to get accurate, document-wide font stats. A malformed
@@ -502,6 +530,11 @@ pub fn extract_pages_markdown_mem(
 
     // Compute font stats from full document (cross-page consistency).
     let font_stats = markdown::analysis::calculate_font_stats_from_items(&filtered_items);
+    let repeated_header_footer_items = if strip_repeated_headers_footers {
+        repeated_header_footer_item_keys(&all_items, &page_thresholds, &chart_regions, page_count)
+    } else {
+        HashSet::new()
+    };
 
     // When caller doesn't specify pages, return every page in document order.
     let all_pages: Vec<u32>;
@@ -537,7 +570,10 @@ pub fn extract_pages_markdown_mem(
         let (page_items, page_number_removal_mask): (Vec<TextItem>, Vec<bool>) = all_items
             .iter()
             .zip(&page_number_removal_mask)
-            .filter(|(item, _)| item.page == page_1idx)
+            .filter(|(item, _)| {
+                item.page == page_1idx
+                    && !repeated_header_footer_items.contains(&HeaderFooterItemKey::from(*item))
+            })
             .map(|(item, remove)| (item.clone(), *remove))
             .unzip();
 
@@ -574,7 +610,7 @@ pub fn extract_pages_markdown_mem(
             base_font_size: Some(font_stats.most_common_size),
             include_page_numbers: false,
             strip_headers_footers: false,
-            ..MarkdownOptions::default()
+            ..markdown_options.clone()
         };
 
         let md = if has_text_quality_issue {
@@ -633,14 +669,129 @@ pub fn extract_pages_markdown_mem(
         });
     }
 
-    Ok(PagesExtractionResult {
-        pages: results,
-        pages_with_tables: complexity.pages_with_tables,
-        pages_with_columns: complexity.pages_with_columns,
-        pages_needing_ocr,
-        ocr_reasons_by_page: page_ocr_reasons_vec(ocr_reasons_by_page),
-        is_complex: complexity.is_complex,
-    })
+    Ok((
+        PagesExtractionResult {
+            pages: results,
+            pages_with_tables: complexity.pages_with_tables,
+            pages_with_columns: complexity.pages_with_columns,
+            pages_needing_ocr,
+            ocr_reasons_by_page: page_ocr_reasons_vec(ocr_reasons_by_page),
+            is_complex: complexity.is_complex,
+        },
+        page_count,
+    ))
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Hash)]
+struct HeaderFooterItemKey {
+    page: u32,
+    x: u32,
+    y: u32,
+    text: String,
+}
+
+impl From<&TextItem> for HeaderFooterItemKey {
+    fn from(item: &TextItem) -> Self {
+        Self {
+            page: item.page,
+            x: item.x.to_bits(),
+            y: item.y.to_bits(),
+            text: item.text.clone(),
+        }
+    }
+}
+
+fn repeated_header_footer_item_keys(
+    items: &[TextItem],
+    page_thresholds: &HashMap<u32, f32>,
+    chart_regions: &HashMap<u32, Vec<(f32, f32, f32, f32)>>,
+    page_count: u32,
+) -> HashSet<HeaderFooterItemKey> {
+    let candidates = items
+        .iter()
+        .filter(|item| {
+            matches!(
+                item.item_type,
+                types::ItemType::Text | types::ItemType::FormField
+            )
+        })
+        .cloned()
+        .collect();
+    let lines = extractor::group_prefiltered_items_into_lines_with_thresholds_and_charts(
+        candidates,
+        page_thresholds,
+        &HashSet::new(),
+        chart_regions,
+    );
+    let all_items: HashSet<_> = lines
+        .iter()
+        .flat_map(|line| line.items.iter().map(HeaderFooterItemKey::from))
+        .collect();
+    let kept = markdown::strip_repeated_header_footer_lines(lines, page_count);
+    let kept_items: HashSet<_> = kept
+        .iter()
+        .flat_map(|line| line.items.iter().map(HeaderFooterItemKey::from))
+        .collect();
+    all_items.difference(&kept_items).cloned().collect()
+}
+
+#[cfg(all(test, feature = "ocr", not(target_arch = "wasm32")))]
+mod ocr_header_footer_tests {
+    use super::*;
+
+    fn item(page: u32, text: &str, y: f32) -> TextItem {
+        TextItem {
+            text: text.to_string(),
+            x: 10.0,
+            y,
+            width: 120.0,
+            height: 10.0,
+            font: "Test".to_string(),
+            font_size: 10.0,
+            page,
+            is_bold: false,
+            is_italic: false,
+            is_underline: false,
+            is_strikeout: false,
+            item_type: types::ItemType::Text,
+            mcid: None,
+        }
+    }
+
+    #[test]
+    fn local_pipeline_prefilters_document_wide_repeated_headers() {
+        let mut items = Vec::new();
+        let mut thresholds = HashMap::new();
+        for page in 1..=3 {
+            items.push(item(page, "Repeated report header", 800.0));
+            for line in 0..12 {
+                items.push(item(
+                    page,
+                    &format!("Page {page} paragraph {line} unique content"),
+                    700.0 - line as f32 * 40.0,
+                ));
+            }
+            thresholds.insert(page, 0.1);
+        }
+
+        let removed = repeated_header_footer_item_keys(&items, &thresholds, &HashMap::new(), 3);
+        assert_eq!(removed.len(), 2);
+        for page in 1..=3 {
+            assert_eq!(
+                removed.contains(&HeaderFooterItemKey::from(&item(
+                    page,
+                    "Repeated report header",
+                    800.0,
+                ))),
+                page > 1,
+            );
+            assert!(!removed.contains(&HeaderFooterItemKey::from(&item(
+                page,
+                &format!("Page {page} paragraph 5 unique content"),
+                500.0,
+            ))));
+        }
+    }
 }
 
 /// Path-based wrapper for [`extract_pages_markdown_mem`].
@@ -657,6 +808,80 @@ pub fn extract_pages_markdown<P: AsRef<Path>>(
     extract_pages_markdown_mem(&buffer, pages)
 }
 
+// =========================================================================
+// Structure-tree element extraction (tagged PDFs)
+// =========================================================================
+
+/// One structure-tree element reference from a tagged PDF, resolved to a
+/// page and Marked Content ID.
+///
+/// Join `(page, mcid)` against [`TextItem::page`] / [`TextItem::mcid`] from
+/// [`extract_text_with_positions`] to attach semantic roles (heading levels,
+/// paragraphs, table cells, …) to extracted text.
+#[derive(Debug, Clone)]
+pub struct StructureElement {
+    /// 1-indexed page number (matches [`TextItem::page`]).
+    pub page: u32,
+    /// Marked Content ID from the page's content stream (matches
+    /// [`TextItem::mcid`]).
+    pub mcid: i64,
+    /// Standard structure type name ("H1".."H6", "P", "Table", "TD", …).
+    /// Custom tags are resolved through the document's `/RoleMap`; tags
+    /// with no standard mapping are returned verbatim.
+    pub role: String,
+}
+
+/// Extract structure-tree element references from a tagged PDF in memory.
+///
+/// Parses `/StructTreeRoot` (when present) and returns one entry per
+/// marked-content reference, resolved to its 1-indexed page, MCID, and
+/// structure type name. Returns an empty list when the PDF is not tagged.
+///
+/// Pass `Some(&[...])` with 1-indexed page numbers (matching
+/// [`TextItem::page`]) to restrict output to those pages; pass `None` for
+/// the whole document. Entries are sorted by `(page, mcid)`.
+pub fn extract_structure_elements_mem(
+    buffer: &[u8],
+    pages: Option<&[u32]>,
+) -> Result<Vec<StructureElement>, PdfError> {
+    validate_pdf_bytes(buffer)?;
+    let (doc, _page_count) = load_document_from_mem(buffer)?;
+    let Some(tree) = structure_tree::StructTree::from_doc(&doc) else {
+        return Ok(Vec::new());
+    };
+    let page_ids = doc.get_pages();
+    let roles = tree.mcid_to_roles(&page_ids);
+
+    let page_filter: Option<HashSet<u32>> = pages.map(|p| p.iter().copied().collect());
+    let mut elements: Vec<StructureElement> = roles
+        .into_iter()
+        .filter(|(page, _)| page_filter.as_ref().is_none_or(|f| f.contains(page)))
+        .flat_map(|(page, mcids)| {
+            mcids.into_iter().map(move |(mcid, role)| StructureElement {
+                page,
+                mcid,
+                role: role.name().to_string(),
+            })
+        })
+        .collect();
+    elements.sort_unstable_by_key(|e| (e.page, e.mcid));
+    Ok(elements)
+}
+
+/// Path-based wrapper for [`extract_structure_elements_mem`].
+///
+/// Reads the PDF from disk and extracts structure-tree element references.
+/// Pass `None` for `pages` to return the whole document, or `Some(&[...])`
+/// to restrict to specific 1-indexed pages.
+pub fn extract_structure_elements<P: AsRef<Path>>(
+    path: P,
+    pages: Option<&[u32]>,
+) -> Result<Vec<StructureElement>, PdfError> {
+    validate_pdf_file(&path)?;
+    let buffer = std::fs::read(path.as_ref())?;
+    extract_structure_elements_mem(&buffer, pages)
+}
+
 // =========================================================================
 // Region-based text extraction (for hybrid OCR pipelines)
 // =========================================================================
@@ -683,6 +908,23 @@ pub struct PageRegionResult {
     pub regions: Vec<RegionText>,
 }
 
+/// Minimum alphanumeric mass an invisible (Tr 3) text layer must carry for
+/// the OCR-layer fallback in [`extract_text_in_regions_mem`] to adopt it. A
+/// real OCR layer carries far more; a stray watermark or artifact does not.
+const OCR_LAYER_MIN_ALNUM: usize = 40;
+
+/// Alphanumeric mass of extracted items, ignoring raster placeholders.
+/// `[Image: ...]` items (ItemType::Image) are synthesized for image
+/// XObjects — they mark that pixels exist, not that text was read, so they
+/// must not count as coverage.
+fn non_placeholder_alnum(items: &[TextItem]) -> usize {
+    items
+        .iter()
+        .filter(|it| !matches!(it.item_type, types::ItemType::Image))
+        .map(|it| it.text.chars().filter(|c| c.is_alphanumeric()).count())
+        .sum()
+}
+
 /// Extract text within bounding-box regions from a PDF in memory.
 ///
 /// This is designed for hybrid OCR pipelines: a layout model detects regions
@@ -735,8 +977,11 @@ pub fn extract_text_in_regions_mem(
         let height = get_page_height(&doc, page_id).unwrap_or(792.0);
         page_heights.insert(*page_num, height);
 
-        // Extract text items for this page
-        let ((mut items, _rects, _lines), has_gid, coords_rotated) =
+        // Extract text items for this page. The Form XObject budget is shared
+        // with the invisible-layer retry below so one page cannot consume two
+        // full expansion budgets.
+        let mut form_budget = extractor::FormWalkBudget::new();
+        let ((mut items, _rects, _lines), mut has_gid, mut coords_rotated, skipped_invisible) =
             extractor::content_stream::extract_page_text_items(
                 &doc,
                 page_id,
@@ -744,7 +989,54 @@ pub fn extract_text_in_regions_mem(
                 &font_cmaps,
                 false,
                 &mut style_cache,
+                &mut form_budget,
             )?;
+        // OCR-layer fallback: scanned pages often carry their text as an
+        // invisible (Tr 3) layer behind the page raster. The visible-only
+        // pass sees nothing there but `[Image: ...]` placeholders, so every
+        // region on the page reports needs_ocr even though the exact text is
+        // embedded in the PDF — and this extractor then disagrees with the
+        // markdown path, which already retries Mixed PDFs with the invisible
+        // layer included. Retry page-scoped, and only when (a) the first
+        // pass actually SKIPPED invisible text — blank pages and image-only
+        // scans without an OCR layer must not pay a second content-stream
+        // parse (review catch) — and (b) the page has NO visible text item
+        // at all (punctuation counts, whitespace-only artifacts don't): an
+        // invisible OCR layer transcribes the raster, so any visible glyph
+        // has an invisible twin there and adoption would duplicate it
+        // (review catches — strict gate, no fuzzy dedupe). Adopt the retry
+        // only when it contributes real, non-garbage text.
+        let has_visible_text = items.iter().any(|it| {
+            !matches!(it.item_type, types::ItemType::Image) && !it.text.trim().is_empty()
+        });
+        if skipped_invisible && !has_visible_text {
+            if let Ok(((inv_items, _inv_rects, _inv_lines), inv_gid, inv_rotated, _)) =
+                extractor::content_stream::extract_page_text_items(
+                    &doc,
+                    page_id,
+                    *page_num,
+                    &font_cmaps,
+                    true,
+                    &mut style_cache,
+                    &mut form_budget,
+                )
+            {
+                let inv_alnum = non_placeholder_alnum(&inv_items);
+                // Judge the WHOLE recovered layer, not a prefix — a broken
+                // OCR layer can hide its garbage past any fixed sample size
+                // (review catch).
+                let sample: String = inv_items
+                    .iter()
+                    .filter(|it| !matches!(it.item_type, types::ItemType::Image))
+                    .map(|it| it.text.as_str())
+                    .collect();
+                if inv_alnum >= OCR_LAYER_MIN_ALNUM && !is_garbage_text(&sample) {
+                    items = inv_items;
+                    has_gid = inv_gid;
+                    coords_rotated = inv_rotated;
+                }
+            }
+        }
         let threshold = text_utils::fix_letterspaced_items(&mut items);
         if threshold > 0.10 {
             page_thresholds.insert(*page_num, threshold);
@@ -901,7 +1193,7 @@ pub fn extract_tables_in_regions_mem(
         let height = get_page_height(&doc, page_id).unwrap_or(792.0);
         page_heights.insert(*page_num, height);
 
-        let ((mut items, rects, lines), has_gid, coords_rotated) =
+        let ((mut items, rects, lines), has_gid, coords_rotated, _skipped_invisible) =
             extractor::content_stream::extract_page_text_items(
                 &doc,
                 page_id,
@@ -909,6 +1201,7 @@ pub fn extract_tables_in_regions_mem(
                 &font_cmaps,
                 false,
                 &mut style_cache,
+                &mut extractor::FormWalkBudget::new(),
             )?;
         let threshold = text_utils::fix_letterspaced_items(&mut items);
         if threshold > 0.10 {
@@ -1212,7 +1505,7 @@ pub fn detect_vector_grid_in_region_mem(
     let needed_pages = HashSet::from([page_1idx]);
     let font_cmaps = FontCMaps::from_doc_pages_fast(&doc, Some(&needed_pages));
     let page_h = get_page_height(&doc, page_id).unwrap_or(792.0);
-    let ((mut items, rects, lines), _has_gid, coords_rotated) =
+    let ((mut items, rects, lines), _has_gid, coords_rotated, _skipped_invisible) =
         extractor::content_stream::extract_page_text_items(
             &doc,
             page_id,
@@ -1220,6 +1513,7 @@ pub fn detect_vector_grid_in_region_mem(
             &font_cmaps,
             false,
             &mut extractor::FontStyleCache::new(),
+            &mut extractor::FormWalkBudget::new(),
         )?;
     text_utils::fix_letterspaced_items(&mut items);
 
@@ -1406,15 +1700,17 @@ mod vector_grid_tests {
         let &page_id = pages.get(&1).unwrap();
         let needed: HashSet<u32> = HashSet::from([1]);
         let cmaps = FontCMaps::from_doc_pages_fast(&doc, Some(&needed));
-        let ((items, rects, _lines), _has_gid, _rotated) = extract_page_text_items(
-            &doc,
-            page_id,
-            1,
-            &cmaps,
-            false,
-            &mut crate::extractor::FontStyleCache::new(),
-        )
-        .unwrap();
+        let ((items, rects, _lines), _has_gid, _rotated, _skipped_invisible) =
+            extract_page_text_items(
+                &doc,
+                page_id,
+                1,
+                &cmaps,
+                false,
+                &mut crate::extractor::FontStyleCache::new(),
+                &mut crate::extractor::FormWalkBudget::new(),
+            )
+            .unwrap();
 
         let (rect_tables, _) = detect_tables_from_rects(&items, &rects, 1);
         assert_eq!(rect_tables.len(), 1, "expected one rect-detected table");
@@ -1448,15 +1744,17 @@ mod vector_grid_tests {
         let &page_id = pages.get(&page_num).unwrap();
         let needed: HashSet<u32> = HashSet::from([page_num]);
         let cmaps = FontCMaps::from_doc_pages_fast(&doc, Some(&needed));
-        let ((items, rects, _lines), _has_gid, _rotated) = extract_page_text_items(
-            &doc,
-            page_id,
-            page_num,
-            &cmaps,
-            false,
-            &mut crate::extractor::FontStyleCache::new(),
-        )
-        .unwrap();
+        let ((items, rects, _lines), _has_gid, _rotated, _skipped_invisible) =
+            extract_page_text_items(
+                &doc,
+                page_id,
+                page_num,
+                &cmaps,
+                false,
+                &mut crate::extractor::FontStyleCache::new(),
+                &mut crate::extractor::FormWalkBudget::new(),
+            )
+            .unwrap();
 
         let (rect_tables, _) = detect_tables_from_rects(&items, &rects, page_num);
         rect_tables
@@ -2184,7 +2482,7 @@ pub fn extract_tables_with_structure_cells_mem(
         let height = get_page_height(&doc, page_id).unwrap_or(792.0);
         page_heights.insert(*page_num, height);
 
-        let ((mut items, _rects, _lines), _has_gid, coords_rotated) =
+        let ((mut items, _rects, _lines), _has_gid, coords_rotated, _skipped_invisible) =
             extractor::content_stream::extract_page_text_items(
                 &doc,
                 page_id,
@@ -2192,6 +2490,7 @@ pub fn extract_tables_with_structure_cells_mem(
                 &font_cmaps,
                 false,
                 &mut style_cache,
+                &mut extractor::FormWalkBudget::new(),
             )?;
         let threshold = text_utils::fix_letterspaced_items(&mut items);
         if threshold > 0.10 {
@@ -2986,7 +3285,7 @@ fn detect_tsr_quality_issue(
     let mut needed: HashSet<u32> = HashSet::new();
     needed.insert(page_1idx);
     let font_cmaps = FontCMaps::from_doc_pages_fast(&doc, Some(&needed));
-    let ((mut items, _rects, _lines), _has_gid, coords_rotated) =
+    let ((mut items, _rects, _lines), _has_gid, coords_rotated, _skipped_invisible) =
         extractor::content_stream::extract_page_text_items(
             &doc,
             page_id,
@@ -2994,6 +3293,7 @@ fn detect_tsr_quality_issue(
             &font_cmaps,
             false,
             &mut extractor::FontStyleCache::new(),
+            &mut extractor::FormWalkBudget::new(),
         )?;
     let adaptive_threshold = text_utils::fix_letterspaced_items(&mut items);
     let coords = if coords_rotated {

src/markdown/mod.rs

@@ -453,6 +453,110 @@ fn merged_retry_skips_body_font(detected_columns: bool, has_chart_regions: bool)
     detected_columns && !has_chart_regions
 }
 
+/// Identity of a piece of page furniture: the same trimmed text drawn at the
+/// same position (quantized to 0.5pt) — page numbers excluded by construction
+/// because their text differs per page.
+type FurnitureKey = (String, i32, i32);
+
+fn furniture_key(item: &TextItem) -> FurnitureKey {
+    (
+        item.text.trim().to_string(),
+        (item.x * 2.0).round() as i32,
+        (item.y * 2.0).round() as i32,
+    )
+}
+
+/// Minimum distinct pages an identical (text, position) must appear on before
+/// it counts as a running header/footer rather than coincidence.
+const RUNNING_FURNITURE_MIN_PAGES: usize = 3;
+
+/// Fraction of each page's vertical content extent, at the top and at the
+/// bottom, where running furniture may live. Repetition alone is not enough:
+/// a form template repeated per record carries identical labels at identical
+/// mid-page coordinates on every page, and those are real table cells. What
+/// makes a header/footer is repetition *at the page edge*.
+const RUNNING_FURNITURE_BAND: f32 = 0.2;
+
+/// Collect the keys of items that repeat verbatim at the same position on at
+/// least [`RUNNING_FURNITURE_MIN_PAGES`] distinct pages, restricted to the
+/// top/bottom [`RUNNING_FURNITURE_BAND`] of each page's content extent —
+/// running headers and footers. Single- and two-page documents produce an
+/// empty set.
+fn running_furniture_keys(items: &[TextItem]) -> HashSet<FurnitureKey> {
+    // Vertical content extent per page, so the edge bands adapt to the
+    // document's real margins instead of assuming a media box.
+    let mut page_extent: HashMap<u32, (f32, f32)> = HashMap::new();
+    for item in items {
+        if item.text.trim().is_empty() {
+            continue;
+        }
+        let entry = page_extent.entry(item.page).or_insert((item.y, item.y));
+        entry.0 = entry.0.min(item.y);
+        entry.1 = entry.1.max(item.y);
+    }
+
+    let mut pages_by_key: HashMap<FurnitureKey, HashSet<u32>> = HashMap::new();
+    for item in items {
+        if item.text.trim().is_empty() {
+            continue;
+        }
+        let Some(&(min_y, max_y)) = page_extent.get(&item.page) else {
+            continue;
+        };
+        // A page whose text has no vertical span gives no evidence of where
+        // its edges are — without this guard, a zero band would classify its
+        // every item as edge furniture.
+        let extent = max_y - min_y;
+        if extent <= 0.0 {
+            continue;
+        }
+        let band = extent * RUNNING_FURNITURE_BAND;
+        if item.y > min_y + band && item.y < max_y - band {
+            continue; // mid-page: never furniture, however often it repeats
+        }
+        pages_by_key
+            .entry(furniture_key(item))
+            .or_default()
+            .insert(item.page);
+    }
+    pages_by_key
+        .into_iter()
+        .filter(|(_, pages)| pages.len() >= RUNNING_FURNITURE_MIN_PAGES)
+        .map(|(key, _)| key)
+        .collect()
+}
+
+/// Reject a heuristic table whose items are almost entirely running
+/// headers/footers. A wrapped document title repeated at the bottom of every
+/// page aligns well enough to read as a grid, but it is page furniture, not
+/// data — vetoing the table lets the text flow as prose instead. Real tables
+/// carry per-page content, so even a repeated *header row* stays under the
+/// threshold once its body rows differ.
+fn is_running_furniture_table(
+    detection_items: &[TextItem],
+    table: &crate::tables::Table,
+    running: &HashSet<FurnitureKey>,
+) -> bool {
+    if running.is_empty() {
+        return false;
+    }
+    let mut total = 0usize;
+    let mut furniture = 0usize;
+    for &idx in &table.item_indices {
+        let Some(item) = detection_items.get(idx) else {
+            continue;
+        };
+        if item.text.trim().is_empty() {
+            continue;
+        }
+        total += 1;
+        if running.contains(&furniture_key(item)) {
+            furniture += 1;
+        }
+    }
+    total > 0 && (furniture as f32) >= (total as f32) * 0.8
+}
+
 /// Reject a heuristic table only when its cells are overwhelmingly parallel
 /// prose fragments. This is deliberately narrower than disabling body-font
 /// detection for the whole page: numeric, compact, headed, and otherwise
@@ -1040,6 +1144,14 @@ pub fn to_markdown(text: &str, options: MarkdownOptions) -> String {
     output
 }
 
+/// Applies the document-wide repeated header/footer classifier to grouped lines.
+pub(crate) fn strip_repeated_header_footer_lines(
+    lines: Vec<crate::types::TextLine>,
+    page_count: u32,
+) -> Vec<crate::types::TextLine> {
+    preprocess::strip_repeated_lines(lines, page_count)
+}
+
 /// Convert positioned text items to markdown with structure detection
 pub fn to_markdown_from_items(items: Vec<TextItem>, options: MarkdownOptions) -> String {
     to_markdown_from_items_with_rects(items, options, &[])
@@ -1188,6 +1300,12 @@ pub(crate) fn to_markdown_from_items_with_rects_and_lines(
     let mut table_items: HashSet<usize> = HashSet::new();
     let mut page_tables: HashMap<u32, Vec<PositionedMarkdown>> = HashMap::new();
 
+    // Running headers/footers repeat verbatim at the same position on many
+    // pages. When such a block wraps a long title over aligned lines, the
+    // heuristic detector reads it as a table. Knowing which items are page
+    // furniture is a document-wide question, so answer it once here.
+    let running_furniture = running_furniture_keys(&text_items);
+
     // Pre-group items by page with their global indices (O(n) instead of O(pages*n))
     let mut page_groups: HashMap<u32, Vec<(usize, &TextItem)>> = HashMap::new();
     for (global_idx, item) in text_items.iter().enumerate() {
@@ -1517,6 +1635,15 @@ pub(crate) fn to_markdown_from_items_with_rects_and_lines(
                             );
                             continue;
                         }
+                        if is_running_furniture_table(subset_items, &table, &running_furniture) {
+                            log::debug!(
+                                "page {}: rejected {}x{} running header/footer table hypothesis",
+                                page,
+                                table.rows.len(),
+                                table.columns.len()
+                            );
+                            continue;
+                        }
                         for &idx in &table.item_indices {
                             if let Some(&band_idx) = index_map.get(idx) {
                                 if let Some(&page_idx) = band_index_map.get(band_idx) {
@@ -1703,6 +1830,15 @@ pub(crate) fn to_markdown_from_items_with_rects_and_lines(
                     );
                     continue;
                 }
+                if is_running_furniture_table(&chart_free, table, &running_furniture) {
+                    log::debug!(
+                        "page {}: rejected {}x{} merged-band running header/footer table hypothesis",
+                        page,
+                        table.rows.len(),
+                        table.columns.len()
+                    );
+                    continue;
+                }
                 for &idx in &table.item_indices {
                     if let Some(&page_idx) = chart_free_map
                         .get(idx)
@@ -2058,6 +2194,174 @@ mod tests {
         assert!(md.contains("- Second item"));
     }
 
+    fn furniture_item(text: &str, x: f32, y: f32, page: u32) -> TextItem {
+        let mut it = make_item(x, y, page);
+        it.text = text.into();
+        it
+    }
+
+    /// Items repeating verbatim at the same position on 3+ pages are running
+    /// furniture; the same text on fewer pages, or at different positions, is
+    /// not.
+    #[test]
+    fn running_furniture_requires_three_pages_at_same_position() {
+        let mut items = Vec::new();
+        for page in 1..=3 {
+            // Body content so each page has a real vertical extent.
+            items.push(furniture_item("body", 85.0, 700.0, page));
+            items.push(furniture_item("TITULAR DEL", 85.0, 68.0, page));
+        }
+        // Same text but only two pages.
+        for page in 1..=2 {
+            items.push(furniture_item("SECRETARÍA", 200.0, 68.0, page));
+        }
+        // Same text on three pages but at drifting positions.
+        for (page, x) in [(1, 300.0), (2, 320.0), (3, 340.0)] {
+            items.push(furniture_item("MÉXICO", x, 68.0, page));
+        }
+
+        let running = running_furniture_keys(&items);
+        assert!(running.contains(&furniture_key(&furniture_item(
+            "TITULAR DEL",
+            85.0,
+            68.0,
+            1
+        ))));
+        assert!(!running.contains(&furniture_key(&furniture_item(
+            "SECRETARÍA",
+            200.0,
+            68.0,
+            1
+        ))));
+        assert!(!running.contains(&furniture_key(&furniture_item("MÉXICO", 300.0, 68.0, 1))));
+    }
+
+    /// A table made of running-footer items is vetoed; a table whose body rows
+    /// carry per-page content is kept even when its header row repeats.
+    #[test]
+    fn running_furniture_table_veto() {
+        // The footer block, present identically on pages 1-3.
+        let mut items = Vec::new();
+        for page in 1..=3 {
+            items.push(furniture_item("PROPOSICIÓN CON PUNTO", 85.0, 78.5, page));
+            items.push(furniture_item("EL SENADO", 286.6, 78.5, page));
+            items.push(furniture_item("TITULAR DEL", 85.0, 68.0, page));
+            items.push(furniture_item("A TRAVÉS DE LA", 243.4, 68.0, page));
+        }
+        // A real table on page 1: repeated header row, per-page data rows.
+        let header = [
+            furniture_item("Year", 85.0, 500.0, 1),
+            furniture_item("Total", 200.0, 500.0, 1),
+        ];
+        let data = [
+            furniture_item("2023", 85.0, 488.0, 1),
+            furniture_item("1,204", 200.0, 488.0, 1),
+            furniture_item("2024", 85.0, 476.0, 1),
+            furniture_item("1,377", 200.0, 476.0, 1),
+        ];
+        // Header repeats on every page (like a continued table's header).
+        for page in 2..=3 {
+            items.push(furniture_item("Year", 85.0, 500.0, page));
+            items.push(furniture_item("Total", 200.0, 500.0, page));
+        }
+        items.extend(header.iter().cloned());
+        items.extend(data.iter().cloned());
+
+        let running = running_furniture_keys(&items);
+
+        let table_of = |detection_items: &[TextItem]| crate::tables::Table {
+            columns: vec![],
+            rows: vec![],
+            cells: vec![],
+            item_indices: (0..detection_items.len()).collect(),
+            kind: crate::tables::TableKind::Data,
+        };
+
+        // Footer-only candidate: every item is furniture -> vetoed.
+        let footer_items: Vec<TextItem> = (1..=1)
+            .flat_map(|page| {
+                vec![
+                    furniture_item("PROPOSICIÓN CON PUNTO", 85.0, 78.5, page),
+                    furniture_item("EL SENADO", 286.6, 78.5, page),
+                    furniture_item("TITULAR DEL", 85.0, 68.0, page),
+                    furniture_item("A TRAVÉS DE LA", 243.4, 68.0, page),
+                ]
+            })
+            .collect();
+        assert!(is_running_furniture_table(
+            &footer_items,
+            &table_of(&footer_items),
+            &running
+        ));
+
+        // Real table: header row repeats across pages, body rows do not ->
+        // 2 furniture of 6 items (33%) stays under the 80% threshold.
+        let real_items: Vec<TextItem> =
+            header.iter().cloned().chain(data.iter().cloned()).collect();
+        assert!(!is_running_furniture_table(
+            &real_items,
+            &table_of(&real_items),
+            &running
+        ));
+    }
+
+    /// A form template repeated per record carries identical labels at
+    /// identical mid-page coordinates on every page — those are real table
+    /// cells, not furniture. Only the page-edge bands qualify.
+    #[test]
+    fn mid_page_repetition_is_not_furniture() {
+        let mut items = Vec::new();
+        for page in 1..=4 {
+            // Content spanning the page: y 60 (bottom) to 740 (top).
+            items.push(furniture_item("body top", 85.0, 740.0, page));
+            items.push(furniture_item("body bottom", 85.0, 60.0, page));
+            // Form labels repeated dead centre on every page.
+            items.push(furniture_item("Name of creditor", 85.0, 400.0, page));
+            items.push(furniture_item("Amount of claim", 300.0, 400.0, page));
+            // A genuine footer inside the bottom band.
+            items.push(furniture_item("FORM 78 — page footer", 85.0, 70.0, page));
+        }
+
+        let running = running_furniture_keys(&items);
+        assert!(
+            !running.contains(&furniture_key(&furniture_item(
+                "Name of creditor",
+                85.0,
+                400.0,
+                1
+            ))),
+            "mid-page form labels must not be furniture"
+        );
+        assert!(running.contains(&furniture_key(&furniture_item(
+            "FORM 78 — page footer",
+            85.0,
+            70.0,
+            1
+        ))));
+    }
+
+    #[test]
+    fn running_furniture_empty_on_short_documents() {
+        let mut items = Vec::new();
+        for page in 1..=2 {
+            items.push(furniture_item("body", 85.0, 700.0, page));
+            items.push(furniture_item("FOOTER", 85.0, 68.0, page));
+        }
+        assert!(running_furniture_keys(&items).is_empty());
+    }
+
+    /// A page whose text has no vertical span (a single line) gives no
+    /// evidence of where its edges are; its items never become furniture.
+    #[test]
+    fn zero_span_page_contributes_no_furniture() {
+        let mut items = Vec::new();
+        for page in 1..=4 {
+            items.push(furniture_item("ROW LABEL", 85.0, 400.0, page));
+            items.push(furniture_item("ROW VALUE", 300.0, 400.0, page));
+        }
+        assert!(running_furniture_keys(&items).is_empty());
+    }
+
     fn make_item(x: f32, y: f32, page: u32) -> TextItem {
         TextItem {
             text: "A".into(),

src/markdown/postprocess.rs

@@ -13,7 +13,11 @@ pub(crate) fn clean_markdown(mut text: String, options: &MarkdownOptions) -> Str
         text = collapse_dot_leaders(&text);
     }
 
-    // Fix hyphenation first (before other processing)
+    // Collapse runs of spaces first: double-spaced breaks ("de-  fendant")
+    // must look like single-spaced ones before the hyphenation passes.
+    collapse_consecutive_spaces(&mut text);
+
+    // Fix hyphenation (before other processing)
     if options.fix_hyphenation {
         text = fix_hyphenation(&text);
     }
@@ -143,7 +147,213 @@ fn fix_hyphenation(text: &str) -> String {
         })
         .to_string();
 
-    result
+    dehyphenate_line_breaks(&result)
+}
+
+/// What a line-break hyphen pair should become. Policy output only — how the
+/// decision is rendered (plain text vs. inside split emphasis markers) is the
+/// caller's business.
+#[derive(Clone, Copy, PartialEq, Eq, Debug)]
+enum Join {
+    /// The fragments are one word: "de- fendant" -> "defendant".
+    Plain,
+    /// The fragments are a hyphenated compound: "six- month" -> "six-month".
+    Hyphen,
+    /// No evidence (or a suspended hyphen): leave the break as it is.
+    Keep,
+}
+
+/// Decide what a line-break hyphen pair becomes, given the document's
+/// vocabulary evidence. The whole dehyphenation policy lives here so it can
+/// be reasoned about (and tested) apart from the Markdown scanning around it;
+/// see [`dehyphenate_line_breaks`] for the rule rationale.
+fn join_decision(
+    a: &str,
+    b: &str,
+    words: &std::collections::HashSet<String>,
+    hyphenated: &std::collections::HashSet<String>,
+) -> Join {
+    // Word-length sanity: syllable fragments are short, and even long German
+    // compounds stay under this. Fragments beyond it are already fused
+    // reading-order noise (interleaved columns); joining would compound the
+    // damage.
+    if a.chars().count() + b.chars().count() > 40 {
+        return Join::Keep;
+    }
+    let key_plain = format!("{}{}", a.to_lowercase(), b.to_lowercase());
+    let key_hyphen = format!("{}-{}", a.to_lowercase(), b.to_lowercase());
+    if words.contains(&key_plain) {
+        Join::Plain
+    } else if hyphenated.contains(&key_hyphen) || b.chars().next().is_some_and(|c| c.is_uppercase())
+    {
+        Join::Hyphen
+    } else if b.chars().count() >= 4
+        && words.contains(&a.to_lowercase())
+        && words.contains(&b.to_lowercase())
+    {
+        // No direct evidence, but both fragments are themselves words the
+        // document uses ("commercial- type"): hyphenated compounds are made
+        // of words, while syllable fragments ("evi", "judg", "mo") are not.
+        //
+        // The continuation must be at least four letters. Suspended hyphens
+        // ("mid- and long-term", "klein- und mittelgroß") put a conjunction
+        // after the hyphen, and conjunctions are near-universally one to
+        // three letters in any language — the length floor keeps this rule
+        // off them without a hard-coded conjunction list.
+        Join::Hyphen
+    } else {
+        // No evidence at all: leave the break as it is. An unconditional join
+        // here covered only ~1% more breaks on a vocabulary-rich document,
+        // but it was the sole rule able to corrupt output — fusing
+        // interleaved-column fragments ("com- real" -> "comreal") into
+        // unrecoverable tokens. A visible break is honest; a silent fusion
+        // is not.
+        Join::Keep
+    }
+}
+
+/// Rejoin words hyphenated at the original line breaks.
+///
+/// Justified print breaks words at syllables; after paragraph lines are
+/// joined with spaces those breaks survive as "de- fendant" (and, when an
+/// emphasis span was split with the word, "Bap-** **tist"). Whether the
+/// hyphen itself belongs in the word cannot be decided locally — "de-
+/// fendant" is one word but "Third- Party" is a hyphenated compound — so the
+/// document is its own dictionary:
+///
+///   1. fragments appear elsewhere joined plain ("defendant") — join plain;
+///   2. appear elsewhere hyphenated ("six-month"), or the continuation is
+///      capitalized ("Hinds- Radix", "Third- Party") — keep the hyphen;
+///   3. both fragments are words the document uses and the continuation
+///      has four or more letters ("commercial- type" where "commercial"
+///      and "type" appear elsewhere) — a compound, keep the hyphen. The
+///      length floor keeps this rule off suspended hyphens ("mid- and
+///      long-term", "klein- und mittelgroß"): conjunctions are one to
+///      three letters in essentially every language, so no conjunction
+///      list is needed;
+///   4. no evidence — leave the break untouched. Evidence covers ~99% of
+///      breaks on vocabulary-rich documents, and an unconditional join was
+///      the one rule able to corrupt output (fusing interleaved-column
+///      fragments into unrecoverable tokens).
+///
+/// Every rule is either document evidence or script-agnostic typography;
+/// deliberately no hard-coded word lists beyond the three suspension
+/// conjunctions (a curated suffix list was tried and removed — it was
+/// English-only, its membership was unfalsifiable, and it could invent
+/// hyphens: "proto- type" -> "proto-type").
+///
+/// Table rows and fenced code blocks are left untouched.
+fn dehyphenate_line_breaks(text: &str) -> String {
+    use once_cell::sync::Lazy;
+    use std::collections::HashSet;
+
+    const WORD: &str = r"\p{L}";
+
+    // "de- fendant"
+    static BREAK_RE: Lazy<Regex> =
+        Lazy::new(|| Regex::new(&format!("({WORD}{{2,}})- ({WORD}{{2,}})")).unwrap());
+    // "Bap-** **tist" — an emphasis span split together with the word. The
+    // regex crate has no backreferences, so both markers are captured and
+    // compared in the replacement closure.
+    static BREAK_EMPH_RE: Lazy<Regex> = Lazy::new(|| {
+        Regex::new(&format!(
+            r"({WORD}{{2,}})-(\*{{1,2}}) (\*{{1,2}})({WORD}{{2,}})"
+        ))
+        .unwrap()
+    });
+    static PLAIN_WORD_RE: Lazy<Regex> = Lazy::new(|| Regex::new(&format!("{WORD}{{3,}}")).unwrap());
+    static HYPHENATED_RE: Lazy<Regex> =
+        Lazy::new(|| Regex::new(&format!("({WORD}{{2,}})-({WORD}{{2,}})")).unwrap());
+
+    // The document is its own dictionary: whole words and hyphenated
+    // compounds as they appear away from line breaks. Two exclusions keep the
+    // evidence sound:
+    //   - the break pairs themselves are scrubbed first, otherwise every
+    //     broken word donates its own fragments ("evi", "dence") and the
+    //     compound rule would see them as words;
+    //   - fenced code blocks are skipped: identifiers are a different
+    //     language, and one like `comreal` must not justify fusing prose.
+    //     Table rows stay — cells carry genuine document vocabulary.
+    let mut prose = String::with_capacity(text.len());
+    let mut in_code = false;
+    for line in text.split('\n') {
+        if line.trim_start().starts_with("```") {
+            in_code = !in_code;
+            continue;
+        }
+        if !in_code {
+            prose.push_str(line);
+            prose.push('\n');
+        }
+    }
+    let scrubbed = BREAK_EMPH_RE.replace_all(&prose, " ");
+    let scrubbed = BREAK_RE.replace_all(&scrubbed, " ");
+    let mut words: HashSet<String> = HashSet::new();
+    let mut hyphenated: HashSet<String> = HashSet::new();
+    for m in PLAIN_WORD_RE.find_iter(&scrubbed) {
+        words.insert(m.as_str().to_lowercase());
+    }
+    for caps in HYPHENATED_RE.captures_iter(&scrubbed) {
+        hyphenated.insert(format!(
+            "{}-{}",
+            caps[1].to_lowercase(),
+            caps[2].to_lowercase()
+        ));
+    }
+
+    let join = |a: &str, b: &str| join_decision(a, b, &words, &hyphenated);
+
+    let mut in_code_block = false;
+    let mut out = String::with_capacity(text.len());
+    for (i, line) in text.split('\n').enumerate() {
+        if i > 0 {
+            out.push('\n');
+        }
+        if line.trim_start().starts_with("```") {
+            in_code_block = !in_code_block;
+        }
+        if in_code_block || line.trim_start().starts_with('|') {
+            out.push_str(line);
+            continue;
+        }
+        // A break can chain ("unconsti- tu- tional"); each pass joins one
+        // junction, and every successful join removes a break, so running
+        // until the line is stable is bounded by the number of breaks.
+        let mut current = line.to_string();
+        loop {
+            let next = BREAK_EMPH_RE
+                .replace_all(&current, |caps: &regex::Captures| {
+                    // Mismatched markers aren't a split span; a Keep decision
+                    // preserves the original spacing and markers.
+                    if caps[2] != caps[3] {
+                        return caps[0].to_string();
+                    }
+                    let (a, b) = (&caps[1], &caps[4]);
+                    match join(a, b) {
+                        Join::Plain => format!("{a}{b}"),
+                        Join::Hyphen => format!("{a}-{b}"),
+                        Join::Keep => caps[0].to_string(),
+                    }
+                })
+                .to_string();
+            let next = BREAK_RE
+                .replace_all(&next, |caps: &regex::Captures| {
+                    let (a, b) = (&caps[1], &caps[2]);
+                    match join(a, b) {
+                        Join::Plain => format!("{a}{b}"),
+                        Join::Hyphen => format!("{a}-{b}"),
+                        Join::Keep => caps[0].to_string(),
+                    }
+                })
+                .to_string();
+            if next == current {
+                break;
+            }
+            current = next;
+        }
+        out.push_str(&current);
+    }
+    out
 }
 
 /// Remove isolated page-number expressions from Markdown.
@@ -384,6 +594,180 @@ mod tests {
         assert_eq!(t, "version 3 .14 released");
     }
 
+    // --- dehyphenate_line_breaks ---
+
+    #[test]
+    fn line_break_joins_plain_on_vocabulary_evidence() {
+        // "defendant" appears whole elsewhere, so the broken form joins plain.
+        let text = "The defendant appeared. The de- fendant argued.";
+        assert_eq!(
+            dehyphenate_line_breaks(text),
+            "The defendant appeared. The defendant argued."
+        );
+    }
+
+    #[test]
+    fn line_break_keeps_hyphen_on_hyphenated_evidence() {
+        // "six-month" appears hyphenated elsewhere, so the broken form keeps it.
+        let text = "A six-month term. After a six- month delay.";
+        assert_eq!(
+            dehyphenate_line_breaks(text),
+            "A six-month term. After a six-month delay."
+        );
+    }
+
+    #[test]
+    fn no_evidence_leaves_the_break_untouched() {
+        // Without document evidence a join cannot be distinguished from
+        // interleaved-column noise; the visible break is kept.
+        let text = "The evi- dence was clear.";
+        assert_eq!(dehyphenate_line_breaks(text), text);
+    }
+
+    #[test]
+    fn line_break_keeps_hyphen_before_capitalized_continuation() {
+        // Broken compounds: "Third-Party", "Hinds-Radix".
+        let text = "The Third- Party complaint by Hinds- Radix.";
+        assert_eq!(
+            dehyphenate_line_breaks(text),
+            "The Third-Party complaint by Hinds-Radix."
+        );
+    }
+
+    #[test]
+    fn cyrillic_words_join_on_evidence() {
+        // The word class is Unicode-wide, not a hard-coded Latin subset.
+        let text = "Это решение важно. Это реше- ние суда.";
+        assert_eq!(
+            dehyphenate_line_breaks(text),
+            "Это решение важно. Это решение суда."
+        );
+    }
+
+    #[test]
+    fn double_spaced_breaks_join_through_clean_markdown() {
+        // Space collapsing runs before hyphenation, so a break that arrives
+        // with two spaces ("de-  fendant") still rejoins.
+        let options = MarkdownOptions::default();
+        let out = clean_markdown(
+            "The defendant appeared. The de-  fendant argued.".to_string(),
+            &options,
+        );
+        assert_eq!(
+            out.trim_end(),
+            "The defendant appeared. The defendant argued."
+        );
+    }
+
+    #[test]
+    fn code_block_identifiers_are_not_vocabulary_evidence() {
+        // A fused identifier in code must not justify fusing unrelated prose.
+        let text = "```\nlet comreal = 1;\n```\nThe com- real estate story.";
+        assert_eq!(dehyphenate_line_breaks(text), text);
+    }
+
+    #[test]
+    fn fused_column_noise_is_not_joined() {
+        // Interleaved-column garbage arrives already fused; joining across
+        // its breaks would compound the damage. Real syllable fragments are
+        // short; fragments this long are left exactly as they are.
+        let text = "spreadswerenegativeintheearlytomid- seriouslyflawedduetoappraisallags";
+        assert_eq!(dehyphenate_line_breaks(text), text);
+        // Long German compounds stay under the length gate and join on
+        // vocabulary evidence.
+        let german =
+            "Das Bundesausbildungsförderungsgesetz. Das Bundesausbildungsförderungs- gesetz gilt.";
+        assert_eq!(
+            dehyphenate_line_breaks(german),
+            "Das Bundesausbildungsförderungsgesetz. Das Bundesausbildungsförderungsgesetz gilt."
+        );
+    }
+
+    #[test]
+    fn no_evidence_compounds_stay_visibly_broken() {
+        // No hard-coded suffix list: without document evidence even a likely
+        // compound keeps its visible break. A curated list was tried and
+        // removed — English-only, unfalsifiable membership, and able to
+        // invent hyphens ("proto- type" -> "proto-type").
+        let text = "Their world- class support and proto- type systems.";
+        assert_eq!(dehyphenate_line_breaks(text), text);
+    }
+
+    #[test]
+    fn both_fragments_being_words_keeps_the_hyphen() {
+        // "commercial-type insurance": no evidence either way, but both
+        // fragments are words the document uses, so this is a compound.
+        let text = "Any commercial firm of this type offering commercial- type insurance.";
+        assert_eq!(
+            dehyphenate_line_breaks(text),
+            "Any commercial firm of this type offering commercial-type insurance."
+        );
+    }
+
+    #[test]
+    fn suspended_hyphen_is_preserved() {
+        // "mid- to long-term": joining would fuse unrelated words. No
+        // conjunction list is involved — conjunctions are 1-3 letters in
+        // essentially every language, and the compound rule requires a
+        // 4-letter continuation, so suspended hyphens fall through to Keep.
+        let text = "Planned over the mid- to long-term horizon, in- and out-of-possession.";
+        assert_eq!(dehyphenate_line_breaks(text), text);
+        // Same construction in German, which a hard-coded English list
+        // would have missed. "klein" appears standalone so it IS in the
+        // vocabulary — only the length floor (continuation "und" has three
+        // letters) keeps the compound rule from fusing "klein-und".
+        let german = "Das klein geschriebene Wort und die klein- und mittelgroßen Betriebe.";
+        assert_eq!(dehyphenate_line_breaks(german), german);
+    }
+
+    #[test]
+    fn split_emphasis_span_joins_inside_markers() {
+        // Vocabulary evidence ("Baptist", "Consolidated" elsewhere) drives
+        // the join; the split emphasis markers collapse with it.
+        let text = "The Baptist and Consolidated cases. By **Bap-** **tist** pastors and *Consoli-* *dated* Edison.";
+        assert_eq!(
+            dehyphenate_line_breaks(text),
+            "The Baptist and Consolidated cases. By **Baptist** pastors and *Consolidated* Edison."
+        );
+    }
+
+    #[test]
+    fn mismatched_emphasis_markers_are_left_alone() {
+        let text = "Odd **Bap-** *tist* markers.";
+        assert_eq!(dehyphenate_line_breaks(text), text);
+    }
+
+    #[test]
+    fn chained_breaks_join_stepwise_with_evidence() {
+        // A word broken twice joins across passes when each junction has
+        // vocabulary evidence for its intermediate form.
+        let text = "The word unconstitutional, and unconstitu appears too: unconsti- tu- tional.";
+        assert_eq!(
+            dehyphenate_line_breaks(text),
+            "The word unconstitutional, and unconstitu appears too: unconstitutional."
+        );
+    }
+
+    #[test]
+    fn table_rows_and_code_blocks_are_untouched() {
+        let text =
+            "The defendant.\n|de- fendant|value|\n```\nlet x = de- fendant;\n```\nThe de- fendant won.";
+        assert_eq!(
+            dehyphenate_line_breaks(text),
+            "The defendant.\n|de- fendant|value|\n```\nlet x = de- fendant;\n```\nThe defendant won."
+        );
+    }
+
+    #[test]
+    fn accented_words_join() {
+        // Spanish syllable break with accented continuation, evidence-backed.
+        let text = "Una resolución firme. La resolu- ción fue clara.";
+        assert_eq!(
+            dehyphenate_line_breaks(text),
+            "Una resolución firme. La resolución fue clara."
+        );
+    }
+
     // --- fix_hyphenation ---
 
     #[test]

src/python.rs

@@ -271,6 +271,12 @@ pub struct PyTextItem {
     pub is_strikeout: bool,
     #[pyo3(get)]
     pub item_type: String,
+    /// Marked Content ID from the content stream's BDC/BMC operator, None
+    /// when the text is not part of marked content. Join with the
+    /// (page, mcid) pairs from extract_structure_elements to attach
+    /// structure-tree roles (headings, paragraphs, ...) in tagged PDFs.
+    #[pyo3(get)]
+    pub mcid: Option<i64>,
 }
 
 #[pymethods]
@@ -286,6 +292,32 @@ impl PyTextItem {
     }
 }
 
+/// One structure-tree element reference from a tagged PDF.
+#[pyclass(name = "StructureElement")]
+#[derive(Clone)]
+pub struct PyStructureElement {
+    /// 1-indexed page number (matches TextItem.page).
+    #[pyo3(get)]
+    pub page: u32,
+    /// Marked Content ID from the page's content stream (matches
+    /// TextItem.mcid).
+    #[pyo3(get)]
+    pub mcid: i64,
+    /// Standard structure type name ("H1".."H6", "P", "Table", "TD", ...).
+    #[pyo3(get)]
+    pub role: String,
+}
+
+#[pymethods]
+impl PyStructureElement {
+    fn __repr__(&self) -> String {
+        format!(
+            "StructureElement(page={}, mcid={}, role='{}')",
+            self.page, self.mcid, self.role
+        )
+    }
+}
+
 // ---------------------------------------------------------------------------
 // Helpers
 // ---------------------------------------------------------------------------
@@ -356,6 +388,18 @@ fn convert_text_items(items: Vec<crate::TextItem>) -> Vec<PyTextItem> {
             is_underline: item.is_underline,
             is_strikeout: item.is_strikeout,
             item_type: item_type_str(&item.item_type),
+            mcid: item.mcid,
+        })
+        .collect()
+}
+
+fn convert_structure_elements(elements: Vec<crate::StructureElement>) -> Vec<PyStructureElement> {
+    elements
+        .into_iter()
+        .map(|e| PyStructureElement {
+            page: e.page,
+            mcid: e.mcid,
+            role: e.role,
         })
         .collect()
 }
@@ -613,6 +657,48 @@ fn extract_pages_markdown_bytes(
     Ok(to_py_pages_result(result))
 }
 
+/// Extract structure-tree element references from a tagged PDF file.
+///
+/// Parses the document's structure tree (when present) and returns one
+/// entry per marked-content reference, resolved to its 1-indexed page,
+/// MCID, and structure type name ("H1".."H6", "P", "Table", ...). Returns
+/// an empty list when the PDF is not tagged.
+///
+/// Join (page, mcid) against the page/mcid attributes from
+/// [`extract_text_with_positions`] to attach heading levels and other
+/// semantic roles to extracted text.
+///
+/// Args:
+///     path: Path to the PDF file.
+///     pages: Optional list of 1-indexed pages (matching TextItem.page).
+///         When None (default), the whole document is returned.
+///
+/// Returns:
+///     List of StructureElement sorted by (page, mcid).
+#[pyfunction]
+#[pyo3(signature = (path, pages=None))]
+fn extract_structure_elements(
+    path: &str,
+    pages: Option<Vec<u32>>,
+) -> PyResult<Vec<PyStructureElement>> {
+    let elements = crate::extract_structure_elements(path, pages.as_deref()).map_err(to_py_err)?;
+    Ok(convert_structure_elements(elements))
+}
+
+/// Extract structure-tree element references from tagged PDF bytes.
+///
+/// See [`extract_structure_elements`] for details.
+#[pyfunction]
+#[pyo3(signature = (data, pages=None))]
+fn extract_structure_elements_bytes(
+    data: &[u8],
+    pages: Option<Vec<u32>>,
+) -> PyResult<Vec<PyStructureElement>> {
+    let elements =
+        crate::extract_structure_elements_mem(data, pages.as_deref()).map_err(to_py_err)?;
+    Ok(convert_structure_elements(elements))
+}
+
 /// Python module definition.
 #[pymodule]
 fn pdf_inspector(m: &Bound<'_, PyModule>) -> PyResult<()> {
@@ -620,6 +706,7 @@ fn pdf_inspector(m: &Bound<'_, PyModule>) -> PyResult<()> {
     m.add_class::<PyPageOcrReasons>()?;
     m.add_class::<PyPdfClassification>()?;
     m.add_class::<PyTextItem>()?;
+    m.add_class::<PyStructureElement>()?;
     m.add_class::<PyRegionText>()?;
     m.add_class::<PyPageRegionTexts>()?;
     m.add_class::<PyPageMarkdown>()?;
@@ -634,6 +721,8 @@ fn pdf_inspector(m: &Bound<'_, PyModule>) -> PyResult<()> {
     m.add_function(wrap_pyfunction!(extract_text_bytes, m)?)?;
     m.add_function(wrap_pyfunction!(extract_text_with_positions, m)?)?;
     m.add_function(wrap_pyfunction!(extract_text_with_positions_bytes, m)?)?;
+    m.add_function(wrap_pyfunction!(extract_structure_elements, m)?)?;
+    m.add_function(wrap_pyfunction!(extract_structure_elements_bytes, m)?)?;
     m.add_function(wrap_pyfunction!(extract_text_in_regions, m)?)?;
     m.add_function(wrap_pyfunction!(extract_text_in_regions_bytes, m)?)?;
     m.add_function(wrap_pyfunction!(extract_pages_markdown, m)?)?;

src/tables/detect_rects.rs

@@ -1,6 +1,6 @@
 //! Rectangle-based table detection using union-find clustering.
 
-use std::collections::HashMap;
+use std::collections::{BTreeMap, HashMap, HashSet};
 
 use log::debug;
 
@@ -78,19 +78,111 @@ pub(crate) fn rects_overlap(a: &(f32, f32, f32, f32), b: &(f32, f32, f32, f32),
     !(a_right < b_left || b_right < a_left || a_top < b_bottom || b_top < a_bottom)
 }
 
+fn grid_coord(value: f32, cell: f32) -> i32 {
+    (value / cell).floor().clamp(-1_000_000.0, 1_000_000.0) as i32
+}
+
+/// Inclusive grid range. `None` if the rect covers more cells than we will
+/// materialize — those rects are clustered via a bounded fallback.
+fn grid_span(lo: f32, hi: f32, cell: f32) -> Option<std::ops::RangeInclusive<i32>> {
+    let a = grid_coord(lo.min(hi), cell);
+    let b = grid_coord(lo.max(hi), cell);
+    let span = b.saturating_sub(a);
+    if span > 64 {
+        return None;
+    }
+    Some(a..=b)
+}
+
+fn union_bucket_pairs(
+    uf: &mut UnionFind,
+    rects: &[(f32, f32, f32, f32)],
+    bucket: &[usize],
+    tolerance: f32,
+) {
+    let m = bucket.len();
+    let mut pairs = 0usize;
+    'cell: for a in 0..m {
+        let i = bucket[a];
+        if uf.component_size(i) >= MAX_CLUSTER_RECTS {
+            continue;
+        }
+        for &j in &bucket[a + 1..] {
+            if pairs >= MAX_CLUSTER_PAIRS_PER_CELL {
+                break 'cell;
+            }
+            if uf.component_size(j) >= MAX_CLUSTER_RECTS {
+                continue;
+            }
+            pairs += 1;
+            if rects_overlap(&rects[i], &rects[j], tolerance) {
+                uf.union(i, j);
+                if uf.component_size(i) >= MAX_CLUSTER_RECTS {
+                    break;
+                }
+            }
+        }
+    }
+}
+
+fn union_rect_against_bands(
+    uf: &mut UnionFind,
+    rects: &[(f32, f32, f32, f32)],
+    i: usize,
+    bands: &BTreeMap<i32, Vec<usize>>,
+    lo: i32,
+    hi: i32,
+    tolerance: f32,
+) {
+    if uf.component_size(i) >= MAX_CLUSTER_RECTS {
+        return;
+    }
+    let mut pairs = 0usize;
+    let mut seen = HashSet::new();
+    for (_, bucket) in bands.range(lo..=hi) {
+        for &j in bucket {
+            if !seen.insert(j) {
+                continue;
+            }
+            if pairs >= MAX_CLUSTER_PAIRS_PER_CELL {
+                return;
+            }
+            if i == j || uf.component_size(j) >= MAX_CLUSTER_RECTS {
+                continue;
+            }
+            pairs += 1;
+            if rects_overlap(&rects[i], &rects[j], tolerance) {
+                uf.union(i, j);
+                if uf.component_size(i) >= MAX_CLUSTER_RECTS {
+                    return;
+                }
+            }
+        }
+    }
+}
+
 /// Maximum component size for rect clustering.  No real table has thousands
 /// of cell rects — once a component exceeds this, it is a vector drawing or
 /// page-spanning clipping path.  We skip overlap checks for rects already in
-/// an oversized component, keeping the original O(n²) loop but making it
-/// effectively O(n) for pathological pages.
+/// an oversized component.
 const MAX_CLUSTER_RECTS: usize = 2000;
 
+/// Pairwise-disjoint rects never merge, so a component-size cap does not
+/// stop an all-pairs loop. Rects are hashed into this many points of grid
+/// and compared only against others in the same cell.
+const CLUSTER_GRID_CELL: f32 = 64.0;
+
+/// All-pairs AABB tests allowed inside one grid cell. A real table cell is
+/// tens of points wide, so a 64-pt cell holds a handful of neighbors — not
+/// thousands of stacked drawings.
+const MAX_CLUSTER_PAIRS_PER_CELL: usize = 16_384;
+
 /// Cluster rects by spatial overlap using union-find.
 /// Returns groups of rect indices; only groups with ≥ `min_size` rects are returned.
 ///
-/// Skips overlap checks for rects whose component has already exceeded
-/// [`MAX_CLUSTER_RECTS`], so pages with tens of thousands of vector-drawing
-/// rects complete in milliseconds instead of minutes.
+/// Overlap tests run inside a uniform grid so far-apart rects are never
+/// compared, and each cell is pair-capped so a dense stack cannot go
+/// quadratic or starve an independent table in another cell.
 pub(crate) fn cluster_rects(
     rects: &[(f32, f32, f32, f32)],
     tolerance: f32,
@@ -98,23 +190,144 @@ pub(crate) fn cluster_rects(
 ) -> Vec<Vec<usize>> {
     let n = rects.len();
     let mut uf = UnionFind::new(n);
+    let cell = CLUSTER_GRID_CELL.max(tolerance * 4.0);
 
-    for i in 0..n {
-        // If rect i is already in an oversized component, no point comparing
-        // it against further rects — the component won't be used for table
-        // detection anyway.
+    let mut grid: HashMap<(i32, i32), Vec<usize>> = HashMap::new();
+    let mut large: Vec<usize> = Vec::new();
+    for (idx, &(x, y, w, h)) in rects.iter().enumerate() {
+        match (
+            grid_span(x - tolerance, x + w + tolerance, cell),
+            grid_span(y - tolerance, y + h + tolerance, cell),
+        ) {
+            (Some(xs), Some(ys)) => {
+                for gx in xs {
+                    for gy in ys.clone() {
+                        grid.entry((gx, gy)).or_default().push(idx);
+                    }
+                }
+            }
+            _ => large.push(idx),
+        }
+    }
+
+    let mut keys: Vec<_> = grid.keys().copied().collect();
+    keys.sort_unstable();
+    let mut keys_by_y: BTreeMap<i32, Vec<i32>> = BTreeMap::new();
+    for &key in &keys {
+        union_bucket_pairs(&mut uf, rects, &grid[&key], tolerance);
+        keys_by_y.entry(key.1).or_default().push(key.0);
+    }
+
+    // Oversized spans skip insert. Range-query occupied cells they cover so
+    // later X-ranges are not starved and we do not scan unrelated rows.
+    for &i in &large {
         if uf.component_size(i) >= MAX_CLUSTER_RECTS {
             continue;
         }
-        for j in (i + 1)..n {
-            if rects_overlap(&rects[i], &rects[j], tolerance) {
-                uf.union(i, j);
-                // Check if the merged component just exceeded the cap —
-                // if so, no need to test more pairs for rect i.
+        let (x, y, w, h) = rects[i];
+        let x_lo = grid_coord(x - tolerance, cell);
+        let x_hi = grid_coord(x + w + tolerance, cell);
+        let y_lo = grid_coord(y - tolerance, cell);
+        let y_hi = grid_coord(y + h + tolerance, cell);
+        for (&gy, gxs) in keys_by_y.range(y_lo..=y_hi) {
+            let start = gxs.partition_point(|&gx| gx < x_lo);
+            for &gx in &gxs[start..] {
+                if gx > x_hi {
+                    break;
+                }
+                let bucket = &grid[&(gx, gy)];
+                let mut pairs = 0usize;
+                for &j in bucket {
+                    if pairs >= MAX_CLUSTER_PAIRS_PER_CELL {
+                        break;
+                    }
+                    if uf.component_size(j) >= MAX_CLUSTER_RECTS {
+                        continue;
+                    }
+                    pairs += 1;
+                    if rects_overlap(&rects[i], &rects[j], tolerance) {
+                        uf.union(i, j);
+                        if uf.component_size(i) >= MAX_CLUSTER_RECTS {
+                            break;
+                        }
+                    }
+                }
                 if uf.component_size(i) >= MAX_CLUSTER_RECTS {
                     break;
                 }
             }
+            if uf.component_size(i) >= MAX_CLUSTER_RECTS {
+                break;
+            }
+        }
+    }
+
+    // Oversized-vs-oversized: band on the short axis so stacked or side-by-side
+    // page-spanning rules stay linear. Wide vs tall pairs are matched by
+    // querying the tall X-index; dual-oversized rects occupy every coarse-Y
+    // cell they span.
+    let mut large_x: BTreeMap<i32, Vec<usize>> = BTreeMap::new();
+    let mut large_y: BTreeMap<i32, Vec<usize>> = BTreeMap::new();
+    let mut large_coarse_y: BTreeMap<i32, Vec<usize>> = BTreeMap::new();
+    let mut wide: Vec<usize> = Vec::new();
+    let mut dual: Vec<usize> = Vec::new();
+    for &i in &large {
+        let (x, y, w, h) = rects[i];
+        let xs = grid_span(x - tolerance, x + w + tolerance, cell);
+        let ys = grid_span(y - tolerance, y + h + tolerance, cell);
+        match (xs, ys) {
+            (Some(xs), _) => {
+                for gx in xs {
+                    large_x.entry(gx).or_default().push(i);
+                }
+            }
+            (_, Some(ys)) => {
+                wide.push(i);
+                for gy in ys {
+                    large_y.entry(gy).or_default().push(i);
+                }
+            }
+            _ => {
+                dual.push(i);
+                let coarse = cell * 64.0;
+                match grid_span(y - tolerance, y + h + tolerance, coarse) {
+                    Some(ys) => {
+                        for gy in ys {
+                            large_coarse_y.entry(gy).or_default().push(i);
+                        }
+                    }
+                    None => {
+                        large_coarse_y.entry(i32::MIN).or_default().push(i);
+                    }
+                }
+            }
+        }
+    }
+    for bands in [&large_x, &large_y, &large_coarse_y] {
+        for bucket in bands.values() {
+            union_bucket_pairs(&mut uf, rects, bucket, tolerance);
+        }
+    }
+    // Cross-orientation is |wide|×|tall| if every wide rule spans the page.
+    // Skip that pass when the product cannot be a table (a few rules).
+    let tall_n = large
+        .len()
+        .saturating_sub(wide.len())
+        .saturating_sub(dual.len());
+    let cross_n =
+        (wide.len() + dual.len()).saturating_mul(tall_n) + dual.len().saturating_mul(wide.len());
+    if cross_n > 0 && cross_n <= MAX_CLUSTER_PAIRS_PER_CELL {
+        for &i in wide.iter().chain(&dual) {
+            let (x, _, w, _) = rects[i];
+            let x_lo = grid_coord(x - tolerance, cell);
+            let x_hi = grid_coord(x + w + tolerance, cell);
+            union_rect_against_bands(&mut uf, rects, i, &large_x, x_lo, x_hi, tolerance);
+        }
+        for &i in &dual {
+            let (_, y, _, h) = rects[i];
+            let y_lo = grid_coord(y - tolerance, cell);
+            let y_hi = grid_coord(y + h + tolerance, cell);
+            union_rect_against_bands(&mut uf, rects, i, &large_y, y_lo, y_hi, tolerance);
         }
     }
 
@@ -3810,6 +4023,113 @@ mod tests {
         assert_eq!(groups[0].len(), 2);
     }
 
+    #[test]
+    fn test_cluster_rects_overlapping_grid_still_clusters() {
+        // Neighboring cells overlap; the grid must still union the whole table.
+        let mut rects = Vec::new();
+        for row in 0..4 {
+            for col in 0..4 {
+                rects.push((col as f32 * 9.0, row as f32 * 9.0, 10.0, 10.0));
+            }
+        }
+        let groups = cluster_rects(&rects, 0.0, 1);
+        assert_eq!(groups.len(), 1);
+        assert_eq!(groups[0].len(), 16);
+    }
+
+    #[test]
+    fn test_cluster_rects_many_disjoint_stays_subquadratic() {
+        // Pairwise-disjoint rects never merge, so a component-size cap does
+        // not stop all-pairs overlap tests. Spread in X so they land in
+        // different grid cells; 8k is enough that n² tests would dominate.
+        let n = 8_000usize;
+        let rects: Vec<(f32, f32, f32, f32)> =
+            (0..n).map(|i| (i as f32 * 20.0, 0.0, 10.0, 10.0)).collect();
+        let groups = cluster_rects(&rects, 0.0, 2);
+        assert!(groups.is_empty());
+    }
+
+    #[test]
+    fn test_cluster_rects_stacked_disjoint_does_not_starve_later_table() {
+        // Same X, spread in Y: a spatial grid must still union an overlapping
+        // pair in another region of the page.
+        let n = 8_000usize;
+        let mut rects: Vec<(f32, f32, f32, f32)> =
+            (0..n).map(|i| (0.0, i as f32 * 20.0, 10.0, 10.0)).collect();
+        rects.push((500.0, 0.0, 10.0, 10.0));
+        rects.push((508.0, 0.0, 10.0, 10.0));
+        let groups = cluster_rects(&rects, 0.0, 2);
+        assert_eq!(groups.len(), 1);
+        assert_eq!(groups[0].len(), 2);
+    }
+
+    #[test]
+    fn test_cluster_rects_oversized_span_still_unions() {
+        // Wider than 64 grid cells; must still union the small overlapping rect.
+        let rects = vec![(0.0, 0.0, 5000.0, 10.0), (4900.0, 0.0, 10.0, 10.0)];
+        let groups = cluster_rects(&rects, 0.0, 1);
+        assert_eq!(groups.len(), 1);
+        assert_eq!(groups[0].len(), 2);
+    }
+
+    #[test]
+    fn test_cluster_rects_many_oversized_spans_all_get_a_pass() {
+        // More than 32 huge rects: the last one must still union its overlap.
+        let mut rects: Vec<(f32, f32, f32, f32)> = (0..40)
+            .map(|i| (0.0, i as f32 * 20.0, 5000.0, 10.0))
+            .collect();
+        rects.push((4900.0, 39.0 * 20.0, 10.0, 10.0));
+        let groups = cluster_rects(&rects, 0.0, 2);
+        assert_eq!(groups.len(), 1);
+        assert_eq!(groups[0].len(), 2);
+    }
+
+    #[test]
+    fn test_cluster_rects_oversized_not_starved_by_earlier_disjoint() {
+        // 9k earlier disjoint drawings would exhaust an index-order cap of
+        // 8,192 before the overlapping cell is visited.
+        let mut rects: Vec<(f32, f32, f32, f32)> = (0..9_000)
+            .map(|i| (10_000.0, i as f32 * 20.0, 10.0, 10.0))
+            .collect();
+        let wide = rects.len();
+        rects.push((0.0, 0.0, 5000.0, 10.0));
+        let target = rects.len();
+        rects.push((4900.0, 0.0, 10.0, 10.0));
+        let groups = cluster_rects(&rects, 0.0, 2);
+        assert!(
+            groups
+                .iter()
+                .any(|g| g.contains(&wide) && g.contains(&target)),
+            "wide rule and far-end cell must share a cluster"
+        );
+    }
+
+    #[test]
+    fn test_cluster_rects_wide_and_tall_oversized_union() {
+        let rects = vec![(0.0, 0.0, 5000.0, 10.0), (0.0, 0.0, 10.0, 5000.0)];
+        let groups = cluster_rects(&rects, 0.0, 2);
+        assert_eq!(groups.len(), 1);
+        assert_eq!(groups[0].len(), 2);
+    }
+
+    #[test]
+    fn test_cluster_rects_dual_oversized_spans_coarse_y() {
+        let rects = vec![(0.0, 0.0, 5000.0, 5000.0), (0.0, 4500.0, 5000.0, 5000.0)];
+        let groups = cluster_rects(&rects, 0.0, 2);
+        assert_eq!(groups.len(), 1);
+        assert_eq!(groups[0].len(), 2);
+    }
+
+    #[test]
+    fn test_cluster_rects_many_wide_and_tall_stays_subquadratic() {
+        let mut rects = Vec::with_capacity(4_000);
+        for i in 0..2_000 {
+            rects.push((0.0, i as f32 * 20.0, 5000.0, 10.0));
+            rects.push((i as f32 * 20.0, 0.0, 10.0, 5000.0));
+        }
+        let _groups = cluster_rects(&rects, 0.0, 2);
+    }
+
     // --- snap_edges ---
 
     #[test]

src/tables/grid.rs

@@ -604,7 +604,7 @@ mod tests {
         let items: Vec<(usize, &TextItem)> = vec![];
         assert_eq!(
             find_column_boundaries(&items, TableDetectionMode::SmallFont),
-            vec![]
+            Vec::<f32>::new()
         );
     }
 
@@ -661,7 +661,7 @@ mod tests {
     #[test]
     fn test_find_row_boundaries_empty() {
         let items: Vec<(usize, &TextItem)> = vec![];
-        assert_eq!(find_row_boundaries(&items), vec![]);
+        assert_eq!(find_row_boundaries(&items), Vec::<f32>::new());
     }
 
     #[test]

src/tounicode.rs

@@ -540,18 +540,14 @@ impl ToUnicodeCMap {
 
     /// Remap a CMap that references pre-subsetting GIDs to sequential post-subsetting GIDs.
     /// Collects all source CIDs, sorts them, and reassigns to 1, 2, 3, ...
+    ///
+    /// Range expansion stops after `MAX_CID_W_EXPANSION` CID visits, counting
+    /// overwrites, so repeated full-width `bfrange`s cannot re-expand the
+    /// 16-bit domain. Later overlapping ranges that would have introduced new
+    /// CIDs after that many visits are truncated.
     pub fn remap_to_sequential(&self) -> ToUnicodeCMap {
         let mut cid_to_unicode: HashMap<u16, String> = HashMap::new();
-
-        // Expand ranges first
-        for &(start, end, base) in &self.ranges {
-            for cid in start..=end {
-                let unicode_cp = base + (cid - start) as u32;
-                if let Some(ch) = char::from_u32(unicode_cp) {
-                    cid_to_unicode.insert(cid, ch.to_string());
-                }
-            }
-        }
+        expand_bfranges_for_remap(&self.ranges, &mut cid_to_unicode, MAX_CID_W_EXPANSION);
 
         // char_map entries override range entries
         for (&cid, unicode) in &self.char_map {
@@ -576,6 +572,33 @@ impl ToUnicodeCMap {
     }
 }
 
+/// Expand `bfrange` entries into individual CID→Unicode inserts.
+/// Returns how many CIDs were visited. Counts overwrites so a repeated
+/// full-width range cannot keep working after `max_assignments`.
+fn expand_bfranges_for_remap(
+    ranges: &[(u16, u16, u32)],
+    cid_to_unicode: &mut HashMap<u16, String>,
+    max_assignments: usize,
+) -> usize {
+    let mut assigned = 0usize;
+    'ranges: for &(start, end, base) in ranges {
+        if start > end {
+            continue;
+        }
+        for cid in start..=end {
+            if assigned >= max_assignments {
+                break 'ranges;
+            }
+            assigned += 1;
+            let unicode_cp = base + (cid - start) as u32;
+            if let Some(ch) = char::from_u32(unicode_cp) {
+                cid_to_unicode.insert(cid, ch.to_string());
+            }
+        }
+    }
+    assigned
+}
+
 /// Parse a hex string to u16
 fn parse_hex_u16(hex: &str) -> Option<u16> {
     u16::from_str_radix(hex.trim(), 16).ok()
@@ -594,7 +617,7 @@ fn hex_to_unicode_string(hex: &str) -> Option<String> {
 
     let bytes: Option<Vec<u8>> = (0..hex.len())
         .step_by(2)
-        .map(|i| u8::from_str_radix(&hex[i..i + 2], 16).ok())
+        .map(|i| u8::from_str_radix(hex.get(i..i + 2)?, 16).ok())
         .collect();
     let bytes = bytes?;
 
@@ -1561,23 +1584,31 @@ fn parse_encoding_cmap_stream(data: &[u8]) -> Option<EncodingCMap> {
     }
 
     let mut map = HashMap::new();
+    let mut assigned = 0usize;
     let mut pos = 0;
     while let Some(start) = text[pos..].find("begincidchar") {
         let section_start = pos + start + "begincidchar".len();
         if let Some(end) = text[section_start..].find("endcidchar") {
             let section = &text[section_start..section_start + end];
-            parse_cidchar_section(section, &mut map, &mut src_hex_lengths);
+            if !parse_cidchar_section(section, &mut map, &mut src_hex_lengths, &mut assigned) {
+                break;
+            }
             pos = section_start + end;
         } else {
             break;
         }
     }
     pos = 0;
-    while let Some(start) = text[pos..].find("begincidrange") {
+    while assigned < MAX_CID_W_EXPANSION {
+        let Some(start) = text[pos..].find("begincidrange") else {
+            break;
+        };
         let section_start = pos + start + "begincidrange".len();
         if let Some(end) = text[section_start..].find("endcidrange") {
             let section = &text[section_start..section_start + end];
-            parse_cidrange_section(section, &mut map, &mut src_hex_lengths);
+            if !parse_cidrange_section(section, &mut map, &mut src_hex_lengths, &mut assigned) {
+                break;
+            }
             pos = section_start + end;
         } else {
             break;
@@ -1612,7 +1643,8 @@ fn parse_cidchar_section(
     section: &str,
     map: &mut HashMap<u16, u16>,
     src_hex_lengths: &mut Vec<usize>,
-) {
+    assigned: &mut usize,
+) -> bool {
     let mut chars = section.chars().peekable();
     loop {
         while chars.peek().is_some_and(|c| c.is_whitespace()) {
@@ -1643,16 +1675,20 @@ fn parse_cidchar_section(
             }
         }
         if let (Some(code), Ok(cid)) = (parse_hex_u16(&src_hex), cid_str.parse::<u16>()) {
-            map.insert(code, cid);
+            if !assign_encoding_cid(map, code, cid, assigned) {
+                return false;
+            }
         }
     }
+    true
 }
 
 fn parse_cidrange_section(
     section: &str,
     map: &mut HashMap<u16, u16>,
     src_hex_lengths: &mut Vec<usize>,
-) {
+    assigned: &mut usize,
+) -> bool {
     let mut chars = section.chars().peekable();
     loop {
         while chars.peek().is_some_and(|c| c.is_whitespace()) {
@@ -1703,12 +1739,34 @@ fn parse_cidrange_section(
         ) else {
             continue;
         };
+        if start > end {
+            continue;
+        }
         let mut cid = start_cid;
         for code in start..=end {
-            map.insert(code, cid);
+            if !assign_encoding_cid(map, code, cid, assigned) {
+                return false;
+            }
             cid = cid.saturating_add(1);
         }
     }
+    true
+}
+
+fn assign_encoding_cid(
+    map: &mut HashMap<u16, u16>,
+    code: u16,
+    cid: u16,
+    assigned: &mut usize,
+) -> bool {
+    // Count overwrites: unique-key coverage alone would not stop a repeated
+    // full-width range from re-inserting all 65,536 codes.
+    if *assigned >= MAX_CID_W_EXPANSION {
+        return false;
+    }
+    map.insert(code, cid);
+    *assigned += 1;
+    true
 }
 
 fn parse_binary_cmap_encoding(data: &[u8]) -> Result<EncodingCMap, String> {
@@ -1832,6 +1890,13 @@ fn merge_cmaps(mut base: ToUnicodeCMap, overlay: ToUnicodeCMap) -> ToUnicodeCMap
     base
 }
 
+/// Shared 16-bit CID expansion cap (65,536).
+/// Encoding `begincidrange`, `/W` width assignment, and ToUnicode sequential
+/// remap count every insert, including overwrites, so a repeated full-width
+/// range cannot keep working after the domain is filled. The `/W` unicode
+/// heuristic caps unique CIDs with the same number.
+pub(crate) const MAX_CID_W_EXPANSION: usize = 65_536;
+
 /// Check if a CIDFont's /W (widths) array contains CID values that look like
 /// Unicode codepoints rather than low-value GIDs.
 ///
@@ -1843,20 +1908,23 @@ pub(crate) fn cid_values_look_like_unicode(cid_font_dict: &lopdf::Dictionary) ->
         _ => return false,
     };
 
-    // The /W array format: [cid [w1 w2 ...]] or [cid_start cid_end w]
-    // We extract all CID values (the first element of each group).
-    let mut cids: Vec<u16> = Vec::new();
+    // The /W array format: [cid [w1 w2 ...]] or [cid_start cid_end w].
+    // Collect unique CIDs only: repeating a full-width range must not grow a
+    // temporary vector (or the sort) with the range length on every copy.
+    let mut seen = HashSet::new();
     let mut i = 0;
-    while i < w_arr.len() {
+    while i < w_arr.len() && seen.len() < MAX_CID_W_EXPANSION {
         if let Ok(cid) = w_arr[i].as_i64() {
-            cids.push(cid as u16);
-            // Skip the width data
+            let start = cid as u16;
             if i + 1 < w_arr.len() {
                 match &w_arr[i + 1] {
                     Object::Array(widths) => {
                         // [cid [w1 w2 ...]] — CIDs are cid, cid+1, ..., cid+len-1
-                        for j in 1..widths.len() {
-                            cids.push((cid as u16).wrapping_add(j as u16));
+                        for j in 0..widths.len() {
+                            if seen.len() >= MAX_CID_W_EXPANSION {
+                                break;
+                            }
+                            seen.insert(start.wrapping_add(j as u16));
                         }
                         i += 2;
                     }
@@ -1864,9 +1932,7 @@ pub(crate) fn cid_values_look_like_unicode(cid_font_dict: &lopdf::Dictionary) ->
                         // [cid_start cid_end w] — range of CIDs
                         if i + 2 < w_arr.len() {
                             if let Ok(cid_end) = w_arr[i + 1].as_i64() {
-                                for c in (cid as u16)..=(cid_end as u16) {
-                                    cids.push(c);
-                                }
+                                record_unique_cid_range(start, cid_end as u16, &mut seen);
                             }
                             i += 3;
                         } else {
@@ -1875,6 +1941,7 @@ pub(crate) fn cid_values_look_like_unicode(cid_font_dict: &lopdf::Dictionary) ->
                     }
                 }
             } else {
+                seen.insert(start);
                 i += 1;
             }
         } else {
@@ -1882,10 +1949,11 @@ pub(crate) fn cid_values_look_like_unicode(cid_font_dict: &lopdf::Dictionary) ->
         }
     }
 
-    if cids.is_empty() {
+    if seen.is_empty() {
         return false;
     }
 
+    let mut cids: Vec<u16> = seen.into_iter().collect();
     cids.sort_unstable();
     let median = cids[cids.len() / 2];
     // Unicode text CIDs are typically >= 0x20 (space) with letters at 0x41+.
@@ -1894,6 +1962,18 @@ pub(crate) fn cid_values_look_like_unicode(cid_font_dict: &lopdf::Dictionary) ->
     median >= 0x41
 }
 
+fn record_unique_cid_range(start: u16, end: u16, seen: &mut HashSet<u16>) {
+    if start > end {
+        return;
+    }
+    for cid in start..=end {
+        if seen.len() >= MAX_CID_W_EXPANSION {
+            return;
+        }
+        seen.insert(cid);
+    }
+}
+
 /// Build a ToUnicodeCMap from predefined CID→Unicode mapping based on CIDSystemInfo.
 ///
 /// Supports Adobe-Korea1 (Korean) character collection. Can be extended for
@@ -2606,6 +2686,24 @@ endcmap
         assert_eq!(cmap.lookup(0x0025), Some("B".to_string()));
     }
 
+    #[test]
+    fn test_hex_to_unicode_non_ascii_no_panic() {
+        // A destination containing a multi-byte char makes the byte length even
+        // while a byte offset can land inside a char. Slicing must not panic;
+        // it should be rejected gracefully.
+        assert_eq!(hex_to_unicode_string("XéY"), None);
+        assert_eq!(hex_to_unicode_string("\u{fffd}0"), None);
+    }
+
+    #[test]
+    fn test_parse_bfchar_non_ascii_destination_no_panic() {
+        // Crafted /ToUnicode CMap: a non-hex, non-ASCII destination previously
+        // triggered a char-boundary panic in hex_to_unicode_string.
+        let cmap_content = "beginbfchar <0041> <XéY> endbfchar";
+        // Must not panic; the malformed entry is simply skipped.
+        let _ = ToUnicodeCMap::parse(cmap_content.as_bytes());
+    }
+
     #[test]
     fn test_parse_bfchar_1byte() {
         // This is the pattern that caused the CJK bug: codespace is <0000><FFFF>
@@ -2849,6 +2947,33 @@ endbfrange
         assert!(remapped.ranges.is_empty());
     }
 
+    #[test]
+    fn remap_to_sequential_repeated_full_bfranges_stay_bounded() {
+        // 5,000 copies of `<0003> <ffff>` must stop after 65,536 CID visits,
+        // not 5,000 × ~65,533 expansions.
+        let ranges = vec![(3u16, 65535u16, 0x41u32); 5_000];
+        let mut map = std::collections::HashMap::new();
+        let assigned = expand_bfranges_for_remap(&ranges, &mut map, MAX_CID_W_EXPANSION);
+        assert_eq!(assigned, MAX_CID_W_EXPANSION);
+        assert!(map.len() <= MAX_CID_W_EXPANSION);
+
+        let mut body = String::new();
+        let mut remaining = 5_000usize;
+        while remaining > 0 {
+            let n = remaining.min(100);
+            body.push_str(&format!("{n} beginbfrange\n"));
+            for _ in 0..n {
+                body.push_str("<0003> <ffff> <0041>\n");
+            }
+            body.push_str("endbfrange\n");
+            remaining -= n;
+        }
+        let data = format!("1 begincodespacerange\n<0000> <ffff>\nendcodespacerange\n{body}");
+        let cmap = ToUnicodeCMap::parse(data.as_bytes()).unwrap();
+        let remapped = cmap.remap_to_sequential();
+        assert_eq!(remapped.lookup(1), Some("A".to_string()));
+    }
+
     #[test]
     fn test_min_source_cid() {
         let cmap_content = r#"
@@ -3280,4 +3405,79 @@ endbfrange
             "An indirect /Subtype naming CIDFontType2 must still reach the remap"
         );
     }
+
+    #[test]
+    fn cid_values_look_like_unicode_letter_range() {
+        let mut dict = lopdf::Dictionary::new();
+        dict.set(
+            "W",
+            Object::Array(vec![
+                Object::Integer(0x41),
+                Object::Integer(0x5A),
+                Object::Integer(500),
+            ]),
+        );
+        assert!(cid_values_look_like_unicode(&dict));
+    }
+
+    #[test]
+    fn cid_values_look_like_unicode_low_gids() {
+        let mut dict = lopdf::Dictionary::new();
+        dict.set(
+            "W",
+            Object::Array(vec![
+                Object::Integer(0),
+                Object::Array(vec![Object::Integer(500); 10]),
+            ]),
+        );
+        assert!(!cid_values_look_like_unicode(&dict));
+    }
+
+    #[test]
+    fn cid_values_look_like_unicode_repeated_full_ranges_stay_bounded() {
+        // Repeating `[0 65535 w]` must not materialize 65,536 CIDs per copy.
+        let mut w = Vec::new();
+        for _ in 0..5_000 {
+            w.push(Object::Integer(0));
+            w.push(Object::Integer(65535));
+            w.push(Object::Integer(500));
+        }
+        let mut dict = lopdf::Dictionary::new();
+        dict.set("W", Object::Array(w));
+        assert!(cid_values_look_like_unicode(&dict));
+    }
+
+    #[test]
+    fn encoding_cidrange_maps_a_normal_range() {
+        let data = b"1 begincodespacerange\n<0000> <FFFF>\nendcodespacerange\n\
+1 begincidrange\n<0041> <0043> 65\nendcidrange\n";
+        let enc = parse_encoding_cmap_stream(data).unwrap();
+        assert_eq!(enc.map.get(&0x41), Some(&65));
+        assert_eq!(enc.map.get(&0x42), Some(&66));
+        assert_eq!(enc.map.get(&0x43), Some(&67));
+        assert_eq!(enc.map.len(), 3);
+        assert_eq!(enc.code_byte_length, 2);
+    }
+
+    #[test]
+    fn encoding_cidrange_repeated_full_ranges_stay_bounded() {
+        // 5,000 copies of `<0000> <ffff> 0` must not re-expand the 16-bit
+        // domain on every declaration.
+        let mut body = String::new();
+        let mut remaining = 5_000usize;
+        while remaining > 0 {
+            let n = remaining.min(100);
+            body.push_str(&format!("{n} begincidrange\n"));
+            for _ in 0..n {
+                body.push_str("<0000> <ffff> 0\n");
+            }
+            body.push_str("endcidrange\n");
+            remaining -= n;
+        }
+        let data = format!("1 begincodespacerange\n<0000> <FFFF>\nendcodespacerange\n{body}");
+        let enc = parse_encoding_cmap_stream(data.as_bytes()).unwrap();
+        assert!(enc.map.len() <= MAX_CID_W_EXPANSION);
+        assert_eq!(enc.map.get(&0), Some(&0));
+        assert_eq!(enc.map.get(&65535), Some(&65535));
+    }
 }

src/vision/contracts.rs

@@ -0,0 +1,414 @@
+//! Public contracts between rendering, OCR, layout, and orchestration.
+
+use std::error::Error;
+use std::path::PathBuf;
+
+use super::{RenderOptions, RenderedPage};
+
+/// Selects when OCR may run.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+#[non_exhaustive]
+pub enum OcrMode {
+    /// Never run OCR. This is the default and preserves existing behavior.
+    #[default]
+    Off,
+    /// Run OCR only on pages selected by pdf-inspector's OCR routing signals.
+    Auto,
+    /// Run OCR on every selected page, including pages with native text.
+    Force,
+}
+
+/// Resource/quality profile for the OCR engine.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+#[non_exhaustive]
+pub enum OcrProfile {
+    /// Lowest latency and memory footprint.
+    Edge,
+    /// OCR-oriented balance of quality and CPU cost.
+    #[default]
+    Balanced,
+    /// Highest quality within the lightweight model family.
+    Quality,
+}
+
+/// Controls whether missing model artifacts may be fetched.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+#[non_exhaustive]
+pub enum ModelDownloadPolicy {
+    /// Fetch a pinned artifact only after OCR has actually been selected.
+    #[default]
+    IfMissing,
+    /// Never access the network; require an override or a warm model cache.
+    Offline,
+}
+
+/// OCR engine configuration independent of a particular runtime.
+#[derive(Debug, Clone, PartialEq)]
+pub struct OcrOptions {
+    /// Page-level routing behavior.
+    pub mode: OcrMode,
+    /// Local quality/resource profile.
+    pub profile: OcrProfile,
+    /// Drop recognition spans below this confidence threshold.
+    pub minimum_confidence: f32,
+    /// Optional language hints understood by the selected engine.
+    pub languages: Vec<String>,
+    /// Optional directory containing an offline model set.
+    pub model_directory: Option<PathBuf>,
+    /// Whether a missing pinned artifact may be downloaded.
+    pub model_downloads: ModelDownloadPolicy,
+}
+
+impl Default for OcrOptions {
+    fn default() -> Self {
+        Self {
+            mode: OcrMode::Off,
+            profile: OcrProfile::Balanced,
+            minimum_confidence: 0.0,
+            languages: Vec::new(),
+            model_directory: None,
+            model_downloads: ModelDownloadPolicy::IfMissing,
+        }
+    }
+}
+
+impl OcrOptions {
+    /// Creates OCR options with OCR disabled.
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// Sets page-level OCR routing.
+    pub fn mode(mut self, mode: OcrMode) -> Self {
+        self.mode = mode;
+        self
+    }
+
+    /// Sets the local resource/quality profile.
+    pub fn profile(mut self, profile: OcrProfile) -> Self {
+        self.profile = profile;
+        self
+    }
+
+    /// Sets the minimum accepted recognition confidence.
+    pub fn minimum_confidence(mut self, minimum_confidence: f32) -> Self {
+        self.minimum_confidence = minimum_confidence;
+        self
+    }
+
+    /// Replaces the language hints passed to the OCR engine.
+    pub fn languages(mut self, languages: impl IntoIterator<Item = impl Into<String>>) -> Self {
+        self.languages = languages.into_iter().map(Into::into).collect();
+        self
+    }
+
+    /// Uses an explicit model directory, suitable for offline packaging.
+    pub fn model_directory(mut self, directory: impl Into<PathBuf>) -> Self {
+        self.model_directory = Some(directory.into());
+        self
+    }
+
+    /// Sets the missing-model download policy.
+    pub fn model_downloads(mut self, policy: ModelDownloadPolicy) -> Self {
+        self.model_downloads = policy;
+        self
+    }
+}
+
+/// Configuration for an optional learned layout engine.
+///
+/// Layout inference is disabled by default. Existing deterministic layout,
+/// table, and Markdown logic remains the assembly path when this is disabled.
+#[derive(Debug, Clone, PartialEq)]
+pub struct LayoutOptions {
+    /// Whether the learned layout extension may run.
+    pub enabled: bool,
+    /// Drop layout regions below this confidence threshold.
+    pub minimum_confidence: f32,
+    /// Optional directory containing an offline layout model set.
+    pub model_directory: Option<PathBuf>,
+}
+
+impl Default for LayoutOptions {
+    fn default() -> Self {
+        Self {
+            enabled: false,
+            minimum_confidence: 0.0,
+            model_directory: None,
+        }
+    }
+}
+
+impl LayoutOptions {
+    /// Creates layout options with learned layout disabled.
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// Enables or disables learned layout inference.
+    pub fn enabled(mut self, enabled: bool) -> Self {
+        self.enabled = enabled;
+        self
+    }
+
+    /// Sets the minimum accepted region confidence.
+    pub fn minimum_confidence(mut self, minimum_confidence: f32) -> Self {
+        self.minimum_confidence = minimum_confidence;
+        self
+    }
+
+    /// Uses an explicit layout model directory.
+    pub fn model_directory(mut self, directory: impl Into<PathBuf>) -> Self {
+        self.model_directory = Some(directory.into());
+        self
+    }
+}
+
+/// A point in bitmap space, measured from the top-left in pixels.
+#[derive(Debug, Clone, Copy, Default, PartialEq)]
+pub struct ImagePoint {
+    /// Horizontal pixel coordinate.
+    pub x: f32,
+    /// Vertical pixel coordinate, increasing downward.
+    pub y: f32,
+}
+
+impl ImagePoint {
+    /// Creates a bitmap-space point.
+    pub fn new(x: f32, y: f32) -> Self {
+        Self { x, y }
+    }
+}
+
+/// Four-point polygon in bitmap coordinates.
+#[derive(Debug, Clone, Copy, Default, PartialEq)]
+pub struct ImageQuad {
+    /// Polygon points in engine-provided order.
+    pub points: [ImagePoint; 4],
+}
+
+impl ImageQuad {
+    /// Creates a four-point bitmap polygon.
+    pub fn new(points: [ImagePoint; 4]) -> Self {
+        Self { points }
+    }
+}
+
+/// Stable identity for an inference model used in output provenance.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct ModelIdentity {
+    /// Model family/name, for example `pp-ocrv6-small`.
+    pub name: String,
+    /// Immutable model or artifact-set revision.
+    pub revision: String,
+}
+
+impl ModelIdentity {
+    /// Creates a model identity.
+    pub fn new(name: impl Into<String>, revision: impl Into<String>) -> Self {
+        Self {
+            name: name.into(),
+            revision: revision.into(),
+        }
+    }
+}
+
+/// One positioned OCR recognition result in bitmap coordinates.
+#[derive(Debug, Clone, PartialEq)]
+pub struct OcrSpan {
+    /// Recognized text.
+    pub text: String,
+    /// Detection polygon in the original rendered page's pixel space.
+    pub polygon: ImageQuad,
+    /// Recognition confidence in the inclusive range 0–1.
+    pub confidence: f32,
+    /// Optional text-line orientation in clockwise degrees.
+    pub orientation_degrees: Option<f32>,
+}
+
+/// OCR output for one 1-indexed page.
+#[derive(Debug, Clone, PartialEq)]
+pub struct OcrPage {
+    /// 1-indexed PDF page number.
+    pub page: u32,
+    /// Positioned recognition spans.
+    pub spans: Vec<OcrSpan>,
+    /// Mean confidence across accepted spans, when available.
+    pub mean_confidence: Option<f32>,
+    /// Exact model identity used for this result.
+    pub model: ModelIdentity,
+    /// OCR wall time for this page.
+    pub processing_time_ms: u64,
+    /// Non-fatal engine warnings.
+    pub warnings: Vec<String>,
+}
+
+/// Normalized semantic class emitted by a learned layout engine.
+#[derive(Debug, Clone, PartialEq, Eq)]
+#[non_exhaustive]
+pub enum LayoutRegionKind {
+    /// Body or other prose text.
+    Text,
+    /// Document heading or title.
+    Heading,
+    /// Table region.
+    Table,
+    /// Figure/image region.
+    Figure,
+    /// Figure or table caption.
+    Caption,
+    /// Header/footer/page furniture.
+    Furniture,
+    /// Model-specific class retained without changing the common taxonomy.
+    Other(String),
+}
+
+/// One learned layout region in bitmap coordinates.
+#[derive(Debug, Clone, PartialEq)]
+pub struct LayoutRegion {
+    /// Normalized semantic class.
+    pub kind: LayoutRegionKind,
+    /// Region polygon in the original rendered page's pixel space.
+    pub polygon: ImageQuad,
+    /// Model confidence in the inclusive range 0–1.
+    pub confidence: f32,
+    /// Optional model-provided reading-order position.
+    pub reading_order: Option<u32>,
+}
+
+/// Learned layout output for one 1-indexed page.
+#[derive(Debug, Clone, PartialEq)]
+pub struct LayoutPage {
+    /// 1-indexed PDF page number.
+    pub page: u32,
+    /// Semantic regions.
+    pub regions: Vec<LayoutRegion>,
+    /// Exact model identity used for this result.
+    pub model: ModelIdentity,
+    /// Layout inference wall time for this page.
+    pub processing_time_ms: u64,
+    /// Non-fatal engine warnings.
+    pub warnings: Vec<String>,
+}
+
+/// How final page content was sourced.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+#[non_exhaustive]
+pub enum PageContentSource {
+    /// Trusted native PDF text only.
+    Native,
+    /// OCR output only.
+    Ocr,
+    /// Native and OCR spans were fused.
+    Fused,
+}
+
+/// Per-page local processing timings.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct VisionTimings {
+    /// Rasterization wall time.
+    pub render_ms: u64,
+    /// OCR wall time.
+    pub ocr_ms: u64,
+    /// Optional learned layout wall time.
+    pub layout_ms: u64,
+    /// Native/OCR fusion and assembly wall time.
+    pub assembly_ms: u64,
+}
+
+/// Source and model metadata retained for one processed page.
+#[derive(Debug, Clone, PartialEq)]
+pub struct PageProvenance {
+    /// 1-indexed PDF page number.
+    pub page: u32,
+    /// Final page-content source.
+    pub source: PageContentSource,
+    /// OCR model, when OCR ran.
+    pub ocr_model: Option<ModelIdentity>,
+    /// Learned layout model, when layout inference ran.
+    pub layout_model: Option<ModelIdentity>,
+    /// Render resolution used for local vision.
+    pub render_dpi: Option<f32>,
+    /// Mean accepted OCR confidence, when available.
+    pub ocr_confidence: Option<f32>,
+    /// Stage timings.
+    pub timings: VisionTimings,
+    /// Non-fatal warnings surfaced to downstream users.
+    pub warnings: Vec<String>,
+    /// True when this lightweight local path detected a case better suited to
+    /// Firecrawl's hosted document pipeline.
+    pub hosted_recommended: bool,
+}
+
+/// Converts selected PDF pages into renderer-neutral owned bitmaps.
+pub trait PageRenderer: Send + Sync {
+    /// Renderer-specific failure type.
+    type Error: Error + Send + Sync + 'static;
+
+    /// Renders selected 1-indexed pages in the same order as `pages`.
+    fn render_pages(
+        &self,
+        pdf_bytes: &[u8],
+        pages: &[u32],
+        password: Option<&str>,
+        options: &RenderOptions,
+    ) -> Result<Vec<RenderedPage>, Self::Error>;
+}
+
+/// Recognizes positioned text from rendered pages.
+pub trait OcrEngine: Send + Sync {
+    /// Engine-specific failure type.
+    type Error: Error + Send + Sync + 'static;
+
+    /// Exact model identity used by this engine instance.
+    fn model(&self) -> &ModelIdentity;
+
+    /// Recognizes pages in batch and returns results in input order.
+    fn recognize(
+        &self,
+        pages: &[RenderedPage],
+        options: &OcrOptions,
+    ) -> Result<Vec<OcrPage>, Self::Error>;
+}
+
+/// Optional learned semantic layout extension.
+pub trait LayoutEngine: Send + Sync {
+    /// Engine-specific failure type.
+    type Error: Error + Send + Sync + 'static;
+
+    /// Exact model identity used by this engine instance.
+    fn model(&self) -> &ModelIdentity;
+
+    /// Analyzes rendered pages, optionally using their OCR spans.
+    fn analyze(
+        &self,
+        pages: &[RenderedPage],
+        ocr: &[OcrPage],
+        options: &LayoutOptions,
+    ) -> Result<Vec<LayoutPage>, Self::Error>;
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn ocr_defaults_never_enable_recognition() {
+        let options = OcrOptions::default();
+        assert_eq!(options.mode, OcrMode::Off);
+    }
+
+    #[test]
+    fn offline_model_override_is_explicit() {
+        let options = OcrOptions::new()
+            .mode(OcrMode::Auto)
+            .model_directory("/models/pp-ocr")
+            .model_downloads(ModelDownloadPolicy::Offline);
+        assert_eq!(options.mode, OcrMode::Auto);
+        assert_eq!(options.model_downloads, ModelDownloadPolicy::Offline);
+        assert_eq!(
+            options.model_directory,
+            Some(PathBuf::from("/models/pp-ocr"))
+        );
+    }
+}

src/vision/download.rs

@@ -0,0 +1,84 @@
+//! HTTPS acquisition for pinned local model artifacts.
+
+use std::fmt;
+use std::io::Read;
+use std::time::Duration;
+
+use thiserror::Error;
+
+use super::{ModelArtifact, ModelDownloader};
+
+/// Default end-to-end timeout for one model artifact request.
+pub const DEFAULT_MODEL_DOWNLOAD_TIMEOUT: Duration = Duration::from_secs(5 * 60);
+
+/// Streaming HTTPS downloader used by lazy model resolution.
+#[derive(Clone)]
+pub struct HttpModelDownloader {
+    agent: ureq::Agent,
+}
+
+impl fmt::Debug for HttpModelDownloader {
+    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
+        formatter
+            .debug_struct("HttpModelDownloader")
+            .finish_non_exhaustive()
+    }
+}
+
+impl Default for HttpModelDownloader {
+    fn default() -> Self {
+        Self::new(DEFAULT_MODEL_DOWNLOAD_TIMEOUT)
+    }
+}
+
+impl HttpModelDownloader {
+    /// Creates an HTTPS-only downloader with an end-to-end request timeout.
+    pub fn new(timeout: Duration) -> Self {
+        let config = ureq::Agent::config_builder()
+            .https_only(true)
+            .timeout_global(Some(timeout))
+            .user_agent(concat!("pdf-inspector/", env!("CARGO_PKG_VERSION")))
+            .build();
+        Self {
+            agent: ureq::Agent::new_with_config(config),
+        }
+    }
+}
+
+impl ModelDownloader for HttpModelDownloader {
+    type Error = HttpModelDownloadError;
+
+    fn open(&self, artifact: &ModelArtifact) -> Result<Box<dyn Read + Send>, Self::Error> {
+        let response = self.agent.get(artifact.url).call()?;
+        if let Some(actual) = response.body().content_length() {
+            if actual != artifact.size {
+                return Err(HttpModelDownloadError::ContentLength {
+                    expected: artifact.size,
+                    actual,
+                });
+            }
+        }
+
+        // One extra byte lets ModelStore report an exact size mismatch while
+        // preventing a malicious or broken server from filling the disk.
+        let limit = artifact.size.saturating_add(1);
+        Ok(Box::new(response.into_body().into_reader().take(limit)))
+    }
+}
+
+/// Failures before a response stream reaches [`super::ModelStore`].
+#[derive(Debug, Error)]
+#[non_exhaustive]
+pub enum HttpModelDownloadError {
+    /// DNS, TLS, redirect, HTTP status, or response-stream setup failed.
+    #[error(transparent)]
+    Request(#[from] ureq::Error),
+    /// The server declared a size that disagrees with the pinned manifest.
+    #[error("server declared {actual} bytes; manifest requires {expected}")]
+    ContentLength {
+        /// Pinned artifact size.
+        expected: u64,
+        /// Server-declared size.
+        actual: u64,
+    },
+}

src/vision/fusion.rs

@@ -0,0 +1,797 @@
+//! Geometry-aware OCR Markdown assembly and native-text fusion.
+
+use std::collections::{BTreeMap, BTreeSet};
+use std::time::Instant;
+
+use thiserror::Error;
+
+use crate::markdown::{to_markdown_from_items_with_rects_and_page_count, MarkdownOptions};
+use crate::types::{ItemType, TextItem};
+use crate::PageMarkdown;
+
+use super::{
+    OcrRun, PageContentSource, PageProvenance, RoutedOcrPage, VisionTimings, DEFAULT_RENDER_DPI,
+};
+
+/// OCR assembly and hosted-fallback policy.
+#[derive(Debug, Clone)]
+pub struct OcrFusionOptions {
+    /// Markdown conversion options applied to positioned OCR spans.
+    pub markdown: MarkdownOptions,
+    /// Render resolution recorded in page provenance.
+    pub render_dpi: f32,
+    /// Recommend the hosted pipeline below this mean confidence when native
+    /// extraction already marked the page as requiring OCR.
+    pub hosted_recommendation_confidence: f32,
+}
+
+impl Default for OcrFusionOptions {
+    fn default() -> Self {
+        Self {
+            markdown: MarkdownOptions {
+                include_page_numbers: false,
+                strip_headers_footers: false,
+                ..MarkdownOptions::default()
+            },
+            render_dpi: DEFAULT_RENDER_DPI,
+            hosted_recommendation_confidence: 0.5,
+        }
+    }
+}
+
+impl OcrFusionOptions {
+    /// Creates OCR fusion options with local defaults.
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// Replaces Markdown conversion options.
+    pub fn markdown(mut self, markdown: MarkdownOptions) -> Self {
+        self.markdown = markdown;
+        self
+    }
+
+    /// Records the renderer resolution in provenance.
+    pub fn render_dpi(mut self, render_dpi: f32) -> Self {
+        self.render_dpi = render_dpi;
+        self
+    }
+
+    /// Sets the weak-OCR confidence threshold for recommending hosted parsing.
+    pub fn hosted_recommendation_confidence(mut self, confidence: f32) -> Self {
+        self.hosted_recommendation_confidence = confidence;
+        self
+    }
+
+    /// Validates rendering and hosted-fallback thresholds without doing work.
+    pub fn validate(&self) -> Result<(), OcrFusionError> {
+        validate_options(self)
+    }
+}
+
+/// Final Markdown and provenance for one page.
+#[derive(Debug, Clone, PartialEq)]
+pub struct FusedPageMarkdown {
+    /// 1-indexed document page number, matching OCR and provenance fields.
+    pub page: u32,
+    /// Final page Markdown.
+    pub markdown: String,
+    /// Native/OCR source, model, timing, and fallback metadata.
+    pub provenance: PageProvenance,
+}
+
+/// Output of fusing a selective OCR run into native page extraction.
+#[derive(Debug, Clone, PartialEq)]
+pub struct FusedPages {
+    /// Pages in the same order as the native input.
+    pub pages: Vec<FusedPageMarkdown>,
+    /// Batch rendering wall time from the OCR run.
+    pub render_time_ms: u64,
+    /// Batch OCR wall time from the OCR run.
+    pub ocr_time_ms: u64,
+}
+
+/// Converts positioned OCR spans to Markdown through pdf-inspector's existing
+/// deterministic geometry, reading-order, table, and Markdown pipeline.
+///
+/// The page number is 1-indexed. `document_page_count` prevents a selected
+/// page from being mistaken for a one-page document during page-number logic.
+pub fn ocr_page_to_markdown(
+    page: &RoutedOcrPage,
+    document_page_count: u32,
+    options: &MarkdownOptions,
+) -> String {
+    let (items, _) = ocr_text_items(page);
+    to_markdown_from_items_with_rects_and_page_count(
+        items,
+        options.clone(),
+        &[],
+        document_page_count,
+    )
+}
+
+/// Fuses a selective OCR run into per-page native Markdown.
+///
+/// OCR replaces pages whose native extraction was already rejected. On clean
+/// native pages (for example in `Force` mode), normalized duplicate OCR blocks
+/// are removed and only genuinely additional blocks are appended. Pages that
+/// needed OCR but still have no credible local result recommend the hosted
+/// document pipeline instead of silently presenting an empty result as final.
+pub fn fuse_ocr_pages(
+    native_pages: &[PageMarkdown],
+    ocr_run: &OcrRun,
+    document_page_count: u32,
+    options: &OcrFusionOptions,
+) -> Result<FusedPages, OcrFusionError> {
+    validate_options(options)?;
+
+    let mut native_numbers = BTreeSet::new();
+    for page in native_pages {
+        let page_number = page
+            .page
+            .checked_add(1)
+            .ok_or(OcrFusionError::PageOverflow)?;
+        if !native_numbers.insert(page_number) {
+            return Err(OcrFusionError::DuplicateNativePage { page: page_number });
+        }
+    }
+
+    let mut ocr_by_page = BTreeMap::new();
+    for page in &ocr_run.pages {
+        let page_number = page.rendered.page();
+        if !native_numbers.contains(&page_number) {
+            return Err(OcrFusionError::UnexpectedOcrPage { page: page_number });
+        }
+        if ocr_by_page.insert(page_number, page).is_some() {
+            return Err(OcrFusionError::DuplicateOcrPage { page: page_number });
+        }
+    }
+
+    let render_shares = equal_time_shares(ocr_run.render_time_ms, ocr_run.pages.len());
+    let render_by_page: BTreeMap<u32, u64> = ocr_run
+        .pages
+        .iter()
+        .zip(render_shares)
+        .map(|(page, share)| (page.rendered.page(), share))
+        .collect();
+
+    let mut pages = Vec::with_capacity(native_pages.len());
+    for native in native_pages {
+        let page_number = native.page + 1;
+        let assembly_started = Instant::now();
+        let mut warnings = Vec::new();
+
+        let (markdown, source, ocr_model, ocr_confidence, ocr_ms, hosted_recommended) =
+            if let Some(local) = ocr_by_page.get(&page_number) {
+                let (ocr_items, discarded_spans) = ocr_text_items(local);
+                if discarded_spans > 0 {
+                    warnings.push(format!(
+                        "discarded {discarded_spans} OCR spans with unusable text or geometry"
+                    ));
+                }
+                warnings.extend(local.ocr.warnings.iter().cloned());
+                let ocr_markdown = to_markdown_from_items_with_rects_and_page_count(
+                    ocr_items,
+                    options.markdown.clone(),
+                    &[],
+                    document_page_count,
+                );
+                let (markdown, source) = if native.markdown.trim().is_empty() || native.needs_ocr {
+                    (ocr_markdown, PageContentSource::Ocr)
+                } else {
+                    merge_native_and_ocr(&native.markdown, &ocr_markdown)
+                };
+                let weak_ocr = local
+                    .ocr
+                    .mean_confidence
+                    .is_none_or(|confidence| confidence < options.hosted_recommendation_confidence);
+                let recommend_hosted = native.needs_ocr && (markdown.trim().is_empty() || weak_ocr);
+                if native.needs_ocr && markdown.trim().is_empty() {
+                    warnings.push("OCR produced no usable text".to_string());
+                }
+                (
+                    markdown,
+                    source,
+                    Some(local.ocr.model.clone()),
+                    local.ocr.mean_confidence,
+                    local.ocr.processing_time_ms,
+                    recommend_hosted,
+                )
+            } else {
+                if native.needs_ocr {
+                    warnings.push("page was recommended for OCR but was not processed".to_string());
+                }
+                (
+                    native.markdown.clone(),
+                    PageContentSource::Native,
+                    None,
+                    None,
+                    0,
+                    native.needs_ocr,
+                )
+            };
+
+        pages.push(FusedPageMarkdown {
+            page: page_number,
+            markdown,
+            provenance: PageProvenance {
+                page: page_number,
+                source,
+                ocr_model,
+                layout_model: None,
+                render_dpi: ocr_by_page
+                    .contains_key(&page_number)
+                    .then_some(options.render_dpi),
+                ocr_confidence,
+                timings: VisionTimings {
+                    render_ms: render_by_page.get(&page_number).copied().unwrap_or(0),
+                    ocr_ms,
+                    layout_ms: 0,
+                    assembly_ms: elapsed_ms(assembly_started),
+                },
+                warnings,
+                hosted_recommended,
+            },
+        });
+    }
+
+    Ok(FusedPages {
+        pages,
+        render_time_ms: ocr_run.render_time_ms,
+        ocr_time_ms: ocr_run.ocr_time_ms,
+    })
+}
+
+/// Converts recognized line polygons to ordinary PDF-space text items.
+fn ocr_text_items(page: &RoutedOcrPage) -> (Vec<TextItem>, usize) {
+    let mut discarded = 0usize;
+    let mut items = Vec::with_capacity(page.ocr.spans.len());
+    for span in &page.ocr.spans {
+        if span.text.trim().is_empty() {
+            discarded += 1;
+            continue;
+        }
+        let Some((left, top, right, bottom)) = image_quad_bounds(
+            &span.polygon.points,
+            page.rendered.width(),
+            page.rendered.height(),
+        ) else {
+            discarded += 1;
+            continue;
+        };
+        let rect = page.rendered.pixel_rect_to_pdf_rect(
+            f64::from(left),
+            f64::from(top),
+            f64::from(right - left),
+            f64::from(bottom - top),
+        );
+        items.push(TextItem {
+            text: span.text.trim().to_string(),
+            x: rect.x,
+            y: rect.y,
+            width: rect.width,
+            height: rect.height,
+            font: "OCR".to_string(),
+            font_size: rect.height.max(1.0),
+            page: page.rendered.page(),
+            is_bold: false,
+            is_italic: false,
+            is_underline: false,
+            is_strikeout: false,
+            item_type: ItemType::Text,
+            mcid: None,
+        });
+    }
+    // OCR engines do not share an ordering contract. Geometry gives the
+    // deterministic top-to-bottom seed expected by the existing layout
+    // pipeline, which can still replace it with column-aware reading order.
+    items.sort_by(|first, second| {
+        first
+            .page
+            .cmp(&second.page)
+            .then(second.y.total_cmp(&first.y))
+            .then(first.x.total_cmp(&second.x))
+    });
+    (items, discarded)
+}
+
+fn image_quad_bounds(
+    points: &[super::ImagePoint; 4],
+    width: u32,
+    height: u32,
+) -> Option<(f32, f32, f32, f32)> {
+    if points
+        .iter()
+        .any(|point| !point.x.is_finite() || !point.y.is_finite())
+    {
+        return None;
+    }
+    let left = points
+        .iter()
+        .map(|point| point.x)
+        .fold(f32::INFINITY, f32::min)
+        .clamp(0.0, width as f32);
+    let right = points
+        .iter()
+        .map(|point| point.x)
+        .fold(f32::NEG_INFINITY, f32::max)
+        .clamp(0.0, width as f32);
+    let top = points
+        .iter()
+        .map(|point| point.y)
+        .fold(f32::INFINITY, f32::min)
+        .clamp(0.0, height as f32);
+    let bottom = points
+        .iter()
+        .map(|point| point.y)
+        .fold(f32::NEG_INFINITY, f32::max)
+        .clamp(0.0, height as f32);
+    (right > left && bottom > top).then_some((left, top, right, bottom))
+}
+
+fn merge_native_and_ocr(native: &str, ocr: &str) -> (String, PageContentSource) {
+    let native_keys = comparison_units(native);
+    let mut addition_keys = Vec::new();
+    let mut additions: Vec<String> = Vec::new();
+    for block in markdown_blocks(ocr) {
+        let key = ComparisonKey::new(block);
+        if key.is_empty()
+            || native_keys.iter().any(|native| native.same_content(&key))
+            || addition_keys
+                .iter()
+                .any(|existing: &ComparisonKey| existing.same_content(&key))
+        {
+            continue;
+        }
+
+        let addition = if let Some(native) = best_partial_overlap(&key, &native_keys) {
+            let Some(novel) = novel_fragments(block, native) else {
+                continue;
+            };
+            novel
+        } else {
+            block.trim_end().to_string()
+        };
+        let addition_key = ComparisonKey::new(&addition);
+        if addition_key.is_empty()
+            || native_keys
+                .iter()
+                .any(|native| native.same_content(&addition_key))
+            || addition_keys
+                .iter()
+                .any(|existing: &ComparisonKey| existing.same_content(&addition_key))
+        {
+            continue;
+        }
+        addition_keys.push(addition_key);
+        additions.push(addition);
+    }
+
+    if additions.is_empty() {
+        (ensure_trailing_newline(native), PageContentSource::Native)
+    } else {
+        let mut result = native.trim_end_matches('\n').to_string();
+        if !result.is_empty() {
+            result.push_str("\n\n");
+        }
+        result.push_str(&additions.join("\n\n"));
+        result.push('\n');
+        (result, PageContentSource::Fused)
+    }
+}
+
+fn markdown_blocks(markdown: &str) -> impl Iterator<Item = &str> {
+    markdown
+        .split("\n\n")
+        .map(str::trim_end)
+        .filter(|block| !block.trim().is_empty())
+}
+
+#[derive(Debug)]
+struct ComparisonKey {
+    tokens: Vec<String>,
+    punctuation: String,
+}
+
+impl ComparisonKey {
+    fn new(text: &str) -> Self {
+        let mut tokens = Vec::new();
+        let mut token = String::new();
+        let mut punctuation = String::new();
+        for character in text.chars().flat_map(char::to_lowercase) {
+            if character.is_alphanumeric() {
+                token.push(character);
+            } else {
+                if !token.is_empty() {
+                    tokens.push(std::mem::take(&mut token));
+                }
+                if !character.is_whitespace() {
+                    punctuation.push(character);
+                }
+            }
+        }
+        if !token.is_empty() {
+            tokens.push(token);
+        }
+        Self {
+            tokens,
+            punctuation,
+        }
+    }
+
+    fn is_empty(&self) -> bool {
+        self.tokens.is_empty() && self.punctuation.is_empty()
+    }
+
+    fn same_content(&self, other: &Self) -> bool {
+        if self.tokens.is_empty() || other.tokens.is_empty() {
+            self.tokens.is_empty()
+                && other.tokens.is_empty()
+                && self.punctuation == other.punctuation
+        } else {
+            token_counts(&self.tokens) == token_counts(&other.tokens)
+        }
+    }
+}
+
+fn comparison_units(markdown: &str) -> Vec<ComparisonKey> {
+    let mut units: Vec<_> = markdown_blocks(markdown).map(ComparisonKey::new).collect();
+    units.extend(
+        markdown
+            .lines()
+            .filter(|line| !line.trim().is_empty())
+            .map(ComparisonKey::new),
+    );
+    units
+}
+
+fn token_counts(tokens: &[String]) -> BTreeMap<&str, usize> {
+    let mut counts = BTreeMap::new();
+    for token in tokens {
+        *counts.entry(token.as_str()).or_insert(0) += 1;
+    }
+    counts
+}
+
+fn best_partial_overlap<'a>(
+    ocr: &ComparisonKey,
+    native: &'a [ComparisonKey],
+) -> Option<&'a ComparisonKey> {
+    if ocr.tokens.len() < 2 {
+        return None;
+    }
+    native
+        .iter()
+        .filter(|key| !key.tokens.is_empty())
+        .filter_map(|key| {
+            let overlap = token_overlap(&ocr.tokens, &key.tokens);
+            let substantial =
+                overlap >= 2 && overlap * 2 >= ocr.tokens.len() && overlap * 2 >= key.tokens.len();
+            substantial.then_some((overlap, key))
+        })
+        .max_by_key(|(overlap, _)| *overlap)
+        .map(|(_, key)| key)
+}
+
+fn token_overlap(first: &[String], second: &[String]) -> usize {
+    let first = token_counts(first);
+    let second = token_counts(second);
+    first
+        .iter()
+        .map(|(token, count)| (*count).min(second.get(token).copied().unwrap_or(0)))
+        .sum()
+}
+
+fn novel_fragments(block: &str, native: &ComparisonKey) -> Option<String> {
+    let mut available = token_counts(&native.tokens);
+    let mut novel = Vec::new();
+    for fragment in block.split_whitespace() {
+        let key = ComparisonKey::new(fragment);
+        let mut adds_content = false;
+        for token in &key.tokens {
+            match available.get_mut(token.as_str()) {
+                Some(count) if *count > 0 => *count -= 1,
+                _ => adds_content = true,
+            }
+        }
+        if adds_content {
+            novel.push(fragment);
+        }
+    }
+    (!novel.is_empty()).then(|| novel.join(" "))
+}
+
+fn ensure_trailing_newline(markdown: &str) -> String {
+    let mut result = markdown.trim_end_matches('\n').to_string();
+    result.push('\n');
+    result
+}
+
+fn equal_time_shares(total: u64, count: usize) -> Vec<u64> {
+    if count == 0 {
+        return Vec::new();
+    }
+    let count = count as u64;
+    let base = total / count;
+    let remainder = total % count;
+    (0..count)
+        .map(|index| base + u64::from(index < remainder))
+        .collect()
+}
+
+fn validate_options(options: &OcrFusionOptions) -> Result<(), OcrFusionError> {
+    if !options.render_dpi.is_finite() || options.render_dpi <= 0.0 {
+        return Err(OcrFusionError::InvalidRenderDpi {
+            value: options.render_dpi,
+        });
+    }
+    let confidence = options.hosted_recommendation_confidence;
+    if !confidence.is_finite() || !(0.0..=1.0).contains(&confidence) {
+        return Err(OcrFusionError::InvalidHostedConfidence { value: confidence });
+    }
+    Ok(())
+}
+
+fn elapsed_ms(started: Instant) -> u64 {
+    u64::try_from(started.elapsed().as_millis()).unwrap_or(u64::MAX)
+}
+
+/// Invalid page sets or fusion options.
+#[derive(Debug, Error, PartialEq)]
+#[non_exhaustive]
+pub enum OcrFusionError {
+    /// A 0-indexed native page could not be converted to 1-indexed form.
+    #[error("native page number cannot be converted to a 1-indexed page")]
+    PageOverflow,
+    /// Native input repeated a page number.
+    #[error("native page {page} appears more than once")]
+    DuplicateNativePage {
+        /// Repeated 1-indexed page number.
+        page: u32,
+    },
+    /// OCR input repeated a page number.
+    #[error("OCR page {page} appears more than once")]
+    DuplicateOcrPage {
+        /// Repeated 1-indexed page number.
+        page: u32,
+    },
+    /// OCR returned a page that was not part of native extraction.
+    #[error("OCR page {page} is not present in native page extraction")]
+    UnexpectedOcrPage {
+        /// Unexpected 1-indexed page number.
+        page: u32,
+    },
+    /// Render resolution is non-finite or non-positive.
+    #[error("render DPI must be positive and finite, got {value}")]
+    InvalidRenderDpi {
+        /// Invalid value.
+        value: f32,
+    },
+    /// Hosted recommendation confidence is outside 0–1.
+    #[error("hosted recommendation confidence must be between 0 and 1, got {value}")]
+    InvalidHostedConfidence {
+        /// Invalid value.
+        value: f32,
+    },
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::vision::{
+        ImagePoint, ImageQuad, ModelIdentity, OcrPage, OcrSpan, PageTransform, RenderPixelFormat,
+        RenderedPage,
+    };
+
+    fn native(page: u32, markdown: &str, needs_ocr: bool) -> PageMarkdown {
+        PageMarkdown {
+            page,
+            markdown: markdown.to_string(),
+            needs_ocr,
+            ocr_reason: needs_ocr.then(|| "scanned".to_string()),
+        }
+    }
+
+    fn rendered_page(page: u32) -> RenderedPage {
+        let transform =
+            PageTransform::from_corners(200, 100, (0.0, 100.0), (200.0, 100.0), (0.0, 0.0))
+                .unwrap();
+        RenderedPage::new(
+            page,
+            200.0,
+            100.0,
+            200,
+            100,
+            600,
+            RenderPixelFormat::Rgb8,
+            vec![255; 60_000],
+            transform,
+        )
+        .unwrap()
+    }
+
+    fn span(text: &str, top: f32, confidence: f32) -> OcrSpan {
+        OcrSpan {
+            text: text.to_string(),
+            polygon: ImageQuad::new([
+                ImagePoint::new(10.0, top),
+                ImagePoint::new(190.0, top),
+                ImagePoint::new(190.0, top + 10.0),
+                ImagePoint::new(10.0, top + 10.0),
+            ]),
+            confidence,
+            orientation_degrees: None,
+        }
+    }
+
+    fn routed_page(page: u32, spans: Vec<OcrSpan>, confidence: Option<f32>) -> RoutedOcrPage {
+        RoutedOcrPage {
+            rendered: rendered_page(page),
+            ocr: OcrPage {
+                page,
+                spans,
+                mean_confidence: confidence,
+                model: ModelIdentity::new("test-ocr", "v1"),
+                processing_time_ms: 7,
+                warnings: Vec::new(),
+            },
+        }
+    }
+
+    fn run(pages: Vec<RoutedOcrPage>) -> OcrRun {
+        OcrRun {
+            pages,
+            render_time_ms: 5,
+            ocr_time_ms: 7,
+        }
+    }
+
+    #[test]
+    fn scanned_page_uses_geometry_ordered_ocr_and_provenance() {
+        let native = [native(0, "", true)];
+        let run = run(vec![routed_page(
+            1,
+            vec![
+                span("Second line", 30.0, 0.9),
+                span("First line", 10.0, 0.9),
+            ],
+            Some(0.9),
+        )]);
+
+        let result = fuse_ocr_pages(&native, &run, 1, &OcrFusionOptions::new()).unwrap();
+
+        assert!(
+            result.pages[0].markdown.find("First").unwrap()
+                < result.pages[0].markdown.find("Second").unwrap()
+        );
+        assert_eq!(result.pages[0].provenance.source, PageContentSource::Ocr);
+        assert_eq!(result.pages[0].page, 1);
+        assert_eq!(result.pages[0].page, result.pages[0].provenance.page);
+        assert_eq!(
+            result.pages[0].provenance.ocr_model.as_ref().unwrap().name,
+            "test-ocr"
+        );
+        assert!(!result.pages[0].provenance.hosted_recommended);
+    }
+
+    #[test]
+    fn force_mode_deduplicates_native_content() {
+        let native = [native(0, "Hello, world!\n", false)];
+        let run = run(vec![routed_page(
+            1,
+            vec![span("Hello world", 10.0, 0.9)],
+            Some(0.9),
+        )]);
+
+        let result = fuse_ocr_pages(&native, &run, 1, &OcrFusionOptions::new()).unwrap();
+
+        assert_eq!(result.pages[0].markdown, "Hello, world!\n");
+        assert_eq!(result.pages[0].provenance.source, PageContentSource::Native);
+    }
+
+    #[test]
+    fn force_mode_appends_only_additional_ocr_blocks() {
+        let native = [native(0, "Native title\n", false)];
+        let run = run(vec![routed_page(
+            1,
+            vec![
+                span("Native title", 10.0, 0.9),
+                span("Image-only label", 30.0, 0.9),
+            ],
+            Some(0.9),
+        )]);
+
+        let result = fuse_ocr_pages(&native, &run, 1, &OcrFusionOptions::new()).unwrap();
+
+        assert_eq!(result.pages[0].provenance.source, PageContentSource::Fused);
+        assert_eq!(result.pages[0].markdown.matches("Native title").count(), 1);
+        assert!(result.pages[0].markdown.contains("Image-only label"));
+    }
+
+    #[test]
+    fn force_mode_preserves_punctuation_and_short_distinct_blocks() {
+        let (markdown, source) = merge_native_and_ocr("Figure 1\n", "1\n\n***");
+        assert_eq!(source, PageContentSource::Fused);
+        assert_eq!(markdown, "Figure 1\n\n1\n\n***\n");
+    }
+
+    #[test]
+    fn force_mode_extracts_only_novel_tokens_from_overlapping_blocks() {
+        let (markdown, source) = merge_native_and_ocr("A C\n", "## A B C");
+        assert_eq!(source, PageContentSource::Fused);
+        assert_eq!(markdown, "A C\n\nB\n");
+        assert_eq!(markdown.matches('A').count(), 1);
+        assert_eq!(markdown.matches('C').count(), 1);
+    }
+
+    #[test]
+    fn force_mode_deduplicates_reordered_tokens() {
+        let native = [native(0, "Alpha Beta Gamma\n", false)];
+        let run = run(vec![routed_page(
+            1,
+            vec![span("Gamma Alpha Beta", 10.0, 0.9)],
+            Some(0.9),
+        )]);
+
+        let result = fuse_ocr_pages(&native, &run, 1, &OcrFusionOptions::new()).unwrap();
+
+        assert_eq!(result.pages[0].provenance.source, PageContentSource::Native);
+        assert_eq!(result.pages[0].markdown, "Alpha Beta Gamma\n");
+    }
+
+    #[test]
+    fn fusion_preserves_native_leading_indentation() {
+        let native = [native(0, "    indented code\n", false)];
+        let duplicate = run(vec![routed_page(
+            1,
+            vec![span("indented code", 10.0, 0.9)],
+            Some(0.9),
+        )]);
+        let result = fuse_ocr_pages(&native, &duplicate, 1, &OcrFusionOptions::new()).unwrap();
+        assert_eq!(result.pages[0].markdown, "    indented code\n");
+
+        let addition = run(vec![routed_page(
+            1,
+            vec![span("image label", 10.0, 0.9)],
+            Some(0.9),
+        )]);
+        let result = fuse_ocr_pages(&native, &addition, 1, &OcrFusionOptions::new()).unwrap();
+        assert_eq!(
+            result.pages[0].markdown,
+            "    indented code\n\nimage label\n"
+        );
+    }
+
+    #[test]
+    fn missing_or_weak_required_ocr_recommends_hosted() {
+        let native = [native(0, "", true), native(1, "", true)];
+        let run = run(vec![routed_page(
+            2,
+            vec![span("uncertain", 10.0, 0.3)],
+            Some(0.3),
+        )]);
+
+        let result = fuse_ocr_pages(&native, &run, 2, &OcrFusionOptions::new()).unwrap();
+
+        assert!(result.pages[0].provenance.hosted_recommended);
+        assert!(result.pages[1].provenance.hosted_recommended);
+    }
+
+    #[test]
+    fn rejects_ocr_pages_outside_native_selection() {
+        let error = fuse_ocr_pages(
+            &[native(0, "text", false)],
+            &run(vec![routed_page(2, Vec::new(), None)]),
+            2,
+            &OcrFusionOptions::new(),
+        )
+        .unwrap_err();
+        assert_eq!(error, OcrFusionError::UnexpectedOcrPage { page: 2 });
+    }
+
+    #[test]
+    fn time_shares_preserve_batch_total() {
+        assert_eq!(equal_time_shares(8, 3), vec![3, 3, 2]);
+        assert_eq!(equal_time_shares(8, 0), Vec::<u64>::new());
+    }
+}

src/vision/mod.rs

@@ -0,0 +1,66 @@
+//! Optional native vision primitives used by OCR pipelines.
+//!
+//! The existing lopdf extractor remains the default path. Native page
+//! rendering is available only with the `render-pdfium` feature. Engine
+//! contracts are available with `vision`, while checksum-verified model
+//! resolution is a separate `model-cache` feature. The `ocr-oar` feature adds
+//! a CPU PP-OCRv6 Small implementation of [`OcrEngine`]. These remain separate
+//! so browser WASM, text-only consumers, and renderer-only users take on no
+//! model-management or inference dependencies.
+
+#[cfg(all(feature = "vision", not(target_arch = "wasm32")))]
+mod contracts;
+#[cfg(all(feature = "model-download", not(target_arch = "wasm32")))]
+mod download;
+#[cfg(all(feature = "vision", not(target_arch = "wasm32")))]
+mod fusion;
+#[cfg(all(feature = "model-cache", not(target_arch = "wasm32")))]
+mod models;
+#[cfg(all(feature = "ocr-oar", not(target_arch = "wasm32")))]
+mod oar;
+#[cfg(all(feature = "ocr", not(target_arch = "wasm32")))]
+mod pipeline;
+#[cfg(all(feature = "vision", not(target_arch = "wasm32")))]
+mod render;
+#[cfg(all(feature = "vision", not(target_arch = "wasm32")))]
+mod routing;
+
+#[cfg(all(feature = "render-pdfium", not(target_arch = "wasm32")))]
+mod pdfium;
+
+#[cfg(all(feature = "vision", not(target_arch = "wasm32")))]
+pub use contracts::{
+    ImagePoint, ImageQuad, LayoutEngine, LayoutOptions, LayoutPage, LayoutRegion, LayoutRegionKind,
+    ModelDownloadPolicy, ModelIdentity, OcrEngine, OcrMode, OcrOptions, OcrPage, OcrProfile,
+    OcrSpan, PageContentSource, PageProvenance, PageRenderer, VisionTimings,
+};
+#[cfg(all(feature = "model-download", not(target_arch = "wasm32")))]
+pub use download::{HttpModelDownloadError, HttpModelDownloader, DEFAULT_MODEL_DOWNLOAD_TIMEOUT};
+#[cfg(all(feature = "vision", not(target_arch = "wasm32")))]
+pub use fusion::{
+    fuse_ocr_pages, ocr_page_to_markdown, FusedPageMarkdown, FusedPages, OcrFusionError,
+    OcrFusionOptions,
+};
+#[cfg(all(feature = "model-cache", not(target_arch = "wasm32")))]
+pub use models::{
+    ModelAcquireError, ModelArtifact, ModelArtifactKind, ModelDownloader, ModelManifest,
+    ModelPaths, ModelStore, ModelStoreError, PP_OCR_V6_SMALL,
+};
+#[cfg(all(feature = "ocr-oar", not(target_arch = "wasm32")))]
+pub use oar::{OarOcrEngine, OarOcrError, ONNX_RUNTIME_LIBRARY_ENV};
+#[cfg(all(feature = "ocr", not(target_arch = "wasm32")))]
+pub use pipeline::{
+    process_pdf_with_ocr, process_pdf_with_ocr_mem, OcrPdfOptions, OcrPdfResult, OcrPipelineError,
+};
+#[cfg(all(feature = "vision", not(target_arch = "wasm32")))]
+pub use render::{
+    PagePoint, PageTransform, RenderBufferError, RenderOptions, RenderPixelFormat, RenderedPage,
+    DEFAULT_RENDER_DPI,
+};
+#[cfg(all(feature = "vision", not(target_arch = "wasm32")))]
+pub use routing::{
+    route_ocr_pages, run_ocr_pages, OcrRoutingError, OcrRun, OcrRunError, RoutedOcrPage,
+};
+
+#[cfg(all(feature = "render-pdfium", not(target_arch = "wasm32")))]
+pub use pdfium::{PdfiumRenderer, RenderError};

src/vision/oar.rs

@@ -0,0 +1,469 @@
+//! PP-OCRv6 Small implementation backed by OAR and ONNX Runtime.
+
+use std::path::PathBuf;
+use std::time::Instant;
+
+use image::RgbImage;
+use oar_ocr::oarocr::{OAROCRBuilder, OAROCR};
+use oar_ocr::processors::BoundingBox;
+use thiserror::Error;
+
+use super::{
+    ImagePoint, ImageQuad, ModelArtifactKind, ModelIdentity, ModelPaths, OcrEngine, OcrMode,
+    OcrOptions, OcrPage, OcrSpan, RenderPixelFormat, RenderedPage,
+};
+
+/// Environment variable selecting the ONNX Runtime shared library.
+pub const ONNX_RUNTIME_LIBRARY_ENV: &str = "ORT_DYLIB_PATH";
+
+/// Failures while constructing or running the OAR OCR backend.
+#[derive(Debug, Error)]
+#[non_exhaustive]
+pub enum OarOcrError {
+    /// A required file is missing from the resolved model set.
+    #[error("resolved OCR model set is missing {kind:?}")]
+    MissingModelArtifact {
+        /// Missing artifact role.
+        kind: ModelArtifactKind,
+    },
+    /// OCR was invoked while the caller explicitly disabled it.
+    #[error("OCR is disabled; select Auto or Force before invoking the engine")]
+    OcrDisabled,
+    /// Confidence thresholds must match the normalized engine output range.
+    #[error("minimum OCR confidence must be finite and between 0 and 1, got {value}")]
+    InvalidMinimumConfidence {
+        /// Invalid threshold.
+        value: f32,
+    },
+    /// Bitmap dimension arithmetic exceeded the host address space.
+    #[error("rendered page {page} bitmap dimensions overflow the host address space")]
+    ImageSizeOverflow {
+        /// 1-indexed page number.
+        page: u32,
+    },
+    /// A validated renderer buffer could not be represented as an RGB image.
+    #[error("rendered page {page} could not be converted to an RGB image")]
+    InvalidImageBuffer {
+        /// 1-indexed page number.
+        page: u32,
+    },
+    /// The external ONNX Runtime shared library could not be loaded.
+    #[error("failed to load ONNX Runtime from {path}: {source}")]
+    OnnxRuntimeLoad {
+        /// Requested shared-library path or platform library name.
+        path: PathBuf,
+        /// Dynamic-loader failure.
+        #[source]
+        source: ort::LoadDynamicError,
+    },
+    /// OAR returned no result for a submitted page.
+    #[error("OAR returned no result for rendered page {page}")]
+    MissingPageResult {
+        /// 1-indexed page number.
+        page: u32,
+    },
+    /// OAR or ONNX Runtime rejected the models or failed during inference.
+    #[error(transparent)]
+    Backend(#[from] oar_ocr::core::OCRError),
+}
+
+/// CPU PP-OCRv6 Small engine using OAR's detection and recognition pipeline.
+///
+/// Construction accepts only [`ModelPaths`] that have already passed
+/// pdf-inspector's manifest size and SHA-256 verification. OAR's independent
+/// model auto-download feature is deliberately not enabled.
+#[derive(Debug)]
+pub struct OarOcrEngine {
+    pipeline: OAROCR,
+    model: ModelIdentity,
+}
+
+impl OarOcrEngine {
+    /// Loads PP-OCRv6 Small from a resolved, verified model set.
+    pub fn from_models(models: &ModelPaths) -> Result<Self, OarOcrError> {
+        load_onnx_runtime()?;
+        let detection = required_model(models, ModelArtifactKind::TextDetection)?;
+        let recognition = required_model(models, ModelArtifactKind::TextRecognition)?;
+        let dictionary = required_model(models, ModelArtifactKind::CharacterDictionary)?;
+
+        let pipeline = OAROCRBuilder::new(detection, recognition, dictionary).build()?;
+        let model = ModelIdentity::new(models.manifest_id(), models.revision());
+        Ok(Self { pipeline, model })
+    }
+
+    fn recognize_page(
+        &self,
+        page: &RenderedPage,
+        options: &OcrOptions,
+    ) -> Result<OcrPage, OarOcrError> {
+        let started = Instant::now();
+        let image = rendered_page_to_rgb(page)?;
+        let result = self
+            .pipeline
+            .predict(vec![image])?
+            .into_iter()
+            .next()
+            .ok_or(OarOcrError::MissingPageResult { page: page.page() })?;
+
+        let mut spans = Vec::with_capacity(result.text_regions.len());
+        let mut invalid_geometry = 0usize;
+        let mut missing_recognition = 0usize;
+        for region in result.text_regions {
+            let (Some(text), Some(confidence)) = (region.text, region.confidence) else {
+                missing_recognition += 1;
+                continue;
+            };
+            if text.trim().is_empty() || !confidence.is_finite() {
+                missing_recognition += 1;
+                continue;
+            }
+            let confidence = confidence.clamp(0.0, 1.0);
+            if confidence < options.minimum_confidence {
+                continue;
+            }
+
+            let polygon = region.dt_poly.as_ref().unwrap_or(&region.bounding_box);
+            let Some(polygon) = bounding_box_to_quad(polygon, page.width(), page.height()) else {
+                invalid_geometry += 1;
+                continue;
+            };
+            spans.push(OcrSpan {
+                text: text.to_string(),
+                polygon,
+                confidence,
+                orientation_degrees: region.orientation_angle,
+            });
+        }
+
+        let mut warnings = Vec::new();
+        if !options.languages.is_empty() {
+            warnings
+                .push("language hints are not used by the PP-OCRv6 Small OAR backend".to_string());
+        }
+        if missing_recognition > 0 {
+            warnings.push(format!(
+                "discarded {missing_recognition} regions without usable recognition output"
+            ));
+        }
+        if invalid_geometry > 0 {
+            warnings.push(format!(
+                "discarded {invalid_geometry} recognized regions with invalid geometry"
+            ));
+        }
+
+        let mean_confidence = if spans.is_empty() {
+            None
+        } else {
+            Some(spans.iter().map(|span| span.confidence).sum::<f32>() / spans.len() as f32)
+        };
+        let processing_time_ms = u64::try_from(started.elapsed().as_millis()).unwrap_or(u64::MAX);
+
+        Ok(OcrPage {
+            page: page.page(),
+            spans,
+            mean_confidence,
+            model: self.model.clone(),
+            processing_time_ms,
+            warnings,
+        })
+    }
+}
+
+fn load_onnx_runtime() -> Result<(), OarOcrError> {
+    let path = std::env::var_os(ONNX_RUNTIME_LIBRARY_ENV)
+        .filter(|path| !path.is_empty())
+        .map(PathBuf::from)
+        .unwrap_or_else(default_onnx_runtime_library);
+    drop(
+        ort::init_from(&path).map_err(|source| OarOcrError::OnnxRuntimeLoad {
+            path: path.clone(),
+            source,
+        })?,
+    );
+    Ok(())
+}
+
+fn default_onnx_runtime_library() -> PathBuf {
+    #[cfg(target_os = "windows")]
+    const NAME: &str = "onnxruntime.dll";
+    #[cfg(any(target_os = "linux", target_os = "android", target_os = "freebsd"))]
+    const NAME: &str = "libonnxruntime.so";
+    #[cfg(any(target_os = "macos", target_os = "ios"))]
+    const NAME: &str = "libonnxruntime.dylib";
+    PathBuf::from(NAME)
+}
+
+impl OcrEngine for OarOcrEngine {
+    type Error = OarOcrError;
+
+    fn model(&self) -> &ModelIdentity {
+        &self.model
+    }
+
+    fn recognize(
+        &self,
+        pages: &[RenderedPage],
+        options: &OcrOptions,
+    ) -> Result<Vec<OcrPage>, Self::Error> {
+        validate_options(options)?;
+
+        pages
+            .iter()
+            .map(|page| self.recognize_page(page, options))
+            .collect()
+    }
+}
+
+fn validate_options(options: &OcrOptions) -> Result<(), OarOcrError> {
+    if options.mode == OcrMode::Off {
+        return Err(OarOcrError::OcrDisabled);
+    }
+    if !options.minimum_confidence.is_finite() || !(0.0..=1.0).contains(&options.minimum_confidence)
+    {
+        return Err(OarOcrError::InvalidMinimumConfidence {
+            value: options.minimum_confidence,
+        });
+    }
+    Ok(())
+}
+
+fn required_model(
+    models: &ModelPaths,
+    kind: ModelArtifactKind,
+) -> Result<&std::path::Path, OarOcrError> {
+    models
+        .get(kind)
+        .ok_or(OarOcrError::MissingModelArtifact { kind })
+}
+
+fn rendered_page_to_rgb(page: &RenderedPage) -> Result<RgbImage, OarOcrError> {
+    let width = usize::try_from(page.width())
+        .map_err(|_| OarOcrError::ImageSizeOverflow { page: page.page() })?;
+    let height = usize::try_from(page.height())
+        .map_err(|_| OarOcrError::ImageSizeOverflow { page: page.page() })?;
+    let output_len = width
+        .checked_mul(height)
+        .and_then(|pixels| pixels.checked_mul(3))
+        .ok_or(OarOcrError::ImageSizeOverflow { page: page.page() })?;
+    let input_bpp = page.format().bytes_per_pixel();
+    let active_input_row = width
+        .checked_mul(input_bpp)
+        .ok_or(OarOcrError::ImageSizeOverflow { page: page.page() })?;
+    let output_row = width
+        .checked_mul(3)
+        .ok_or(OarOcrError::ImageSizeOverflow { page: page.page() })?;
+
+    let mut rgb = vec![0u8; output_len];
+    for row in 0..height {
+        let input_start = row * page.stride();
+        let input = &page.pixels()[input_start..input_start + active_input_row];
+        let output_start = row * output_row;
+        let output = &mut rgb[output_start..output_start + output_row];
+        match page.format() {
+            RenderPixelFormat::Rgb8 => output.copy_from_slice(input),
+            RenderPixelFormat::Rgba8 => {
+                for (rgba, rgb) in input.chunks_exact(4).zip(output.chunks_exact_mut(3)) {
+                    rgb.copy_from_slice(&rgba[..3]);
+                }
+            }
+            RenderPixelFormat::Gray8 => {
+                for (&gray, rgb) in input.iter().zip(output.chunks_exact_mut(3)) {
+                    rgb.fill(gray);
+                }
+            }
+        }
+    }
+
+    RgbImage::from_raw(page.width(), page.height(), rgb)
+        .ok_or(OarOcrError::InvalidImageBuffer { page: page.page() })
+}
+
+fn bounding_box_to_quad(bounding_box: &BoundingBox, width: u32, height: u32) -> Option<ImageQuad> {
+    let points: Vec<ImagePoint> = bounding_box
+        .points
+        .iter()
+        .filter(|point| point.x.is_finite() && point.y.is_finite())
+        .map(|point| {
+            ImagePoint::new(
+                point.x.clamp(0.0, width as f32),
+                point.y.clamp(0.0, height as f32),
+            )
+        })
+        .collect();
+
+    if bounding_box.points.len() == 4 && points.len() == 4 && is_ordered_convex_quad(&points) {
+        return Some(ImageQuad::new([points[0], points[1], points[2], points[3]]));
+    }
+    if points.len() < 3 {
+        return None;
+    }
+
+    let min_x = points
+        .iter()
+        .map(|point| point.x)
+        .fold(f32::INFINITY, f32::min);
+    let max_x = points
+        .iter()
+        .map(|point| point.x)
+        .fold(f32::NEG_INFINITY, f32::max);
+    let min_y = points
+        .iter()
+        .map(|point| point.y)
+        .fold(f32::INFINITY, f32::min);
+    let max_y = points
+        .iter()
+        .map(|point| point.y)
+        .fold(f32::NEG_INFINITY, f32::max);
+    if max_x <= min_x || max_y <= min_y {
+        return None;
+    }
+    Some(ImageQuad::new([
+        ImagePoint::new(min_x, min_y),
+        ImagePoint::new(max_x, min_y),
+        ImagePoint::new(max_x, max_y),
+        ImagePoint::new(min_x, max_y),
+    ]))
+}
+
+fn is_ordered_convex_quad(points: &[ImagePoint]) -> bool {
+    if points.len() != 4 {
+        return false;
+    }
+    let mut orientation = 0.0_f32;
+    for index in 0..4 {
+        let first = points[index];
+        let second = points[(index + 1) % 4];
+        let third = points[(index + 2) % 4];
+        let cross = (second.x - first.x) * (third.y - second.y)
+            - (second.y - first.y) * (third.x - second.x);
+        if cross.abs() <= f32::EPSILON {
+            return false;
+        }
+        if orientation == 0.0 {
+            orientation = cross.signum();
+        } else if cross.signum() != orientation {
+            return false;
+        }
+    }
+    true
+}
+
+#[cfg(test)]
+mod tests {
+    use oar_ocr::processors::Point;
+
+    use super::*;
+    use crate::vision::PageTransform;
+
+    fn page(format: RenderPixelFormat, stride: usize, pixels: Vec<u8>) -> RenderedPage {
+        let transform =
+            PageTransform::from_corners(2, 2, (0.0, 2.0), (2.0, 2.0), (0.0, 0.0)).unwrap();
+        RenderedPage::new(1, 2.0, 2.0, 2, 2, stride, format, pixels, transform).unwrap()
+    }
+
+    #[test]
+    fn converts_padded_rgb_without_exposing_padding() {
+        let page = page(
+            RenderPixelFormat::Rgb8,
+            8,
+            vec![1, 2, 3, 4, 5, 6, 99, 99, 7, 8, 9, 10, 11, 12, 99, 99],
+        );
+        let image = rendered_page_to_rgb(&page).unwrap();
+        assert_eq!(image.as_raw(), &[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]);
+    }
+
+    #[test]
+    fn converts_rgba_and_gray_to_rgb() {
+        let rgba = page(
+            RenderPixelFormat::Rgba8,
+            8,
+            vec![1, 2, 3, 44, 4, 5, 6, 55, 7, 8, 9, 66, 10, 11, 12, 77],
+        );
+        assert_eq!(
+            rendered_page_to_rgb(&rgba).unwrap().as_raw(),
+            &[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]
+        );
+
+        let gray = page(RenderPixelFormat::Gray8, 2, vec![1, 2, 3, 4]);
+        assert_eq!(
+            rendered_page_to_rgb(&gray).unwrap().as_raw(),
+            &[1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4]
+        );
+    }
+
+    #[test]
+    fn preserves_quads_and_clamps_them_to_the_bitmap() {
+        let bbox = BoundingBox::new(vec![
+            Point::new(-1.0, 2.0),
+            Point::new(11.0, 2.0),
+            Point::new(11.0, 9.0),
+            Point::new(-1.0, 9.0),
+        ]);
+        let quad = bounding_box_to_quad(&bbox, 10, 8).unwrap();
+        assert_eq!(quad.points[0], ImagePoint::new(0.0, 2.0));
+        assert_eq!(quad.points[2], ImagePoint::new(10.0, 8.0));
+    }
+
+    #[test]
+    fn reduces_polygons_to_a_stable_axis_aligned_quad() {
+        let bbox = BoundingBox::new(vec![
+            Point::new(2.0, 1.0),
+            Point::new(7.0, 2.0),
+            Point::new(8.0, 6.0),
+            Point::new(5.0, 9.0),
+            Point::new(1.0, 5.0),
+        ]);
+        let quad = bounding_box_to_quad(&bbox, 10, 10).unwrap();
+        assert_eq!(quad.points[0], ImagePoint::new(1.0, 1.0));
+        assert_eq!(quad.points[2], ImagePoint::new(8.0, 9.0));
+    }
+
+    #[test]
+    fn normalizes_unordered_or_partially_invalid_quads() {
+        let unordered = BoundingBox::new(vec![
+            Point::new(1.0, 1.0),
+            Point::new(8.0, 8.0),
+            Point::new(8.0, 1.0),
+            Point::new(1.0, 8.0),
+        ]);
+        let quad = bounding_box_to_quad(&unordered, 10, 10).unwrap();
+        assert_eq!(quad.points[0], ImagePoint::new(1.0, 1.0));
+        assert_eq!(quad.points[1], ImagePoint::new(8.0, 1.0));
+        assert_eq!(quad.points[2], ImagePoint::new(8.0, 8.0));
+
+        let partially_invalid = BoundingBox::new(vec![
+            Point::new(8.0, 8.0),
+            Point::new(f32::NAN, 4.0),
+            Point::new(1.0, 8.0),
+            Point::new(8.0, 1.0),
+            Point::new(1.0, 1.0),
+        ]);
+        let quad = bounding_box_to_quad(&partially_invalid, 10, 10).unwrap();
+        assert_eq!(quad.points[0], ImagePoint::new(1.0, 1.0));
+        assert_eq!(quad.points[1], ImagePoint::new(8.0, 1.0));
+        assert_eq!(quad.points[2], ImagePoint::new(8.0, 8.0));
+    }
+
+    #[test]
+    fn refuses_disabled_or_invalid_options_before_inference() {
+        assert!(matches!(
+            validate_options(&OcrOptions::new()),
+            Err(OarOcrError::OcrDisabled)
+        ));
+        for value in [-0.1, 1.1, f32::NAN, f32::INFINITY] {
+            let options = OcrOptions::new()
+                .mode(OcrMode::Force)
+                .minimum_confidence(value);
+            assert!(matches!(
+                validate_options(&options),
+                Err(OarOcrError::InvalidMinimumConfidence { .. })
+            ));
+        }
+        assert!(validate_options(
+            &OcrOptions::new()
+                .mode(OcrMode::Auto)
+                .minimum_confidence(1.0)
+        )
+        .is_ok());
+    }
+}

src/vision/pdfium.rs

@@ -0,0 +1,266 @@
+//! PDFium-backed implementation of the renderer-neutral page contract.
+
+use std::path::Path;
+
+use firecrawl_pdfium::{Pdfium, PixelFormat, PixelPoint, RenderConfig};
+use thiserror::Error;
+
+use super::{
+    PageRenderer, PageTransform, RenderBufferError, RenderOptions, RenderPixelFormat, RenderedPage,
+};
+
+impl RenderPixelFormat {
+    fn pdfium_format(self) -> PixelFormat {
+        match self {
+            // PDFium produces BGR directly; `rendered_page_from_pdfium`
+            // swaps the red and blue channels in place.
+            Self::Rgb8 => PixelFormat::Bgr8,
+            Self::Rgba8 => PixelFormat::Rgba8,
+            Self::Gray8 => PixelFormat::Gray8,
+        }
+    }
+}
+
+impl RenderOptions {
+    fn pdfium_config(&self) -> RenderConfig {
+        RenderConfig::new()
+            .dpi(self.dpi)
+            .pixel_format(self.pixel_format.pdfium_format())
+            .annotations(self.annotations)
+            .form_fields(self.form_fields)
+            .max_output_bytes(self.max_output_bytes_per_page)
+    }
+}
+
+/// Errors produced by the optional local renderer.
+#[derive(Debug, Error)]
+#[non_exhaustive]
+pub enum RenderError {
+    /// Page numbers in pdf-inspector APIs are 1-indexed, so zero is invalid.
+    #[error("page numbers are 1-indexed; page 0 is invalid")]
+    InvalidPageNumber,
+    /// The requested 1-indexed page is not present in the document.
+    #[error("page {page} is out of bounds for a {page_count}-page document")]
+    PageOutOfBounds {
+        /// Requested 1-indexed page.
+        page: u32,
+        /// Number of pages in the document.
+        page_count: usize,
+    },
+    /// PDFium loading, document parsing, form setup, or rendering failed.
+    #[error(transparent)]
+    Pdfium(#[from] firecrawl_pdfium::Error),
+    /// PDFium returned an internally inconsistent bitmap or transform.
+    #[error(transparent)]
+    Buffer(#[from] RenderBufferError),
+}
+
+/// Loaded PDFium renderer used to prepare pages for OCR.
+///
+/// PDFium calls are safe from concurrent threads but serialize inside the
+/// underlying binding. Returned [`RenderedPage`] values are ordinary owned
+/// data and can be processed concurrently after rendering.
+#[derive(Debug, Clone, Copy)]
+pub struct PdfiumRenderer {
+    pdfium: Pdfium,
+}
+
+impl PdfiumRenderer {
+    /// Loads PDFium using `firecrawl-pdfium`'s documented discovery chain.
+    pub fn load() -> Result<Self, RenderError> {
+        Ok(Self {
+            pdfium: Pdfium::load()?,
+        })
+    }
+
+    /// Loads PDFium from an explicit native library path.
+    pub fn load_from_path(path: impl AsRef<Path>) -> Result<Self, RenderError> {
+        Ok(Self {
+            pdfium: Pdfium::load_from_path(path)?,
+        })
+    }
+
+    /// Path of the active PDFium library, if it was loaded from a concrete
+    /// file rather than through the system loader.
+    pub fn loaded_from(&self) -> Option<&Path> {
+        self.pdfium.loaded_from()
+    }
+
+    /// Renders selected 1-indexed pages in the same order as `pages`.
+    ///
+    /// This inherent method mirrors [`PageRenderer`] so existing callers do
+    /// not need to import the trait.
+    pub fn render_pages(
+        &self,
+        pdf_bytes: &[u8],
+        pages: &[u32],
+        password: Option<&str>,
+        options: &RenderOptions,
+    ) -> Result<Vec<RenderedPage>, RenderError> {
+        self.render_pages_impl(pdf_bytes, pages, password, options)
+    }
+
+    fn render_pages_impl(
+        &self,
+        pdf_bytes: &[u8],
+        pages: &[u32],
+        password: Option<&str>,
+        options: &RenderOptions,
+    ) -> Result<Vec<RenderedPage>, RenderError> {
+        if pages.is_empty() {
+            return Ok(Vec::new());
+        }
+
+        if pages.contains(&0) {
+            return Err(RenderError::InvalidPageNumber);
+        }
+
+        let document = self.pdfium.load_document(pdf_bytes.to_vec(), password)?;
+        let page_count = document.page_count();
+
+        if let Some(&page) = pages.iter().find(|&&page| page as usize > page_count) {
+            return Err(RenderError::PageOutOfBounds { page, page_count });
+        }
+
+        if options.form_fields {
+            document.enable_form_rendering()?;
+        }
+
+        let config = options.pdfium_config();
+        let mut rendered_pages = Vec::with_capacity(pages.len());
+        for &page_number in pages {
+            let page = document.page(page_number as usize - 1)?;
+            let size = page.size();
+            let rendered = page.render(&config)?;
+            rendered_pages.push(rendered_page_from_pdfium(
+                page_number,
+                size.width,
+                size.height,
+                options.pixel_format,
+                rendered,
+            )?);
+        }
+
+        Ok(rendered_pages)
+    }
+}
+
+impl PageRenderer for PdfiumRenderer {
+    type Error = RenderError;
+
+    fn render_pages(
+        &self,
+        pdf_bytes: &[u8],
+        pages: &[u32],
+        password: Option<&str>,
+        options: &RenderOptions,
+    ) -> Result<Vec<RenderedPage>, Self::Error> {
+        self.render_pages_impl(pdf_bytes, pages, password, options)
+    }
+}
+
+fn rendered_page_from_pdfium(
+    page: u32,
+    page_width: f32,
+    page_height: f32,
+    format: RenderPixelFormat,
+    rendered: firecrawl_pdfium::RenderedPage,
+) -> Result<RenderedPage, RenderBufferError> {
+    let width = rendered.width();
+    let height = rendered.height();
+    let stride = rendered.stride();
+    let pdfium_transform = *rendered.transform();
+    let corner = |x, y| {
+        let point = pdfium_transform.pixel_to_page(PixelPoint::new(x, y));
+        (point.x, point.y)
+    };
+    let transform = PageTransform::from_corners(
+        width,
+        height,
+        corner(0.0, 0.0),
+        corner(f64::from(width), 0.0),
+        corner(0.0, f64::from(height)),
+    )
+    .ok_or(RenderBufferError::InvalidTransform)?;
+    let mut pixels = rendered.into_pixels();
+
+    if format == RenderPixelFormat::Rgb8 {
+        bgr_to_rgb_in_place(&mut pixels, width, height, stride)?;
+    }
+
+    RenderedPage::new(
+        page,
+        page_width,
+        page_height,
+        width,
+        height,
+        stride,
+        format,
+        pixels,
+        transform,
+    )
+}
+
+fn bgr_to_rgb_in_place(
+    pixels: &mut [u8],
+    width: u32,
+    height: u32,
+    stride: usize,
+) -> Result<(), RenderBufferError> {
+    let row_bytes = (width as usize)
+        .checked_mul(RenderPixelFormat::Rgb8.bytes_per_pixel())
+        .ok_or(RenderBufferError::SizeOverflow)?;
+    if stride < row_bytes {
+        return Err(RenderBufferError::InvalidStride {
+            stride,
+            minimum: row_bytes,
+        });
+    }
+    let expected = stride
+        .checked_mul(height as usize)
+        .ok_or(RenderBufferError::SizeOverflow)?;
+    if pixels.len() != expected {
+        return Err(RenderBufferError::InvalidBufferLength {
+            actual: pixels.len(),
+            expected,
+        });
+    }
+
+    for row in pixels.chunks_exact_mut(stride) {
+        for pixel in row[..row_bytes].chunks_exact_mut(3) {
+            pixel.swap(0, 2);
+        }
+    }
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn bgr_pixels_are_converted_to_rgb_in_place() {
+        let mut pixels = vec![1, 2, 3, 4, 5, 6];
+        bgr_to_rgb_in_place(&mut pixels, 2, 1, 6).unwrap();
+        assert_eq!(pixels, [3, 2, 1, 6, 5, 4]);
+    }
+
+    #[test]
+    fn bgr_conversion_skips_row_padding() {
+        let mut pixels = vec![1, 2, 3, 9, 7, 8, 9, 6];
+        bgr_to_rgb_in_place(&mut pixels, 1, 2, 4).unwrap();
+        assert_eq!(pixels, [3, 2, 1, 9, 9, 8, 7, 6]);
+    }
+
+    #[test]
+    fn malformed_bgr_buffers_return_errors() {
+        assert!(matches!(
+            bgr_to_rgb_in_place(&mut [0; 6], 2, 1, 5),
+            Err(RenderBufferError::InvalidStride { .. })
+        ));
+        assert!(matches!(
+            bgr_to_rgb_in_place(&mut [0; 5], 1, 2, 3),
+            Err(RenderBufferError::InvalidBufferLength { .. })
+        ));
+    }
+}

src/vision/pipeline.rs

@@ -0,0 +1,426 @@
+//! One-call native extraction and OCR pipeline.
+
+use std::collections::BTreeSet;
+use std::path::Path;
+use std::time::Instant;
+
+use thiserror::Error;
+
+use crate::{MarkdownOptions, PageOcrReasons, PdfError};
+
+use super::{
+    fuse_ocr_pages, route_ocr_pages, run_ocr_pages, FusedPageMarkdown, HttpModelDownloadError,
+    HttpModelDownloader, ModelAcquireError, ModelStore, ModelStoreError, OarOcrEngine, OarOcrError,
+    OcrFusionError, OcrFusionOptions, OcrMode, OcrOptions, OcrRoutingError, OcrRun, OcrRunError,
+    PdfiumRenderer, RenderError, RenderOptions, PP_OCR_V6_SMALL,
+};
+
+/// Options for native extraction with optional OCR.
+#[derive(Clone)]
+pub struct OcrPdfOptions {
+    /// Page rasterization settings used when OCR is routed.
+    pub render: RenderOptions,
+    /// OCR routing, model, and recognition settings.
+    pub ocr: OcrOptions,
+    /// Markdown formatting shared by native and OCR assembly.
+    pub markdown: MarkdownOptions,
+    /// Optional 1-indexed page selection. `None` processes the full document.
+    pub page_filter: Option<BTreeSet<u32>>,
+    /// Password for an encrypted PDF.
+    pub password: Option<String>,
+    /// Weak OCR threshold for recommending the hosted pipeline.
+    pub hosted_recommendation_confidence: f32,
+}
+
+impl Default for OcrPdfOptions {
+    fn default() -> Self {
+        Self {
+            render: RenderOptions::default(),
+            ocr: OcrOptions::default(),
+            markdown: MarkdownOptions::default(),
+            page_filter: None,
+            password: None,
+            hosted_recommendation_confidence: 0.5,
+        }
+    }
+}
+
+impl std::fmt::Debug for OcrPdfOptions {
+    fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        formatter
+            .debug_struct("OcrPdfOptions")
+            .field("render", &self.render)
+            .field("ocr", &self.ocr)
+            .field("markdown", &self.markdown)
+            .field("page_filter", &self.page_filter)
+            .field("password", &self.password.as_ref().map(|_| "[REDACTED]"))
+            .field(
+                "hosted_recommendation_confidence",
+                &self.hosted_recommendation_confidence,
+            )
+            .finish()
+    }
+}
+
+impl OcrPdfOptions {
+    /// Creates options with OCR disabled, preserving the native-only path.
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// Replaces page rasterization settings.
+    pub fn render(mut self, render: RenderOptions) -> Self {
+        self.render = render;
+        self
+    }
+
+    /// Replaces OCR routing and recognition settings.
+    pub fn ocr(mut self, ocr: OcrOptions) -> Self {
+        self.ocr = ocr;
+        self
+    }
+
+    /// Sets OCR routing without changing the remaining OCR settings.
+    pub fn mode(mut self, mode: OcrMode) -> Self {
+        self.ocr.mode = mode;
+        self
+    }
+
+    /// Replaces Markdown formatting options.
+    pub fn markdown(mut self, markdown: MarkdownOptions) -> Self {
+        self.markdown = markdown;
+        self
+    }
+
+    /// Restricts processing to 1-indexed pages in ascending order.
+    pub fn pages(mut self, pages: impl IntoIterator<Item = u32>) -> Self {
+        self.page_filter = Some(pages.into_iter().collect());
+        self
+    }
+
+    /// Sets the password used to decrypt the PDF.
+    pub fn password(mut self, password: impl Into<String>) -> Self {
+        self.password = Some(password.into());
+        self
+    }
+
+    /// Sets the weak-OCR threshold for recommending hosted document parsing.
+    pub fn hosted_recommendation_confidence(mut self, confidence: f32) -> Self {
+        self.hosted_recommendation_confidence = confidence;
+        self
+    }
+}
+
+/// Complete native/OCR Markdown output for a PDF request.
+#[derive(Debug, Clone, PartialEq)]
+pub struct OcrPdfResult {
+    /// Final document Markdown in selected-page order.
+    pub markdown: String,
+    /// Final per-page Markdown and provenance, using 1-indexed page numbers.
+    pub pages: Vec<FusedPageMarkdown>,
+    /// Total pages in the PDF, independent of page selection.
+    pub page_count: u32,
+    /// 1-indexed selected pages recommended for OCR by native extraction.
+    pub pages_recommended_for_ocr: Vec<u32>,
+    /// 1-indexed pages actually rendered and recognized.
+    pub pages_routed_to_ocr: Vec<u32>,
+    /// 1-indexed pages whose OCR result recommends hosted document parsing.
+    pub pages_recommending_hosted: Vec<u32>,
+    /// Original machine-readable OCR reasons for selected pages.
+    pub ocr_reasons_by_page: Vec<PageOcrReasons>,
+    /// Selected pages where deterministic table detection found tables.
+    pub pages_with_tables: Vec<u32>,
+    /// Selected pages where deterministic layout found multiple columns.
+    pub pages_with_columns: Vec<u32>,
+    /// Whether deterministic extraction found tables or columns.
+    pub is_complex: bool,
+    /// End-to-end processing time.
+    pub processing_time_ms: u64,
+    /// Batch page-rendering time; zero when no OCR work was routed.
+    pub render_time_ms: u64,
+    /// Batch OCR time; zero when no OCR work was routed.
+    pub ocr_time_ms: u64,
+}
+
+/// Processes a PDF file through native extraction and selective OCR.
+pub fn process_pdf_with_ocr(
+    path: impl AsRef<Path>,
+    options: OcrPdfOptions,
+) -> Result<OcrPdfResult, OcrPipelineError> {
+    let bytes = std::fs::read(path).map_err(PdfError::from)?;
+    process_pdf_with_ocr_mem(&bytes, options)
+}
+
+/// Processes PDF bytes through native extraction and selective OCR.
+///
+/// Native extraction always runs first. `Auto` initializes PDFium, downloads
+/// models, and starts OAR only if the detector selected at least one page.
+/// `Off` therefore has no renderer, model-cache, network, or inference side
+/// effects even though the complete feature is compiled into the application.
+pub fn process_pdf_with_ocr_mem(
+    buffer: &[u8],
+    options: OcrPdfOptions,
+) -> Result<OcrPdfResult, OcrPipelineError> {
+    OcrFusionOptions::new()
+        .render_dpi(options.render.dpi)
+        .hosted_recommendation_confidence(options.hosted_recommendation_confidence)
+        .validate()?;
+    let minimum_confidence = options.ocr.minimum_confidence;
+    if !minimum_confidence.is_finite() || !(0.0..=1.0).contains(&minimum_confidence) {
+        return Err(OcrPipelineError::InvalidMinimumConfidence {
+            value: minimum_confidence,
+        });
+    }
+    if options
+        .page_filter
+        .as_ref()
+        .is_some_and(|pages| pages.contains(&0))
+    {
+        return Err(OcrPipelineError::InvalidSelectedPage { page: 0 });
+    }
+
+    let started = Instant::now();
+    let selected_pages: Option<Vec<u32>> = options
+        .page_filter
+        .as_ref()
+        .map(|pages| pages.iter().copied().collect());
+    let selected_pages_zero_indexed: Option<Vec<u32>> = selected_pages
+        .as_ref()
+        .map(|pages| pages.iter().map(|page| page - 1).collect());
+
+    let mut page_markdown_options = options.markdown.clone();
+    page_markdown_options.include_page_numbers = false;
+    let (native, page_count) = crate::extract_pages_markdown_mem_for_ocr(
+        buffer,
+        selected_pages_zero_indexed.as_deref(),
+        options.password.as_deref(),
+        &page_markdown_options,
+    )?;
+    if let Some(invalid) = selected_pages
+        .as_ref()
+        .and_then(|pages| pages.iter().copied().find(|page| *page > page_count))
+    {
+        return Err(OcrPipelineError::InvalidSelectedPage { page: invalid });
+    }
+
+    let routed = route_ocr_pages(
+        options.ocr.mode,
+        page_count,
+        &native.pages_needing_ocr,
+        selected_pages.as_deref(),
+    )?;
+
+    let ocr_run = if routed.is_empty() {
+        OcrRun {
+            pages: Vec::new(),
+            render_time_ms: 0,
+            ocr_time_ms: 0,
+        }
+    } else {
+        // Resolve the native renderer before any network request so a missing
+        // PDFium installation cannot trigger a model download it cannot use.
+        let renderer = PdfiumRenderer::load()?;
+        let store = ModelStore::from_options(&options.ocr)?;
+        let models = store.resolve_or_download(
+            &PP_OCR_V6_SMALL,
+            options.ocr.model_downloads,
+            &HttpModelDownloader::default(),
+        )?;
+        let engine = OarOcrEngine::from_models(&models)?;
+        run_ocr_pages(
+            &renderer,
+            &engine,
+            buffer,
+            &routed,
+            options.password.as_deref(),
+            &options.render,
+            &options.ocr,
+        )?
+    };
+
+    let fusion_options = OcrFusionOptions::new()
+        .markdown(page_markdown_options)
+        .render_dpi(options.render.dpi)
+        .hosted_recommendation_confidence(options.hosted_recommendation_confidence);
+    let fused = fuse_ocr_pages(&native.pages, &ocr_run, page_count, &fusion_options)?;
+    let pages_recommending_hosted = fused
+        .pages
+        .iter()
+        .filter(|page| page.provenance.hosted_recommended)
+        .map(|page| page.provenance.page)
+        .collect();
+    let markdown = assemble_document_markdown(&fused.pages, options.markdown.include_page_numbers);
+
+    Ok(OcrPdfResult {
+        markdown,
+        pages: fused.pages,
+        page_count,
+        pages_recommended_for_ocr: native.pages_needing_ocr,
+        pages_routed_to_ocr: routed,
+        pages_recommending_hosted,
+        ocr_reasons_by_page: native.ocr_reasons_by_page,
+        pages_with_tables: native.pages_with_tables,
+        pages_with_columns: native.pages_with_columns,
+        is_complex: native.is_complex,
+        processing_time_ms: elapsed_ms(started),
+        render_time_ms: fused.render_time_ms,
+        ocr_time_ms: fused.ocr_time_ms,
+    })
+}
+
+fn assemble_document_markdown(pages: &[FusedPageMarkdown], include_page_numbers: bool) -> String {
+    let mut document = String::new();
+    for (index, page) in pages.iter().enumerate() {
+        if index > 0 {
+            document.push_str("\n\n");
+        }
+        if include_page_numbers {
+            document.push_str(&format!("<!-- Page {} -->\n\n", page.page));
+        }
+        document.push_str(page.markdown.trim());
+    }
+    if !document.is_empty() {
+        document.push('\n');
+    }
+    document
+}
+
+fn elapsed_ms(started: Instant) -> u64 {
+    u64::try_from(started.elapsed().as_millis()).unwrap_or(u64::MAX)
+}
+
+/// Failures from the complete OCR pipeline.
+#[derive(Debug, Error)]
+#[non_exhaustive]
+pub enum OcrPipelineError {
+    /// PDF loading or native extraction failed.
+    #[error(transparent)]
+    Pdf(#[from] PdfError),
+    /// Page routing rejected an invalid request.
+    #[error(transparent)]
+    Routing(#[from] OcrRoutingError),
+    /// The model cache could not be located or initialized.
+    #[error(transparent)]
+    ModelStore(#[from] ModelStoreError),
+    /// A pinned model set could not be resolved or acquired.
+    #[error(transparent)]
+    ModelAcquire(#[from] ModelAcquireError<HttpModelDownloadError>),
+    /// PDFium could not load or rasterize the request.
+    #[error(transparent)]
+    Render(#[from] RenderError),
+    /// The OAR engine could not initialize.
+    #[error(transparent)]
+    Oar(#[from] OarOcrError),
+    /// Selective rendering or OCR execution failed.
+    #[error(transparent)]
+    Run(#[from] OcrRunError),
+    /// OCR/native Markdown fusion failed.
+    #[error(transparent)]
+    Fusion(#[from] OcrFusionError),
+    /// Page zero is invalid because public page selections are 1-indexed.
+    #[error("selected page {page} is invalid; page numbers are 1-indexed")]
+    InvalidSelectedPage {
+        /// Invalid page number.
+        page: u32,
+    },
+    /// OCR span confidence is outside the inclusive 0–1 range.
+    #[error("minimum OCR confidence must be between 0 and 1, got {value}")]
+    InvalidMinimumConfidence {
+        /// Invalid value.
+        value: f32,
+    },
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn off_mode_extracts_native_text_without_runtime_side_effects() {
+        let bytes = std::fs::read("tests/fixtures/thermo-freon12.pdf").unwrap();
+        let result = process_pdf_with_ocr_mem(&bytes, OcrPdfOptions::new()).unwrap();
+
+        assert_eq!(result.page_count, 3);
+        assert_eq!(result.pages.len(), 3);
+        assert!(result.markdown.contains("Thermodynamic Properties"));
+        assert!(result.pages_recommended_for_ocr.is_empty());
+        assert!(result.pages_routed_to_ocr.is_empty());
+        assert!(result.pages_recommending_hosted.is_empty());
+        assert_eq!(result.render_time_ms, 0);
+        assert_eq!(result.ocr_time_ms, 0);
+    }
+
+    #[test]
+    fn auto_mode_does_not_load_pdfium_or_models_for_clean_pdf() {
+        let bytes = std::fs::read("tests/fixtures/thermo-freon12.pdf").unwrap();
+        let result =
+            process_pdf_with_ocr_mem(&bytes, OcrPdfOptions::new().mode(OcrMode::Auto)).unwrap();
+
+        assert!(result.pages_routed_to_ocr.is_empty());
+        assert!(result.markdown.contains("Freon 12"));
+    }
+
+    #[test]
+    fn selection_and_page_markers_use_public_one_indexed_pages() {
+        let bytes = std::fs::read("tests/fixtures/thermo-freon12.pdf").unwrap();
+        let mut markdown = MarkdownOptions::default();
+        markdown.include_page_numbers = true;
+        let result =
+            process_pdf_with_ocr_mem(&bytes, OcrPdfOptions::new().pages([2]).markdown(markdown))
+                .unwrap();
+
+        assert_eq!(result.pages.len(), 1);
+        assert_eq!(result.pages[0].page, 2);
+        assert_eq!(result.pages[0].page, result.pages[0].provenance.page);
+        assert!(result.markdown.starts_with("<!-- Page 2 -->"));
+    }
+
+    #[test]
+    fn off_mode_marks_unprocessed_scan_for_hosted_fallback() {
+        let bytes = std::fs::read("tests/fixtures/scan_with_native_header_text.pdf").unwrap();
+        let result = process_pdf_with_ocr_mem(&bytes, OcrPdfOptions::new()).unwrap();
+
+        assert!(result.pages_routed_to_ocr.is_empty());
+        assert_eq!(result.pages_recommending_hosted, vec![1]);
+    }
+
+    #[test]
+    fn password_is_redacted_and_used_for_native_extraction() {
+        let options = OcrPdfOptions::new().password("secret123");
+        assert!(!format!("{options:?}").contains("secret123"));
+
+        let bytes = std::fs::read("tests/fixtures/encrypted-secret123.pdf").unwrap();
+        let result = process_pdf_with_ocr_mem(&bytes, options).unwrap();
+        assert!(result.markdown.contains("Procurement"));
+    }
+
+    #[test]
+    fn rejects_out_of_range_selection_even_with_ocr_off() {
+        let bytes = std::fs::read("tests/fixtures/thermo-freon12.pdf").unwrap();
+        let error = process_pdf_with_ocr_mem(&bytes, OcrPdfOptions::new().pages([4])).unwrap_err();
+        assert!(matches!(
+            error,
+            OcrPipelineError::InvalidSelectedPage { page: 4 }
+        ));
+    }
+
+    #[test]
+    fn invalid_expensive_options_fail_before_pdf_or_runtime_access() {
+        let mut invalid_dpi = OcrPdfOptions::new();
+        invalid_dpi.render.dpi = f32::NAN;
+        assert!(matches!(
+            process_pdf_with_ocr_mem(b"not a PDF", invalid_dpi),
+            Err(OcrPipelineError::Fusion(
+                OcrFusionError::InvalidRenderDpi { .. }
+            ))
+        ));
+
+        let invalid_hosted = OcrPdfOptions::new().hosted_recommendation_confidence(1.1);
+        assert!(matches!(
+            process_pdf_with_ocr_mem(b"not a PDF", invalid_hosted),
+            Err(OcrPipelineError::Fusion(
+                OcrFusionError::InvalidHostedConfidence { .. }
+            ))
+        ));
+    }
+}

src/vision/render.rs

@@ -0,0 +1,552 @@
+//! Renderer-neutral page bitmap and coordinate types.
+
+use thiserror::Error;
+
+use crate::PdfRect;
+
+/// Default rendering resolution for OCR.
+pub const DEFAULT_RENDER_DPI: f32 = 150.0;
+
+/// Default maximum size of one rendered page: 256 MiB.
+pub const DEFAULT_MAX_OUTPUT_BYTES: u64 = 256 * 1024 * 1024;
+
+/// Pixel layout returned by [`RenderedPage`].
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+#[non_exhaustive]
+pub enum RenderPixelFormat {
+    /// Three bytes per pixel in red, green, blue order. This is the default
+    /// because OCR preprocessors conventionally consume RGB images.
+    #[default]
+    Rgb8,
+    /// Four bytes per pixel in red, green, blue, alpha order.
+    Rgba8,
+    /// One luminance byte per pixel.
+    Gray8,
+}
+
+impl RenderPixelFormat {
+    /// Number of bytes used by one pixel.
+    pub fn bytes_per_pixel(self) -> usize {
+        match self {
+            Self::Rgb8 => 3,
+            Self::Rgba8 => 4,
+            Self::Gray8 => 1,
+        }
+    }
+}
+
+/// Configuration for pages rendered as input to a local vision pipeline.
+#[derive(Debug, Clone, PartialEq)]
+pub struct RenderOptions {
+    /// Output resolution. Defaults to 150 DPI.
+    pub dpi: f32,
+    /// Pixel layout. Defaults to three-channel RGB.
+    pub pixel_format: RenderPixelFormat,
+    /// Include PDF annotations in the rendered bitmap.
+    pub annotations: bool,
+    /// Include visible static AcroForm field appearances.
+    pub form_fields: bool,
+    /// Maximum allocation for each rendered page.
+    pub max_output_bytes_per_page: u64,
+}
+
+impl Default for RenderOptions {
+    fn default() -> Self {
+        Self {
+            dpi: DEFAULT_RENDER_DPI,
+            pixel_format: RenderPixelFormat::Rgb8,
+            annotations: true,
+            form_fields: true,
+            max_output_bytes_per_page: DEFAULT_MAX_OUTPUT_BYTES,
+        }
+    }
+}
+
+impl RenderOptions {
+    /// Creates local-rendering options with OCR-oriented defaults.
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// Sets the output resolution in dots per inch.
+    pub fn dpi(mut self, dpi: f32) -> Self {
+        self.dpi = dpi;
+        self
+    }
+
+    /// Sets the output pixel layout.
+    pub fn pixel_format(mut self, pixel_format: RenderPixelFormat) -> Self {
+        self.pixel_format = pixel_format;
+        self
+    }
+
+    /// Toggles annotation rendering.
+    pub fn annotations(mut self, annotations: bool) -> Self {
+        self.annotations = annotations;
+        self
+    }
+
+    /// Toggles visible static form-field rendering.
+    pub fn form_fields(mut self, form_fields: bool) -> Self {
+        self.form_fields = form_fields;
+        self
+    }
+
+    /// Sets the maximum allocation for each rendered page.
+    pub fn max_output_bytes_per_page(mut self, bytes: u64) -> Self {
+        self.max_output_bytes_per_page = bytes;
+        self
+    }
+}
+
+/// A point in PDF page space, measured in points from the bottom-left.
+#[derive(Debug, Clone, Copy, PartialEq)]
+pub struct PagePoint {
+    /// Horizontal position in PDF points.
+    pub x: f32,
+    /// Vertical position in PDF points, increasing upward.
+    pub y: f32,
+}
+
+/// Affine transform between top-left pixel space and PDF page space.
+///
+/// Renderers create this from the page-space images of the bitmap corners.
+/// Keeping the coefficients in pdf-inspector makes [`RenderedPage`] neutral
+/// to the renderer implementation that produced it.
+#[derive(Debug, Clone, Copy, PartialEq)]
+pub struct PageTransform {
+    forward: [f64; 6],
+    inverse: [f64; 6],
+    pixel_width: u32,
+    pixel_height: u32,
+}
+
+impl PageTransform {
+    /// Builds a transform from the PDF-space images of device corners
+    /// `(0, 0)`, `(pixel_width, 0)`, and `(0, pixel_height)`.
+    pub fn from_corners(
+        pixel_width: u32,
+        pixel_height: u32,
+        origin: (f64, f64),
+        x_axis: (f64, f64),
+        y_axis: (f64, f64),
+    ) -> Option<Self> {
+        if pixel_width == 0 || pixel_height == 0 {
+            return None;
+        }
+
+        let values = [origin.0, origin.1, x_axis.0, x_axis.1, y_axis.0, y_axis.1];
+        if values.iter().any(|value| !value.is_finite()) {
+            return None;
+        }
+
+        let width = f64::from(pixel_width);
+        let height = f64::from(pixel_height);
+        let a = (x_axis.0 - origin.0) / width;
+        let c = (x_axis.1 - origin.1) / width;
+        let b = (y_axis.0 - origin.0) / height;
+        let d = (y_axis.1 - origin.1) / height;
+        let (e, f) = origin;
+        let forward = [a, b, c, d, e, f];
+        if forward.iter().any(|coefficient| !coefficient.is_finite()) {
+            return None;
+        }
+        let determinant = a * d - b * c;
+        if determinant == 0.0 || !determinant.is_finite() {
+            return None;
+        }
+
+        let inverse_a = d / determinant;
+        let inverse_b = -b / determinant;
+        let inverse_c = -c / determinant;
+        let inverse_d = a / determinant;
+        let inverse_e = -(inverse_a * e + inverse_b * f);
+        let inverse_f = -(inverse_c * e + inverse_d * f);
+        let inverse = [
+            inverse_a, inverse_b, inverse_c, inverse_d, inverse_e, inverse_f,
+        ];
+        if inverse.iter().any(|coefficient| !coefficient.is_finite()) {
+            return None;
+        }
+
+        Some(Self {
+            forward,
+            inverse,
+            pixel_width,
+            pixel_height,
+        })
+    }
+
+    /// Width of the bitmap this transform describes.
+    pub fn pixel_width(&self) -> u32 {
+        self.pixel_width
+    }
+
+    /// Height of the bitmap this transform describes.
+    pub fn pixel_height(&self) -> u32 {
+        self.pixel_height
+    }
+
+    /// Converts a bitmap point to PDF page space.
+    pub fn pixel_to_page(&self, x: f64, y: f64) -> PagePoint {
+        let [a, b, c, d, e, f] = self.forward;
+        PagePoint {
+            x: (a * x + b * y + e) as f32,
+            y: (c * x + d * y + f) as f32,
+        }
+    }
+
+    /// Converts a PDF page-space point to bitmap coordinates.
+    pub fn page_to_pixel(&self, x: f64, y: f64) -> (f64, f64) {
+        let [a, b, c, d, e, f] = self.inverse;
+        (a * x + b * y + e, c * x + d * y + f)
+    }
+}
+
+/// Invalid renderer output rejected by [`RenderedPage::new`].
+#[derive(Debug, Error, Clone, PartialEq, Eq)]
+#[non_exhaustive]
+pub enum RenderBufferError {
+    /// Page numbers are 1-indexed.
+    #[error("rendered page number must be at least 1")]
+    InvalidPageNumber,
+    /// Bitmap dimensions must be non-zero.
+    #[error("rendered bitmap dimensions must be non-zero")]
+    InvalidDimensions,
+    /// Page dimensions must be positive finite numbers.
+    #[error("rendered PDF page dimensions must be positive and finite")]
+    InvalidPageDimensions,
+    /// Transform dimensions must match the bitmap dimensions.
+    #[error("coordinate transform dimensions do not match the rendered bitmap")]
+    TransformDimensions,
+    /// Renderer did not provide an invertible finite coordinate transform.
+    #[error("renderer returned an invalid coordinate transform")]
+    InvalidTransform,
+    /// The stride cannot hold one active row of pixels.
+    #[error("pixel stride {stride} is shorter than the active row size {minimum}")]
+    InvalidStride {
+        /// Supplied bytes per row.
+        stride: usize,
+        /// Minimum bytes required for one row.
+        minimum: usize,
+    },
+    /// Pixel buffer size is inconsistent with height and stride.
+    #[error("pixel buffer has {actual} bytes; expected {expected}")]
+    InvalidBufferLength {
+        /// Actual byte count.
+        actual: usize,
+        /// Required byte count.
+        expected: usize,
+    },
+    /// Dimension arithmetic overflowed the host address space.
+    #[error("rendered bitmap dimensions overflow the host address space")]
+    SizeOverflow,
+}
+
+/// One rendered page with owned pixels and its pixel-to-PDF transform.
+///
+/// The value contains no live renderer, page, or document handles. It can be
+/// moved to an OCR worker and retained after rendering returns.
+#[derive(Debug, Clone)]
+pub struct RenderedPage {
+    page: u32,
+    page_width: f32,
+    page_height: f32,
+    width: u32,
+    height: u32,
+    stride: usize,
+    format: RenderPixelFormat,
+    pixels: Vec<u8>,
+    transform: PageTransform,
+}
+
+impl RenderedPage {
+    /// Creates a renderer-neutral owned page after validating its buffer.
+    #[allow(clippy::too_many_arguments)]
+    pub fn new(
+        page: u32,
+        page_width: f32,
+        page_height: f32,
+        width: u32,
+        height: u32,
+        stride: usize,
+        format: RenderPixelFormat,
+        pixels: Vec<u8>,
+        transform: PageTransform,
+    ) -> Result<Self, RenderBufferError> {
+        if page == 0 {
+            return Err(RenderBufferError::InvalidPageNumber);
+        }
+        if width == 0 || height == 0 {
+            return Err(RenderBufferError::InvalidDimensions);
+        }
+        if page_width <= 0.0
+            || page_height <= 0.0
+            || !page_width.is_finite()
+            || !page_height.is_finite()
+        {
+            return Err(RenderBufferError::InvalidPageDimensions);
+        }
+        if transform.pixel_width() != width || transform.pixel_height() != height {
+            return Err(RenderBufferError::TransformDimensions);
+        }
+
+        let row_bytes = (width as usize)
+            .checked_mul(format.bytes_per_pixel())
+            .ok_or(RenderBufferError::SizeOverflow)?;
+        if stride < row_bytes {
+            return Err(RenderBufferError::InvalidStride {
+                stride,
+                minimum: row_bytes,
+            });
+        }
+        let expected = stride
+            .checked_mul(height as usize)
+            .ok_or(RenderBufferError::SizeOverflow)?;
+        if pixels.len() != expected {
+            return Err(RenderBufferError::InvalidBufferLength {
+                actual: pixels.len(),
+                expected,
+            });
+        }
+
+        Ok(Self {
+            page,
+            page_width,
+            page_height,
+            width,
+            height,
+            stride,
+            format,
+            pixels,
+            transform,
+        })
+    }
+
+    /// 1-indexed page number.
+    pub fn page(&self) -> u32 {
+        self.page
+    }
+
+    /// Page width in PDF points after applying the page's rotation.
+    pub fn page_width(&self) -> f32 {
+        self.page_width
+    }
+
+    /// Page height in PDF points after applying the page's rotation.
+    pub fn page_height(&self) -> f32 {
+        self.page_height
+    }
+
+    /// Bitmap width in pixels.
+    pub fn width(&self) -> u32 {
+        self.width
+    }
+
+    /// Bitmap height in pixels.
+    pub fn height(&self) -> u32 {
+        self.height
+    }
+
+    /// Number of bytes between adjacent bitmap rows.
+    pub fn stride(&self) -> usize {
+        self.stride
+    }
+
+    /// Pixel layout of [`pixels`](Self::pixels).
+    pub fn format(&self) -> RenderPixelFormat {
+        self.format
+    }
+
+    /// Owned bitmap bytes, with rows ordered top-to-bottom.
+    pub fn pixels(&self) -> &[u8] {
+        &self.pixels
+    }
+
+    /// Consumes the page and returns its pixel buffer.
+    pub fn into_pixels(self) -> Vec<u8> {
+        self.pixels
+    }
+
+    /// Coordinate transform associated with the rendered page.
+    pub fn transform(&self) -> PageTransform {
+        self.transform
+    }
+
+    /// Converts a bitmap point (top-left origin, y-down) to PDF page space
+    /// (bottom-left origin, y-up).
+    pub fn pixel_to_page(&self, x: f64, y: f64) -> PagePoint {
+        self.transform.pixel_to_page(x, y)
+    }
+
+    /// Converts a bitmap rectangle to the repository's existing PDF-space
+    /// rectangle type. The returned page number remains 1-indexed.
+    pub fn pixel_rect_to_pdf_rect(&self, x: f64, y: f64, width: f64, height: f64) -> PdfRect {
+        let points = [
+            self.transform.pixel_to_page(x, y),
+            self.transform.pixel_to_page(x + width, y),
+            self.transform.pixel_to_page(x, y + height),
+            self.transform.pixel_to_page(x + width, y + height),
+        ];
+        let left = points
+            .iter()
+            .map(|point| point.x)
+            .fold(f32::INFINITY, f32::min);
+        let right = points
+            .iter()
+            .map(|point| point.x)
+            .fold(f32::NEG_INFINITY, f32::max);
+        let bottom = points
+            .iter()
+            .map(|point| point.y)
+            .fold(f32::INFINITY, f32::min);
+        let top = points
+            .iter()
+            .map(|point| point.y)
+            .fold(f32::NEG_INFINITY, f32::max);
+        PdfRect {
+            x: left,
+            y: bottom,
+            width: right - left,
+            height: top - bottom,
+            page: self.page,
+        }
+    }
+
+    /// Converts a PDF-space rectangle to bitmap coordinates
+    /// `(x, y, width, height)` with a top-left origin.
+    pub fn pdf_rect_to_pixel(&self, rect: &PdfRect) -> (f64, f64, f64, f64) {
+        let left = f64::from(rect.x);
+        let right = f64::from(rect.x + rect.width);
+        let bottom = f64::from(rect.y);
+        let top = f64::from(rect.y + rect.height);
+        let points = [
+            self.transform.page_to_pixel(left, bottom),
+            self.transform.page_to_pixel(right, bottom),
+            self.transform.page_to_pixel(left, top),
+            self.transform.page_to_pixel(right, top),
+        ];
+        let min_x = points
+            .iter()
+            .map(|point| point.0)
+            .fold(f64::INFINITY, f64::min);
+        let max_x = points
+            .iter()
+            .map(|point| point.0)
+            .fold(f64::NEG_INFINITY, f64::max);
+        let min_y = points
+            .iter()
+            .map(|point| point.1)
+            .fold(f64::INFINITY, f64::min);
+        let max_y = points
+            .iter()
+            .map(|point| point.1)
+            .fold(f64::NEG_INFINITY, f64::max);
+        (min_x, min_y, max_x - min_x, max_y - min_y)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn transform() -> PageTransform {
+        PageTransform::from_corners(400, 200, (0.0, 100.0), (200.0, 100.0), (0.0, 0.0)).unwrap()
+    }
+
+    #[test]
+    fn transform_maps_both_directions_at_non_identity_scale() {
+        let transform = transform();
+        let point = transform.pixel_to_page(100.0, 50.0);
+        assert!((point.x - 50.0).abs() < 1e-6);
+        assert!((point.y - 75.0).abs() < 1e-6);
+        let pixel = transform.page_to_pixel(f64::from(point.x), f64::from(point.y));
+        assert!((pixel.0 - 100.0).abs() < 1e-6);
+        assert!((pixel.1 - 50.0).abs() < 1e-6);
+    }
+
+    #[test]
+    fn rendered_page_accepts_padding_and_validates_length() {
+        let page = RenderedPage::new(
+            1,
+            200.0,
+            100.0,
+            400,
+            200,
+            1_204,
+            RenderPixelFormat::Rgb8,
+            vec![0; 1_204 * 200],
+            transform(),
+        )
+        .unwrap();
+        assert_eq!(page.stride(), 1_204);
+
+        assert!(matches!(
+            RenderedPage::new(
+                1,
+                200.0,
+                100.0,
+                400,
+                200,
+                1_204,
+                RenderPixelFormat::Rgb8,
+                vec![0; 5],
+                transform(),
+            ),
+            Err(RenderBufferError::InvalidBufferLength { .. })
+        ));
+    }
+
+    #[test]
+    fn rotated_transform_round_trips_rectangles() {
+        let transform =
+            PageTransform::from_corners(100, 200, (0.0, 0.0), (0.0, 100.0), (200.0, 0.0)).unwrap();
+        let page = RenderedPage::new(
+            1,
+            200.0,
+            100.0,
+            100,
+            200,
+            300,
+            RenderPixelFormat::Rgb8,
+            vec![0; 300 * 200],
+            transform,
+        )
+        .unwrap();
+        let pdf = page.pixel_rect_to_pdf_rect(10.0, 20.0, 30.0, 40.0);
+        let pixel = page.pdf_rect_to_pixel(&pdf);
+        assert!((pixel.0 - 10.0).abs() < 1e-5);
+        assert!((pixel.1 - 20.0).abs() < 1e-5);
+        assert!((pixel.2 - 30.0).abs() < 1e-5);
+        assert!((pixel.3 - 40.0).abs() < 1e-5);
+    }
+
+    #[test]
+    fn skewed_transform_bounds_all_rectangle_corners() {
+        let transform =
+            PageTransform::from_corners(100, 100, (0.0, 100.0), (100.0, 125.0), (25.0, 0.0))
+                .unwrap();
+        let page = RenderedPage::new(
+            1,
+            125.0,
+            125.0,
+            100,
+            100,
+            300,
+            RenderPixelFormat::Rgb8,
+            vec![0; 30_000],
+            transform,
+        )
+        .unwrap();
+
+        let pdf = page.pixel_rect_to_pdf_rect(10.0, 20.0, 30.0, 40.0);
+        assert!((pdf.x - 15.0).abs() < 1e-5);
+        assert!((pdf.y - 42.5).abs() < 1e-5);
+        assert!((pdf.width - 40.0).abs() < 1e-5);
+        assert!((pdf.height - 47.5).abs() < 1e-5);
+
+        let pixels = page.pdf_rect_to_pixel(&pdf);
+        assert!(pixels.0 <= 10.0 && pixels.1 <= 20.0);
+        assert!(pixels.0 + pixels.2 >= 40.0 && pixels.1 + pixels.3 >= 60.0);
+    }
+}

src/vision/routing.rs

@@ -0,0 +1,380 @@
+//! Page routing and renderer/OCR orchestration without Markdown fusion.
+
+use std::collections::BTreeSet;
+use std::error::Error;
+use std::time::Instant;
+
+use thiserror::Error;
+
+use super::{OcrEngine, OcrMode, OcrOptions, OcrPage, PageRenderer, RenderOptions, RenderedPage};
+
+/// A rendered page paired with OCR output in the same bitmap coordinate space.
+#[derive(Debug)]
+pub struct RoutedOcrPage {
+    /// Renderer-owned bitmap and pixel↔PDF transform.
+    pub rendered: RenderedPage,
+    /// Positioned OCR spans for the bitmap.
+    pub ocr: OcrPage,
+}
+
+/// Output of one selective OCR invocation.
+#[derive(Debug)]
+pub struct OcrRun {
+    /// Pages processed in ascending document order.
+    pub pages: Vec<RoutedOcrPage>,
+    /// Total page-rendering wall time.
+    pub render_time_ms: u64,
+    /// Total engine wall time.
+    pub ocr_time_ms: u64,
+}
+
+/// Selects 1-indexed pages for OCR.
+///
+/// `recommended_pages` comes from pdf-inspector's existing detector/text
+/// quality signals. `selected_pages` is an optional user page filter. Results
+/// are validated, deduplicated, and returned in document order.
+pub fn route_ocr_pages(
+    mode: OcrMode,
+    page_count: u32,
+    recommended_pages: &[u32],
+    selected_pages: Option<&[u32]>,
+) -> Result<Vec<u32>, OcrRoutingError> {
+    match mode {
+        OcrMode::Off => Ok(Vec::new()),
+        OcrMode::Auto => {
+            let mut routed = validated_page_set("recommended", recommended_pages, page_count)?;
+            if let Some(selected) = selected_pages {
+                let selected = validated_page_set("selected", selected, page_count)?;
+                routed.retain(|page| selected.contains(page));
+            }
+            Ok(routed.into_iter().collect())
+        }
+        OcrMode::Force => {
+            let routed = selected_pages
+                .map(|pages| validated_page_set("selected", pages, page_count))
+                .transpose()?
+                .unwrap_or_else(|| (1..=page_count).collect());
+            Ok(routed.into_iter().collect())
+        }
+    }
+}
+
+/// Renders and recognizes already-routed pages while retaining transforms for
+/// the following fusion layer.
+///
+/// An empty page list returns without calling either dependency, which keeps
+/// model resolution and inference lazy when Auto routing finds no OCR work.
+#[allow(clippy::too_many_arguments)]
+pub fn run_ocr_pages<R, O>(
+    renderer: &R,
+    engine: &O,
+    pdf_bytes: &[u8],
+    pages: &[u32],
+    password: Option<&str>,
+    render_options: &RenderOptions,
+    ocr_options: &OcrOptions,
+) -> Result<OcrRun, OcrRunError>
+where
+    R: PageRenderer,
+    O: OcrEngine,
+{
+    if pages.is_empty() {
+        return Ok(OcrRun {
+            pages: Vec::new(),
+            render_time_ms: 0,
+            ocr_time_ms: 0,
+        });
+    }
+    if ocr_options.mode == OcrMode::Off {
+        return Err(OcrRunError::OcrDisabled);
+    }
+
+    let render_started = Instant::now();
+    let rendered = renderer
+        .render_pages(pdf_bytes, pages, password, render_options)
+        .map_err(|source| OcrRunError::Render {
+            source: Box::new(source),
+        })?;
+    let render_time_ms = elapsed_ms(render_started);
+    validate_page_order("renderer", pages, rendered.iter().map(RenderedPage::page))?;
+
+    let ocr_started = Instant::now();
+    let recognized =
+        engine
+            .recognize(&rendered, ocr_options)
+            .map_err(|source| OcrRunError::Ocr {
+                source: Box::new(source),
+            })?;
+    let ocr_time_ms = elapsed_ms(ocr_started);
+    validate_page_order("OCR engine", pages, recognized.iter().map(|page| page.page))?;
+
+    Ok(OcrRun {
+        pages: rendered
+            .into_iter()
+            .zip(recognized)
+            .map(|(rendered, ocr)| RoutedOcrPage { rendered, ocr })
+            .collect(),
+        render_time_ms,
+        ocr_time_ms,
+    })
+}
+
+/// Invalid page routing or renderer/engine contract output.
+#[derive(Debug, Error)]
+#[non_exhaustive]
+pub enum OcrRoutingError {
+    /// A page list contained zero or a page beyond the document.
+    #[error("{source_name} OCR page {page} is outside the valid range 1..={page_count}")]
+    InvalidPage {
+        /// Page-list source.
+        source_name: &'static str,
+        /// Invalid 1-indexed page.
+        page: u32,
+        /// Document page count.
+        page_count: u32,
+    },
+}
+
+/// Failures while rendering and recognizing a routed page set.
+#[derive(Debug, Error)]
+#[non_exhaustive]
+pub enum OcrRunError {
+    /// A non-empty route cannot execute with OCR disabled.
+    #[error("cannot process routed pages while OCR mode is Off")]
+    OcrDisabled,
+    /// Page rasterization failed.
+    #[error("page rendering failed: {source}")]
+    Render {
+        /// Renderer-specific failure.
+        #[source]
+        source: Box<dyn Error + Send + Sync>,
+    },
+    /// OCR inference failed.
+    #[error("OCR inference failed: {source}")]
+    Ocr {
+        /// Engine-specific failure.
+        #[source]
+        source: Box<dyn Error + Send + Sync>,
+    },
+    /// A dependency returned the wrong count or order.
+    #[error("{stage} returned pages {actual:?}; expected {expected:?}")]
+    PageOrderMismatch {
+        /// Dependency boundary that violated the contract.
+        stage: &'static str,
+        /// Requested 1-indexed pages.
+        expected: Vec<u32>,
+        /// Returned 1-indexed pages.
+        actual: Vec<u32>,
+    },
+}
+
+fn validated_page_set(
+    source_name: &'static str,
+    pages: &[u32],
+    page_count: u32,
+) -> Result<BTreeSet<u32>, OcrRoutingError> {
+    let mut result = BTreeSet::new();
+    for &page in pages {
+        if page == 0 || page > page_count {
+            return Err(OcrRoutingError::InvalidPage {
+                source_name,
+                page,
+                page_count,
+            });
+        }
+        result.insert(page);
+    }
+    Ok(result)
+}
+
+fn validate_page_order(
+    stage: &'static str,
+    expected: &[u32],
+    actual: impl IntoIterator<Item = u32>,
+) -> Result<(), OcrRunError> {
+    let actual: Vec<u32> = actual.into_iter().collect();
+    if actual == expected {
+        Ok(())
+    } else {
+        Err(OcrRunError::PageOrderMismatch {
+            stage,
+            expected: expected.to_vec(),
+            actual,
+        })
+    }
+}
+
+fn elapsed_ms(started: Instant) -> u64 {
+    u64::try_from(started.elapsed().as_millis()).unwrap_or(u64::MAX)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::vision::{
+        ImagePoint, ImageQuad, ModelIdentity, OcrSpan, PageTransform, RenderBufferError,
+        RenderPixelFormat,
+    };
+
+    #[derive(Debug, Error)]
+    #[error("fake failure")]
+    struct FakeError;
+
+    struct FakeRenderer;
+
+    impl PageRenderer for FakeRenderer {
+        type Error = FakeError;
+
+        fn render_pages(
+            &self,
+            _pdf_bytes: &[u8],
+            pages: &[u32],
+            _password: Option<&str>,
+            _options: &RenderOptions,
+        ) -> Result<Vec<RenderedPage>, Self::Error> {
+            pages
+                .iter()
+                .copied()
+                .map(rendered_page)
+                .collect::<Result<_, _>>()
+                .map_err(|_| FakeError)
+        }
+    }
+
+    struct FakeEngine {
+        model: ModelIdentity,
+    }
+
+    impl FakeEngine {
+        fn new() -> Self {
+            Self {
+                model: ModelIdentity::new("fake", "v1"),
+            }
+        }
+    }
+
+    impl OcrEngine for FakeEngine {
+        type Error = FakeError;
+
+        fn model(&self) -> &ModelIdentity {
+            &self.model
+        }
+
+        fn recognize(
+            &self,
+            pages: &[RenderedPage],
+            _options: &OcrOptions,
+        ) -> Result<Vec<OcrPage>, Self::Error> {
+            Ok(pages
+                .iter()
+                .map(|page| OcrPage {
+                    page: page.page(),
+                    spans: vec![OcrSpan {
+                        text: format!("page {}", page.page()),
+                        polygon: ImageQuad::new([
+                            ImagePoint::new(0.0, 0.0),
+                            ImagePoint::new(1.0, 0.0),
+                            ImagePoint::new(1.0, 1.0),
+                            ImagePoint::new(0.0, 1.0),
+                        ]),
+                        confidence: 0.9,
+                        orientation_degrees: None,
+                    }],
+                    mean_confidence: Some(0.9),
+                    model: self.model.clone(),
+                    processing_time_ms: 1,
+                    warnings: Vec::new(),
+                })
+                .collect())
+        }
+    }
+
+    fn rendered_page(page: u32) -> Result<RenderedPage, RenderBufferError> {
+        let transform =
+            PageTransform::from_corners(1, 1, (0.0, 1.0), (1.0, 1.0), (0.0, 0.0)).unwrap();
+        RenderedPage::new(
+            page,
+            1.0,
+            1.0,
+            1,
+            1,
+            3,
+            RenderPixelFormat::Rgb8,
+            vec![255; 3],
+            transform,
+        )
+    }
+
+    #[test]
+    fn off_auto_and_force_route_expected_pages() {
+        assert_eq!(
+            route_ocr_pages(OcrMode::Off, 0, &[99], Some(&[0])).unwrap(),
+            Vec::<u32>::new()
+        );
+        assert_eq!(
+            route_ocr_pages(OcrMode::Auto, 5, &[5, 3, 3, 1], Some(&[2, 3, 5])).unwrap(),
+            vec![3, 5]
+        );
+        assert_eq!(
+            route_ocr_pages(OcrMode::Force, 4, &[], None).unwrap(),
+            vec![1, 2, 3, 4]
+        );
+        assert_eq!(
+            route_ocr_pages(OcrMode::Force, 4, &[], Some(&[4, 2, 2])).unwrap(),
+            vec![2, 4]
+        );
+    }
+
+    #[test]
+    fn routing_rejects_invalid_page_numbers() {
+        assert!(matches!(
+            route_ocr_pages(OcrMode::Auto, 2, &[0], None),
+            Err(OcrRoutingError::InvalidPage { page: 0, .. })
+        ));
+        assert!(matches!(
+            route_ocr_pages(OcrMode::Force, 2, &[], Some(&[3])),
+            Err(OcrRoutingError::InvalidPage { page: 3, .. })
+        ));
+    }
+
+    #[test]
+    fn run_retains_render_transforms_and_input_order() {
+        let options = OcrOptions::new().mode(OcrMode::Auto);
+        let run = run_ocr_pages(
+            &FakeRenderer,
+            &FakeEngine::new(),
+            b"pdf",
+            &[2, 4],
+            None,
+            &RenderOptions::new(),
+            &options,
+        )
+        .unwrap();
+        assert_eq!(
+            run.pages
+                .iter()
+                .map(|page| page.rendered.page())
+                .collect::<Vec<_>>(),
+            vec![2, 4]
+        );
+        assert_eq!(run.pages[1].ocr.spans[0].text, "page 4");
+        assert_eq!(run.pages[1].rendered.pixel_to_page(0.0, 0.0).y, 1.0);
+    }
+
+    #[test]
+    fn empty_route_is_a_noop_even_when_ocr_is_off() {
+        let run = run_ocr_pages(
+            &FakeRenderer,
+            &FakeEngine::new(),
+            b"pdf",
+            &[],
+            None,
+            &RenderOptions::new(),
+            &OcrOptions::new(),
+        )
+        .unwrap();
+        assert!(run.pages.is_empty());
+        assert_eq!(run.render_time_ms, 0);
+        assert_eq!(run.ocr_time_ms, 0);
+    }
+}

tests/integration_tests.rs

@@ -1429,6 +1429,77 @@ fn test_firecrawl_tagged_pdf_struct_tree() {
     assert_eq!(fence_count % 2, 0, "Code fences should be balanced");
 }
 
+#[test]
+fn test_tagged_pdf_text_items_carry_mcid() {
+    let buf = std::fs::read("tests/fixtures/firecrawl_docs_tagged.pdf").unwrap();
+    let items = pdf_inspector::extractor::extract_text_with_positions_mem(&buf).unwrap();
+    assert!(
+        items.iter().any(|i| i.mcid.is_some()),
+        "Tagged PDF text items should carry Marked Content IDs"
+    );
+}
+
+#[test]
+fn test_extract_structure_elements_tagged_pdf() {
+    let buf = std::fs::read("tests/fixtures/firecrawl_docs_tagged.pdf").unwrap();
+    let elements = pdf_inspector::extract_structure_elements_mem(&buf, None).unwrap();
+    assert!(!elements.is_empty(), "Tagged PDF should yield elements");
+    assert!(
+        elements.iter().any(|e| e.role == "H1"),
+        "Should surface H1 heading roles"
+    );
+    assert!(
+        elements.iter().all(|e| !e.role.is_empty()),
+        "Every element should carry a role name"
+    );
+
+    // Sorted by (page, mcid) for deterministic output
+    assert!(
+        elements
+            .windows(2)
+            .all(|w| (w[0].page, w[0].mcid) <= (w[1].page, w[1].mcid)),
+        "Elements should be sorted by (page, mcid)"
+    );
+
+    // The advertised join: (page, mcid) pairs must line up with the
+    // mcid-carrying TextItems from positioned extraction, and joining the
+    // H1 entries must recover non-empty heading text.
+    let items = pdf_inspector::extractor::extract_text_with_positions_mem(&buf).unwrap();
+    let h1_refs: std::collections::HashSet<(u32, i64)> = elements
+        .iter()
+        .filter(|e| e.role == "H1")
+        .map(|e| (e.page, e.mcid))
+        .collect();
+    let h1_text: String = items
+        .iter()
+        .filter(|i| i.mcid.is_some_and(|mcid| h1_refs.contains(&(i.page, mcid))))
+        .map(|i| i.text.as_str())
+        .collect();
+    assert!(
+        !h1_text.trim().is_empty(),
+        "Joining H1 structure elements to text items should recover heading text"
+    );
+
+    // Page filter is 1-indexed (matching TextItem.page) and equals the
+    // corresponding subset of the full document result.
+    let page1 = pdf_inspector::extract_structure_elements_mem(&buf, Some(&[1])).unwrap();
+    assert!(!page1.is_empty(), "Page 1 should have elements");
+    assert!(page1.iter().all(|e| e.page == 1));
+    let full_page1_count = elements.iter().filter(|e| e.page == 1).count();
+    assert_eq!(page1.len(), full_page1_count);
+}
+
+#[test]
+fn test_extract_structure_elements_untagged_pdf_empty() {
+    let buf = std::fs::read("tests/fixtures/thermo-freon12.pdf").unwrap();
+    let elements = pdf_inspector::extract_structure_elements_mem(&buf, None).unwrap();
+    assert!(
+        elements.is_empty(),
+        "Untagged PDF should yield no structure elements, got {:?}",
+        elements
+    );
+}
+
 #[test]
 fn test_identity_h_no_tounicode_suppresses_garbage() {
     // shinagawa_identity_h.pdf uses YuGothic with Identity-H encoding and no
@@ -1583,6 +1654,282 @@ fn test_extract_regions_mem_basic_text_pdf() {
     assert_eq!(regions[0].page, 0);
 }
 
+/// Build a synthetic "scanned page" PDF: a full-page image XObject with a
+/// text layer drawn in the given render mode (3 = invisible OCR overlay,
+/// 0 = normal visible fill). `visible_extra` optionally adds a normally
+/// rendered line so double-layer behavior can be tested; `layer_lines`
+/// overrides the layer content (default: three pangram lines);
+/// `quote_ops` shows every layer line via the `'` operator instead of Tj
+/// (both are standard show-text encodings for OCR layers).
+fn make_pdf_with_custom_text_layer(
+    text_render_mode: i32,
+    visible_extra: Option<&str>,
+    layer_lines: Option<&[&str]>,
+    quote_ops: bool,
+) -> Vec<u8> {
+    let mut pdf = b"%PDF-1.4\n".to_vec();
+    let mut offsets = vec![0usize];
+
+    fn add_object(pdf: &mut Vec<u8>, offsets: &mut Vec<usize>, id: usize, body: &str) {
+        offsets.push(pdf.len());
+        pdf.extend_from_slice(format!("{id} 0 obj\n").as_bytes());
+        pdf.extend_from_slice(body.as_bytes());
+        pdf.extend_from_slice(b"\nendobj\n");
+    }
+    fn add_stream_object(
+        pdf: &mut Vec<u8>,
+        offsets: &mut Vec<usize>,
+        id: usize,
+        dict: &str,
+        stream_bytes: &[u8],
+    ) {
+        offsets.push(pdf.len());
+        pdf.extend_from_slice(format!("{id} 0 obj\n").as_bytes());
+        pdf.extend_from_slice(
+            format!("<< {} /Length {} >>\nstream\n", dict, stream_bytes.len()).as_bytes(),
+        );
+        pdf.extend_from_slice(stream_bytes);
+        pdf.extend_from_slice(b"\nendstream\nendobj\n");
+    }
+
+    add_object(
+        &mut pdf,
+        &mut offsets,
+        1,
+        "<< /Type /Catalog /Pages 2 0 R >>",
+    );
+    add_object(
+        &mut pdf,
+        &mut offsets,
+        2,
+        "<< /Type /Pages /Kids [3 0 R] /Count 1 >>",
+    );
+    add_object(
+        &mut pdf,
+        &mut offsets,
+        3,
+        "<< /Type /Page /Parent 2 0 R /MediaBox [0 0 612 792] \
+         /Resources << /Font << /F1 5 0 R >> /XObject << /Im0 6 0 R >> >> \
+         /Contents 4 0 R >>",
+    );
+    // Full-page raster, then the text layer in the requested render mode —
+    // several lines so the OCR-layer gate's alnum floor (40) is well cleared.
+    let mut content = String::from("q 612 0 0 792 0 0 cm /Im0 Do Q\n");
+    let default_layer = [
+        "The quick brown fox jumps over the lazy dog",
+        "Pack my box with five dozen liquor jugs tonight",
+        "Sphinx of black quartz judge my vow carefully",
+    ];
+    let layer: &[&str] = layer_lines.unwrap_or(&default_layer);
+    if quote_ops {
+        // Every line shown via `'` (move-to-next-line + show) — nothing on
+        // this layer goes through Tj, pinning the `'` suppression path.
+        content.push_str(&format!(
+            "BT /F1 12 Tf {text_render_mode} Tr 16 TL 72 716 Td "
+        ));
+        for line in layer {
+            content.push_str(&format!("({line}) ' "));
+        }
+    } else {
+        content.push_str(&format!("BT /F1 12 Tf {text_render_mode} Tr 72 700 Td "));
+        for (i, line) in layer.iter().enumerate() {
+            if i > 0 {
+                content.push_str("0 -16 Td ");
+            }
+            content.push_str(&format!("({line}) Tj "));
+        }
+    }
+    content.push_str("ET\n");
+    if let Some(extra) = visible_extra {
+        content.push_str(&format!("BT /F1 12 Tf 0 Tr 72 500 Td ({extra}) Tj ET\n"));
+    }
+    add_stream_object(&mut pdf, &mut offsets, 4, "", content.as_bytes());
+    add_object(
+        &mut pdf,
+        &mut offsets,
+        5,
+        "<< /Type /Font /Subtype /Type1 /BaseFont /Helvetica >>",
+    );
+    let image_pixel = [128u8];
+    add_stream_object(
+        &mut pdf,
+        &mut offsets,
+        6,
+        "/Type /XObject /Subtype /Image /Width 1 /Height 1 \
+         /ColorSpace /DeviceGray /BitsPerComponent 8",
+        &image_pixel,
+    );
+
+    let xref_start = pdf.len();
+    pdf.extend_from_slice(format!("xref\n0 {}\n", offsets.len()).as_bytes());
+    pdf.extend_from_slice(b"0000000000 65535 f \n");
+    for offset in offsets.iter().skip(1) {
+        pdf.extend_from_slice(format!("{offset:010} 00000 n \n").as_bytes());
+    }
+    pdf.extend_from_slice(
+        format!(
+            "trailer\n<< /Size {} /Root 1 0 R >>\nstartxref\n{}\n%%EOF",
+            offsets.len(),
+            xref_start
+        )
+        .as_bytes(),
+    );
+    pdf
+}
+
+fn make_pdf_with_text_layer(text_render_mode: i32, visible_extra: Option<&str>) -> Vec<u8> {
+    make_pdf_with_custom_text_layer(text_render_mode, visible_extra, None, false)
+}
+
+/// A scanned page whose only text is an invisible (Tr 3) OCR layer behind
+/// the raster must serve that layer from the region extractor instead of
+/// reporting the region as needs_ocr — the exact text is already in the PDF.
+#[test]
+fn test_extract_regions_mem_recovers_invisible_ocr_layer() {
+    let buf = make_pdf_with_text_layer(3, None);
+    let regions = extract_text_in_regions_mem(&buf, &full_page_regions(1)).unwrap();
+    assert_eq!(regions.len(), 1);
+    let region = &regions[0].regions[0];
+    assert!(
+        region.text.contains("quick brown fox"),
+        "invisible OCR layer should be served as region text, got: {:?}",
+        region.text
+    );
+    assert!(
+        !region.needs_ocr,
+        "recovered OCR layer must not fall back to GPU OCR"
+    );
+}
+
+/// ANY visible text on the page — even a single short line — must block the
+/// invisible-layer adoption entirely: the invisible pass returns visible
+/// items too, so adopting it alongside visible text would duplicate the
+/// visible words. Strict zero-visible gate, no fuzzy dedupe.
+#[test]
+fn test_extract_regions_mem_visible_text_blocks_invisible_layer() {
+    let buf = make_pdf_with_text_layer(3, Some("Folio 142"));
+    let regions = extract_text_in_regions_mem(&buf, &full_page_regions(1)).unwrap();
+    let region = &regions[0].regions[0];
+    assert!(
+        region.text.contains("Folio 142"),
+        "visible text should be extracted, got: {:?}",
+        region.text
+    );
+    assert!(
+        !region.text.contains("quick brown fox"),
+        "invisible layer must not be adopted when any visible text exists, got: {:?}",
+        region.text
+    );
+    assert_eq!(
+        region.text.matches("Folio 142").count(),
+        1,
+        "visible text must appear exactly once, got: {:?}",
+        region.text
+    );
+}
+
+/// An invisible OCR layer shown entirely via the `'` show-text operator
+/// (move-to-next-line + show) must also be recovered — the skipped_invisible
+/// signal has to fire on every show-text path, not just Tj/TJ.
+#[test]
+fn test_extract_regions_mem_recovers_quote_operator_layer() {
+    let buf = make_pdf_with_custom_text_layer(3, None, None, true);
+    let regions = extract_text_in_regions_mem(&buf, &full_page_regions(1)).unwrap();
+    let region = &regions[0].regions[0];
+    assert!(
+        region.text.contains("quick brown fox"),
+        "'-operator OCR layer should be recovered, got: {:?}",
+        region.text
+    );
+    assert!(!region.needs_ocr);
+}
+
+/// An invisible layer below the 40-alnum floor (a stray watermark line)
+/// must NOT be adopted — the region keeps its needs_ocr fallback.
+#[test]
+fn test_extract_regions_mem_tiny_invisible_layer_not_adopted() {
+    let buf = make_pdf_with_custom_text_layer(3, None, Some(&["Scanned by ACME"]), false);
+    let regions = extract_text_in_regions_mem(&buf, &full_page_regions(1)).unwrap();
+    let region = &regions[0].regions[0];
+    assert!(
+        !region.text.contains("Scanned by ACME"),
+        "below-floor invisible layer must not be adopted, got: {:?}",
+        region.text
+    );
+    // Only the raster placeholder remains — needs_ocr stays whatever main
+    // reports for placeholder-only regions (false today; downstream
+    // pipelines route placeholder-only text to OCR themselves, and this PR
+    // deliberately does not change that contract).
+    assert!(
+        region.text.trim().starts_with("[Image:"),
+        "region should hold only the raster placeholder, got: {:?}",
+        region.text
+    );
+}
+
+/// An invisible layer that clears the alnum floor but is mostly symbol
+/// garbage (a broken OCR run) must be rejected by the garbage gate.
+#[test]
+fn test_extract_regions_mem_garbage_invisible_layer_not_adopted() {
+    // Each line: 5 alphanumerics among 15 symbol chars. Ten lines clear the
+    // 40-alnum floor (50 alnum) while staying well under the half-alnum
+    // ratio is_garbage_text requires.
+    let garbage_lines: Vec<&str> = vec!["a@@b%%c&&d==e~~"; 10];
+    let buf = make_pdf_with_custom_text_layer(3, None, Some(&garbage_lines), false);
+    let regions = extract_text_in_regions_mem(&buf, &full_page_regions(1)).unwrap();
+    let region = &regions[0].regions[0];
+    assert!(
+        !region.text.contains("a@@b"),
+        "garbage invisible layer must not be adopted, got: {:?}",
+        region.text
+    );
+    assert!(
+        region.text.trim().starts_with("[Image:"),
+        "region should hold only the raster placeholder, got: {:?}",
+        region.text
+    );
+}
+
+/// Punctuation-only visible text (zero alphanumerics) must ALSO block
+/// adoption — the gate is item-presence, not alphanumeric mass. (Real-world
+/// rationale: an invisible OCR layer transcribes the raster, so visible
+/// glyphs typically have invisible twins there; this fixture's layers are
+/// disjoint, so it pins the gate itself, not the duplication scenario.)
+#[test]
+fn test_extract_regions_mem_punctuation_visible_blocks_invisible_layer() {
+    let buf = make_pdf_with_text_layer(3, Some("... --- ..."));
+    let regions = extract_text_in_regions_mem(&buf, &full_page_regions(1)).unwrap();
+    let region = &regions[0].regions[0];
+    assert!(
+        !region.text.contains("quick brown fox"),
+        "invisible layer must not be adopted over punctuation-only visible text, got: {:?}",
+        region.text
+    );
+    assert_eq!(
+        region.text.matches("... --- ...").count(),
+        1,
+        "visible punctuation must be preserved exactly once, got: {:?}",
+        region.text
+    );
+}
+
+/// Regression guard: a normal visible-text page (render mode 0) is served
+/// once and only once — if the fallback ever mis-fired here and merged a
+/// second pass, the phrase would duplicate.
+#[test]
+fn test_extract_regions_mem_visible_layer_unchanged() {
+    let buf = make_pdf_with_text_layer(0, None);
+    let regions = extract_text_in_regions_mem(&buf, &full_page_regions(1)).unwrap();
+    let region = &regions[0].regions[0];
+    assert_eq!(
+        region.text.matches("quick brown fox").count(),
+        1,
+        "visible text must appear exactly once, got: {:?}",
+        region.text
+    );
+    assert!(!region.needs_ocr);
+}
+
 #[test]
 fn test_extract_regions_mem_identity_h_needs_ocr() {
     let buf = std::fs::read("tests/fixtures/shinagawa_identity_h.pdf").unwrap();

tests/local_render_tests.rs

@@ -0,0 +1,65 @@
+#![cfg(all(feature = "render-pdfium", not(target_arch = "wasm32")))]
+
+use pdf_inspector::vision::{PdfiumRenderer, RenderError, RenderOptions, RenderPixelFormat};
+
+fn load_renderer() -> Option<PdfiumRenderer> {
+    match PdfiumRenderer::load() {
+        Ok(renderer) => Some(renderer),
+        Err(RenderError::Pdfium(firecrawl_pdfium::Error::Load(
+            firecrawl_pdfium::LoadError::LibraryNotFound { .. },
+        ))) => {
+            eprintln!("skipping PDFium runtime test because no native library is installed");
+            None
+        }
+        Err(error) => panic!("failed to load PDFium: {error}"),
+    }
+}
+
+#[test]
+fn renders_owned_rgb_page_and_round_trips_coordinates() {
+    let Some(renderer) = load_renderer() else {
+        return;
+    };
+    let bytes = std::fs::read("tests/fixtures/thermo-freon12.pdf").unwrap();
+    let pages = renderer
+        .render_pages(
+            &bytes,
+            &[1],
+            None,
+            &RenderOptions::new().dpi(150.0).form_fields(false),
+        )
+        .unwrap();
+
+    assert_eq!(pages.len(), 1);
+    let page = &pages[0];
+    assert_eq!(page.page(), 1);
+    assert_eq!(page.format(), RenderPixelFormat::Rgb8);
+    assert_eq!(page.stride(), page.width() as usize * 3);
+    assert_eq!(page.pixels().len(), page.stride() * page.height() as usize);
+    assert!((page.width() as f32 - page.page_width()).abs() > 1.0);
+
+    let pdf_rect = page.pixel_rect_to_pdf_rect(10.0, 10.0, 20.0, 12.0);
+    let pixel_rect = page.pdf_rect_to_pixel(&pdf_rect);
+    assert!((pixel_rect.0 - 10.0).abs() < 0.01);
+    assert!((pixel_rect.1 - 10.0).abs() < 0.01);
+    assert!((pixel_rect.2 - 20.0).abs() < 0.01);
+    assert!((pixel_rect.3 - 12.0).abs() < 0.01);
+}
+
+#[test]
+fn rejects_zero_and_out_of_range_page_numbers() {
+    let Some(renderer) = load_renderer() else {
+        return;
+    };
+    let bytes = std::fs::read("tests/fixtures/thermo-freon12.pdf").unwrap();
+
+    assert!(matches!(
+        renderer.render_pages(&bytes, &[0], None, &RenderOptions::new()),
+        Err(RenderError::InvalidPageNumber)
+    ));
+
+    assert!(matches!(
+        renderer.render_pages(&bytes, &[u32::MAX], None, &RenderOptions::new()),
+        Err(RenderError::PageOutOfBounds { .. })
+    ));
+}

tests/ocr_tests.rs

@@ -0,0 +1,159 @@
+#![cfg(all(feature = "ocr-oar", not(target_arch = "wasm32")))]
+
+#[cfg(feature = "ocr")]
+use pdf_inspector::vision::{
+    process_pdf_with_ocr_mem, ModelDownloadPolicy, OcrPdfOptions, PageContentSource,
+};
+use pdf_inspector::vision::{
+    ModelStore, OarOcrEngine, OcrEngine, OcrMode, OcrOptions, PageTransform, RenderPixelFormat,
+    RenderedPage, PP_OCR_V6_SMALL,
+};
+#[cfg(feature = "render-pdfium")]
+use pdf_inspector::vision::{PdfiumRenderer, RenderError, RenderOptions};
+
+const MODEL_DIRECTORY_ENV: &str = "PDF_INSPECTOR_OCR_TEST_MODELS";
+const IMAGE_ENV: &str = "PDF_INSPECTOR_OCR_TEST_IMAGE";
+const EXPECTED_TEXT_ENV: &str = "PDF_INSPECTOR_OCR_TEST_EXPECTED";
+
+#[cfg(feature = "render-pdfium")]
+fn load_renderer() -> Option<PdfiumRenderer> {
+    match PdfiumRenderer::load() {
+        Ok(renderer) => Some(renderer),
+        Err(RenderError::Pdfium(firecrawl_pdfium::Error::Load(
+            firecrawl_pdfium::LoadError::LibraryNotFound { .. },
+        ))) => {
+            eprintln!("skipping OCR runtime test because no native PDFium library is installed");
+            None
+        }
+        Err(error) => panic!("failed to load PDFium: {error}"),
+    }
+}
+
+#[test]
+fn recognizes_an_rgb_image_with_verified_models() {
+    let Some(model_directory) = std::env::var_os(MODEL_DIRECTORY_ENV) else {
+        eprintln!("skipping OCR runtime test because {MODEL_DIRECTORY_ENV} is not set");
+        return;
+    };
+    let Some(image_path) = std::env::var_os(IMAGE_ENV) else {
+        eprintln!("skipping OCR runtime test because {IMAGE_ENV} is not set");
+        return;
+    };
+
+    let image = image::open(image_path).unwrap().into_rgb8();
+    let (width, height) = image.dimensions();
+    let transform = PageTransform::from_corners(
+        width,
+        height,
+        (0.0, f64::from(height)),
+        (f64::from(width), f64::from(height)),
+        (0.0, 0.0),
+    )
+    .unwrap();
+    let page = RenderedPage::new(
+        1,
+        width as f32,
+        height as f32,
+        width,
+        height,
+        width as usize * 3,
+        RenderPixelFormat::Rgb8,
+        image.into_raw(),
+        transform,
+    )
+    .unwrap();
+    let results = recognize(&model_directory, &[page]);
+    assert_usable_result(&results);
+
+    let text = results[0]
+        .spans
+        .iter()
+        .map(|span| span.text.as_str())
+        .collect::<Vec<_>>()
+        .join(" ");
+    eprintln!("recognized: {text}");
+    if let Ok(expected) = std::env::var(EXPECTED_TEXT_ENV) {
+        assert!(
+            text.to_lowercase().contains(&expected.to_lowercase()),
+            "expected OCR output to contain {expected:?}, got {text:?}"
+        );
+    }
+}
+
+#[cfg(feature = "render-pdfium")]
+#[test]
+fn recognizes_a_pdfium_rendered_fixture_with_verified_models() {
+    let Some(model_directory) = std::env::var_os(MODEL_DIRECTORY_ENV) else {
+        eprintln!("skipping OCR runtime test because {MODEL_DIRECTORY_ENV} is not set");
+        return;
+    };
+    let Some(renderer) = load_renderer() else {
+        return;
+    };
+
+    let bytes = std::fs::read("tests/fixtures/thermo-freon12.pdf").unwrap();
+    let pages = renderer
+        .render_pages(
+            &bytes,
+            &[1],
+            None,
+            &RenderOptions::new().dpi(150.0).form_fields(false),
+        )
+        .unwrap();
+    let results = recognize(&model_directory, &pages);
+    assert_usable_result(&results);
+}
+
+#[cfg(all(feature = "ocr", feature = "render-pdfium"))]
+#[test]
+fn complete_ocr_pipeline_routes_and_assembles_a_scanned_fixture() {
+    let Some(model_directory) = std::env::var_os(MODEL_DIRECTORY_ENV) else {
+        eprintln!("skipping OCR runtime test because {MODEL_DIRECTORY_ENV} is not set");
+        return;
+    };
+    let Some(_renderer) = load_renderer() else {
+        return;
+    };
+
+    let bytes = std::fs::read("tests/fixtures/scan_with_native_header_text.pdf").unwrap();
+    let ocr = OcrOptions::new()
+        .mode(OcrMode::Auto)
+        .minimum_confidence(0.3)
+        .model_directory(model_directory)
+        .model_downloads(ModelDownloadPolicy::Offline);
+    let result = process_pdf_with_ocr_mem(&bytes, OcrPdfOptions::new().ocr(ocr)).unwrap();
+
+    assert_eq!(result.pages_routed_to_ocr, vec![1]);
+    assert!(!result.markdown.trim().is_empty());
+    assert_eq!(result.pages[0].provenance.source, PageContentSource::Ocr);
+    assert_eq!(
+        result.pages[0].provenance.ocr_model.as_ref().unwrap().name,
+        PP_OCR_V6_SMALL.id
+    );
+}
+
+fn recognize(
+    model_directory: &std::ffi::OsStr,
+    pages: &[RenderedPage],
+) -> Vec<pdf_inspector::vision::OcrPage> {
+    let store = ModelStore::new(model_directory).override_root(model_directory);
+    let models = store.resolve(&PP_OCR_V6_SMALL).unwrap();
+    let engine = OarOcrEngine::from_models(&models).unwrap();
+    engine
+        .recognize(
+            pages,
+            &OcrOptions::new()
+                .mode(OcrMode::Force)
+                .minimum_confidence(0.3),
+        )
+        .unwrap()
+}
+
+fn assert_usable_result(results: &[pdf_inspector::vision::OcrPage]) {
+    assert_eq!(results.len(), 1);
+    assert_eq!(results[0].page, 1);
+    assert_eq!(results[0].model.name, PP_OCR_V6_SMALL.id);
+    assert_eq!(results[0].model.revision, PP_OCR_V6_SMALL.revision);
+    assert!(!results[0].spans.is_empty());
+    assert!(results[0].spans.iter().all(|span| span.confidence >= 0.3));
+}

tests/snapshots/real-estate-pricing.md

@@ -46,9 +46,9 @@ more about investing in tax losses than burst, cap rates spreads steadily com- r
 
 8 6 Z E L L / L U R I E R E A L E S T A T E C E N T E R
 
-ingdebtspreadswerepartoforiginalpro formamodels.Thiscapratespreadcom- pressionoffsetweakcashflowsinapost- recessionary economy from 2002 to 2005, while continued compression, combined with improved cash flows, pushed property values skyward in 2006 throughmid-2007. Cap rate compression reduced the importance of the ability to add value. After all, if all you had to do to make moneywastoleveragetothehiltwhilecap ratesfell,whytakeontheextraworkand riskofattemptingtoaddvalue?Stateddif- ferently: Why print money if it is laying everywhereonthestreets? In Tables III and IV, we demonstrate thepowerofcapratecompressionviavery simple pro forma cash flow analyses that assume Year 1 NOI of $100; a going-in cap rate of 9 percent; an LTV of 70 per- cent; and an interest rate of 7 percent. Withineachfigure,wedisplaytwoscenar- ios, which vary based on NOI growth assumptions.ScenarioIassumesthatNOI growsby3percentperyear,whileScenario IIassumesavalue-addNOIgrowthof20 percentbetweenyearstwoandthree. The only other difference between TablesIIIandIVisinresidualcaprates, which are assumed to be 6 percent and 9 percent, respectively. Based on these assumptions, we calculate the equity IRRs. It is clear that cap rate compres- sion is a significant factor in driving
+ingdebtspreadswerepartoforiginalpro formamodels.Thiscapratespreadcom- pressionoffsetweakcashflowsinapost- recessionary economy from 2002 to 2005, while continued compression, combined with improved cash flows, pushed property values skyward in 2006 throughmid-2007. Cap rate compression reduced the importance of the ability to add value. After all, if all you had to do to make moneywastoleveragetothehiltwhilecap ratesfell,whytakeontheextraworkand riskofattemptingtoaddvalue?Stateddif- ferently: Why print money if it is laying everywhereonthestreets? In Tables III and IV, we demonstrate thepowerofcapratecompressionviavery simple pro forma cash flow analyses that assume Year 1 NOI of $100; a going-in cap rate of 9 percent; an LTV of 70 percent; and an interest rate of 7 percent. Withineachfigure,wedisplaytwoscenar- ios, which vary based on NOI growth assumptions.ScenarioIassumesthatNOI growsby3percentperyear,whileScenario IIassumesavalue-addNOIgrowthof20 percentbetweenyearstwoandthree. The only other difference between TablesIIIandIVisinresidualcaprates, which are assumed to be 6 percent and 9 percent, respectively. Based on these assumptions, we calculate the equity IRRs. It is clear that cap rate compression is a significant factor in driving
 
-returns. That is, cap rate compression from 9 percent to 6 percent increased IRR on leveraged stabilized properties by 250 percent, to a staggering 57 per- cent. Who needs to take on value add riskatthisreturnforstabilizedassets? Intheearly1980s,moneywasmadein real estate by mastering the creation and syndication of tax gimmicks. In the late 1980s, one made money by mastering bank and S&L connections to over-lever- age.Intheearly1990s,onemademoneyin realestatebyhavingaccesstoequity—the morethebetter.Duringthelate1990s,one made money from real estate by realizing large spreads between cap rates and debt costs.And,overthepastfiveyears,theway to make money in real estate was to own realestateonahighlyleveragedbasisascap ratesplunged. Theclassicassetpricingmodelisthe capital asset pricing model (CAPM). CAPM is a simple, yet elegant, model that relates asset pricing to the risk-free rate(F),theabilityofanassettoreduce portfolio variance (B), and the expected rate of return on the market bundle of investableassets(M).CAPMisfarfrom perfect,butprovidesacrudebenchmark for asset pricing, around which discrep- ancies and novelties arise. Specifically, CAPM states that an asset’s price is set suchthattheexpectedreturnforanasset
+returns. That is, cap rate compression from 9 percent to 6 percent increased IRR on leveraged stabilized properties by 250 percent, to a staggering 57 percent. Who needs to take on value add riskatthisreturnforstabilizedassets? Intheearly1980s,moneywasmadein real estate by mastering the creation and syndication of tax gimmicks. In the late 1980s, one made money by mastering bank and S&L connections to over-leverage.Intheearly1990s,onemademoneyin realestatebyhavingaccesstoequity—the morethebetter.Duringthelate1990s,one made money from real estate by realizing large spreads between cap rates and debt costs.And,overthepastfiveyears,theway to make money in real estate was to own realestateonahighlyleveragedbasisascap ratesplunged. Theclassicassetpricingmodelisthe capital asset pricing model (CAPM). CAPM is a simple, yet elegant, model that relates asset pricing to the risk-free rate(F),theabilityofanassettoreduce portfolio variance (B), and the expected rate of return on the market bundle of investableassets(M).CAPMisfarfrom perfect,butprovidesacrudebenchmark for asset pricing, around which discrep- ancies and novelties arise. Specifically, CAPM states that an asset’s price is set suchthattheexpectedreturnforanasset
 
 (R)is R=F+ β(M-F).
 R E V I E W 8 7

tests/snapshots/shannon-entropy-p1-2.md

@@ -14,7 +14,7 @@ basis for such a theory is contained in the important papers of Nyquist¹ and Ha
 
 1. It is practically more useful. Parameters of engineering importance such as time, bandwidth, number of relays, etc., tend to vary linearly with the logarithm of the number of possibilities. For example, adding one relay to a group doubles the number of possible states of the relays. It adds 1 to the base 2 logarithm of this number. Doubling the time roughly squares the number of possible messages, or doubles the logarithm, etc.
 2. It is nearer to our intuitive feeling as to the proper measure. This is closely related to (1) since we in- tuitively measures entities by linear comparison with common standards. One feels, for example, that two punched cards should have twice the capacity of one for information storage, and two identical channels twice the capacity of one for transmitting information.
-3. It is mathematically more suitable. Many of the limiting operations are simple in terms of the loga- rithm but would require clumsy restatement in terms of the number of possibilities. The choice of a logarithmic base corresponds to the choice of a unit for measuring information. If the
+3. It is mathematically more suitable. Many of the limiting operations are simple in terms of the logarithm but would require clumsy restatement in terms of the number of possibilities. The choice of a logarithmic base corresponds to the choice of a unit for measuring information. If the
 base 2 is used the resulting units may be called binary digits, or more briefly *bits,* a word suggested by
 
 J. W. Tukey. A device with two stable positions, such as a relay or a flip-flop circuit, can store one bit of information. *N* such devices can store*N* bits, since the total number of possible states is 2
@@ -34,8 +34,8 @@ Fig. 1 — Schematic diagram of a general communication system.
 
 a decimal digit is about 3 13 bits. A digit wheel on a desk computing machine has ten stable positions and therefore has a storage capacity of one decimal digit. In analytical work where integration and differentiation are involved the base *e* is sometimes useful. The resulting units of information will be called natural units. Change from the base *a* to base *b* merely requires multiplication by log*ba*. By a communication system we will mean a system of the type indicated schematically in Fig. 1. It consists of essentially five parts:
 
-1. An *information source* which produces a message or sequence of messages to be communicated to the receiving terminal. The message may be of various types: (a) A sequence of letters as in a telegraph of teletype system; (b) A single function of time *f* (*t*) as in radio or telephony; (c) A function of time and other variables as in black and white television — here the message may be thought of as a function *f* (*x*; *y*;*t*) of two space coordinates and time, the light intensity at point (*x*; *y*) and time *t* on a pickup tube plate; (d) Two or more functions of time, say *f* (*t*), *g*(*t*), *h*(*t*) — this is the case in “three- dimensional” sound transmission or if the system is intended to service several individual channels in multiplex; (e) Several functions of several variables — in color television the message consists of three functions *f* (*x*; *y*;*t*), *g*(*x*; *y*;*t*), *h*(*x*; *y*;*t*) defined in a three-dimensional continuum — we may also think of these three functions as components of a vector field defined in the region — similarly, several black and white television sources would produce “messages” consisting of a number of functions of three variables; (f) Various combinations also occur, for example in television with an associated audio channel.
-2. A *transmitter* which operates on the message in some way to produce a signal suitable for trans- mission over the channel. In telephony this operation consists merely of changing sound pressure into a proportional electrical current. In telegraphy we have an encoding operation which produces a sequence of dots, dashes and spaces on the channel corresponding to the message. In a multiplex PCM system the different speech functions must be sampled, compressed, quantized and encoded, and finally interleaved properly to construct the signal. Vocoder systems, television and frequency modulation are other examples of complex operations applied to the message to obtain the signal.
+1. An *information source* which produces a message or sequence of messages to be communicated to the receiving terminal. The message may be of various types: (a) A sequence of letters as in a telegraph of teletype system; (b) A single function of time *f* (*t*) as in radio or telephony; (c) A function of time and other variables as in black and white television — here the message may be thought of as a function *f* (*x*; *y*;*t*) of two space coordinates and time, the light intensity at point (*x*; *y*) and time *t* on a pickup tube plate; (d) Two or more functions of time, say *f* (*t*), *g*(*t*), *h*(*t*) — this is the case in “three-dimensional” sound transmission or if the system is intended to service several individual channels in multiplex; (e) Several functions of several variables — in color television the message consists of three functions *f* (*x*; *y*;*t*), *g*(*x*; *y*;*t*), *h*(*x*; *y*;*t*) defined in a three-dimensional continuum — we may also think of these three functions as components of a vector field defined in the region — similarly, several black and white television sources would produce “messages” consisting of a number of functions of three variables; (f) Various combinations also occur, for example in television with an associated audio channel.
+2. A *transmitter* which operates on the message in some way to produce a signal suitable for transmission over the channel. In telephony this operation consists merely of changing sound pressure into a proportional electrical current. In telegraphy we have an encoding operation which produces a sequence of dots, dashes and spaces on the channel corresponding to the message. In a multiplex PCM system the different speech functions must be sampled, compressed, quantized and encoded, and finally interleaved properly to construct the signal. Vocoder systems, television and frequency modulation are other examples of complex operations applied to the message to obtain the signal.
 3. The *channel* is merely the medium used to transmit the signal from transmitter to receiver. It may be a pair of wires, a coaxial cable, a band of radio frequencies, a beam of light, etc.
 4. The *receiver* ordinarily performs the inverse operation of that done by the transmitter, reconstructing the message from the signal.
 5. The *destination* is the person (or thing) for whom the message is intended. We wish to consider certain general problems involving communication systems. To do this it is first

tests/test_python.py

@@ -203,6 +203,79 @@ class TestExtractTextWithPositions:
         assert len(items) > 0
         assert all(item.page == 1 for item in items)
 
+    def test_mcid(self):
+        # Untagged fixture: mcid is None or int, never anything else
+        items = pdf_inspector.extract_text_with_positions(
+            fixture_path("thermo-freon12.pdf")
+        )
+        assert all(item.mcid is None or isinstance(item.mcid, int) for item in items)
+        # Tagged fixture: marked content carries MCIDs
+        tagged = pdf_inspector.extract_text_with_positions(
+            fixture_path("firecrawl_docs_tagged.pdf")
+        )
+        assert any(item.mcid is not None for item in tagged)
+
+
+# ---------------------------------------------------------------------------
+# extract_structure_elements / extract_structure_elements_bytes
+# ---------------------------------------------------------------------------
+
+
+class TestExtractStructureElements:
+    def test_tagged_file(self):
+        elements = pdf_inspector.extract_structure_elements(
+            fixture_path("firecrawl_docs_tagged.pdf")
+        )
+        assert len(elements) > 0
+        assert all(isinstance(e.page, int) for e in elements)
+        assert all(isinstance(e.mcid, int) for e in elements)
+        assert all(isinstance(e.role, str) and len(e.role) > 0 for e in elements)
+        assert any(e.role == "H1" for e in elements)
+
+    def test_join_with_text_items(self):
+        # (page, mcid) joins against extract_text_with_positions to recover
+        # heading text
+        path = fixture_path("firecrawl_docs_tagged.pdf")
+        elements = pdf_inspector.extract_structure_elements(path)
+        items = pdf_inspector.extract_text_with_positions(path)
+        h1_refs = {(e.page, e.mcid) for e in elements if e.role == "H1"}
+        h1_text = "".join(
+            item.text
+            for item in items
+            if item.mcid is not None and (item.page, item.mcid) in h1_refs
+        )
+        assert len(h1_text.strip()) > 0
+
+    def test_with_pages(self):
+        # pages filter is 1-indexed, matching TextItem.page
+        elements = pdf_inspector.extract_structure_elements(
+            fixture_path("firecrawl_docs_tagged.pdf"), pages=[1]
+        )
+        assert len(elements) > 0
+        assert all(e.page == 1 for e in elements)
+
+    def test_bytes(self):
+        data = fixture_bytes("firecrawl_docs_tagged.pdf")
+        elements = pdf_inspector.extract_structure_elements_bytes(data)
+        assert len(elements) > 0
+        assert any(e.role == "H1" for e in elements)
+
+    def test_untagged_returns_empty(self):
+        elements = pdf_inspector.extract_structure_elements(
+            fixture_path("thermo-freon12.pdf")
+        )
+        assert elements == []
+
+    def test_repr(self):
+        elements = pdf_inspector.extract_structure_elements(
+            fixture_path("firecrawl_docs_tagged.pdf")
+        )
+        assert "StructureElement" in repr(elements[0])
+
+    def test_not_a_pdf(self):
+        with pytest.raises(ValueError):
+            pdf_inspector.extract_structure_elements_bytes(b"not a pdf")
+
 
 # ---------------------------------------------------------------------------
 # extract_text_in_regions / extract_text_in_regions_bytes

wasm/Cargo.lock

@@ -724,7 +724,7 @@ checksum = "d6790f58c7ff633d8771f42965289203411a5e5c68388703c06e14f24770b41e"
 
 [[package]]
 name = "pdf-inspector"
-version = "0.1.7"
+version = "1.14.2"
 dependencies = [
  "env_logger",
  "include_dir",
@@ -740,7 +740,7 @@ dependencies = [
 
 [[package]]
 name = "pdf-inspector-wasm"
-version = "0.1.3"
+version = "1.14.2"
 dependencies = [
  "console_error_panic_hook",
  "js-sys",

wasm/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "pdf-inspector-wasm"
-version = "0.1.3"
+version = "1.14.2"
 edition = "2021"
 authors = ["Firecrawl Team"]
 description = "Browser WebAssembly bindings for pdf-inspector"