docs(structure-tree): mention traversal budget in truncation warning

The one-shot truncation warning listed the node budget, cycle, and depth
as causes but not the new traversal (work) budget, so a work-budget
truncation printed a misleading message. Include MAX_STRUCT_WORK so
malformed-PDF debugging identifies the actual limit hit.

Co-authored-by: Abimael Martell <abimaelmartell@users.noreply.github.com>

.github/workflows/ci.yml

@@ -14,44 +14,54 @@ jobs:
     name: Test
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
       - name: Install Rust
-        uses: dtolnay/rust-toolchain@stable
+        uses: dtolnay/rust-toolchain@4cda84d5c5c54efe2404f9d843567869ab1699d4 # stable
+        with:
+          toolchain: stable
 
       - name: Cache cargo
-        uses: Swatinem/rust-cache@v2
+        uses: Swatinem/rust-cache@c19371144df3bb44fab255c43d04cbc2ab54d1c4 # v2.9.1
 
       - name: Run tests
         run: cargo test --verbose
 
+      - name: Test developer scripts
+        run: python3 -m unittest discover -s scripts/tests
+
   fmt:
     name: Format
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
       - name: Install Rust
-        uses: dtolnay/rust-toolchain@stable
+        uses: dtolnay/rust-toolchain@4cda84d5c5c54efe2404f9d843567869ab1699d4 # stable
         with:
+          toolchain: stable
           components: rustfmt
 
       - name: Check formatting
         run: cargo fmt --all -- --check
 
+      - name: Check WASM formatting
+        run: cargo fmt --manifest-path wasm/Cargo.toml -- --check
+
   clippy:
     name: Clippy
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
       - name: Install Rust
-        uses: dtolnay/rust-toolchain@stable
+        uses: dtolnay/rust-toolchain@4cda84d5c5c54efe2404f9d843567869ab1699d4 # stable
         with:
+          toolchain: stable
           components: clippy
 
       - name: Cache cargo
-        uses: Swatinem/rust-cache@v2
+        uses: Swatinem/rust-cache@c19371144df3bb44fab255c43d04cbc2ab54d1c4 # v2.9.1
         with:
           key: clippy
 
@@ -65,15 +75,52 @@ jobs:
       matrix:
         os: [ubuntu-latest, macos-latest]
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
       - name: Install Rust
-        uses: dtolnay/rust-toolchain@stable
+        uses: dtolnay/rust-toolchain@4cda84d5c5c54efe2404f9d843567869ab1699d4 # stable
+        with:
+          toolchain: stable
 
       - name: Cache cargo
-        uses: Swatinem/rust-cache@v2
+        uses: Swatinem/rust-cache@c19371144df3bb44fab255c43d04cbc2ab54d1c4 # v2.9.1
         with:
           key: build
 
       - name: Build
         run: cargo build --release --verbose
+
+  wasm:
+    name: WebAssembly
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+
+      - name: Install Rust
+        uses: dtolnay/rust-toolchain@4cda84d5c5c54efe2404f9d843567869ab1699d4 # stable
+        with:
+          toolchain: stable
+          targets: wasm32-unknown-unknown
+          components: clippy
+
+      - name: Cache cargo
+        uses: Swatinem/rust-cache@c19371144df3bb44fab255c43d04cbc2ab54d1c4 # v2.9.1
+        with:
+          workspaces: |
+            wasm -> target
+          key: wasm
+
+      - name: Check WebAssembly bindings
+        run: cargo check --manifest-path wasm/Cargo.toml --target wasm32-unknown-unknown
+
+      - name: Check root package for WebAssembly
+        run: cargo check --target wasm32-unknown-unknown
+
+      - name: Lint WebAssembly bindings
+        run: cargo clippy --manifest-path wasm/Cargo.toml --target wasm32-unknown-unknown -- -D warnings
+
+      - name: Install wasm-pack
+        run: cargo install wasm-pack --version 0.15.0 --locked
+
+      - name: Test WebAssembly package
+        run: wasm-pack test --node --release wasm

.github/workflows/pages.yml

@@ -24,13 +24,13 @@ jobs:
       name: github-pages
       url: ${{ steps.deploy.outputs.page_url }}
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
       - name: Upload site artifact
-        uses: actions/upload-pages-artifact@v3
+        uses: actions/upload-pages-artifact@fc324d3547104276b827a68afc52ff2a11cc49c9 # v5.0.0
         with:
           path: site
 
       - name: Deploy to GitHub Pages
         id: deploy
-        uses: actions/deploy-pages@v4
+        uses: actions/deploy-pages@cd2ce8fcbc39b97be8ca5fce6e763baed58fa128 # v5.0.0

.github/workflows/publish-crate.yml

@@ -4,6 +4,9 @@ on:
   push:
     branches: [main]
     paths: ['Cargo.toml']
+  # Manual fallback: retry a publish that failed after the version was
+  # already merged (a plain re-push won't register as a version change).
+  workflow_dispatch:
 
 permissions:
   contents: read
@@ -14,13 +17,17 @@ env:
 jobs:
   check-version:
     name: Check version change
+    # Guard manual dispatches: crates.io trusted publishing matches
+    # repo+workflow+environment but NOT branch, so without this a
+    # workflow_dispatch from any branch could publish unmerged code.
+    if: github.ref == 'refs/heads/main'
     runs-on: ubuntu-latest
     outputs:
       changed: ${{ steps.check.outputs.changed }}
       published: ${{ steps.check.outputs.published }}
       version: ${{ steps.check.outputs.version }}
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           fetch-depth: 2
 
@@ -28,17 +35,26 @@ jobs:
         id: check
         run: |
           NEW_VERSION=$(python3 -c 'import pathlib, tomllib; print(tomllib.loads(pathlib.Path("Cargo.toml").read_text())["package"]["version"])')
-          OLD_VERSION=$(git show HEAD~1:Cargo.toml | python3 -c 'import sys, tomllib; print(tomllib.loads(sys.stdin.read())["package"]["version"])')
-          echo "old=$OLD_VERSION new=$NEW_VERSION"
           echo "version=$NEW_VERSION" >> "$GITHUB_OUTPUT"
 
-          if [ "$NEW_VERSION" = "$OLD_VERSION" ]; then
-            echo "changed=false" >> "$GITHUB_OUTPUT"
-            echo "published=false" >> "$GITHUB_OUTPUT"
-            exit 0
-          fi
+          if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
+            # Manual dispatch publishes the current version regardless of the
+            # previous commit; the crates.io check below still prevents
+            # double-publishing an already-released version.
+            echo "manual dispatch: publishing v$NEW_VERSION"
+            echo "changed=true" >> "$GITHUB_OUTPUT"
+          else
+            OLD_VERSION=$(git show HEAD~1:Cargo.toml | python3 -c 'import sys, tomllib; print(tomllib.loads(sys.stdin.read())["package"]["version"])')
+            echo "old=$OLD_VERSION new=$NEW_VERSION"
 
-          echo "changed=true" >> "$GITHUB_OUTPUT"
+            if [ "$NEW_VERSION" = "$OLD_VERSION" ]; then
+              echo "changed=false" >> "$GITHUB_OUTPUT"
+              echo "published=false" >> "$GITHUB_OUTPUT"
+              exit 0
+            fi
+
+            echo "changed=true" >> "$GITHUB_OUTPUT"
+          fi
 
           HTTP_STATUS=$(curl --silent --show-error --output /tmp/crate-version.json --write-out "%{http_code}" \
             -H "User-Agent: firecrawl/pdf-inspector publish workflow (https://github.com/firecrawl/pdf-inspector)" \
@@ -69,17 +85,19 @@ jobs:
       contents: read
       id-token: write
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
       - name: Install Rust
-        uses: dtolnay/rust-toolchain@stable
+        uses: dtolnay/rust-toolchain@4cda84d5c5c54efe2404f9d843567869ab1699d4 # stable
+        with:
+          toolchain: stable
 
       - name: Verify package
         run: cargo publish --dry-run
 
       - name: Authenticate with crates.io
         id: auth
-        uses: rust-lang/crates-io-auth-action@v1
+        uses: rust-lang/crates-io-auth-action@c6f97d42243bad5fab37ca0427f495c86d5b1a18 # v1.0.5
 
       - name: Publish crate
         run: cargo publish

.github/workflows/publish-pypi.yml

@@ -0,0 +1,166 @@
+name: Publish Python package
+
+on:
+  push:
+    branches: [main]
+    paths: ['pyproject.toml']
+  # Manual fallback: re-publish the current version without a version bump
+  # (e.g. first run after PyPI trusted publishing is configured).
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  check-version:
+    name: Check version change
+    # Guard manual dispatches too: PyPI trusted publishing matches
+    # repo+workflow+environment but NOT branch, so without this a
+    # workflow_dispatch from any branch could publish unmerged code.
+    if: github.ref == 'refs/heads/main'
+    runs-on: ubuntu-latest
+    outputs:
+      changed: ${{ steps.check.outputs.changed }}
+      published: ${{ steps.check.outputs.published }}
+      version: ${{ steps.check.outputs.version }}
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+        with:
+          fetch-depth: 2
+
+      - name: Check if version changed
+        id: check
+        run: |
+          NEW_VERSION=$(python3 -c 'import pathlib, tomllib; print(tomllib.loads(pathlib.Path("pyproject.toml").read_text())["project"]["version"])')
+          echo "version=$NEW_VERSION" >> "$GITHUB_OUTPUT"
+
+          if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
+            # Manual dispatch always rebuilds and publishes. Combined with
+            # skip-existing on the publish step, this repairs partial releases
+            # (PyPI's version endpoint returns 200 even when only some of the
+            # expected wheels were uploaded).
+            echo "manual dispatch: publishing v$NEW_VERSION (skip-existing handles uploaded files)"
+            echo "changed=true" >> "$GITHUB_OUTPUT"
+            echo "published=false" >> "$GITHUB_OUTPUT"
+            exit 0
+          fi
+
+          # .get(): the parent commit may predate the static version field
+          # (pyproject.toml used dynamic = ["version"]) — treat that as a change
+          # so the very first merge of this workflow publishes.
+          OLD_VERSION=$(git show HEAD~1:pyproject.toml | python3 -c 'import sys, tomllib; print(tomllib.loads(sys.stdin.read())["project"].get("version", ""))')
+          echo "old=$OLD_VERSION new=$NEW_VERSION"
+
+          if [ "$NEW_VERSION" = "$OLD_VERSION" ]; then
+            echo "changed=false" >> "$GITHUB_OUTPUT"
+            echo "published=false" >> "$GITHUB_OUTPUT"
+            exit 0
+          fi
+
+          echo "changed=true" >> "$GITHUB_OUTPUT"
+
+          HTTP_STATUS=$(curl --silent --show-error --output /tmp/pypi-version.json --write-out "%{http_code}" \
+            "https://pypi.org/pypi/pdf-inspector/$NEW_VERSION/json")
+
+          case "$HTTP_STATUS" in
+            200)
+              echo "published=true" >> "$GITHUB_OUTPUT"
+              echo "pdf-inspector v$NEW_VERSION is already published to PyPI"
+              ;;
+            404)
+              echo "published=false" >> "$GITHUB_OUTPUT"
+              ;;
+            *)
+              cat /tmp/pypi-version.json
+              echo "Unexpected PyPI response: $HTTP_STATUS" >&2
+              exit 1
+              ;;
+          esac
+
+  build:
+    needs: check-version
+    if: needs.check-version.outputs.changed == 'true' && needs.check-version.outputs.published == 'false'
+    name: Build ${{ matrix.target }}
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        include:
+          - os: ubuntu-latest
+            target: x86_64-unknown-linux-gnu
+          - os: ubuntu-latest
+            target: aarch64-unknown-linux-gnu
+          # macos-13 was retired by GitHub; macos-15-intel is the remaining
+          # Intel runner label (available through 2027).
+          - os: macos-15-intel
+            target: x86_64-apple-darwin
+          - os: macos-14
+            target: aarch64-apple-darwin
+          - os: windows-latest
+            target: x86_64-pc-windows-msvc
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+
+      - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
+        with:
+          python-version: '3.12'
+
+      - name: Build wheel
+        uses: PyO3/maturin-action@e83996d129638aa358a18fbd1dfb82f0b0fb5d3b # v1.51.0
+        with:
+          target: ${{ matrix.target }}
+          args: --release --out dist
+          manylinux: auto
+
+      - name: Upload wheel
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
+        with:
+          name: wheels-${{ matrix.target }}
+          path: dist/*.whl
+          if-no-files-found: error
+
+  sdist:
+    needs: check-version
+    if: needs.check-version.outputs.changed == 'true' && needs.check-version.outputs.published == 'false'
+    name: Build sdist
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+
+      - name: Build sdist
+        uses: PyO3/maturin-action@e83996d129638aa358a18fbd1dfb82f0b0fb5d3b # v1.51.0
+        with:
+          command: sdist
+          args: --out dist
+
+      - name: Upload sdist
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
+        with:
+          name: sdist
+          path: dist/*.tar.gz
+          if-no-files-found: error
+
+  publish:
+    name: Publish to PyPI
+    needs: [check-version, build, sdist]
+    runs-on: ubuntu-latest
+    environment: pypi
+    permissions:
+      contents: read
+      id-token: write
+    steps:
+      - name: Download all artifacts
+        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8.0.1
+        with:
+          path: dist
+          merge-multiple: true
+
+      - name: List artifacts
+        run: ls -la dist/
+
+      - name: Publish to PyPI
+        uses: pypa/gh-action-pypi-publish@dc37677b2e1c63e2034f94d8a5b11f265b73ba33 # release/v1
+        with:
+          packages-dir: dist
+          # Tolerate already-uploaded files so a manual re-run can complete
+          # a release that previously failed partway through.
+          skip-existing: true

.github/workflows/publish-wasm.yml

@@ -0,0 +1,112 @@
+name: Publish WebAssembly package
+
+on:
+  push:
+    branches: [main]
+    paths: ['wasm/Cargo.toml']
+  workflow_dispatch:
+
+permissions:
+  contents: read
+  id-token: write
+
+env:
+  CARGO_TERM_COLOR: always
+
+jobs:
+  check-version:
+    name: Check version change
+    if: github.ref == 'refs/heads/main'
+    runs-on: ubuntu-latest
+    outputs:
+      changed: ${{ steps.check.outputs.changed }}
+      package_exists: ${{ steps.check.outputs.package_exists }}
+      published: ${{ steps.check.outputs.published }}
+      version: ${{ steps.check.outputs.version }}
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+        with:
+          fetch-depth: 2
+
+      - name: Check package version
+        id: check
+        run: |
+          NEW_VERSION=$(python3 -c 'import pathlib, tomllib; print(tomllib.loads(pathlib.Path("wasm/Cargo.toml").read_text())["package"]["version"])')
+          echo "version=$NEW_VERSION" >> "$GITHUB_OUTPUT"
+
+          if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
+            echo "changed=true" >> "$GITHUB_OUTPUT"
+          elif git cat-file -e HEAD~1:wasm/Cargo.toml 2>/dev/null; then
+            OLD_VERSION=$(git show HEAD~1:wasm/Cargo.toml | python3 -c 'import sys, tomllib; print(tomllib.loads(sys.stdin.read())["package"]["version"])')
+            if [ "$NEW_VERSION" = "$OLD_VERSION" ]; then
+              echo "changed=false" >> "$GITHUB_OUTPUT"
+              echo "published=false" >> "$GITHUB_OUTPUT"
+              exit 0
+            fi
+            echo "changed=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "changed=true" >> "$GITHUB_OUTPUT"
+          fi
+
+          if ! npm view "@firecrawl/pdf-inspector-wasm" name >/dev/null 2>&1; then
+            echo "package_exists=false" >> "$GITHUB_OUTPUT"
+            echo "published=false" >> "$GITHUB_OUTPUT"
+            echo "The initial package must be published once before trusted publishing can be configured."
+            exit 0
+          fi
+
+          echo "package_exists=true" >> "$GITHUB_OUTPUT"
+          if npm view "@firecrawl/pdf-inspector-wasm@$NEW_VERSION" version >/dev/null 2>&1; then
+            echo "published=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "published=false" >> "$GITHUB_OUTPUT"
+          fi
+
+  publish:
+    name: Build and publish
+    needs: check-version
+    if: needs.check-version.outputs.changed == 'true' && needs.check-version.outputs.package_exists == 'true' && needs.check-version.outputs.published == 'false'
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+      id-token: write
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+
+      - name: Install Rust
+        uses: dtolnay/rust-toolchain@4cda84d5c5c54efe2404f9d843567869ab1699d4 # stable
+        with:
+          toolchain: stable
+          targets: wasm32-unknown-unknown
+
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
+        with:
+          node-version: '24'
+          registry-url: 'https://registry.npmjs.org'
+
+      - name: Install wasm-pack
+        run: cargo install wasm-pack --version 0.15.0 --locked
+
+      - name: Build browser package
+        run: wasm-pack build wasm --target web --scope firecrawl --out-dir pkg --release
+
+      - name: Prepare package metadata
+        run: |
+          node -e '
+            const fs = require("fs")
+            const path = "wasm/pkg/package.json"
+            const pkg = JSON.parse(fs.readFileSync(path, "utf8"))
+            pkg.name = "@firecrawl/pdf-inspector-wasm"
+            pkg.description = "Browser WebAssembly bindings for the pdf-inspector Rust PDF parser"
+            pkg.keywords = ["pdf", "pdf-parser", "webassembly", "wasm", "markdown", "rust", "firecrawl"]
+            pkg.repository = { type: "git", url: "https://github.com/firecrawl/pdf-inspector" }
+            pkg.homepage = "https://github.com/firecrawl/pdf-inspector/tree/main/wasm"
+            pkg.publishConfig = { access: "public" }
+            fs.writeFileSync(path, JSON.stringify(pkg, null, 2) + "\n")
+          '
+
+      - name: Inspect package contents
+        run: npm pack --dry-run ./wasm/pkg
+
+      - name: Publish package
+        run: npm publish ./wasm/pkg --provenance --access public

.github/workflows/publish.yml

@@ -4,6 +4,9 @@ on:
   push:
     branches: [main]
     paths: ['napi/package.json']
+  # Manual fallback: retry a publish that failed partway (per-package
+  # already-published checks make re-runs idempotent).
+  workflow_dispatch:
 
 permissions:
   contents: read
@@ -12,12 +15,16 @@ permissions:
 jobs:
   check-version:
     name: Check version change
+    # Guard manual dispatches: npm trusted publishing matches
+    # repo+workflow+environment but NOT branch, so without this a
+    # workflow_dispatch from any branch could publish unmerged code.
+    if: github.ref == 'refs/heads/main'
     runs-on: ubuntu-latest
     outputs:
       changed: ${{ steps.check.outputs.changed }}
       version: ${{ steps.check.outputs.version }}
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           fetch-depth: 2
 
@@ -25,11 +32,21 @@ jobs:
         id: check
         run: |
           NEW_VERSION=$(node -p "require('./napi/package.json').version")
+          echo "version=$NEW_VERSION" >> "$GITHUB_OUTPUT"
+
+          if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
+            # Manual dispatch rebuilds and publishes the current version; the
+            # per-package already-published checks in the publish job skip
+            # anything that made it out in a previous partial run.
+            echo "manual dispatch: publishing v$NEW_VERSION"
+            echo "changed=true" >> "$GITHUB_OUTPUT"
+            exit 0
+          fi
+
           OLD_VERSION=$(git show HEAD~1:napi/package.json | node -p "JSON.parse(require('fs').readFileSync('/dev/stdin','utf8')).version")
           echo "old=$OLD_VERSION new=$NEW_VERSION"
           if [ "$NEW_VERSION" != "$OLD_VERSION" ]; then
             echo "changed=true" >> "$GITHUB_OUTPUT"
-            echo "version=$NEW_VERSION" >> "$GITHUB_OUTPUT"
           else
             echo "changed=false" >> "$GITHUB_OUTPUT"
           fi
@@ -44,31 +61,61 @@ jobs:
         include:
           - os: ubuntu-latest
             target: x86_64-unknown-linux-gnu
+          # napi-cross builds gnu targets against an old glibc sysroot for
+          # broad distro compatibility; musl targets cross-compile with
+          # zig via cargo-zigbuild (napi's -x flag).
+          - os: ubuntu-latest
+            target: aarch64-unknown-linux-gnu
+            build-flags: --use-napi-cross
+          - os: ubuntu-latest
+            target: x86_64-unknown-linux-musl
+            build-flags: -x
+          - os: ubuntu-latest
+            target: aarch64-unknown-linux-musl
+            build-flags: -x
           - os: macos-14
             target: aarch64-apple-darwin
           - os: windows-latest
             target: x86_64-pc-windows-msvc
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
       - name: Install Rust
-        uses: dtolnay/rust-toolchain@stable
+        uses: dtolnay/rust-toolchain@4cda84d5c5c54efe2404f9d843567869ab1699d4 # stable
+        with:
+          toolchain: stable
+          targets: ${{ matrix.target }}
 
-      - uses: oven-sh/setup-bun@v2
+      - uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0
         with:
           bun-version: latest
 
+      - name: Install zig
+        if: contains(matrix.target, 'musl')
+        uses: mlugg/setup-zig@d1434d08867e3ee9daa34448df10607b98908d29 # v2.2.1
+        with:
+          version: 0.14.1
+
+      - name: Install cargo-zigbuild
+        if: contains(matrix.target, 'musl')
+        uses: taiki-e/install-action@67729d5c413db75907f0ad1e39bb04b9c868ff60 # v2.85.7
+        env:
+          GITHUB_TOKEN: ${{ github.token }}
+        with:
+          tool: cargo-zigbuild
+
       - name: Cache cargo
-        uses: actions/cache@v4
+        uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
         with:
           path: |
             ~/.cargo/registry/index/
             ~/.cargo/registry/cache/
             ~/.cargo/git/db/
+            ~/.napi-rs
             napi/target/
-          key: ${{ runner.os }}-cargo-napi-${{ hashFiles('**/Cargo.lock') }}
+          key: ${{ matrix.target }}-cargo-napi-${{ hashFiles('**/Cargo.lock') }}
           restore-keys: |
-            ${{ runner.os }}-cargo-napi-
+            ${{ matrix.target }}-cargo-napi-
 
       - name: Install dependencies
         working-directory: napi
@@ -76,10 +123,10 @@ jobs:
 
       - name: Build native addon
         working-directory: napi
-        run: bunx napi build --platform --release
+        run: bunx napi build --platform --release --target ${{ matrix.target }} ${{ matrix.build-flags }}
 
       - name: Upload native binary
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
         with:
           name: bindings-${{ matrix.target }}
           path: napi/*.node
@@ -87,7 +134,7 @@ jobs:
 
       - name: Upload generated JS bindings
         if: matrix.target == 'x86_64-unknown-linux-gnu'
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
         with:
           name: js-bindings
           path: |
@@ -95,34 +142,149 @@ jobs:
             napi/index.d.ts
           if-no-files-found: error
 
+  smoke-test:
+    name: Smoke test ${{ matrix.target }}
+    needs: [check-version, build]
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - os: ubuntu-latest
+            target: x86_64-unknown-linux-gnu
+          - os: ubuntu-latest
+            target: x86_64-unknown-linux-musl
+          - os: ubuntu-24.04-arm
+            target: aarch64-unknown-linux-gnu
+          - os: ubuntu-24.04-arm
+            target: aarch64-unknown-linux-musl
+          - os: macos-14
+            target: aarch64-apple-darwin
+          - os: windows-latest
+            target: x86_64-pc-windows-msvc
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+
+      - name: Download native binary
+        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8.0.1
+        with:
+          name: bindings-${{ matrix.target }}
+          path: napi
+
+      - name: Download generated JS bindings
+        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8.0.1
+        with:
+          name: js-bindings
+          path: napi
+
+      # musl binaries must load under a real musl libc, so run inside Alpine.
+      - name: Run smoke test (Alpine)
+        if: contains(matrix.target, 'musl')
+        run: docker run --rm -v "$PWD:/repo" -w /repo/napi node:24-alpine node test.mjs
+
+      - name: Run smoke test
+        if: ${{ !contains(matrix.target, 'musl') }}
+        working-directory: napi
+        run: node test.mjs
+
   publish:
     name: Publish to npm
-    needs: [check-version, build]
+    needs: [check-version, build, smoke-test]
     runs-on: ubuntu-latest
     permissions:
       contents: read
       id-token: write
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
-      - uses: actions/setup-node@v6
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
         with:
           node-version: '24'
           registry-url: 'https://registry.npmjs.org'
 
       - name: Download all artifacts
-        uses: actions/download-artifact@v4
+        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8.0.1
         with:
           path: napi/artifacts
 
-      - name: Collect binaries and publish
+      - name: Publish platform packages
         working-directory: napi
         run: |
-          cp artifacts/bindings-*/*.node .
+          VERSION="${{ needs.check-version.outputs.version }}"
+
+          for node_file in artifacts/bindings-*/pdf-inspector.*.node; do
+            base=$(basename "$node_file")
+            suffix=${base#pdf-inspector.}
+            suffix=${suffix%.node}
+            pkg="@firecrawl/pdf-inspector-$suffix"
+
+            if npm view "$pkg@$VERSION" version >/dev/null 2>&1; then
+              echo "$pkg@$VERSION already published — skipping"
+              continue
+            fi
+
+            dir="npm-dist/$suffix"
+            mkdir -p "$dir"
+            cp "$node_file" "$dir/"
+            node -e '
+              const [suffix, version] = process.argv.slice(1)
+              const meta = {
+                "linux-x64-gnu":    { os: ["linux"],  cpu: ["x64"],   libc: ["glibc"] },
+                "linux-x64-musl":   { os: ["linux"],  cpu: ["x64"],   libc: ["musl"] },
+                "linux-arm64-gnu":  { os: ["linux"],  cpu: ["arm64"], libc: ["glibc"] },
+                "linux-arm64-musl": { os: ["linux"],  cpu: ["arm64"], libc: ["musl"] },
+                "darwin-arm64":     { os: ["darwin"], cpu: ["arm64"] },
+                "win32-x64-msvc":   { os: ["win32"],  cpu: ["x64"] },
+              }[suffix]
+              if (!meta) {
+                console.error(`unknown platform suffix: ${suffix} — add it to the meta map`)
+                process.exit(1)
+              }
+              const pkg = {
+                name: `@firecrawl/pdf-inspector-${suffix}`,
+                version,
+                description: `Prebuilt ${suffix} binary for @firecrawl/pdf-inspector`,
+                main: `pdf-inspector.${suffix}.node`,
+                files: [`pdf-inspector.${suffix}.node`],
+                license: "MIT",
+                engines: { node: ">= 10" },
+                repository: { type: "git", url: "https://github.com/firecrawl/pdf-inspector" },
+                publishConfig: { access: "public" },
+                ...meta,
+              }
+              require("fs").writeFileSync(`npm-dist/${suffix}/package.json`, JSON.stringify(pkg, null, 2) + "\n")
+            ' "$suffix" "$VERSION"
+
+            echo "=== $pkg@$VERSION ==="
+            ls -la "$dir"
+            (cd "$dir" && npm publish --provenance --access public)
+          done
+
+      - name: Publish main package
+        working-directory: napi
+        run: |
+          VERSION="${{ needs.check-version.outputs.version }}"
+
+          if npm view "@firecrawl/pdf-inspector@$VERSION" version >/dev/null 2>&1; then
+            echo "@firecrawl/pdf-inspector@$VERSION already published — skipping"
+            exit 0
+          fi
+
           cp artifacts/js-bindings/index.js .
           cp artifacts/js-bindings/index.d.ts .
 
-          echo "=== Package contents ==="
-          ls -la *.node index.js index.d.ts
+          # Stamp optionalDependencies to this exact version so the platform
+          # pins can never drift from the main package version.
+          node -e '
+            const fs = require("fs")
+            const pkg = JSON.parse(fs.readFileSync("package.json", "utf8"))
+            for (const dep of Object.keys(pkg.optionalDependencies ?? {})) {
+              pkg.optionalDependencies[dep] = pkg.version
+            }
+            fs.writeFileSync("package.json", JSON.stringify(pkg, null, 2) + "\n")
+          '
+
+          echo "=== Main package contents ==="
+          npm pack --dry-run
 
           npm publish --provenance --access public

.gitignore

@@ -1,10 +1,13 @@
 # Rust build artifacts
 /target/
+/wasm/target/
+/wasm/pkg/
 debug/
 *.pdb
 
 # Cargo lock (optional for libraries)
 Cargo.lock
+!/wasm/Cargo.lock
 
 # IDE
 .idea/
@@ -34,9 +37,9 @@ scripts/
 
 # Test output
 test_output/
+.firecrawl/
 
 # Python
 __pycache__/
 *.pyc
 .pytest_cache/
-

AGENTS.md

@@ -61,7 +61,8 @@ src/
 
 - **Unit tests**: inline `#[cfg(test)] mod tests` in each module with synthetic data.
 - **Integration tests**: `tests/integration_tests.rs` with fixture PDFs in `tests/fixtures/`.
-- **Regression suite**: sibling repo `pdf-evals` with 179+ snapshot PDFs. Run `cargo build --release` then `bench.py test` in that repo before committing.
+- **Regression suite**: sibling repo `pdf-evals` with ~200 snapshot PDFs. Run `cargo build --release` then `bench.py test` in that repo before committing. While iterating, prefer a subset run (`bench.py test -q` for the quick set, or `-s <name>` for a named test set) and save the full `bench.py test` for the final pre-commit check.
+- **Semantic quality**: run `bench.py score` in `pdf-evals` for the semantic verdict (TEDS + MHS + reading order + char/word + list preservation, composited). Character-level diff alone misclassifies structural improvements (e.g., column-detection rewrites) as regressions — `score` is the tie-breaker. See `pdf-evals/CLAUDE.md` "Semantic scoring".
 
 ## Debugging
 

CLAUDE.md

@@ -61,7 +61,7 @@ src/
 
 - **Unit tests**: inline `#[cfg(test)] mod tests` in each module with synthetic data.
 - **Integration tests**: `tests/integration_tests.rs` with fixture PDFs in `tests/fixtures/`.
-- **Regression suite**: sibling repo `pdf-evals` with 187+ snapshot PDFs. Run `cargo build --release` then `bench.py test` in that repo before committing.
+- **Regression suite**: sibling repo `pdf-evals` with ~200 snapshot PDFs. Run `cargo build --release` then `bench.py test` in that repo before committing. While iterating, prefer a subset run (`bench.py test -q` for the quick set, or `-s <name>` for a named test set) and save the full `bench.py test` for the final pre-commit check.
 - **Semantic quality**: run `bench.py score` in `pdf-evals` for the semantic verdict (TEDS + MHS + reading order + char/word + list preservation, composited). Character-level diff alone misclassifies structural improvements (e.g., column-detection rewrites) as regressions — `score` is the tie-breaker. See `pdf-evals/CLAUDE.md` "Semantic scoring".
 
 ## Debugging

Cargo.toml

@@ -1,12 +1,25 @@
 [package]
 name = "pdf-inspector"
-version = "0.1.4"
+version = "0.1.7"
 edition = "2021"
 autobins = false
 authors = ["Firecrawl Team"]
 description = "Fast PDF inspection, classification, and text extraction with smart scanned vs text-based detection"
 license = "MIT"
 repository = "https://github.com/firecrawl/pdf-inspector"
+readme = "docs/rust-api.md"
+# Explicit allowlist: crates.io caps uploads at 10 MiB and tests/fixtures
+# alone exceeds that. external/bcmaps ships in the crate — tounicode.rs
+# loads it at runtime relative to CARGO_MANIFEST_DIR.
+include = [
+    "/src/**",
+    "/external/bcmaps/**",
+    "/docs/rust-api.md",
+    "/LICENSE",
+    # maturin derives the sdist file list from this allowlist; the stub must
+    # ship so wheels built from the sdist keep their type hints.
+    "/pdf_inspector.pyi",
+]
 
 [lib]
 name = "pdf_inspector"
@@ -14,20 +27,13 @@ crate-type = ["lib", "cdylib"]
 
 [dependencies]
 # Python bindings
-pyo3 = { version = "0.25", features = ["extension-module"], optional = true }
-
-# PDF parsing
-lopdf = { version = "0.41.0", features = ["rayon"] }
+pyo3 = { version = "0.25", features = ["extension-module", "abi3-py38"], optional = true }
 
 # Error handling
 thiserror = "2.0"
 
-# Parallel processing
-rayon = "1.10"
-
 # Logging
 log = "0.4"
-env_logger = "0.11"
 
 # Text processing
 regex = "1.10"
@@ -37,6 +43,19 @@ unicode-normalization = "0.1"
 # TrueType font parsing (for Identity-H CID font cmap extraction)
 ttf-parser = "0.25"
 
+# Native builds keep lopdf's parallel parser and CLI logging. Browser WASM is
+# deliberately single-threaded so it works without cross-origin isolation.
+[target.'cfg(not(target_arch = "wasm32"))'.dependencies]
+lopdf = { version = "0.42.0", features = ["rayon"] }
+rayon = "1.10"
+env_logger = "0.11"
+
+# Browser builds use JavaScript randomness for encrypted PDFs and embed the
+# bundled CMaps because there is no filesystem at runtime.
+[target.'cfg(target_arch = "wasm32")'.dependencies]
+lopdf = { version = "0.42.0", default-features = false, features = ["wasm_js"] }
+include_dir = "0.7"
+
 [dev-dependencies]
 tempfile = "3.3"
 

README.md

@@ -2,9 +2,10 @@
 
 [![Crates.io](https://img.shields.io/crates/v/pdf-inspector.svg)](https://crates.io/crates/pdf-inspector)
 [![npm](https://img.shields.io/npm/v/@firecrawl/pdf-inspector.svg)](https://www.npmjs.com/package/@firecrawl/pdf-inspector)
+[![PyPI](https://img.shields.io/pypi/v/pdf-inspector.svg)](https://pypi.org/project/pdf-inspector/)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)
 
-Fast Rust library for 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. Includes bindings for [Python](docs/python.md) and [Node.js](napi/README.md).
+Fast Rust library for 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. Includes bindings for [Python](docs/python.md), [Node.js](napi/README.md), and [browser WebAssembly](wasm/README.md).
 
 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.
 
@@ -18,24 +19,28 @@ Built by [Firecrawl](https://firecrawl.dev) to handle text-based PDFs locally in
 - **Multi-column layout** — Automatic detection of newspaper-style columns, sequential reading order, and RTL text support.
 - **Encoding issue detection** — Automatically flags broken font encodings so callers can fall back to OCR.
 - **Single document load** — The document is parsed once and shared between detection and extraction, avoiding redundant I/O.
+- **Browser WebAssembly** — Run the same Rust parser locally in browsers and Web Workers, with embedded CMaps and no server round trip.
 - **Lightweight** — Pure Rust, no ML models, no external services. Single dependency on `lopdf` for PDF parsing.
 
 ## Benchmark
 
-Evaluated on the [opendataloader-bench](https://github.com/opendataloader-project/opendataloader-bench) corpus (200 PDFs). Only direct text extraction engines are shown — no OCR, no ML models. Scores are 0-1, higher is better.
+Evaluated on the [opendataloader-bench](https://github.com/opendataloader-project/opendataloader-bench) corpus (200 PDFs). Only local engines without model-based PDF parsing are shown; OCR was disabled. Scores are 0-1, higher is better.
 
 | Engine | Overall | Reading Order (NID) | Tables (TEDS) | Headings (MHS) | Speed (200 docs) |
 |---|---|---|---|---|---|
-| pdf-inspector | 0.83 | 0.88 | 0.66 | 0.74 | 4s |
-| opendataloader | 0.84 | 0.91 | 0.49 | 0.74 | 11s |
-| pymupdf4llm | 0.73 | 0.89 | 0.40 | 0.41 | 18s |
-| markitdown | 0.58 | 0.88 | 0.00 | 0.00 | 8s |
+| pdf-inspector | **0.875** | **0.915** | **0.814** | 0.788 | **0.470s** |
+| liteparse | 0.873 | 0.913 | 0.693 | **0.811** | 0.750s |
+| opendataloader | 0.831 | 0.902 | 0.489 | 0.739 | 2.569s |
+| pymupdf4llm | 0.735 | 0.886 | 0.401 | 0.424 | 17.117s |
+| markitdown | 0.589 | 0.844 | 0.273 | 0.000 | 16.165s |
 
-For context, engines that use OCR/ML (docling, marker, mineru) score 0.83-0.88 overall but take 2-180 minutes on the same corpus — pdf-inspector reaches the low end of that range without any OCR, in 4 seconds.
+Results were refreshed on July 31, 2026, on an Apple M4 Pro. Engine versions were pdf-inspector 0.2.6, LiteParse 2.10.1, OpenDataLoader 2.2.1, PyMuPDF4LLM 0.2.0, and MarkItDown 0.1.5. Speed is the median of five alternating or rotating complete corpus runs after an excluded warm-up run, with each parser processing documents sequentially in a single process.
 
-**Where we do well:** Speed (fastest of all engines), the best table detection of any engine shown, and heading detection now on par with opendataloader. Overall lands within 0.01 of opendataloader at roughly 2.5× the speed.
+The complete parser configuration, per-document predictions, evaluator output, and generated charts are available in the [reproducible results branch](https://github.com/firecrawl/opendataloader-bench/tree/abi/pdf-parser-benchmark-results).
 
-**Where we lag:** Reading order still trails opendataloader slightly, and table structure trails OCR-based engines that can see visual layout.
+**Best fit:** Native-text PDFs where speed, reading order, and table structure matter. In this comparison, pdf-inspector delivered the higher overall, reading-order, and table scores, along with the fastest complete run. That makes it a strong local default for reports, research papers, financial documents, invoices, and legal PDFs that need clean, structured Markdown without adding OCR latency or infrastructure.
+
+Use the [paired benchmark harness](docs/benchmarking.md) to compare two local builds against the exact same corpus and evaluator revision.
 
 ## Quick start
 
@@ -73,6 +78,26 @@ console.log(result.markdown);  // Markdown string or null
 
 > Full API reference: [napi/README.md](napi/README.md)
 
+### Browser WebAssembly
+
+```bash
+npm install @firecrawl/pdf-inspector-wasm
+```
+
+```javascript
+import init, { processPdf } from '@firecrawl/pdf-inspector-wasm';
+
+await init();
+const response = await fetch('/document.pdf');
+const pdf = new Uint8Array(await response.arrayBuffer());
+const result = processPdf(pdf);
+
+console.log(result.pdfType);
+console.log(result.markdown);
+```
+
+> Full API reference: [wasm/README.md](wasm/README.md)
+
 ### Rust
 
 Install from [crates.io](https://crates.io/crates/pdf-inspector):
@@ -118,6 +143,9 @@ pdf2md document.pdf --items-json
 # Raw markdown only (no headers)
 pdf2md document.pdf --raw
 
+# Token-efficient output (collapses long dot leaders and similar source padding)
+pdf2md document.pdf --compact
+
 # Insert page break markers (<!-- Page N -->)
 pdf2md document.pdf --pages
 
@@ -181,6 +209,7 @@ src/
   markdown/             — Markdown conversion and structure detection
   bin/                  — CLI tools (pdf2md, detect_pdf)
 napi/                   — Node.js/Bun bindings (napi-rs)
+wasm/                   — Browser bindings (wasm-bindgen)
 ```
 
 ## How classification works
@@ -209,7 +238,7 @@ The converter handles:
 |---|---|
 | Headings (H1-H4) | Font size tiers relative to body text, with 0.5pt clustering |
 | Bold/italic | Font name patterns (Bold, Italic, Oblique) |
-| Bullet lists | `*`, `-`, `*`, `○`, `●`, `◦` prefixes |
+| Bullet lists | `•`, `-`, `*`, `○`, `●`, `◦` prefixes |
 | Numbered lists | `1.`, `1)`, `(1)` patterns |
 | Letter lists | `a.`, `a)`, `(a)` patterns |
 | Code blocks | Monospace fonts (Courier, Consolas, Monaco, Menlo, Fira Code, JetBrains Mono) and keyword detection |

docs/benchmarking.md

@@ -0,0 +1,62 @@
+# Benchmarking against OpenDataLoader
+
+The paired harness runs two `pdf2md` binaries through the same local
+OpenDataLoader corpus, evaluates both outputs, and reports aggregate and
+per-document deltas. This avoids comparing results produced from different
+corpus revisions or evaluator versions.
+
+Build a candidate and provide a released or worktree build as the baseline:
+
+```bash
+cargo build --release
+python3 scripts/bench_opendataloader.py \
+  --bench-dir ../opendataloader-bench \
+  --baseline ../pdf-inspector-main/target/release/pdf2md \
+  --candidate target/release/pdf2md \
+  --max-document-regression 0.02 \
+  --json-output /tmp/pdf-inspector-benchmark.json
+```
+
+Pass `--reference-evaluation path/to/evaluation.json` to report the candidate
+delta against another evaluation, and add `--require-reference-lead` to make a
+negative reference delta fail the run. By default, the candidate must not
+regress the baseline overall score or introduce missing predictions. Use
+`--min-overall-delta` to require a specific aggregate gain.
+
+The OpenDataLoader repository is external and keeps its normal
+`prediction/pdf-inspector` output. Paired evaluation copies each run into a
+temporary directory before evaluating it, so the baseline and candidate cannot
+overwrite one another.
+
+## Published comparison protocol
+
+The public benchmark table was refreshed on July 31, 2026, on an Apple M4 Pro
+using pdf-inspector 0.2.6, LiteParse 2.10.1, OpenDataLoader 2.2.1,
+PyMuPDF4LLM 0.2.0, and MarkItDown 0.1.5. Every engine processed the same 200
+PDFs sequentially in a single process with OCR disabled. Reported speed is the
+median of five alternating or rotating complete corpus runs after an excluded
+warm-up run; quality scores come from the benchmark evaluator over all 200
+outputs. Raw timings, predictions, evaluations, and charts are available in the
+[results branch](https://github.com/firecrawl/opendataloader-bench/tree/abi/pdf-parser-benchmark-results).
+
+## Optional backend evidence probe
+
+The evidence probe compares positioned `pdf2md` items with MuPDF structured
+text on the same pages. It is intended to find deterministic extraction or
+layout evidence that could justify a future native implementation; it does not
+merge MuPDF output into Markdown, invoke OCR, or add a runtime dependency.
+
+Install MuPDF's `mutool`, build `pdf2md`, then run:
+
+```bash
+python3 scripts/probe_backend_evidence.py document.pdf \
+  --pdf2md target/release/pdf2md \
+  --json-output /tmp/backend-evidence.json
+```
+
+The report flags pages when MuPDF exposes a material net token gain, repeated
+alignment anchors absent from local evidence, or additional image blocks. The
+JSON includes bounded token samples and page-level counts so promising cases
+can be inspected without treating backend disagreement as automatically
+correct. Thresholds are configurable with `--min-token-gain`,
+`--min-alternate-only-ratio`, and `--min-anchor-gain`.

docs/publishing.md

@@ -19,3 +19,20 @@ The workflow uses `rust-lang/crates-io-auth-action@v1` to exchange GitHub's OIDC
 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.
 
 If `Cargo.toml` changes without a package version bump, the workflow exits without publishing.
+
+## Browser WebAssembly package
+
+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 npm package must exist before a trusted publisher can be configured. For the first release only:
+
+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`
+
+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.

docs/python.md

@@ -1,9 +1,39 @@
-# Python API
+# pdf-inspector
 
-Python bindings via [PyO3](https://pyo3.rs). Requires Rust toolchain for building from source.
+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. Python bindings via [PyO3](https://pyo3.rs) for the [pdf-inspector](https://github.com/firecrawl/pdf-inspector) Rust library.
+
+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.
+
+## Features
+
+- **Smart classification** — `text_based` / `scanned` / `image_based` / `mixed` in ~10–50ms, with a confidence score and per-page OCR routing.
+- **Markdown conversion** — headings, lists, code blocks, bold/italic, URL linking, and dual-mode table detection (PDF drawing ops + text-alignment heuristics).
+- **Layout-aware extraction** — multi-column reading order, position and font info per text item, RTL support.
+- **Robust text decoding** — CID/Type0 fonts via ToUnicode CMaps, plus automatic flagging of broken encodings so callers can fall back to OCR.
+- **Lightweight** — native Rust core, no ML models, no external services; ships type stubs.
+
+## Benchmark
+
+[opendataloader-bench](https://github.com/opendataloader-project/opendataloader-bench) corpus (200 PDFs), local engines without model-based PDF parsing; OCR disabled. Scores 0–1, higher is better:
+
+| Engine | Overall | Reading order | Tables (TEDS) | Headings | Speed |
+|---|---|---|---|---|---|
+| **pdf-inspector** | **0.875** | **0.915** | **0.814** | 0.788 | **0.470s** |
+| liteparse | 0.873 | 0.913 | 0.693 | **0.811** | 0.750s |
+| opendataloader | 0.831 | 0.902 | 0.489 | 0.739 | 2.569s |
+| pymupdf4llm | 0.735 | 0.886 | 0.401 | 0.424 | 17.117s |
+| markitdown | 0.589 | 0.844 | 0.273 | 0.000 | 16.165s |
+
+Refreshed July 31, 2026, on Apple M4 Pro; speed is the median of five complete corpus runs after an excluded warm-up. Full methodology and versions are in the [repo README](https://github.com/firecrawl/pdf-inspector#benchmark), with raw timings and artifacts in the [results branch](https://github.com/firecrawl/opendataloader-bench/tree/abi/pdf-parser-benchmark-results).
 
 ## Install
 
+```bash
+pip install pdf-inspector
+```
+
+Prebuilt wheels cover CPython ≥3.8 on Linux (x86_64, aarch64), macOS (Intel, Apple Silicon), and Windows (x64). Other platforms build from source, which requires a Rust toolchain. For local development in a repo checkout:
+
 ```bash
 pip install maturin
 maturin develop --release
@@ -73,16 +103,62 @@ result = pdf_inspector.extract_pages_markdown("document.pdf", pages=[0, 2])
 
 ## Types
 
-**`PdfResult` fields:** `pdf_type`, `markdown`, `page_count`, `processing_time_ms`, `pages_needing_ocr`, `title`, `confidence`, `is_complex_layout`, `pages_with_tables`, `pages_with_columns`, `has_encoding_issues`
+Type stubs (`pdf_inspector.pyi`) ship with the package. Result types at a glance:
 
-**`PdfClassification` fields:** `pdf_type`, `page_count`, `pages_needing_ocr` (0-indexed), `confidence`
+```python
+class PdfResult:                     # process_pdf / detect_pdf
+    pdf_type: str                    # "text_based" | "scanned" | "image_based" | "mixed"
+    markdown: str | None             # extracted Markdown (None for detect_pdf)
+    page_count: int
+    processing_time_ms: int
+    pages_needing_ocr: list[int]     # 1-indexed
+    ocr_reasons_by_page: list[PageOcrReasons]
+    title: str | None
+    confidence: float                # 0.0 - 1.0
+    is_complex_layout: bool
+    pages_with_tables: list[int]
+    pages_with_columns: list[int]
+    has_encoding_issues: bool        # broken font encodings — consider OCR fallback
 
-**`TextItem` fields:** `text`, `x`, `y`, `width`, `height`, `font`, `font_size`, `page`, `is_bold`, `is_italic`, `item_type`
+class PageOcrReasons:                # per-page OCR diagnostics
+    page: int                        # 1-indexed
+    reasons: list[str]               # machine-readable reason identifiers
 
-**`RegionText` fields:** `text`, `needs_ocr`
+class PdfClassification:             # classify_pdf
+    pdf_type: str
+    page_count: int
+    pages_needing_ocr: list[int]     # 0-indexed
+    confidence: float
 
-**`PageRegionTexts` fields:** `page` (0-indexed), `regions` (list of RegionText)
+class TextItem:                      # extract_text_with_positions
+    text: str
+    x: float
+    y: float
+    width: float
+    height: float
+    font: str
+    font_size: float
+    page: int
+    is_bold: bool
+    is_italic: bool
+    is_underline: bool
+    is_strikeout: bool
+    item_type: str
 
-**`PageMarkdown` fields:** `page` (0-indexed), `markdown`, `needs_ocr`
+class RegionText:                    # extract_text_in_regions
+    text: str
+    needs_ocr: bool
+    ocr_reason: str | None           # machine-readable OCR reason
 
-**`PagesExtractionResult` fields:** `pages` (list of PageMarkdown), `pages_with_tables` (1-indexed), `pages_with_columns` (1-indexed), `pages_needing_ocr` (1-indexed), `is_complex`
+class PageRegionTexts:               # extract_text_in_regions
+    page: int                        # 0-indexed
+    regions: list[RegionText]
+
+class PagesExtractionResult:         # extract_pages_markdown
+    pages: list[PageMarkdown]        # PageMarkdown: page (0-indexed), markdown, needs_ocr, ocr_reason
+    pages_with_tables: list[int]     # 1-indexed
+    pages_with_columns: list[int]    # 1-indexed
+    pages_needing_ocr: list[int]     # 1-indexed
+    ocr_reasons_by_page: list[PageOcrReasons]
+    is_complex: bool                 # any page has tables or multi-column layout
+```

docs/rust-api.md

@@ -1,12 +1,50 @@
-# Rust API
+# pdf-inspector
 
-Add to your `Cargo.toml`:
+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).
+
+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.
+
+## Features
+
+- **Smart classification** — TextBased / Scanned / ImageBased / Mixed in ~10–50ms, with a confidence score and per-page OCR routing.
+- **Markdown conversion** — headings, lists, code blocks, bold/italic, URL linking, and dual-mode table detection (PDF drawing ops + text-alignment heuristics).
+- **Layout-aware extraction** — multi-column reading order, position and font info per text item, RTL support.
+- **Robust text decoding** — CID/Type0 fonts via ToUnicode CMaps, plus automatic flagging of broken encodings so callers can fall back to OCR.
+- **Lightweight** — pure Rust, no ML models, no external services; single PDF dependency ([lopdf](https://crates.io/crates/lopdf)).
+
+## Benchmark
+
+[opendataloader-bench](https://github.com/opendataloader-project/opendataloader-bench) corpus (200 PDFs), local engines without model-based PDF parsing; OCR disabled. Scores 0–1, higher is better:
+
+| Engine | Overall | Reading order | Tables (TEDS) | Headings | Speed |
+|---|---|---|---|---|---|
+| **pdf-inspector** | **0.875** | **0.915** | **0.814** | 0.788 | **0.470s** |
+| liteparse | 0.873 | 0.913 | 0.693 | **0.811** | 0.750s |
+| opendataloader | 0.831 | 0.902 | 0.489 | 0.739 | 2.569s |
+| pymupdf4llm | 0.735 | 0.886 | 0.401 | 0.424 | 17.117s |
+| markitdown | 0.589 | 0.844 | 0.273 | 0.000 | 16.165s |
+
+Refreshed July 31, 2026, on Apple M4 Pro; speed is the median of five complete corpus runs after an excluded warm-up. Full methodology and versions are in the [repo README](https://github.com/firecrawl/pdf-inspector#benchmark), with raw timings and artifacts in the [results branch](https://github.com/firecrawl/opendataloader-bench/tree/abi/pdf-parser-benchmark-results).
+
+## Install
+
+```bash
+cargo add pdf-inspector
+```
+
+For the latest unreleased changes, use the git dependency instead:
 
 ```toml
 [dependencies]
 pdf-inspector = { git = "https://github.com/firecrawl/pdf-inspector" }
 ```
 
+The crate also ships CLI binaries — `pdf2md` (PDF → Markdown, with `--json`, `--pages`, `--select-pages`, and the opt-in token-saving `--compact` profile) and `detect-pdf` (classification, with `--analyze --json`):
+
+```bash
+cargo install pdf-inspector
+```
+
 ## Usage
 
 Detect and extract in one call:

napi/Cargo.lock

@@ -499,11 +499,13 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0de51e6874e94e7bf76d726fc5d13ba782deca734ff60d5bb2fb2607c7406555"
 dependencies = [
  "cfg-if",
+ "js-sys",
  "libc",
  "r-efi",
  "rand_core",
  "wasip2",
  "wasip3",
+ "wasm-bindgen",
 ]
 
 [[package]]
@@ -557,6 +559,25 @@ version = "2.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3d3067d79b975e8844ca9eb072e16b31c3c1c36928edf9c6789548c524d0d954"
 
+[[package]]
+name = "include_dir"
+version = "0.7.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "923d117408f1e49d914f1a379a309cffe4f18c05cf4e3d12e613a15fc81bd0dd"
+dependencies = [
+ "include_dir_macros",
+]
+
+[[package]]
+name = "include_dir_macros"
+version = "0.7.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7cab85a7ed0bd5f0e76d93846e0147172bed2e2d3f859bcc33a8d9699cad1a75"
+dependencies = [
+ "proc-macro2",
+ "quote",
+]
+
 [[package]]
 name = "indexmap"
 version = "2.13.0"
@@ -672,9 +693,9 @@ checksum = "5e5032e24019045c762d3c0f28f5b6b8bbf38563a65908389bf7978758920897"
 
 [[package]]
 name = "lopdf"
-version = "0.41.0"
+version = "0.42.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "67513274c50a2b51e5f75d9e682fcf4ab064a8a9c9ae2c3c59309084882bb24d"
+checksum = "25aab26d99567469098e64a02f42679f8965c6401263eefa31d8f2dcc37a221c"
 dependencies = [
  "aes",
  "bitflags",
@@ -830,9 +851,10 @@ checksum = "384b8ab6d37215f3c5301a95a4accb5d64aa607f1fcb26a11b5303878451b4fe"
 
 [[package]]
 name = "pdf-inspector"
-version = "0.1.4"
+version = "0.1.7"
 dependencies = [
  "env_logger",
+ "include_dir",
  "log",
  "lopdf",
  "once_cell",

napi/README.md

@@ -4,6 +4,28 @@ Fast PDF classification and region-based text extraction for Node.js/Bun. Native
 
 Built by [Firecrawl](https://firecrawl.dev) for hybrid OCR pipelines — extract text from PDF structure where possible, fall back to OCR only when needed.
 
+## Features
+
+- **Smart classification** — text-based / scanned / image-based / mixed in ~10–50ms, with a confidence score and per-page OCR routing.
+- **Region-based extraction** — pull text from bounding boxes with per-region quality checks (`needsOcr`).
+- **Layout-aware** — multi-column reading order, position and font info per text item, RTL support.
+- **Robust text decoding** — CID/Type0 fonts via ToUnicode CMaps, plus automatic flagging of broken encodings so callers can fall back to OCR.
+- **Lightweight** — native Rust core via napi-rs, no ML models, no external services; ~5–6 MB platform binary, TypeScript definitions included.
+
+## Benchmark
+
+[opendataloader-bench](https://github.com/opendataloader-project/opendataloader-bench) corpus (200 PDFs), local engines without model-based PDF parsing; OCR disabled. Scores 0–1, higher is better:
+
+| Engine | Overall | Reading order | Tables (TEDS) | Headings | Speed |
+|---|---|---|---|---|---|
+| **pdf-inspector** | **0.875** | **0.915** | **0.814** | 0.788 | **0.470s** |
+| liteparse | 0.873 | 0.913 | 0.693 | **0.811** | 0.750s |
+| opendataloader | 0.831 | 0.902 | 0.489 | 0.739 | 2.569s |
+| pymupdf4llm | 0.735 | 0.886 | 0.401 | 0.424 | 17.117s |
+| markitdown | 0.589 | 0.844 | 0.273 | 0.000 | 16.165s |
+
+Refreshed July 31, 2026, on Apple M4 Pro; speed is the median of five complete corpus runs after an excluded warm-up. Full methodology and versions are in the [repo README](https://github.com/firecrawl/pdf-inspector#benchmark), with raw timings and artifacts in the [results branch](https://github.com/firecrawl/opendataloader-bench/tree/abi/pdf-parser-benchmark-results).
+
 ## Install
 
 ```bash
@@ -12,7 +34,7 @@ npm install @firecrawl/pdf-inspector
 bun add @firecrawl/pdf-inspector
 ```
 
-Prebuilt binaries included for **linux-x64** and **macOS ARM64**. No Rust toolchain needed.
+Prebuilt binaries for **Linux x64/ARM64** (glibc and musl/Alpine), **macOS ARM64**, and **Windows x64** — npm installs only the one matching your platform. No Rust toolchain needed.
 
 ## API
 
@@ -90,10 +112,16 @@ interface RegionText {
 
 ## Platforms
 
-| Platform | Architecture | Supported |
-|----------|-------------|-----------|
-| Linux    | x64         | Yes       |
-| macOS    | ARM64       | Yes       |
+Prebuilt binaries ship as platform-specific packages installed automatically via `optionalDependencies`:
+
+| Platform | Architecture | Package |
+|----------|-------------|---------|
+| Linux    | x64 (glibc)         | `@firecrawl/pdf-inspector-linux-x64-gnu` |
+| Linux    | x64 (musl/Alpine)   | `@firecrawl/pdf-inspector-linux-x64-musl` |
+| Linux    | ARM64 (glibc)       | `@firecrawl/pdf-inspector-linux-arm64-gnu` |
+| Linux    | ARM64 (musl/Alpine) | `@firecrawl/pdf-inspector-linux-arm64-musl` |
+| macOS    | ARM64               | `@firecrawl/pdf-inspector-darwin-arm64` |
+| Windows  | x64                 | `@firecrawl/pdf-inspector-win32-x64-msvc` |
 
 ## License
 

napi/bun.lock

@@ -7,6 +7,14 @@
       "devDependencies": {
         "@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",
+      },
     },
   },
   "packages": {

napi/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@firecrawl/pdf-inspector",
-  "version": "1.10.1",
+  "version": "1.12.0",
   "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",
@@ -22,7 +22,6 @@
   "files": [
     "index.js",
     "index.d.ts",
-    "*.node",
     "bin/",
     "README.md"
   ],
@@ -38,12 +37,12 @@
     "binaryName": "pdf-inspector",
     "targets": [
       "x86_64-unknown-linux-gnu",
+      "x86_64-unknown-linux-musl",
+      "aarch64-unknown-linux-gnu",
+      "aarch64-unknown-linux-musl",
       "aarch64-apple-darwin",
       "x86_64-pc-windows-msvc"
-    ],
-    "package": {
-      "name": "@firecrawl/pdf-inspector-js"
-    }
+    ]
   },
   "scripts": {
     "build": "napi build --platform --release",
@@ -51,5 +50,13 @@
   },
   "devDependencies": {
     "@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"
   }
 }

pdf_inspector.pyi

@@ -10,6 +10,9 @@ class PdfResult:
     page_count: int
     processing_time_ms: int
     pages_needing_ocr: list[int]
+    """1-indexed page numbers that need OCR."""
+    ocr_reasons_by_page: list["PageOcrReasons"]
+    """Machine-readable OCR reasons by 1-indexed page."""
     title: Optional[str]
     confidence: float
     is_complex_layout: bool
@@ -17,6 +20,13 @@ class PdfResult:
     pages_with_columns: list[int]
     has_encoding_issues: bool
 
+class PageOcrReasons:
+    """OCR reasons for a single 1-indexed page."""
+    page: int
+    """1-indexed page number."""
+    reasons: list[str]
+    """Machine-readable OCR reason identifiers."""
+
 class PdfClassification:
     """Lightweight PDF classification result."""
     pdf_type: str
@@ -47,6 +57,8 @@ class RegionText:
     text: str
     needs_ocr: bool
     """True when the text should not be trusted."""
+    ocr_reason: Optional[str]
+    """Machine-readable OCR reason when the cause is known."""
 
 class PageRegionTexts:
     """Extracted text for one page's regions."""
@@ -62,6 +74,8 @@ class PageMarkdown:
     """Formatted markdown for this page (empty string when needs_ocr is True)."""
     needs_ocr: bool
     """True when text on this page is unreliable and OCR should be used instead."""
+    ocr_reason: Optional[str]
+    """Machine-readable OCR reason when the cause is known."""
 
 class PagesExtractionResult:
     """Per-page markdown output with document-wide layout classification."""
@@ -73,6 +87,8 @@ class PagesExtractionResult:
     """1-indexed pages where multi-column layout was detected."""
     pages_needing_ocr: list[int]
     """1-indexed pages that need OCR."""
+    ocr_reasons_by_page: list[PageOcrReasons]
+    """Machine-readable OCR reasons by 1-indexed page."""
     is_complex: bool
     """True if any page has tables or multi-column layout."""
 

pyproject.toml

@@ -4,10 +4,11 @@ build-backend = "maturin"
 
 [project]
 name = "pdf-inspector"
-# Version is sourced from Cargo.toml [package] version by maturin so the Python
-# artifact always tracks the crate release instead of drifting on its own.
-dynamic = ["version"]
+# 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"
 description = "Fast PDF inspection, classification, and text extraction with smart scanned vs text-based detection"
+readme = "docs/python.md"
 license = { text = "MIT" }
 requires-python = ">=3.8"
 classifiers = [
@@ -19,5 +20,10 @@ classifiers = [
     "Topic :: Text Processing",
 ]
 
+[project.urls]
+Homepage = "https://github.com/firecrawl/pdf-inspector"
+Repository = "https://github.com/firecrawl/pdf-inspector"
+Documentation = "https://github.com/firecrawl/pdf-inspector/blob/main/docs/python.md"
+
 [tool.maturin]
 features = ["python"]

scripts/bench_opendataloader.py

@@ -0,0 +1,351 @@
+#!/usr/bin/env python3
+"""Run a paired pdf-inspector OpenDataLoader benchmark and report deltas."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import math
+import os
+import shutil
+import subprocess
+import sys
+import tempfile
+from pathlib import Path
+from typing import Any
+
+
+SCORE_KEYS = (
+    "overall_mean",
+    "nid_mean",
+    "nid_s_mean",
+    "teds_mean",
+    "teds_s_mean",
+    "mhs_mean",
+    "mhs_s_mean",
+)
+
+
+def _non_negative_int(value: str) -> int:
+    parsed = int(value)
+    if parsed < 0:
+        raise argparse.ArgumentTypeError("must be non-negative")
+    return parsed
+
+
+def _non_negative_float(value: str) -> float:
+    parsed = float(value)
+    if not math.isfinite(parsed) or parsed < 0.0:
+        raise argparse.ArgumentTypeError("must be finite and non-negative")
+    return parsed
+
+
+def _finite_float(value: str) -> float:
+    parsed = float(value)
+    if not math.isfinite(parsed):
+        raise argparse.ArgumentTypeError("must be finite")
+    return parsed
+
+
+def _scores(evaluation: dict[str, Any]) -> dict[str, float]:
+    score = evaluation.get("metrics", {}).get("score", {})
+    return {key: float(score[key]) for key in SCORE_KEYS if score.get(key) is not None}
+
+
+def _documents(evaluation: dict[str, Any]) -> dict[str, float]:
+    documents: dict[str, float] = {}
+    for document in evaluation.get("documents", []):
+        overall = document.get("scores", {}).get("overall")
+        if overall is not None:
+            documents[str(document["document_id"])] = float(overall)
+    return documents
+
+
+def compare_evaluations(
+    baseline: dict[str, Any],
+    candidate: dict[str, Any],
+    reference: dict[str, Any] | None = None,
+    *,
+    top: int = 10,
+) -> dict[str, Any]:
+    """Build aggregate and per-document deltas from evaluator JSON payloads."""
+    baseline_scores = _scores(baseline)
+    candidate_scores = _scores(candidate)
+    metric_deltas = {
+        key: candidate_scores[key] - baseline_scores[key]
+        for key in SCORE_KEYS
+        if key in baseline_scores and key in candidate_scores
+    }
+
+    baseline_documents = _documents(baseline)
+    candidate_documents = _documents(candidate)
+    shared = sorted(baseline_documents.keys() & candidate_documents.keys())
+    document_deltas = [
+        {
+            "document_id": document_id,
+            "baseline": baseline_documents[document_id],
+            "candidate": candidate_documents[document_id],
+            "delta": candidate_documents[document_id] - baseline_documents[document_id],
+        }
+        for document_id in shared
+    ]
+    epsilon = 1e-12
+    improvements = sorted(document_deltas, key=lambda item: item["delta"], reverse=True)
+    regressions = sorted(document_deltas, key=lambda item: item["delta"])
+
+    result: dict[str, Any] = {
+        "baseline": baseline_scores,
+        "candidate": candidate_scores,
+        "deltas": metric_deltas,
+        "missing_predictions": {
+            "baseline": int(baseline.get("metrics", {}).get("missing_predictions", 0)),
+            "candidate": int(candidate.get("metrics", {}).get("missing_predictions", 0)),
+        },
+        "documents": {
+            "shared": len(shared),
+            "improved": sum(item["delta"] > epsilon for item in document_deltas),
+            "regressed": sum(item["delta"] < -epsilon for item in document_deltas),
+            "unchanged": sum(abs(item["delta"]) <= epsilon for item in document_deltas),
+            "largest_improvements": [
+                item for item in improvements if item["delta"] > epsilon
+            ][:top],
+            "largest_regressions": [
+                item for item in regressions if item["delta"] < -epsilon
+            ][:top],
+            "worst_regression": next(
+                (item for item in regressions if item["delta"] < -epsilon), None
+            ),
+        },
+    }
+    if reference is not None:
+        reference_scores = _scores(reference)
+        result["reference"] = reference_scores
+        result["candidate_vs_reference"] = {
+            key: candidate_scores[key] - reference_scores[key]
+            for key in SCORE_KEYS
+            if key in candidate_scores and key in reference_scores
+        }
+    return result
+
+
+def evaluate_gates(
+    comparison: dict[str, Any],
+    *,
+    min_overall_delta: float,
+    max_document_regression: float | None,
+    max_missing: int,
+    require_reference_lead: bool,
+) -> list[str]:
+    """Return human-readable gate failures; an empty list means pass."""
+    failures: list[str] = []
+    overall_delta = comparison["deltas"].get("overall_mean")
+    if overall_delta is None or overall_delta < min_overall_delta:
+        failures.append(
+            f"overall delta {overall_delta!r} is below {min_overall_delta:+.6f}"
+        )
+    candidate_missing = comparison["missing_predictions"]["candidate"]
+    if candidate_missing > max_missing:
+        failures.append(
+            f"candidate has {candidate_missing} missing predictions (maximum {max_missing})"
+        )
+    if max_document_regression is not None:
+        regression = comparison["documents"].get("worst_regression")
+        if regression is not None and regression["delta"] < -max_document_regression:
+            failures.append(
+                "largest document regression "
+                f"{regression['document_id']}={regression['delta']:+.6f} "
+                f"exceeds {-max_document_regression:+.6f}"
+            )
+    if require_reference_lead:
+        reference_delta = comparison.get("candidate_vs_reference", {}).get("overall_mean")
+        if reference_delta is None:
+            failures.append("reference overall score is unavailable")
+        elif reference_delta < 0.0:
+            failures.append(
+                f"candidate trails reference overall by {reference_delta!r}"
+            )
+    return failures
+
+
+def _run(command: list[str], *, cwd: Path, env: dict[str, str] | None = None) -> None:
+    print("+", " ".join(command), flush=True)
+    subprocess.run(command, cwd=cwd, env=env, check=True)
+
+
+def _run_engine(
+    *,
+    bench_dir: Path,
+    python: Path,
+    binary: Path,
+    label: str,
+    scratch_root: Path,
+) -> dict[str, Any]:
+    env = os.environ.copy()
+    env["PDF_INSPECTOR_BINARY"] = str(binary)
+    source = bench_dir / "prediction" / "pdf-inspector"
+    if source.exists():
+        if source.is_dir():
+            shutil.rmtree(source)
+        else:
+            source.unlink()
+    _run(
+        [
+            str(python),
+            "src/pdf_parser.py",
+            "--engine",
+            "pdf-inspector",
+            "--log-level",
+            "WARNING",
+        ],
+        cwd=bench_dir,
+        env=env,
+    )
+
+    if not source.is_dir():
+        raise RuntimeError(f"parser did not produce predictions: {source}")
+    destination = scratch_root / label
+    shutil.copytree(source, destination)
+    _run(
+        [
+            str(python),
+            "src/evaluator.py",
+            "--prediction-root",
+            str(scratch_root),
+            "--engine",
+            label,
+            "--log-level",
+            "WARNING",
+        ],
+        cwd=bench_dir,
+    )
+    with (destination / "evaluation.json").open(encoding="utf-8") as handle:
+        return json.load(handle)
+
+
+def _print_report(comparison: dict[str, Any]) -> None:
+    print("\nMetric                 baseline   candidate       delta")
+    print("--------------------  ----------  ----------  ----------")
+    for key in SCORE_KEYS:
+        if key not in comparison["deltas"]:
+            continue
+        print(
+            f"{key:<20}  {comparison['baseline'][key]:>10.6f}  "
+            f"{comparison['candidate'][key]:>10.6f}  "
+            f"{comparison['deltas'][key]:>+10.6f}"
+        )
+    if "reference" in comparison:
+        delta = comparison["candidate_vs_reference"].get("overall_mean")
+        reference = comparison["reference"].get("overall_mean")
+        reference_display = f"{reference:.6f}" if reference is not None else "n/a"
+        delta_display = f"{delta:+.6f}" if delta is not None else "n/a"
+        print(f"\nReference overall: {reference_display}; candidate delta: {delta_display}")
+
+    documents = comparison["documents"]
+    print(
+        "\nDocuments: "
+        f"{documents['improved']} improved, {documents['regressed']} regressed, "
+        f"{documents['unchanged']} unchanged ({documents['shared']} shared)"
+    )
+    for heading, key in (
+        ("Largest improvements", "largest_improvements"),
+        ("Largest regressions", "largest_regressions"),
+    ):
+        print(f"\n{heading}:")
+        rows = documents[key]
+        if not rows:
+            print("  none")
+        for row in rows:
+            print(
+                f"  {row['document_id']}: {row['delta']:+.6f} "
+                f"({row['baseline']:.6f} -> {row['candidate']:.6f})"
+            )
+
+
+def _arguments(argv: list[str] | None = None) -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--bench-dir", type=Path, required=True)
+    parser.add_argument("--baseline", type=Path, required=True)
+    parser.add_argument("--candidate", type=Path, required=True)
+    parser.add_argument("--python", type=Path)
+    parser.add_argument("--reference-evaluation", type=Path)
+    parser.add_argument("--json-output", type=Path)
+    parser.add_argument("--top", type=_non_negative_int, default=10)
+    parser.add_argument("--min-overall-delta", type=_finite_float, default=0.0)
+    parser.add_argument("--max-document-regression", type=_non_negative_float)
+    parser.add_argument("--max-missing", type=_non_negative_int, default=0)
+    parser.add_argument("--require-reference-lead", action="store_true")
+    return parser.parse_args(argv)
+
+
+def main(argv: list[str] | None = None) -> int:
+    args = _arguments(argv)
+    bench_dir = args.bench_dir.resolve()
+    baseline = args.baseline.resolve()
+    candidate = args.candidate.resolve()
+    # Keep the virtualenv launcher path intact. Resolving its symlink would
+    # invoke the underlying system interpreter without the benchmark's site
+    # packages.
+    python = (args.python or bench_dir / ".venv" / "bin" / "python").absolute()
+    for path, description in (
+        (bench_dir / "src" / "pdf_parser.py", "OpenDataLoader parser"),
+        (bench_dir / "src" / "evaluator.py", "OpenDataLoader evaluator"),
+        (baseline, "baseline binary"),
+        (candidate, "candidate binary"),
+        (python, "Python interpreter"),
+    ):
+        if not path.exists():
+            raise SystemExit(f"{description} not found: {path}")
+
+    with tempfile.TemporaryDirectory(prefix="pdf-inspector-opendataloader-") as temporary:
+        scratch_root = Path(temporary)
+        baseline_evaluation = _run_engine(
+            bench_dir=bench_dir,
+            python=python,
+            binary=baseline,
+            label="baseline",
+            scratch_root=scratch_root,
+        )
+        candidate_evaluation = _run_engine(
+            bench_dir=bench_dir,
+            python=python,
+            binary=candidate,
+            label="candidate",
+            scratch_root=scratch_root,
+        )
+
+        reference = None
+        if args.reference_evaluation is not None:
+            with args.reference_evaluation.resolve().open(encoding="utf-8") as handle:
+                reference = json.load(handle)
+
+        comparison = compare_evaluations(
+            baseline_evaluation,
+            candidate_evaluation,
+            reference,
+            top=args.top,
+        )
+
+    _print_report(comparison)
+    if args.json_output is not None:
+        args.json_output.resolve().write_text(
+            json.dumps(comparison, indent=2) + "\n", encoding="utf-8"
+        )
+
+    failures = evaluate_gates(
+        comparison,
+        min_overall_delta=args.min_overall_delta,
+        max_document_regression=args.max_document_regression,
+        max_missing=args.max_missing,
+        require_reference_lead=args.require_reference_lead,
+    )
+    if failures:
+        print("\nBenchmark gate failed:", file=sys.stderr)
+        for failure in failures:
+            print(f"  - {failure}", file=sys.stderr)
+        return 1
+    print("\nBenchmark gate passed.")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

scripts/probe_backend_evidence.py

@@ -0,0 +1,351 @@
+#!/usr/bin/env python3
+"""Compare pdf-inspector evidence with optional MuPDF structured text.
+
+This is an experiment and diagnostic tool, not an extraction fallback. It runs
+MuPDF's deterministic ``stext.json`` backend without OCR and highlights pages
+where that backend exposes materially different text or layout evidence.
+"""
+
+from __future__ import annotations
+
+import argparse
+from collections import Counter
+import json
+from pathlib import Path
+import re
+import shutil
+import subprocess
+import sys
+from typing import Any, Iterable
+
+
+TOKEN_PATTERN = re.compile(r"[^\W_]+(?:[\u2019'][^\W_]+)*", re.UNICODE)
+
+
+def _tokens(texts: Iterable[str]) -> Counter[str]:
+    tokens: Counter[str] = Counter()
+    for text in texts:
+        for token in TOKEN_PATTERN.findall(text.casefold()):
+            # Lone letters are frequently bullets, chart labels, or fragmented
+            # glyphs. Digits remain useful even when they are one character.
+            if len(token) > 1 or token.isdigit():
+                tokens[token] += 1
+    return tokens
+
+
+def _repeated_x_anchors(xs: Iterable[float], *, tolerance: float = 4.0) -> int:
+    buckets = Counter(round(float(x) / tolerance) for x in xs)
+    return sum(count >= 3 for count in buckets.values())
+
+
+def local_pages(payload: dict[str, Any]) -> dict[int, dict[str, Any]]:
+    """Summarize positioned ``pdf2md --items-json`` evidence by page."""
+    pages: dict[int, dict[str, Any]] = {}
+    for item in payload.get("items", []):
+        page_number = int(item["page"])
+        page = pages.setdefault(
+            page_number,
+            {"texts": [], "xs": [], "text_items": 0, "image_items": 0},
+        )
+        if item.get("item_type") == "image":
+            page["image_items"] += 1
+            continue
+        text = str(item.get("text", ""))
+        if text.strip():
+            page["texts"].append(text)
+            page["xs"].append(float(item.get("x", 0.0)))
+            page["text_items"] += 1
+    return pages
+
+
+def alternate_pages(
+    payload: dict[str, Any] | list[dict[str, Any]],
+) -> dict[int, dict[str, Any]]:
+    """Summarize MuPDF ``stext.json`` evidence by page."""
+    pages: dict[int, dict[str, Any]] = {}
+    raw_pages = payload if isinstance(payload, list) else payload.get("pages", [])
+    for index, raw_page in enumerate(raw_pages, start=1):
+        page_number = int(raw_page.get("number", index))
+        page = {
+            "texts": [],
+            "xs": [],
+            "text_blocks": 0,
+            "text_lines": 0,
+            "image_blocks": 0,
+        }
+        for block in raw_page.get("blocks", []):
+            if block.get("type") == "image":
+                page["image_blocks"] += 1
+                continue
+            if block.get("type") != "text":
+                continue
+            page["text_blocks"] += 1
+            for line in block.get("lines", []):
+                text = str(line.get("text", ""))
+                if text.strip():
+                    page["texts"].append(text)
+                    bbox = line.get("bbox", {})
+                    page["xs"].append(float(bbox.get("x", line.get("x", 0.0))))
+                    page["text_lines"] += 1
+        pages[page_number] = page
+    return pages
+
+
+def compare_page(
+    local: dict[str, Any],
+    alternate: dict[str, Any],
+    *,
+    min_token_gain: int,
+    min_alternate_only_ratio: float,
+    min_anchor_gain: int,
+) -> dict[str, Any]:
+    """Compare semantic and coarse layout evidence for one page."""
+    local_tokens = _tokens(local.get("texts", []))
+    alternate_tokens = _tokens(alternate.get("texts", []))
+    shared = local_tokens & alternate_tokens
+    alternate_only = alternate_tokens - local_tokens
+    local_only = local_tokens - alternate_tokens
+    local_total = sum(local_tokens.values())
+    alternate_total = sum(alternate_tokens.values())
+    shared_total = sum(shared.values())
+    alternate_only_total = sum(alternate_only.values())
+    local_only_total = sum(local_only.values())
+    net_token_gain = alternate_total - local_total
+    alternate_only_ratio = alternate_only_total / max(alternate_total, 1)
+
+    local_anchors = _repeated_x_anchors(local.get("xs", []))
+    alternate_anchors = _repeated_x_anchors(alternate.get("xs", []))
+    anchor_gain = alternate_anchors - local_anchors
+    image_gain = int(alternate.get("image_blocks", 0)) - int(
+        local.get("image_items", 0)
+    )
+
+    reasons: list[str] = []
+    if local_total == 0 and alternate_total >= max(5, min_token_gain // 2):
+        reasons.append("local_text_empty")
+    elif (
+        net_token_gain >= min_token_gain
+        and alternate_only_ratio >= min_alternate_only_ratio
+    ):
+        reasons.append("alternate_has_more_text")
+    if anchor_gain >= min_anchor_gain:
+        reasons.append("alternate_has_more_alignment_anchors")
+    if image_gain > 0:
+        reasons.append("alternate_has_more_image_blocks")
+
+    if reasons:
+        classification = "investigate_alternate_evidence"
+    elif local_total - alternate_total >= min_token_gain:
+        classification = "local_has_more_text"
+    elif alternate_only_total + local_only_total:
+        classification = "different_segmentation_or_decoding"
+    else:
+        classification = "equivalent_text_evidence"
+
+    return {
+        "classification": classification,
+        "reasons": reasons,
+        "tokens": {
+            "local": local_total,
+            "alternate": alternate_total,
+            "shared": shared_total,
+            "net_alternate_gain": net_token_gain,
+            "alternate_only": alternate_only_total,
+            "local_only": local_only_total,
+            "alternate_only_ratio": alternate_only_ratio,
+            "alternate_only_sample": sorted(alternate_only)[:12],
+            "local_only_sample": sorted(local_only)[:12],
+        },
+        "layout": {
+            "local_text_items": int(local.get("text_items", 0)),
+            "local_image_items": int(local.get("image_items", 0)),
+            "local_repeated_x_anchors": local_anchors,
+            "alternate_text_blocks": int(alternate.get("text_blocks", 0)),
+            "alternate_text_lines": int(alternate.get("text_lines", 0)),
+            "alternate_image_blocks": int(alternate.get("image_blocks", 0)),
+            "alternate_repeated_x_anchors": alternate_anchors,
+        },
+    }
+
+
+def compare_documents(
+    local_payload: dict[str, Any],
+    alternate_payload: dict[str, Any] | list[dict[str, Any]],
+    *,
+    min_token_gain: int = 20,
+    min_alternate_only_ratio: float = 0.15,
+    min_anchor_gain: int = 2,
+) -> dict[str, Any]:
+    """Return a page-level evidence report for already extracted payloads."""
+    local = local_pages(local_payload)
+    alternate = alternate_pages(alternate_payload)
+    page_numbers = sorted(local.keys() | alternate.keys())
+    pages = []
+    for page_number in page_numbers:
+        result = compare_page(
+            local.get(page_number, {}),
+            alternate.get(page_number, {}),
+            min_token_gain=min_token_gain,
+            min_alternate_only_ratio=min_alternate_only_ratio,
+            min_anchor_gain=min_anchor_gain,
+        )
+        result["page"] = page_number
+        pages.append(result)
+
+    flagged = [
+        page
+        for page in pages
+        if page["classification"] == "investigate_alternate_evidence"
+    ]
+    return {
+        "summary": {
+            "pages": len(pages),
+            "flagged_pages": len(flagged),
+            "flagged_page_numbers": [page["page"] for page in flagged],
+            "local_tokens": sum(page["tokens"]["local"] for page in pages),
+            "alternate_tokens": sum(page["tokens"]["alternate"] for page in pages),
+            "alternate_only_tokens": sum(
+                page["tokens"]["alternate_only"] for page in pages
+            ),
+        },
+        "pages": pages,
+    }
+
+
+def _json_command(command: list[str]) -> Any:
+    try:
+        completed = subprocess.run(
+            command,
+            check=True,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.PIPE,
+            text=True,
+        )
+    except subprocess.CalledProcessError as error:
+        detail = error.stderr.strip() or error.stdout.strip() or "no diagnostic output"
+        raise RuntimeError(f"command failed: {' '.join(command)}\n{detail}") from error
+    try:
+        return json.loads(completed.stdout)
+    except json.JSONDecodeError as error:
+        raise RuntimeError(
+            f"command did not return JSON: {' '.join(command)}: {error}"
+        ) from error
+
+
+def probe_pdf(
+    pdf: Path,
+    *,
+    pdf2md: Path,
+    mutool: Path,
+    min_token_gain: int,
+    min_alternate_only_ratio: float,
+    min_anchor_gain: int,
+) -> dict[str, Any]:
+    local_payload = _json_command([str(pdf2md), str(pdf), "--items-json"])
+    # `stext.json` is MuPDF's native structured text output. The OCR formats
+    # are intentionally not used so this remains a deterministic no-model
+    # comparison.
+    alternate_payload = _json_command(
+        [str(mutool), "draw", "-q", "-F", "stext.json", "-o", "-", str(pdf)]
+    )
+    report = compare_documents(
+        local_payload,
+        alternate_payload,
+        min_token_gain=min_token_gain,
+        min_alternate_only_ratio=min_alternate_only_ratio,
+        min_anchor_gain=min_anchor_gain,
+    )
+    report["pdf"] = str(pdf)
+    return report
+
+
+def _arguments(argv: list[str] | None = None) -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("pdf", type=Path, nargs="+")
+    parser.add_argument("--pdf2md", type=Path, default=Path("target/release/pdf2md"))
+    parser.add_argument("--mutool", type=Path)
+    parser.add_argument("--json-output", type=Path)
+    parser.add_argument("--min-token-gain", type=int, default=20)
+    parser.add_argument("--min-alternate-only-ratio", type=float, default=0.15)
+    parser.add_argument("--min-anchor-gain", type=int, default=2)
+    return parser.parse_args(argv)
+
+
+def _print_report(result: dict[str, Any]) -> None:
+    summary = result["summary"]
+    print(f"\n{result['pdf']}")
+    print(
+        f"  {summary['flagged_pages']}/{summary['pages']} pages flagged; "
+        f"tokens local={summary['local_tokens']} alternate={summary['alternate_tokens']} "
+        f"alternate-only={summary['alternate_only_tokens']}"
+    )
+    for page in result["pages"]:
+        if page["classification"] != "investigate_alternate_evidence":
+            continue
+        reasons = ", ".join(page["reasons"])
+        tokens = page["tokens"]
+        print(
+            f"  page {page['page']}: {reasons}; "
+            f"net tokens={tokens['net_alternate_gain']:+d}, "
+            f"alternate-only={tokens['alternate_only']}"
+        )
+
+
+def main(argv: list[str] | None = None) -> int:
+    args = _arguments(argv)
+    pdf2md = args.pdf2md.absolute()
+    mutool = args.mutool or (Path(found) if (found := shutil.which("mutool")) else None)
+    if not pdf2md.is_file():
+        print(f"error: pdf2md binary not found: {pdf2md}", file=sys.stderr)
+        return 2
+    if mutool is None or not mutool.is_file():
+        print("error: mutool not found; install MuPDF or pass --mutool", file=sys.stderr)
+        return 2
+    if (
+        args.min_token_gain < 0
+        or args.min_anchor_gain < 0
+        or not 0.0 <= args.min_alternate_only_ratio <= 1.0
+    ):
+        print("error: thresholds must be non-negative and ratio must be in [0, 1]", file=sys.stderr)
+        return 2
+
+    results = []
+    for pdf in args.pdf:
+        path = pdf.absolute()
+        if not path.is_file():
+            print(f"error: PDF not found: {path}", file=sys.stderr)
+            return 2
+        try:
+            result = probe_pdf(
+                path,
+                pdf2md=pdf2md,
+                mutool=mutool,
+                min_token_gain=args.min_token_gain,
+                min_alternate_only_ratio=args.min_alternate_only_ratio,
+                min_anchor_gain=args.min_anchor_gain,
+            )
+        except RuntimeError as error:
+            print(f"error: {error}", file=sys.stderr)
+            return 1
+        results.append(result)
+        _print_report(result)
+
+    payload = {
+        "schema_version": 1,
+        "experiment": "optional_mupdf_stext_evidence",
+        "ocr": False,
+        "thresholds": {
+            "min_token_gain": args.min_token_gain,
+            "min_alternate_only_ratio": args.min_alternate_only_ratio,
+            "min_anchor_gain": args.min_anchor_gain,
+        },
+        "documents": results,
+    }
+    if args.json_output:
+        args.json_output.parent.mkdir(parents=True, exist_ok=True)
+        args.json_output.write_text(json.dumps(payload, indent=2) + "\n", encoding="utf-8")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

scripts/tests/test_bench_opendataloader.py

@@ -0,0 +1,203 @@
+import io
+import json
+import sys
+import tempfile
+import unittest
+from contextlib import redirect_stderr, redirect_stdout
+from pathlib import Path
+from unittest.mock import patch
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
+
+from bench_opendataloader import (
+    _arguments,
+    _print_report,
+    _run_engine,
+    compare_evaluations,
+    evaluate_gates,
+)
+
+
+def evaluation(overall, documents, *, missing=0):
+    return {
+        "metrics": {
+            "score": {
+                "overall_mean": overall,
+                "nid_mean": overall + 0.01,
+            },
+            "missing_predictions": missing,
+        },
+        "documents": [
+            {
+                "document_id": document_id,
+                "scores": {"overall": score},
+            }
+            for document_id, score in documents.items()
+        ],
+    }
+
+
+class ComparisonTests(unittest.TestCase):
+    def test_reports_metric_and_document_deltas(self):
+        baseline = evaluation(0.80, {"a": 0.8, "b": 0.6, "c": 0.7})
+        candidate = evaluation(0.82, {"a": 0.9, "b": 0.5, "c": 0.7})
+
+        result = compare_evaluations(baseline, candidate, top=1)
+
+        self.assertAlmostEqual(result["deltas"]["overall_mean"], 0.02)
+        self.assertEqual(result["documents"]["improved"], 1)
+        self.assertEqual(result["documents"]["regressed"], 1)
+        self.assertEqual(result["documents"]["unchanged"], 1)
+        self.assertEqual(
+            result["documents"]["largest_improvements"][0]["document_id"], "a"
+        )
+        self.assertEqual(
+            result["documents"]["largest_regressions"][0]["document_id"], "b"
+        )
+
+    def test_reference_delta_is_reported(self):
+        baseline = evaluation(0.80, {})
+        candidate = evaluation(0.82, {})
+        reference = evaluation(0.81, {})
+
+        result = compare_evaluations(baseline, candidate, reference)
+
+        self.assertAlmostEqual(
+            result["candidate_vs_reference"]["overall_mean"], 0.01
+        )
+
+    def test_gates_cover_aggregate_document_missing_and_reference(self):
+        comparison = compare_evaluations(
+            evaluation(0.80, {"a": 0.8}),
+            evaluation(0.79, {"a": 0.7}, missing=1),
+            evaluation(0.81, {}),
+        )
+
+        failures = evaluate_gates(
+            comparison,
+            min_overall_delta=0.0,
+            max_document_regression=0.05,
+            max_missing=0,
+            require_reference_lead=True,
+        )
+
+        self.assertEqual(len(failures), 4)
+
+    def test_regression_gate_is_independent_of_report_limit(self):
+        comparison = compare_evaluations(
+            evaluation(0.80, {"a": 0.8}),
+            evaluation(0.80, {"a": 0.7}),
+            top=0,
+        )
+
+        failures = evaluate_gates(
+            comparison,
+            min_overall_delta=0.0,
+            max_document_regression=0.05,
+            max_missing=0,
+            require_reference_lead=False,
+        )
+
+        self.assertEqual(len(failures), 1)
+        self.assertIn("largest document regression", failures[0])
+
+    def test_report_handles_reference_without_overall_score(self):
+        result = compare_evaluations(
+            evaluation(0.80, {}),
+            evaluation(0.82, {}),
+            {"metrics": {"score": {"nid_mean": 0.81}}},
+        )
+
+        output = io.StringIO()
+        with redirect_stdout(output):
+            _print_report(result)
+
+        self.assertIn("Reference overall: n/a; candidate delta: n/a", output.getvalue())
+
+    def test_reference_gate_reports_missing_score_as_unavailable(self):
+        comparison = compare_evaluations(
+            evaluation(0.80, {}),
+            evaluation(0.82, {}),
+        )
+
+        failures = evaluate_gates(
+            comparison,
+            min_overall_delta=0.0,
+            max_document_regression=None,
+            max_missing=0,
+            require_reference_lead=True,
+        )
+
+        self.assertEqual(failures, ["reference overall score is unavailable"])
+
+    def test_arguments_reject_negative_counts_and_allow_zero_top(self):
+        required = [
+            "--bench-dir",
+            ".",
+            "--baseline",
+            "baseline",
+            "--candidate",
+            "candidate",
+        ]
+        self.assertEqual(_arguments(required + ["--top", "0"]).top, 0)
+        for option in ("--top", "--max-document-regression", "--max-missing"):
+            with self.subTest(option=option), redirect_stderr(io.StringIO()):
+                with self.assertRaises(SystemExit):
+                    _arguments(required + [option, "-1"])
+
+    def test_arguments_reject_nonfinite_float_thresholds(self):
+        required = [
+            "--bench-dir",
+            ".",
+            "--baseline",
+            "baseline",
+            "--candidate",
+            "candidate",
+        ]
+        for option in ("--min-overall-delta", "--max-document-regression"):
+            for value in ("nan", "inf", "-inf"):
+                with self.subTest(option=option, value=value), redirect_stderr(
+                    io.StringIO()
+                ):
+                    with self.assertRaises(SystemExit):
+                        _arguments(required + [option, value])
+
+    def test_run_engine_clears_stale_predictions_before_parser(self):
+        with tempfile.TemporaryDirectory() as temporary:
+            root = Path(temporary)
+            bench_dir = root / "bench"
+            source = bench_dir / "prediction" / "pdf-inspector"
+            source.mkdir(parents=True)
+            (source / "stale.md").write_text("stale", encoding="utf-8")
+            scratch = root / "scratch"
+            scratch.mkdir()
+
+            def fake_run(command, *, cwd, env=None):
+                if any(part.endswith("pdf_parser.py") for part in command):
+                    self.assertFalse(source.exists())
+                    (source / "markdown").mkdir(parents=True)
+                    (source / "markdown" / "new.md").write_text(
+                        "new", encoding="utf-8"
+                    )
+                else:
+                    destination = scratch / "candidate"
+                    (destination / "evaluation.json").write_text(
+                        json.dumps(evaluation(0.82, {})), encoding="utf-8"
+                    )
+
+            with patch("bench_opendataloader._run", side_effect=fake_run):
+                result = _run_engine(
+                    bench_dir=bench_dir,
+                    python=Path("python"),
+                    binary=Path("pdf2md"),
+                    label="candidate",
+                    scratch_root=scratch,
+                )
+
+            self.assertEqual(result["metrics"]["score"]["overall_mean"], 0.82)
+            self.assertFalse((source / "stale.md").exists())
+            self.assertFalse((scratch / "candidate" / "stale.md").exists())
+
+
+if __name__ == "__main__":
+    unittest.main()

scripts/tests/test_probe_backend_evidence.py

@@ -0,0 +1,103 @@
+import sys
+import unittest
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
+
+from probe_backend_evidence import compare_documents
+
+
+def local_payload(items):
+    return {"items": items}
+
+
+def item(page, text, x=10, item_type="text"):
+    return {"page": page, "text": text, "x": x, "item_type": item_type}
+
+
+def alternate_payload(pages):
+    return {"pages": pages}
+
+
+def page(lines, *, images=0):
+    blocks = [
+        {
+            "type": "text",
+            "lines": [
+                {"text": text, "bbox": {"x": x, "y": index * 10, "w": 80, "h": 8}}
+                for index, (text, x) in enumerate(lines)
+            ],
+        }
+    ]
+    blocks.extend({"type": "image"} for _ in range(images))
+    return {"blocks": blocks}
+
+
+class EvidenceComparisonTests(unittest.TestCase):
+    def test_accepts_real_top_level_page_array(self):
+        local = local_payload([item(1, "alpha beta")])
+        alternate = [page([("alpha beta gamma", 10)])]
+
+        result = compare_documents(local, alternate)["pages"][0]
+
+        self.assertEqual(result["tokens"]["alternate"], 3)
+        self.assertEqual(result["tokens"]["net_alternate_gain"], 1)
+
+    def test_flags_material_alternate_text_gain(self):
+        local = local_payload([item(1, "alpha beta")])
+        alternate = alternate_payload(
+            [page([("alpha beta gamma delta epsilon zeta", 10)])]
+        )
+
+        report = compare_documents(
+            local,
+            alternate,
+            min_token_gain=3,
+            min_alternate_only_ratio=0.2,
+        )
+
+        result = report["pages"][0]
+        self.assertEqual(result["classification"], "investigate_alternate_evidence")
+        self.assertIn("alternate_has_more_text", result["reasons"])
+        self.assertEqual(result["tokens"]["net_alternate_gain"], 4)
+
+    def test_repeated_alignment_and_image_evidence_are_reported(self):
+        local = local_payload([item(1, "one two", 10)])
+        alternate = alternate_payload(
+            [
+                page(
+                    [
+                        ("one two", 10),
+                        ("row three", 100),
+                        ("row four", 100),
+                        ("row five", 100),
+                    ],
+                    images=1,
+                )
+            ]
+        )
+
+        result = compare_documents(
+            local,
+            alternate,
+            min_token_gain=99,
+            min_anchor_gain=1,
+        )["pages"][0]
+
+        self.assertIn("alternate_has_more_alignment_anchors", result["reasons"])
+        self.assertIn("alternate_has_more_image_blocks", result["reasons"])
+        self.assertEqual(result["layout"]["alternate_repeated_x_anchors"], 1)
+
+    def test_token_segmentation_difference_does_not_imply_more_evidence(self):
+        local = local_payload([item(1, "Revenue 2025")])
+        alternate = alternate_payload([page([("Revenue 2024", 10)])])
+
+        result = compare_documents(local, alternate, min_token_gain=2)["pages"][0]
+
+        self.assertEqual(result["classification"], "different_segmentation_or_decoding")
+        self.assertEqual(result["reasons"], [])
+        self.assertEqual(result["tokens"]["alternate_only_sample"], ["2024"])
+
+
+if __name__ == "__main__":
+    unittest.main()

src/bin/detect_pdf.rs

@@ -63,6 +63,7 @@ fn json_escape(s: &str) -> String {
 }
 
 fn main() {
+    #[cfg(not(target_arch = "wasm32"))]
     env_logger::init();
     let args: Vec<String> = env::args().collect();
 

src/bin/pdf2md.rs

@@ -2,8 +2,8 @@
 
 use pdf_inspector::extractor::ItemType;
 use pdf_inspector::{
-    extract_text_with_positions_pages, process_pdf_with_options, LayoutComplexity, PdfOptions,
-    PdfType, ProcessMode, TextItem,
+    extract_text_with_positions_pages_with_password, process_pdf_with_options, LayoutComplexity,
+    PdfOptions, PdfType, ProcessMode, TextItem,
 };
 use std::collections::HashSet;
 use std::env;
@@ -103,9 +103,18 @@ fn format_items_json(items: &[TextItem]) -> String {
     )
 }
 
+fn extract_items_json(
+    pdf_path: &str,
+    page_filter: Option<&HashSet<u32>>,
+    password: Option<&str>,
+) -> Result<String, pdf_inspector::PdfError> {
+    extract_text_with_positions_pages_with_password(pdf_path, page_filter, password)
+        .map(|items| format_items_json(&items))
+}
+
 #[cfg(test)]
 mod tests {
-    use super::format_items_json;
+    use super::{extract_items_json, format_items_json};
     use pdf_inspector::extractor::ItemType;
     use pdf_inspector::TextItem;
 
@@ -137,6 +146,24 @@ mod tests {
         assert!(json.contains(r#""item_type":"text""#));
         assert!(json.contains(r#""mcid":7"#));
     }
+
+    #[test]
+    fn items_json_uses_supplied_pdf_password() {
+        let path = "tests/fixtures/encrypted-secret123.pdf";
+
+        let without_password = extract_items_json(path, None, None);
+        assert!(
+            without_password.is_err(),
+            "encrypted fixture unexpectedly extracted without a password"
+        );
+
+        let json = extract_items_json(path, None, Some("secret123"))
+            .expect("correct password should decrypt positioned text");
+        assert!(
+            json.contains("Procurement"),
+            "decrypted item JSON should contain fixture text, got {json}"
+        );
+    }
 }
 
 /// Parse a page specification like "1,3,5-10,20" into a HashSet of page numbers.
@@ -190,6 +217,7 @@ fn print_layout_info(layout: &LayoutComplexity) {
 }
 
 fn main() {
+    #[cfg(not(target_arch = "wasm32"))]
     env_logger::init();
     let args: Vec<String> = env::args().collect();
 
@@ -206,6 +234,9 @@ fn main() {
         eprintln!("  --json              Output result as JSON");
         eprintln!("  --items-json        Output positioned TextItem JSON");
         eprintln!("  --raw               Output only markdown (no headers)");
+        eprintln!(
+            "  --compact           Collapse token-heavy source formatting such as dot leaders"
+        );
         eprintln!("  --pages             Insert page break markers (<!-- Page N -->)");
         eprintln!("  --select-pages N    Only process specified pages (e.g. 1,3,5-10)");
         eprintln!("  --password PW       Password for an encrypted PDF");
@@ -218,6 +249,7 @@ fn main() {
     let json_output = args.iter().any(|a| a == "--json");
     let items_json_output = args.iter().any(|a| a == "--items-json");
     let raw_output = args.iter().any(|a| a == "--raw");
+    let compact_output = args.iter().any(|a| a == "--compact");
     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");
@@ -252,8 +284,8 @@ fn main() {
         });
 
     if items_json_output {
-        match extract_text_with_positions_pages(pdf_path, page_filter.as_ref()) {
-            Ok(items) => println!("{}", format_items_json(&items)),
+        match extract_items_json(pdf_path, page_filter.as_ref(), password.as_deref()) {
+            Ok(json) => println!("{}", json),
             Err(e) => {
                 println!(r#"{{"error":"{}"}}"#, json_escape(&e.to_string()));
                 process::exit(1);
@@ -276,6 +308,9 @@ fn main() {
     };
 
     let mut options = PdfOptions::new().mode(process_mode);
+    if compact_output {
+        options.markdown.profile = pdf_inspector::MarkdownProfile::Compact;
+    }
     options.markdown.include_page_numbers = page_numbers;
     if let Some(pages) = page_filter {
         options.page_filter = Some(pages);

src/detector.rs

@@ -1659,7 +1659,15 @@ fn hex_val(b: u8) -> Option<u8> {
 /// Standard page: 612x792 points (US Letter) = ~485,000 sq points
 /// At 2x resolution that's ~1.9M pixels, so we use 250K pixels as threshold
 /// (accounting for varying DPI and page sizes)
-fn analyze_page_images(doc: &Document, page_id: ObjectId) -> (bool, u64, bool) {
+/// Returns `(has_images, total_image_area, has_template_image)` for a page.
+/// `has_template_image` means a single large (>50% page coverage)
+/// background image — the signal `classify_pdf`/`detect_pdf_type` uses to
+/// route a page to OCR regardless of any incidental native text drawn over
+/// it. Exposed at crate visibility so extraction-side per-page `needs_ocr`
+/// computation (`extract_pages_markdown_mem`) can consult the same signal
+/// instead of maintaining its own, independent notion of "needs OCR" that
+/// can silently disagree with detection — see #227.
+pub(crate) fn analyze_page_images(doc: &Document, page_id: ObjectId) -> (bool, u64, bool) {
     // Threshold: image covering roughly half a page at 150+ DPI
     // 612 * 792 / 2 * (150/72)^2 ≈ 1M pixels, but we'll be conservative
     const TEMPLATE_IMAGE_THRESHOLD: u64 = 500_000; // 500K pixels
@@ -1741,6 +1749,62 @@ fn analyze_page_images(doc: &Document, page_id: ObjectId) -> (bool, u64, bool) {
     (has_images, total_area, has_template_image)
 }
 
+/// Computes both `(needs_ocr_for_template_image, has_vector_text)` for a
+/// page from a single shared `analyze_page_content` pass — that call
+/// decompresses and scans every content stream (page + XObjects) plus
+/// image coverage, so `extract_pages_markdown_mem` must not invoke it
+/// twice per page (once per signal) the way `detect_from_document` avoids
+/// by caching its per-page `PageAnalysis`.
+///
+/// `needs_ocr_for_template_image` is true when a page's template image
+/// should be treated as a scan needing OCR — a single full-page background
+/// image with little/no real text — rather than a text page that happens
+/// to carry a watermark, letterhead, or figure. Mirrors the two distinct
+/// signals classification uses to route a template-image page to OCR:
+///
+/// 1. `looks_like_scan`: image_count <= 1, few text operators (<50), and
+///    low alphanumeric diversity in raw string operands (unless decodable
+///    CID/ToUnicode fonts explain that away) — the gate used for
+///    `pages_with_template_images` and Mixed-type per-page routing.
+/// 2. Insufficient real text volume, using `DetectionConfig::default()`'s
+///    `min_text_ops_per_page` (3) — the same threshold Mixed-type per-page
+///    routing applies via `text_operator_count < config.min_text_ops_per_page
+///    && has_images` (simplified here since a template image implies
+///    `has_images`). Deliberately *not* the higher `effective_min_ops`
+///    floor (`min_text_ops_per_page.max(10)`) that whole-document
+///    `PdfType::ImageBased`/`Scanned` classification uses for
+///    `pages_with_text` — that's a cross-page aggregate decision this
+///    per-page function has no way to replicate exactly, and the lower
+///    per-page threshold is the one a single page's own signals can
+///    actually agree with.
+///
+/// `has_vector_text` is true when a page has vector-outlined text (glyphs
+/// drawn as paths rather than shown via text-showing operators) —
+/// `detect_from_document`'s Mixed-type per-page routing always sends
+/// these pages to OCR, independent of any template-image check, since
+/// outlined glyphs can't be extracted as text at all.
+///
+/// Exposed at crate visibility so `extract_pages_markdown_mem` can apply
+/// the same gates classification needs elsewhere instead of treating the
+/// raw signals alone as sufficient — see #227/#231.
+pub(crate) fn page_ocr_signals(doc: &Document, page_id: ObjectId) -> (bool, bool) {
+    let analysis = analyze_page_content(doc, page_id);
+
+    let needs_ocr_for_template_image = if !analysis.has_template_image {
+        false
+    } else {
+        let alphanum_low = analysis.unique_alphanum_chars < 10
+            && !(analysis.has_decodable_text_fonts && analysis.text_operator_count >= 10);
+        let looks_like_scan =
+            analysis.image_count <= 1 && analysis.text_operator_count < 50 && alphanum_low;
+        let insufficient_text =
+            analysis.text_operator_count < DetectionConfig::default().min_text_ops_per_page;
+        looks_like_scan || insufficient_text
+    };
+
+    (needs_ocr_for_template_image, analysis.has_vector_text)
+}
+
 /// Recursively collect image dimensions from XObject resources,
 /// including images nested inside Form XObjects.
 fn collect_images_from_resources(

src/extractor/base14.rs

@@ -0,0 +1,649 @@
+//! Built-in glyph metrics for the 14 standard PDF fonts.
+//!
+//! PDFs may omit `/Widths` for non-embedded base-14 fonts (Times, Helvetica,
+//! Courier, Symbol, ZapfDingbats); per the PDF spec the reader must supply
+//! the metrics. Without them every text item gets width 0, which breaks
+//! space synthesis, sub/superscript detection, and table column detection
+//! (common in 1990s dvips/Distiller output).
+//!
+//! Tables are generated from the Adobe Core 14 AFM files (via reportlab's
+//! `_fontdata`), keyed by Unicode char, sorted for binary search.
+//! Generator: scratchpad/gen_base14.py (session tooling, not checked in).
+
+/// Width in 1000ths of an em for `c` in the given base-14 font, or `None`
+/// if the font is not one of the base 14 (after name normalization) or the
+/// char has no glyph in its AFM.
+pub(crate) fn base14_char_width(base_font: &str, c: char) -> Option<u16> {
+    let table = base14_table(base_font)?;
+    // AFM tables key visible glyphs only; alias the invisible variants the
+    // cp1252 fallback can produce so they get the metric of their visible
+    // counterpart instead of the generic default.
+    let c = match c {
+        '\u{00A0}' => ' ', // no-break space -> space
+        '\u{00AD}' => '-', // soft hyphen -> hyphen
+        _ => c,
+    };
+    table
+        .binary_search_by_key(&c, |&(ch, _)| ch)
+        .ok()
+        .map(|i| table[i].1)
+}
+
+/// True when the base font name normalizes to one of the standard 14 fonts.
+pub(crate) fn is_base14_font(base_font: &str) -> bool {
+    base14_table(base_font).is_some()
+}
+
+/// Code → Unicode through the font's BUILT-IN encoding, for the base-14
+/// fonts whose repertoire is not Latin (Symbol, ZapfDingbats). Their glyphs
+/// live at byte positions that have nothing to do with cp1252 (Symbol 0x61
+/// renders α, Zapf 0x21 renders ✁), so advance widths must be resolved
+/// through this mapping — the renderer draws these glyphs regardless of how
+/// the text decoder transliterates them. Returns `None` for the Latin text
+/// fonts, which follow standard single-byte encodings.
+pub(crate) fn builtin_encoding_char(base_font: &str, code: u8) -> Option<char> {
+    let table = base14_table(base_font)?;
+    let enc: &[(u8, char)] = if std::ptr::eq(table, SYMBOL) {
+        SYMBOL_ENCODING
+    } else if std::ptr::eq(table, ZAPFDINGBATS) {
+        ZAPFDINGBATS_ENCODING
+    } else {
+        return None;
+    };
+    enc.binary_search_by_key(&code, |&(b, _)| b)
+        .ok()
+        .map(|i| enc[i].1)
+}
+
+/// Map a BaseFont name (possibly subset-prefixed, e.g. "ABCDEF+Times-Bold",
+/// or a common alias like "Arial" / "TimesNewRomanPSMT") to its width table.
+fn base14_table(base_font: &str) -> Option<&'static [(char, u16)]> {
+    // Strip subset prefix "ABCDEF+"
+    let name = match base_font.split_once('+') {
+        Some((prefix, rest))
+            if prefix.len() == 6 && prefix.chars().all(|c| c.is_ascii_uppercase()) =>
+        {
+            rest
+        }
+        _ => base_font,
+    };
+    let lower = name.to_ascii_lowercase();
+    let bold = lower.contains("bold");
+    let italic = lower.contains("italic") || lower.contains("oblique");
+    if lower.contains("courier") {
+        return Some(match (bold, italic) {
+            (false, false) => COURIER,
+            (true, false) => COURIER_BOLD,
+            (false, true) => COURIER_OBLIQUE,
+            (true, true) => COURIER_BOLDOBLIQUE,
+        });
+    }
+    if lower.contains("helvetica") || lower.contains("arial") {
+        return Some(match (bold, italic) {
+            (false, false) => HELVETICA,
+            (true, false) => HELVETICA_BOLD,
+            (false, true) => HELVETICA_OBLIQUE,
+            (true, true) => HELVETICA_BOLDOBLIQUE,
+        });
+    }
+    if lower.contains("times") {
+        return Some(match (bold, italic) {
+            (false, false) => TIMES_ROMAN,
+            (true, false) => TIMES_BOLD,
+            (false, true) => TIMES_ITALIC,
+            (true, true) => TIMES_BOLDITALIC,
+        });
+    }
+    // Symbol and ZapfDingbats have unique glyph repertoires, so only exact
+    // names (plus the common MT/ITC aliases) qualify — a custom font that
+    // merely mentions "Symbol" in its name must not get these metrics.
+    match lower.as_str() {
+        "zapfdingbats" | "dingbats" | "itczapfdingbats" | "zapfdingbatsitc" => {
+            return Some(ZAPFDINGBATS)
+        }
+        "symbol" | "symbolmt" | "symbolitc" => return Some(SYMBOL),
+        _ => {}
+    }
+    None
+}
+
+#[rustfmt::skip]
+static COURIER: &[(char, u16)] = &[
+    (' ', 600), ('!', 600), ('"', 600), ('#', 600), ('$', 600), ('%', 600),
+    ('&', 600), ('\'', 600), ('(', 600), (')', 600), ('*', 600), ('+', 600),
+    (',', 600), ('-', 600), ('.', 600), ('/', 600), ('0', 600), ('1', 600),
+    ('2', 600), ('3', 600), ('4', 600), ('5', 600), ('6', 600), ('7', 600),
+    ('8', 600), ('9', 600), (':', 600), (';', 600), ('<', 600), ('=', 600),
+    ('>', 600), ('?', 600), ('@', 600), ('A', 600), ('B', 600), ('C', 600),
+    ('D', 600), ('E', 600), ('F', 600), ('G', 600), ('H', 600), ('I', 600),
+    ('J', 600), ('K', 600), ('L', 600), ('M', 600), ('N', 600), ('O', 600),
+    ('P', 600), ('Q', 600), ('R', 600), ('S', 600), ('T', 600), ('U', 600),
+    ('V', 600), ('W', 600), ('X', 600), ('Y', 600), ('Z', 600), ('[', 600),
+    ('\\', 600), (']', 600), ('^', 600), ('_', 600), ('`', 600), ('a', 600),
+    ('b', 600), ('c', 600), ('d', 600), ('e', 600), ('f', 600), ('g', 600),
+    ('h', 600), ('i', 600), ('j', 600), ('k', 600), ('l', 600), ('m', 600),
+    ('n', 600), ('o', 600), ('p', 600), ('q', 600), ('r', 600), ('s', 600),
+    ('t', 600), ('u', 600), ('v', 600), ('w', 600), ('x', 600), ('y', 600),
+    ('z', 600), ('{', 600), ('|', 600), ('}', 600), ('~', 600), ('\u{00A1}', 600),
+    ('\u{00A2}', 600), ('\u{00A3}', 600), ('\u{00A4}', 600), ('\u{00A5}', 600), ('\u{00A6}', 600), ('\u{00A7}', 600),
+    ('\u{00A8}', 600), ('\u{00A9}', 600), ('\u{00AA}', 600), ('\u{00AB}', 600), ('\u{00AC}', 600), ('\u{00AE}', 600),
+    ('\u{00AF}', 600), ('\u{00B0}', 600), ('\u{00B1}', 600), ('\u{00B2}', 600), ('\u{00B3}', 600), ('\u{00B4}', 600),
+    ('\u{00B5}', 600), ('\u{00B6}', 600), ('\u{00B7}', 600), ('\u{00B8}', 600), ('\u{00B9}', 600), ('\u{00BA}', 600),
+    ('\u{00BB}', 600), ('\u{00BC}', 600), ('\u{00BD}', 600), ('\u{00BE}', 600), ('\u{00BF}', 600), ('\u{00C0}', 600),
+    ('\u{00C1}', 600), ('\u{00C2}', 600), ('\u{00C3}', 600), ('\u{00C4}', 600), ('\u{00C5}', 600), ('\u{00C6}', 600),
+    ('\u{00C7}', 600), ('\u{00C8}', 600), ('\u{00C9}', 600), ('\u{00CA}', 600), ('\u{00CB}', 600), ('\u{00CC}', 600),
+    ('\u{00CD}', 600), ('\u{00CE}', 600), ('\u{00CF}', 600), ('\u{00D0}', 600), ('\u{00D1}', 600), ('\u{00D2}', 600),
+    ('\u{00D3}', 600), ('\u{00D4}', 600), ('\u{00D5}', 600), ('\u{00D6}', 600), ('\u{00D7}', 600), ('\u{00D8}', 600),
+    ('\u{00D9}', 600), ('\u{00DA}', 600), ('\u{00DB}', 600), ('\u{00DC}', 600), ('\u{00DD}', 600), ('\u{00DE}', 600),
+    ('\u{00DF}', 600), ('\u{00E0}', 600), ('\u{00E1}', 600), ('\u{00E2}', 600), ('\u{00E3}', 600), ('\u{00E4}', 600),
+    ('\u{00E5}', 600), ('\u{00E6}', 600), ('\u{00E7}', 600), ('\u{00E8}', 600), ('\u{00E9}', 600), ('\u{00EA}', 600),
+    ('\u{00EB}', 600), ('\u{00EC}', 600), ('\u{00ED}', 600), ('\u{00EE}', 600), ('\u{00EF}', 600), ('\u{00F0}', 600),
+    ('\u{00F1}', 600), ('\u{00F2}', 600), ('\u{00F3}', 600), ('\u{00F4}', 600), ('\u{00F5}', 600), ('\u{00F6}', 600),
+    ('\u{00F7}', 600), ('\u{00F8}', 600), ('\u{00F9}', 600), ('\u{00FA}', 600), ('\u{00FB}', 600), ('\u{00FC}', 600),
+    ('\u{00FD}', 600), ('\u{00FE}', 600), ('\u{00FF}', 600), ('\u{0131}', 600), ('\u{0141}', 600), ('\u{0142}', 600),
+    ('\u{0152}', 600), ('\u{0153}', 600), ('\u{0160}', 600), ('\u{0161}', 600), ('\u{0178}', 600), ('\u{017D}', 600),
+    ('\u{017E}', 600), ('\u{0192}', 600), ('\u{02C6}', 600), ('\u{02C7}', 600), ('\u{02D8}', 600), ('\u{02D9}', 600),
+    ('\u{02DA}', 600), ('\u{02DB}', 600), ('\u{02DC}', 600), ('\u{02DD}', 600), ('\u{2013}', 600), ('\u{2014}', 600),
+    ('\u{2018}', 600), ('\u{2019}', 600), ('\u{201A}', 600), ('\u{201C}', 600), ('\u{201D}', 600), ('\u{201E}', 600),
+    ('\u{2020}', 600), ('\u{2021}', 600), ('\u{2022}', 600), ('\u{2026}', 600), ('\u{2030}', 600), ('\u{2039}', 600),
+    ('\u{203A}', 600), ('\u{2044}', 600), ('\u{20AC}', 600), ('\u{2122}', 600), ('\u{2212}', 600), ('\u{FB01}', 600),
+    ('\u{FB02}', 600),
+];
+
+static COURIER_BOLD: &[(char, u16)] = COURIER;
+
+static COURIER_OBLIQUE: &[(char, u16)] = COURIER;
+
+static COURIER_BOLDOBLIQUE: &[(char, u16)] = COURIER;
+
+#[rustfmt::skip]
+static HELVETICA: &[(char, u16)] = &[
+    (' ', 278), ('!', 278), ('"', 355), ('#', 556), ('$', 556), ('%', 889),
+    ('&', 667), ('\'', 191), ('(', 333), (')', 333), ('*', 389), ('+', 584),
+    (',', 278), ('-', 333), ('.', 278), ('/', 278), ('0', 556), ('1', 556),
+    ('2', 556), ('3', 556), ('4', 556), ('5', 556), ('6', 556), ('7', 556),
+    ('8', 556), ('9', 556), (':', 278), (';', 278), ('<', 584), ('=', 584),
+    ('>', 584), ('?', 556), ('@', 1015), ('A', 667), ('B', 667), ('C', 722),
+    ('D', 722), ('E', 667), ('F', 611), ('G', 778), ('H', 722), ('I', 278),
+    ('J', 500), ('K', 667), ('L', 556), ('M', 833), ('N', 722), ('O', 778),
+    ('P', 667), ('Q', 778), ('R', 722), ('S', 667), ('T', 611), ('U', 722),
+    ('V', 667), ('W', 944), ('X', 667), ('Y', 667), ('Z', 611), ('[', 278),
+    ('\\', 278), (']', 278), ('^', 469), ('_', 556), ('`', 333), ('a', 556),
+    ('b', 556), ('c', 500), ('d', 556), ('e', 556), ('f', 278), ('g', 556),
+    ('h', 556), ('i', 222), ('j', 222), ('k', 500), ('l', 222), ('m', 833),
+    ('n', 556), ('o', 556), ('p', 556), ('q', 556), ('r', 333), ('s', 500),
+    ('t', 278), ('u', 556), ('v', 500), ('w', 722), ('x', 500), ('y', 500),
+    ('z', 500), ('{', 334), ('|', 260), ('}', 334), ('~', 584), ('\u{00A1}', 333),
+    ('\u{00A2}', 556), ('\u{00A3}', 556), ('\u{00A4}', 556), ('\u{00A5}', 556), ('\u{00A6}', 260), ('\u{00A7}', 556),
+    ('\u{00A8}', 333), ('\u{00A9}', 737), ('\u{00AA}', 370), ('\u{00AB}', 556), ('\u{00AC}', 584), ('\u{00AE}', 737),
+    ('\u{00AF}', 333), ('\u{00B0}', 400), ('\u{00B1}', 584), ('\u{00B2}', 333), ('\u{00B3}', 333), ('\u{00B4}', 333),
+    ('\u{00B5}', 556), ('\u{00B6}', 537), ('\u{00B7}', 278), ('\u{00B8}', 333), ('\u{00B9}', 333), ('\u{00BA}', 365),
+    ('\u{00BB}', 556), ('\u{00BC}', 834), ('\u{00BD}', 834), ('\u{00BE}', 834), ('\u{00BF}', 611), ('\u{00C0}', 667),
+    ('\u{00C1}', 667), ('\u{00C2}', 667), ('\u{00C3}', 667), ('\u{00C4}', 667), ('\u{00C5}', 667), ('\u{00C6}', 1000),
+    ('\u{00C7}', 722), ('\u{00C8}', 667), ('\u{00C9}', 667), ('\u{00CA}', 667), ('\u{00CB}', 667), ('\u{00CC}', 278),
+    ('\u{00CD}', 278), ('\u{00CE}', 278), ('\u{00CF}', 278), ('\u{00D0}', 722), ('\u{00D1}', 722), ('\u{00D2}', 778),
+    ('\u{00D3}', 778), ('\u{00D4}', 778), ('\u{00D5}', 778), ('\u{00D6}', 778), ('\u{00D7}', 584), ('\u{00D8}', 778),
+    ('\u{00D9}', 722), ('\u{00DA}', 722), ('\u{00DB}', 722), ('\u{00DC}', 722), ('\u{00DD}', 667), ('\u{00DE}', 667),
+    ('\u{00DF}', 611), ('\u{00E0}', 556), ('\u{00E1}', 556), ('\u{00E2}', 556), ('\u{00E3}', 556), ('\u{00E4}', 556),
+    ('\u{00E5}', 556), ('\u{00E6}', 889), ('\u{00E7}', 500), ('\u{00E8}', 556), ('\u{00E9}', 556), ('\u{00EA}', 556),
+    ('\u{00EB}', 556), ('\u{00EC}', 278), ('\u{00ED}', 278), ('\u{00EE}', 278), ('\u{00EF}', 278), ('\u{00F0}', 556),
+    ('\u{00F1}', 556), ('\u{00F2}', 556), ('\u{00F3}', 556), ('\u{00F4}', 556), ('\u{00F5}', 556), ('\u{00F6}', 556),
+    ('\u{00F7}', 584), ('\u{00F8}', 611), ('\u{00F9}', 556), ('\u{00FA}', 556), ('\u{00FB}', 556), ('\u{00FC}', 556),
+    ('\u{00FD}', 500), ('\u{00FE}', 556), ('\u{00FF}', 500), ('\u{0131}', 278), ('\u{0141}', 556), ('\u{0142}', 222),
+    ('\u{0152}', 1000), ('\u{0153}', 944), ('\u{0160}', 667), ('\u{0161}', 500), ('\u{0178}', 667), ('\u{017D}', 611),
+    ('\u{017E}', 500), ('\u{0192}', 556), ('\u{02C6}', 333), ('\u{02C7}', 333), ('\u{02D8}', 333), ('\u{02D9}', 333),
+    ('\u{02DA}', 333), ('\u{02DB}', 333), ('\u{02DC}', 333), ('\u{02DD}', 333), ('\u{2013}', 556), ('\u{2014}', 1000),
+    ('\u{2018}', 222), ('\u{2019}', 222), ('\u{201A}', 222), ('\u{201C}', 333), ('\u{201D}', 333), ('\u{201E}', 333),
+    ('\u{2020}', 556), ('\u{2021}', 556), ('\u{2022}', 350), ('\u{2026}', 1000), ('\u{2030}', 1000), ('\u{2039}', 333),
+    ('\u{203A}', 333), ('\u{2044}', 167), ('\u{20AC}', 556), ('\u{2122}', 1000), ('\u{2212}', 584), ('\u{FB01}', 500),
+    ('\u{FB02}', 500),
+];
+
+#[rustfmt::skip]
+static HELVETICA_BOLD: &[(char, u16)] = &[
+    (' ', 278), ('!', 333), ('"', 474), ('#', 556), ('$', 556), ('%', 889),
+    ('&', 722), ('\'', 238), ('(', 333), (')', 333), ('*', 389), ('+', 584),
+    (',', 278), ('-', 333), ('.', 278), ('/', 278), ('0', 556), ('1', 556),
+    ('2', 556), ('3', 556), ('4', 556), ('5', 556), ('6', 556), ('7', 556),
+    ('8', 556), ('9', 556), (':', 333), (';', 333), ('<', 584), ('=', 584),
+    ('>', 584), ('?', 611), ('@', 975), ('A', 722), ('B', 722), ('C', 722),
+    ('D', 722), ('E', 667), ('F', 611), ('G', 778), ('H', 722), ('I', 278),
+    ('J', 556), ('K', 722), ('L', 611), ('M', 833), ('N', 722), ('O', 778),
+    ('P', 667), ('Q', 778), ('R', 722), ('S', 667), ('T', 611), ('U', 722),
+    ('V', 667), ('W', 944), ('X', 667), ('Y', 667), ('Z', 611), ('[', 333),
+    ('\\', 278), (']', 333), ('^', 584), ('_', 556), ('`', 333), ('a', 556),
+    ('b', 611), ('c', 556), ('d', 611), ('e', 556), ('f', 333), ('g', 611),
+    ('h', 611), ('i', 278), ('j', 278), ('k', 556), ('l', 278), ('m', 889),
+    ('n', 611), ('o', 611), ('p', 611), ('q', 611), ('r', 389), ('s', 556),
+    ('t', 333), ('u', 611), ('v', 556), ('w', 778), ('x', 556), ('y', 556),
+    ('z', 500), ('{', 389), ('|', 280), ('}', 389), ('~', 584), ('\u{00A1}', 333),
+    ('\u{00A2}', 556), ('\u{00A3}', 556), ('\u{00A4}', 556), ('\u{00A5}', 556), ('\u{00A6}', 280), ('\u{00A7}', 556),
+    ('\u{00A8}', 333), ('\u{00A9}', 737), ('\u{00AA}', 370), ('\u{00AB}', 556), ('\u{00AC}', 584), ('\u{00AE}', 737),
+    ('\u{00AF}', 333), ('\u{00B0}', 400), ('\u{00B1}', 584), ('\u{00B2}', 333), ('\u{00B3}', 333), ('\u{00B4}', 333),
+    ('\u{00B5}', 611), ('\u{00B6}', 556), ('\u{00B7}', 278), ('\u{00B8}', 333), ('\u{00B9}', 333), ('\u{00BA}', 365),
+    ('\u{00BB}', 556), ('\u{00BC}', 834), ('\u{00BD}', 834), ('\u{00BE}', 834), ('\u{00BF}', 611), ('\u{00C0}', 722),
+    ('\u{00C1}', 722), ('\u{00C2}', 722), ('\u{00C3}', 722), ('\u{00C4}', 722), ('\u{00C5}', 722), ('\u{00C6}', 1000),
+    ('\u{00C7}', 722), ('\u{00C8}', 667), ('\u{00C9}', 667), ('\u{00CA}', 667), ('\u{00CB}', 667), ('\u{00CC}', 278),
+    ('\u{00CD}', 278), ('\u{00CE}', 278), ('\u{00CF}', 278), ('\u{00D0}', 722), ('\u{00D1}', 722), ('\u{00D2}', 778),
+    ('\u{00D3}', 778), ('\u{00D4}', 778), ('\u{00D5}', 778), ('\u{00D6}', 778), ('\u{00D7}', 584), ('\u{00D8}', 778),
+    ('\u{00D9}', 722), ('\u{00DA}', 722), ('\u{00DB}', 722), ('\u{00DC}', 722), ('\u{00DD}', 667), ('\u{00DE}', 667),
+    ('\u{00DF}', 611), ('\u{00E0}', 556), ('\u{00E1}', 556), ('\u{00E2}', 556), ('\u{00E3}', 556), ('\u{00E4}', 556),
+    ('\u{00E5}', 556), ('\u{00E6}', 889), ('\u{00E7}', 556), ('\u{00E8}', 556), ('\u{00E9}', 556), ('\u{00EA}', 556),
+    ('\u{00EB}', 556), ('\u{00EC}', 278), ('\u{00ED}', 278), ('\u{00EE}', 278), ('\u{00EF}', 278), ('\u{00F0}', 611),
+    ('\u{00F1}', 611), ('\u{00F2}', 611), ('\u{00F3}', 611), ('\u{00F4}', 611), ('\u{00F5}', 611), ('\u{00F6}', 611),
+    ('\u{00F7}', 584), ('\u{00F8}', 611), ('\u{00F9}', 611), ('\u{00FA}', 611), ('\u{00FB}', 611), ('\u{00FC}', 611),
+    ('\u{00FD}', 556), ('\u{00FE}', 611), ('\u{00FF}', 556), ('\u{0131}', 278), ('\u{0141}', 611), ('\u{0142}', 278),
+    ('\u{0152}', 1000), ('\u{0153}', 944), ('\u{0160}', 667), ('\u{0161}', 556), ('\u{0178}', 667), ('\u{017D}', 611),
+    ('\u{017E}', 500), ('\u{0192}', 556), ('\u{02C6}', 333), ('\u{02C7}', 333), ('\u{02D8}', 333), ('\u{02D9}', 333),
+    ('\u{02DA}', 333), ('\u{02DB}', 333), ('\u{02DC}', 333), ('\u{02DD}', 333), ('\u{2013}', 556), ('\u{2014}', 1000),
+    ('\u{2018}', 278), ('\u{2019}', 278), ('\u{201A}', 278), ('\u{201C}', 500), ('\u{201D}', 500), ('\u{201E}', 500),
+    ('\u{2020}', 556), ('\u{2021}', 556), ('\u{2022}', 350), ('\u{2026}', 1000), ('\u{2030}', 1000), ('\u{2039}', 333),
+    ('\u{203A}', 333), ('\u{2044}', 167), ('\u{20AC}', 556), ('\u{2122}', 1000), ('\u{2212}', 584), ('\u{FB01}', 611),
+    ('\u{FB02}', 611),
+];
+
+static HELVETICA_OBLIQUE: &[(char, u16)] = HELVETICA;
+
+static HELVETICA_BOLDOBLIQUE: &[(char, u16)] = HELVETICA_BOLD;
+
+#[rustfmt::skip]
+static TIMES_ROMAN: &[(char, u16)] = &[
+    (' ', 250), ('!', 333), ('"', 408), ('#', 500), ('$', 500), ('%', 833),
+    ('&', 778), ('\'', 180), ('(', 333), (')', 333), ('*', 500), ('+', 564),
+    (',', 250), ('-', 333), ('.', 250), ('/', 278), ('0', 500), ('1', 500),
+    ('2', 500), ('3', 500), ('4', 500), ('5', 500), ('6', 500), ('7', 500),
+    ('8', 500), ('9', 500), (':', 278), (';', 278), ('<', 564), ('=', 564),
+    ('>', 564), ('?', 444), ('@', 921), ('A', 722), ('B', 667), ('C', 667),
+    ('D', 722), ('E', 611), ('F', 556), ('G', 722), ('H', 722), ('I', 333),
+    ('J', 389), ('K', 722), ('L', 611), ('M', 889), ('N', 722), ('O', 722),
+    ('P', 556), ('Q', 722), ('R', 667), ('S', 556), ('T', 611), ('U', 722),
+    ('V', 722), ('W', 944), ('X', 722), ('Y', 722), ('Z', 611), ('[', 333),
+    ('\\', 278), (']', 333), ('^', 469), ('_', 500), ('`', 333), ('a', 444),
+    ('b', 500), ('c', 444), ('d', 500), ('e', 444), ('f', 333), ('g', 500),
+    ('h', 500), ('i', 278), ('j', 278), ('k', 500), ('l', 278), ('m', 778),
+    ('n', 500), ('o', 500), ('p', 500), ('q', 500), ('r', 333), ('s', 389),
+    ('t', 278), ('u', 500), ('v', 500), ('w', 722), ('x', 500), ('y', 500),
+    ('z', 444), ('{', 480), ('|', 200), ('}', 480), ('~', 541), ('\u{00A1}', 333),
+    ('\u{00A2}', 500), ('\u{00A3}', 500), ('\u{00A4}', 500), ('\u{00A5}', 500), ('\u{00A6}', 200), ('\u{00A7}', 500),
+    ('\u{00A8}', 333), ('\u{00A9}', 760), ('\u{00AA}', 276), ('\u{00AB}', 500), ('\u{00AC}', 564), ('\u{00AE}', 760),
+    ('\u{00AF}', 333), ('\u{00B0}', 400), ('\u{00B1}', 564), ('\u{00B2}', 300), ('\u{00B3}', 300), ('\u{00B4}', 333),
+    ('\u{00B5}', 500), ('\u{00B6}', 453), ('\u{00B7}', 250), ('\u{00B8}', 333), ('\u{00B9}', 300), ('\u{00BA}', 310),
+    ('\u{00BB}', 500), ('\u{00BC}', 750), ('\u{00BD}', 750), ('\u{00BE}', 750), ('\u{00BF}', 444), ('\u{00C0}', 722),
+    ('\u{00C1}', 722), ('\u{00C2}', 722), ('\u{00C3}', 722), ('\u{00C4}', 722), ('\u{00C5}', 722), ('\u{00C6}', 889),
+    ('\u{00C7}', 667), ('\u{00C8}', 611), ('\u{00C9}', 611), ('\u{00CA}', 611), ('\u{00CB}', 611), ('\u{00CC}', 333),
+    ('\u{00CD}', 333), ('\u{00CE}', 333), ('\u{00CF}', 333), ('\u{00D0}', 722), ('\u{00D1}', 722), ('\u{00D2}', 722),
+    ('\u{00D3}', 722), ('\u{00D4}', 722), ('\u{00D5}', 722), ('\u{00D6}', 722), ('\u{00D7}', 564), ('\u{00D8}', 722),
+    ('\u{00D9}', 722), ('\u{00DA}', 722), ('\u{00DB}', 722), ('\u{00DC}', 722), ('\u{00DD}', 722), ('\u{00DE}', 556),
+    ('\u{00DF}', 500), ('\u{00E0}', 444), ('\u{00E1}', 444), ('\u{00E2}', 444), ('\u{00E3}', 444), ('\u{00E4}', 444),
+    ('\u{00E5}', 444), ('\u{00E6}', 667), ('\u{00E7}', 444), ('\u{00E8}', 444), ('\u{00E9}', 444), ('\u{00EA}', 444),
+    ('\u{00EB}', 444), ('\u{00EC}', 278), ('\u{00ED}', 278), ('\u{00EE}', 278), ('\u{00EF}', 278), ('\u{00F0}', 500),
+    ('\u{00F1}', 500), ('\u{00F2}', 500), ('\u{00F3}', 500), ('\u{00F4}', 500), ('\u{00F5}', 500), ('\u{00F6}', 500),
+    ('\u{00F7}', 564), ('\u{00F8}', 500), ('\u{00F9}', 500), ('\u{00FA}', 500), ('\u{00FB}', 500), ('\u{00FC}', 500),
+    ('\u{00FD}', 500), ('\u{00FE}', 500), ('\u{00FF}', 500), ('\u{0131}', 278), ('\u{0141}', 611), ('\u{0142}', 278),
+    ('\u{0152}', 889), ('\u{0153}', 722), ('\u{0160}', 556), ('\u{0161}', 389), ('\u{0178}', 722), ('\u{017D}', 611),
+    ('\u{017E}', 444), ('\u{0192}', 500), ('\u{02C6}', 333), ('\u{02C7}', 333), ('\u{02D8}', 333), ('\u{02D9}', 333),
+    ('\u{02DA}', 333), ('\u{02DB}', 333), ('\u{02DC}', 333), ('\u{02DD}', 333), ('\u{2013}', 500), ('\u{2014}', 1000),
+    ('\u{2018}', 333), ('\u{2019}', 333), ('\u{201A}', 333), ('\u{201C}', 444), ('\u{201D}', 444), ('\u{201E}', 444),
+    ('\u{2020}', 500), ('\u{2021}', 500), ('\u{2022}', 350), ('\u{2026}', 1000), ('\u{2030}', 1000), ('\u{2039}', 333),
+    ('\u{203A}', 333), ('\u{2044}', 167), ('\u{20AC}', 500), ('\u{2122}', 980), ('\u{2212}', 564), ('\u{FB01}', 556),
+    ('\u{FB02}', 556),
+];
+
+#[rustfmt::skip]
+static TIMES_BOLD: &[(char, u16)] = &[
+    (' ', 250), ('!', 333), ('"', 555), ('#', 500), ('$', 500), ('%', 1000),
+    ('&', 833), ('\'', 278), ('(', 333), (')', 333), ('*', 500), ('+', 570),
+    (',', 250), ('-', 333), ('.', 250), ('/', 278), ('0', 500), ('1', 500),
+    ('2', 500), ('3', 500), ('4', 500), ('5', 500), ('6', 500), ('7', 500),
+    ('8', 500), ('9', 500), (':', 333), (';', 333), ('<', 570), ('=', 570),
+    ('>', 570), ('?', 500), ('@', 930), ('A', 722), ('B', 667), ('C', 722),
+    ('D', 722), ('E', 667), ('F', 611), ('G', 778), ('H', 778), ('I', 389),
+    ('J', 500), ('K', 778), ('L', 667), ('M', 944), ('N', 722), ('O', 778),
+    ('P', 611), ('Q', 778), ('R', 722), ('S', 556), ('T', 667), ('U', 722),
+    ('V', 722), ('W', 1000), ('X', 722), ('Y', 722), ('Z', 667), ('[', 333),
+    ('\\', 278), (']', 333), ('^', 581), ('_', 500), ('`', 333), ('a', 500),
+    ('b', 556), ('c', 444), ('d', 556), ('e', 444), ('f', 333), ('g', 500),
+    ('h', 556), ('i', 278), ('j', 333), ('k', 556), ('l', 278), ('m', 833),
+    ('n', 556), ('o', 500), ('p', 556), ('q', 556), ('r', 444), ('s', 389),
+    ('t', 333), ('u', 556), ('v', 500), ('w', 722), ('x', 500), ('y', 500),
+    ('z', 444), ('{', 394), ('|', 220), ('}', 394), ('~', 520), ('\u{00A1}', 333),
+    ('\u{00A2}', 500), ('\u{00A3}', 500), ('\u{00A4}', 500), ('\u{00A5}', 500), ('\u{00A6}', 220), ('\u{00A7}', 500),
+    ('\u{00A8}', 333), ('\u{00A9}', 747), ('\u{00AA}', 300), ('\u{00AB}', 500), ('\u{00AC}', 570), ('\u{00AE}', 747),
+    ('\u{00AF}', 333), ('\u{00B0}', 400), ('\u{00B1}', 570), ('\u{00B2}', 300), ('\u{00B3}', 300), ('\u{00B4}', 333),
+    ('\u{00B5}', 556), ('\u{00B6}', 540), ('\u{00B7}', 250), ('\u{00B8}', 333), ('\u{00B9}', 300), ('\u{00BA}', 330),
+    ('\u{00BB}', 500), ('\u{00BC}', 750), ('\u{00BD}', 750), ('\u{00BE}', 750), ('\u{00BF}', 500), ('\u{00C0}', 722),
+    ('\u{00C1}', 722), ('\u{00C2}', 722), ('\u{00C3}', 722), ('\u{00C4}', 722), ('\u{00C5}', 722), ('\u{00C6}', 1000),
+    ('\u{00C7}', 722), ('\u{00C8}', 667), ('\u{00C9}', 667), ('\u{00CA}', 667), ('\u{00CB}', 667), ('\u{00CC}', 389),
+    ('\u{00CD}', 389), ('\u{00CE}', 389), ('\u{00CF}', 389), ('\u{00D0}', 722), ('\u{00D1}', 722), ('\u{00D2}', 778),
+    ('\u{00D3}', 778), ('\u{00D4}', 778), ('\u{00D5}', 778), ('\u{00D6}', 778), ('\u{00D7}', 570), ('\u{00D8}', 778),
+    ('\u{00D9}', 722), ('\u{00DA}', 722), ('\u{00DB}', 722), ('\u{00DC}', 722), ('\u{00DD}', 722), ('\u{00DE}', 611),
+    ('\u{00DF}', 556), ('\u{00E0}', 500), ('\u{00E1}', 500), ('\u{00E2}', 500), ('\u{00E3}', 500), ('\u{00E4}', 500),
+    ('\u{00E5}', 500), ('\u{00E6}', 722), ('\u{00E7}', 444), ('\u{00E8}', 444), ('\u{00E9}', 444), ('\u{00EA}', 444),
+    ('\u{00EB}', 444), ('\u{00EC}', 278), ('\u{00ED}', 278), ('\u{00EE}', 278), ('\u{00EF}', 278), ('\u{00F0}', 500),
+    ('\u{00F1}', 556), ('\u{00F2}', 500), ('\u{00F3}', 500), ('\u{00F4}', 500), ('\u{00F5}', 500), ('\u{00F6}', 500),
+    ('\u{00F7}', 570), ('\u{00F8}', 500), ('\u{00F9}', 556), ('\u{00FA}', 556), ('\u{00FB}', 556), ('\u{00FC}', 556),
+    ('\u{00FD}', 500), ('\u{00FE}', 556), ('\u{00FF}', 500), ('\u{0131}', 278), ('\u{0141}', 667), ('\u{0142}', 278),
+    ('\u{0152}', 1000), ('\u{0153}', 722), ('\u{0160}', 556), ('\u{0161}', 389), ('\u{0178}', 722), ('\u{017D}', 667),
+    ('\u{017E}', 444), ('\u{0192}', 500), ('\u{02C6}', 333), ('\u{02C7}', 333), ('\u{02D8}', 333), ('\u{02D9}', 333),
+    ('\u{02DA}', 333), ('\u{02DB}', 333), ('\u{02DC}', 333), ('\u{02DD}', 333), ('\u{2013}', 500), ('\u{2014}', 1000),
+    ('\u{2018}', 333), ('\u{2019}', 333), ('\u{201A}', 333), ('\u{201C}', 500), ('\u{201D}', 500), ('\u{201E}', 500),
+    ('\u{2020}', 500), ('\u{2021}', 500), ('\u{2022}', 350), ('\u{2026}', 1000), ('\u{2030}', 1000), ('\u{2039}', 333),
+    ('\u{203A}', 333), ('\u{2044}', 167), ('\u{20AC}', 500), ('\u{2122}', 1000), ('\u{2212}', 570), ('\u{FB01}', 556),
+    ('\u{FB02}', 556),
+];
+
+#[rustfmt::skip]
+static TIMES_ITALIC: &[(char, u16)] = &[
+    (' ', 250), ('!', 333), ('"', 420), ('#', 500), ('$', 500), ('%', 833),
+    ('&', 778), ('\'', 214), ('(', 333), (')', 333), ('*', 500), ('+', 675),
+    (',', 250), ('-', 333), ('.', 250), ('/', 278), ('0', 500), ('1', 500),
+    ('2', 500), ('3', 500), ('4', 500), ('5', 500), ('6', 500), ('7', 500),
+    ('8', 500), ('9', 500), (':', 333), (';', 333), ('<', 675), ('=', 675),
+    ('>', 675), ('?', 500), ('@', 920), ('A', 611), ('B', 611), ('C', 667),
+    ('D', 722), ('E', 611), ('F', 611), ('G', 722), ('H', 722), ('I', 333),
+    ('J', 444), ('K', 667), ('L', 556), ('M', 833), ('N', 667), ('O', 722),
+    ('P', 611), ('Q', 722), ('R', 611), ('S', 500), ('T', 556), ('U', 722),
+    ('V', 611), ('W', 833), ('X', 611), ('Y', 556), ('Z', 556), ('[', 389),
+    ('\\', 278), (']', 389), ('^', 422), ('_', 500), ('`', 333), ('a', 500),
+    ('b', 500), ('c', 444), ('d', 500), ('e', 444), ('f', 278), ('g', 500),
+    ('h', 500), ('i', 278), ('j', 278), ('k', 444), ('l', 278), ('m', 722),
+    ('n', 500), ('o', 500), ('p', 500), ('q', 500), ('r', 389), ('s', 389),
+    ('t', 278), ('u', 500), ('v', 444), ('w', 667), ('x', 444), ('y', 444),
+    ('z', 389), ('{', 400), ('|', 275), ('}', 400), ('~', 541), ('\u{00A1}', 389),
+    ('\u{00A2}', 500), ('\u{00A3}', 500), ('\u{00A4}', 500), ('\u{00A5}', 500), ('\u{00A6}', 275), ('\u{00A7}', 500),
+    ('\u{00A8}', 333), ('\u{00A9}', 760), ('\u{00AA}', 276), ('\u{00AB}', 500), ('\u{00AC}', 675), ('\u{00AE}', 760),
+    ('\u{00AF}', 333), ('\u{00B0}', 400), ('\u{00B1}', 675), ('\u{00B2}', 300), ('\u{00B3}', 300), ('\u{00B4}', 333),
+    ('\u{00B5}', 500), ('\u{00B6}', 523), ('\u{00B7}', 250), ('\u{00B8}', 333), ('\u{00B9}', 300), ('\u{00BA}', 310),
+    ('\u{00BB}', 500), ('\u{00BC}', 750), ('\u{00BD}', 750), ('\u{00BE}', 750), ('\u{00BF}', 500), ('\u{00C0}', 611),
+    ('\u{00C1}', 611), ('\u{00C2}', 611), ('\u{00C3}', 611), ('\u{00C4}', 611), ('\u{00C5}', 611), ('\u{00C6}', 889),
+    ('\u{00C7}', 667), ('\u{00C8}', 611), ('\u{00C9}', 611), ('\u{00CA}', 611), ('\u{00CB}', 611), ('\u{00CC}', 333),
+    ('\u{00CD}', 333), ('\u{00CE}', 333), ('\u{00CF}', 333), ('\u{00D0}', 722), ('\u{00D1}', 667), ('\u{00D2}', 722),
+    ('\u{00D3}', 722), ('\u{00D4}', 722), ('\u{00D5}', 722), ('\u{00D6}', 722), ('\u{00D7}', 675), ('\u{00D8}', 722),
+    ('\u{00D9}', 722), ('\u{00DA}', 722), ('\u{00DB}', 722), ('\u{00DC}', 722), ('\u{00DD}', 556), ('\u{00DE}', 611),
+    ('\u{00DF}', 500), ('\u{00E0}', 500), ('\u{00E1}', 500), ('\u{00E2}', 500), ('\u{00E3}', 500), ('\u{00E4}', 500),
+    ('\u{00E5}', 500), ('\u{00E6}', 667), ('\u{00E7}', 444), ('\u{00E8}', 444), ('\u{00E9}', 444), ('\u{00EA}', 444),
+    ('\u{00EB}', 444), ('\u{00EC}', 278), ('\u{00ED}', 278), ('\u{00EE}', 278), ('\u{00EF}', 278), ('\u{00F0}', 500),
+    ('\u{00F1}', 500), ('\u{00F2}', 500), ('\u{00F3}', 500), ('\u{00F4}', 500), ('\u{00F5}', 500), ('\u{00F6}', 500),
+    ('\u{00F7}', 675), ('\u{00F8}', 500), ('\u{00F9}', 500), ('\u{00FA}', 500), ('\u{00FB}', 500), ('\u{00FC}', 500),
+    ('\u{00FD}', 444), ('\u{00FE}', 500), ('\u{00FF}', 444), ('\u{0131}', 278), ('\u{0141}', 556), ('\u{0142}', 278),
+    ('\u{0152}', 944), ('\u{0153}', 667), ('\u{0160}', 500), ('\u{0161}', 389), ('\u{0178}', 556), ('\u{017D}', 556),
+    ('\u{017E}', 389), ('\u{0192}', 500), ('\u{02C6}', 333), ('\u{02C7}', 333), ('\u{02D8}', 333), ('\u{02D9}', 333),
+    ('\u{02DA}', 333), ('\u{02DB}', 333), ('\u{02DC}', 333), ('\u{02DD}', 333), ('\u{2013}', 500), ('\u{2014}', 889),
+    ('\u{2018}', 333), ('\u{2019}', 333), ('\u{201A}', 333), ('\u{201C}', 556), ('\u{201D}', 556), ('\u{201E}', 556),
+    ('\u{2020}', 500), ('\u{2021}', 500), ('\u{2022}', 350), ('\u{2026}', 889), ('\u{2030}', 1000), ('\u{2039}', 333),
+    ('\u{203A}', 333), ('\u{2044}', 167), ('\u{20AC}', 500), ('\u{2122}', 980), ('\u{2212}', 675), ('\u{FB01}', 500),
+    ('\u{FB02}', 500),
+];
+
+#[rustfmt::skip]
+static TIMES_BOLDITALIC: &[(char, u16)] = &[
+    (' ', 250), ('!', 389), ('"', 555), ('#', 500), ('$', 500), ('%', 833),
+    ('&', 778), ('\'', 278), ('(', 333), (')', 333), ('*', 500), ('+', 570),
+    (',', 250), ('-', 333), ('.', 250), ('/', 278), ('0', 500), ('1', 500),
+    ('2', 500), ('3', 500), ('4', 500), ('5', 500), ('6', 500), ('7', 500),
+    ('8', 500), ('9', 500), (':', 333), (';', 333), ('<', 570), ('=', 570),
+    ('>', 570), ('?', 500), ('@', 832), ('A', 667), ('B', 667), ('C', 667),
+    ('D', 722), ('E', 667), ('F', 667), ('G', 722), ('H', 778), ('I', 389),
+    ('J', 500), ('K', 667), ('L', 611), ('M', 889), ('N', 722), ('O', 722),
+    ('P', 611), ('Q', 722), ('R', 667), ('S', 556), ('T', 611), ('U', 722),
+    ('V', 667), ('W', 889), ('X', 667), ('Y', 611), ('Z', 611), ('[', 333),
+    ('\\', 278), (']', 333), ('^', 570), ('_', 500), ('`', 333), ('a', 500),
+    ('b', 500), ('c', 444), ('d', 500), ('e', 444), ('f', 333), ('g', 500),
+    ('h', 556), ('i', 278), ('j', 278), ('k', 500), ('l', 278), ('m', 778),
+    ('n', 556), ('o', 500), ('p', 500), ('q', 500), ('r', 389), ('s', 389),
+    ('t', 278), ('u', 556), ('v', 444), ('w', 667), ('x', 500), ('y', 444),
+    ('z', 389), ('{', 348), ('|', 220), ('}', 348), ('~', 570), ('\u{00A1}', 389),
+    ('\u{00A2}', 500), ('\u{00A3}', 500), ('\u{00A4}', 500), ('\u{00A5}', 500), ('\u{00A6}', 220), ('\u{00A7}', 500),
+    ('\u{00A8}', 333), ('\u{00A9}', 747), ('\u{00AA}', 266), ('\u{00AB}', 500), ('\u{00AC}', 606), ('\u{00AE}', 747),
+    ('\u{00AF}', 333), ('\u{00B0}', 400), ('\u{00B1}', 570), ('\u{00B2}', 300), ('\u{00B3}', 300), ('\u{00B4}', 333),
+    ('\u{00B5}', 576), ('\u{00B6}', 500), ('\u{00B7}', 250), ('\u{00B8}', 333), ('\u{00B9}', 300), ('\u{00BA}', 300),
+    ('\u{00BB}', 500), ('\u{00BC}', 750), ('\u{00BD}', 750), ('\u{00BE}', 750), ('\u{00BF}', 500), ('\u{00C0}', 667),
+    ('\u{00C1}', 667), ('\u{00C2}', 667), ('\u{00C3}', 667), ('\u{00C4}', 667), ('\u{00C5}', 667), ('\u{00C6}', 944),
+    ('\u{00C7}', 667), ('\u{00C8}', 667), ('\u{00C9}', 667), ('\u{00CA}', 667), ('\u{00CB}', 667), ('\u{00CC}', 389),
+    ('\u{00CD}', 389), ('\u{00CE}', 389), ('\u{00CF}', 389), ('\u{00D0}', 722), ('\u{00D1}', 722), ('\u{00D2}', 722),
+    ('\u{00D3}', 722), ('\u{00D4}', 722), ('\u{00D5}', 722), ('\u{00D6}', 722), ('\u{00D7}', 570), ('\u{00D8}', 722),
+    ('\u{00D9}', 722), ('\u{00DA}', 722), ('\u{00DB}', 722), ('\u{00DC}', 722), ('\u{00DD}', 611), ('\u{00DE}', 611),
+    ('\u{00DF}', 500), ('\u{00E0}', 500), ('\u{00E1}', 500), ('\u{00E2}', 500), ('\u{00E3}', 500), ('\u{00E4}', 500),
+    ('\u{00E5}', 500), ('\u{00E6}', 722), ('\u{00E7}', 444), ('\u{00E8}', 444), ('\u{00E9}', 444), ('\u{00EA}', 444),
+    ('\u{00EB}', 444), ('\u{00EC}', 278), ('\u{00ED}', 278), ('\u{00EE}', 278), ('\u{00EF}', 278), ('\u{00F0}', 500),
+    ('\u{00F1}', 556), ('\u{00F2}', 500), ('\u{00F3}', 500), ('\u{00F4}', 500), ('\u{00F5}', 500), ('\u{00F6}', 500),
+    ('\u{00F7}', 570), ('\u{00F8}', 500), ('\u{00F9}', 556), ('\u{00FA}', 556), ('\u{00FB}', 556), ('\u{00FC}', 556),
+    ('\u{00FD}', 444), ('\u{00FE}', 500), ('\u{00FF}', 444), ('\u{0131}', 278), ('\u{0141}', 611), ('\u{0142}', 278),
+    ('\u{0152}', 944), ('\u{0153}', 722), ('\u{0160}', 556), ('\u{0161}', 389), ('\u{0178}', 611), ('\u{017D}', 611),
+    ('\u{017E}', 389), ('\u{0192}', 500), ('\u{02C6}', 333), ('\u{02C7}', 333), ('\u{02D8}', 333), ('\u{02D9}', 333),
+    ('\u{02DA}', 333), ('\u{02DB}', 333), ('\u{02DC}', 333), ('\u{02DD}', 333), ('\u{2013}', 500), ('\u{2014}', 1000),
+    ('\u{2018}', 333), ('\u{2019}', 333), ('\u{201A}', 333), ('\u{201C}', 500), ('\u{201D}', 500), ('\u{201E}', 500),
+    ('\u{2020}', 500), ('\u{2021}', 500), ('\u{2022}', 350), ('\u{2026}', 1000), ('\u{2030}', 1000), ('\u{2039}', 333),
+    ('\u{203A}', 333), ('\u{2044}', 167), ('\u{20AC}', 500), ('\u{2122}', 1000), ('\u{2212}', 606), ('\u{FB01}', 556),
+    ('\u{FB02}', 556),
+];
+
+#[rustfmt::skip]
+static SYMBOL: &[(char, u16)] = &[
+    (' ', 250), ('!', 333), ('#', 500), ('%', 833), ('&', 778), ('(', 333),
+    (')', 333), ('+', 549), (',', 250), ('.', 250), ('/', 278), ('0', 500),
+    ('1', 500), ('2', 500), ('3', 500), ('4', 500), ('5', 500), ('6', 500),
+    ('7', 500), ('8', 500), ('9', 500), (':', 278), (';', 278), ('<', 549),
+    ('=', 549), ('>', 549), ('?', 444), ('[', 333), (']', 333), ('_', 500),
+    ('{', 480), ('|', 200), ('}', 480), ('\u{00AC}', 713), ('\u{00B0}', 400), ('\u{00B1}', 549),
+    ('\u{00B5}', 576), ('\u{00D7}', 549), ('\u{00F7}', 549), ('\u{0192}', 500), ('\u{0391}', 722), ('\u{0392}', 667),
+    ('\u{0393}', 603), ('\u{0395}', 611), ('\u{0396}', 611), ('\u{0397}', 722), ('\u{0398}', 741), ('\u{0399}', 333),
+    ('\u{039A}', 722), ('\u{039B}', 686), ('\u{039C}', 889), ('\u{039D}', 722), ('\u{039E}', 645), ('\u{039F}', 722),
+    ('\u{03A0}', 768), ('\u{03A1}', 556), ('\u{03A3}', 592), ('\u{03A4}', 611), ('\u{03A5}', 690), ('\u{03A6}', 763),
+    ('\u{03A7}', 722), ('\u{03A8}', 795), ('\u{03B1}', 631), ('\u{03B2}', 549), ('\u{03B3}', 411), ('\u{03B4}', 494),
+    ('\u{03B5}', 439), ('\u{03B6}', 494), ('\u{03B7}', 603), ('\u{03B8}', 521), ('\u{03B9}', 329), ('\u{03BA}', 549),
+    ('\u{03BB}', 549), ('\u{03BD}', 521), ('\u{03BE}', 493), ('\u{03BF}', 549), ('\u{03C0}', 549), ('\u{03C1}', 549),
+    ('\u{03C2}', 439), ('\u{03C3}', 603), ('\u{03C4}', 439), ('\u{03C5}', 576), ('\u{03C6}', 521), ('\u{03C7}', 549),
+    ('\u{03C8}', 686), ('\u{03C9}', 686), ('\u{03D1}', 631), ('\u{03D2}', 620), ('\u{03D5}', 603), ('\u{03D6}', 713),
+    ('\u{2022}', 460), ('\u{2026}', 1000), ('\u{2032}', 247), ('\u{2033}', 411), ('\u{2044}', 167), ('\u{20AC}', 750),
+    ('\u{2111}', 686), ('\u{2118}', 987), ('\u{211C}', 795), ('\u{2126}', 768), ('\u{2135}', 823), ('\u{2190}', 987),
+    ('\u{2191}', 603), ('\u{2192}', 987), ('\u{2193}', 603), ('\u{2194}', 1042), ('\u{21B5}', 658), ('\u{21D0}', 987),
+    ('\u{21D1}', 603), ('\u{21D2}', 987), ('\u{21D3}', 603), ('\u{21D4}', 1042), ('\u{2200}', 713), ('\u{2202}', 494),
+    ('\u{2203}', 549), ('\u{2205}', 823), ('\u{2206}', 612), ('\u{2207}', 713), ('\u{2208}', 713), ('\u{2209}', 713),
+    ('\u{220B}', 439), ('\u{220F}', 823), ('\u{2211}', 713), ('\u{2212}', 549), ('\u{2217}', 500), ('\u{221A}', 549),
+    ('\u{221D}', 713), ('\u{221E}', 713), ('\u{2220}', 768), ('\u{2227}', 603), ('\u{2228}', 603), ('\u{2229}', 768),
+    ('\u{222A}', 768), ('\u{222B}', 274), ('\u{2234}', 863), ('\u{223C}', 549), ('\u{2245}', 549), ('\u{2248}', 549),
+    ('\u{2260}', 549), ('\u{2261}', 549), ('\u{2264}', 549), ('\u{2265}', 549), ('\u{2282}', 713), ('\u{2283}', 713),
+    ('\u{2284}', 713), ('\u{2286}', 713), ('\u{2287}', 713), ('\u{2295}', 768), ('\u{2297}', 768), ('\u{22A5}', 658),
+    ('\u{22C5}', 250), ('\u{2320}', 686), ('\u{2321}', 686), ('\u{2329}', 329), ('\u{232A}', 329), ('\u{25CA}', 494),
+    ('\u{2660}', 753), ('\u{2663}', 753), ('\u{2665}', 753), ('\u{2666}', 753), ('\u{F6D9}', 790), ('\u{F6DA}', 790),
+    ('\u{F6DB}', 890), ('\u{F8E5}', 500), ('\u{F8E6}', 603), ('\u{F8E7}', 1000), ('\u{F8E8}', 790), ('\u{F8E9}', 790),
+    ('\u{F8EA}', 786), ('\u{F8EB}', 384), ('\u{F8EC}', 384), ('\u{F8ED}', 384), ('\u{F8EE}', 384), ('\u{F8EF}', 384),
+    ('\u{F8F0}', 384), ('\u{F8F1}', 494), ('\u{F8F2}', 494), ('\u{F8F3}', 494), ('\u{F8F4}', 494), ('\u{F8F5}', 686),
+    ('\u{F8F6}', 384), ('\u{F8F7}', 384), ('\u{F8F8}', 384), ('\u{F8F9}', 384), ('\u{F8FA}', 384), ('\u{F8FB}', 384),
+    ('\u{F8FC}', 494), ('\u{F8FD}', 494), ('\u{F8FE}', 494), ('\u{F8FF}', 790),
+];
+
+#[rustfmt::skip]
+static ZAPFDINGBATS: &[(char, u16)] = &[
+    (' ', 278), ('\u{2192}', 838), ('\u{2194}', 1016), ('\u{2195}', 458), ('\u{2460}', 788), ('\u{2461}', 788),
+    ('\u{2462}', 788), ('\u{2463}', 788), ('\u{2464}', 788), ('\u{2465}', 788), ('\u{2466}', 788), ('\u{2467}', 788),
+    ('\u{2468}', 788), ('\u{2469}', 788), ('\u{25A0}', 761), ('\u{25B2}', 892), ('\u{25BC}', 892), ('\u{25C6}', 788),
+    ('\u{25CF}', 791), ('\u{25D7}', 438), ('\u{2605}', 816), ('\u{260E}', 719), ('\u{261B}', 960), ('\u{261E}', 939),
+    ('\u{2660}', 626), ('\u{2663}', 776), ('\u{2665}', 694), ('\u{2666}', 595), ('\u{2701}', 974), ('\u{2702}', 961),
+    ('\u{2703}', 974), ('\u{2704}', 980), ('\u{2706}', 789), ('\u{2707}', 790), ('\u{2708}', 791), ('\u{2709}', 690),
+    ('\u{270C}', 549), ('\u{270D}', 855), ('\u{270E}', 911), ('\u{270F}', 933), ('\u{2710}', 911), ('\u{2711}', 945),
+    ('\u{2712}', 974), ('\u{2713}', 755), ('\u{2714}', 846), ('\u{2715}', 762), ('\u{2716}', 761), ('\u{2717}', 571),
+    ('\u{2718}', 677), ('\u{2719}', 763), ('\u{271A}', 760), ('\u{271B}', 759), ('\u{271C}', 754), ('\u{271D}', 494),
+    ('\u{271E}', 552), ('\u{271F}', 537), ('\u{2720}', 577), ('\u{2721}', 692), ('\u{2722}', 786), ('\u{2723}', 788),
+    ('\u{2724}', 788), ('\u{2725}', 790), ('\u{2726}', 793), ('\u{2727}', 794), ('\u{2729}', 823), ('\u{272A}', 789),
+    ('\u{272B}', 841), ('\u{272C}', 823), ('\u{272D}', 833), ('\u{272E}', 816), ('\u{272F}', 831), ('\u{2730}', 923),
+    ('\u{2731}', 744), ('\u{2732}', 723), ('\u{2733}', 749), ('\u{2734}', 790), ('\u{2735}', 792), ('\u{2736}', 695),
+    ('\u{2737}', 776), ('\u{2738}', 768), ('\u{2739}', 792), ('\u{273A}', 759), ('\u{273B}', 707), ('\u{273C}', 708),
+    ('\u{273D}', 682), ('\u{273E}', 701), ('\u{273F}', 826), ('\u{2740}', 815), ('\u{2741}', 789), ('\u{2742}', 789),
+    ('\u{2743}', 707), ('\u{2744}', 687), ('\u{2745}', 696), ('\u{2746}', 689), ('\u{2747}', 786), ('\u{2748}', 787),
+    ('\u{2749}', 713), ('\u{274A}', 791), ('\u{274B}', 785), ('\u{274D}', 873), ('\u{274F}', 762), ('\u{2750}', 762),
+    ('\u{2751}', 759), ('\u{2752}', 759), ('\u{2756}', 784), ('\u{2758}', 138), ('\u{2759}', 277), ('\u{275A}', 415),
+    ('\u{275B}', 392), ('\u{275C}', 392), ('\u{275D}', 668), ('\u{275E}', 668), ('\u{2761}', 732), ('\u{2762}', 544),
+    ('\u{2763}', 544), ('\u{2764}', 910), ('\u{2765}', 667), ('\u{2766}', 760), ('\u{2767}', 760), ('\u{2768}', 390),
+    ('\u{2769}', 390), ('\u{276A}', 317), ('\u{276B}', 317), ('\u{276C}', 276), ('\u{276D}', 276), ('\u{276E}', 509),
+    ('\u{276F}', 509), ('\u{2770}', 410), ('\u{2771}', 410), ('\u{2772}', 234), ('\u{2773}', 234), ('\u{2774}', 334),
+    ('\u{2775}', 334), ('\u{2776}', 788), ('\u{2777}', 788), ('\u{2778}', 788), ('\u{2779}', 788), ('\u{277A}', 788),
+    ('\u{277B}', 788), ('\u{277C}', 788), ('\u{277D}', 788), ('\u{277E}', 788), ('\u{277F}', 788), ('\u{2780}', 788),
+    ('\u{2781}', 788), ('\u{2782}', 788), ('\u{2783}', 788), ('\u{2784}', 788), ('\u{2785}', 788), ('\u{2786}', 788),
+    ('\u{2787}', 788), ('\u{2788}', 788), ('\u{2789}', 788), ('\u{278A}', 788), ('\u{278B}', 788), ('\u{278C}', 788),
+    ('\u{278D}', 788), ('\u{278E}', 788), ('\u{278F}', 788), ('\u{2790}', 788), ('\u{2791}', 788), ('\u{2792}', 788),
+    ('\u{2793}', 788), ('\u{2794}', 894), ('\u{2798}', 748), ('\u{2799}', 924), ('\u{279A}', 748), ('\u{279B}', 918),
+    ('\u{279C}', 927), ('\u{279D}', 928), ('\u{279E}', 928), ('\u{279F}', 834), ('\u{27A0}', 873), ('\u{27A1}', 828),
+    ('\u{27A2}', 924), ('\u{27A3}', 924), ('\u{27A4}', 917), ('\u{27A5}', 930), ('\u{27A6}', 931), ('\u{27A7}', 463),
+    ('\u{27A8}', 883), ('\u{27A9}', 836), ('\u{27AA}', 836), ('\u{27AB}', 867), ('\u{27AC}', 867), ('\u{27AD}', 696),
+    ('\u{27AE}', 696), ('\u{27AF}', 874), ('\u{27B1}', 874), ('\u{27B2}', 760), ('\u{27B3}', 946), ('\u{27B4}', 771),
+    ('\u{27B5}', 865), ('\u{27B6}', 771), ('\u{27B7}', 888), ('\u{27B8}', 967), ('\u{27B9}', 888), ('\u{27BA}', 831),
+    ('\u{27BB}', 873), ('\u{27BC}', 927), ('\u{27BD}', 970), ('\u{27BE}', 918),
+];
+
+#[rustfmt::skip]
+static SYMBOL_ENCODING: &[(u8, char)] = &[
+    (0x20, ' '), (0x21, '!'), (0x22, '\u{2200}'), (0x23, '#'), (0x24, '\u{2203}'), (0x25, '%'),
+    (0x26, '&'), (0x27, '\u{220B}'), (0x28, '('), (0x29, ')'), (0x2A, '\u{2217}'), (0x2B, '+'),
+    (0x2C, ','), (0x2D, '\u{2212}'), (0x2E, '.'), (0x2F, '/'), (0x30, '0'), (0x31, '1'),
+    (0x32, '2'), (0x33, '3'), (0x34, '4'), (0x35, '5'), (0x36, '6'), (0x37, '7'),
+    (0x38, '8'), (0x39, '9'), (0x3A, ':'), (0x3B, ';'), (0x3C, '<'), (0x3D, '='),
+    (0x3E, '>'), (0x3F, '?'), (0x40, '\u{2245}'), (0x41, '\u{0391}'), (0x42, '\u{0392}'), (0x43, '\u{03A7}'),
+    (0x44, '\u{2206}'), (0x45, '\u{0395}'), (0x46, '\u{03A6}'), (0x47, '\u{0393}'), (0x48, '\u{0397}'), (0x49, '\u{0399}'),
+    (0x4A, '\u{03D1}'), (0x4B, '\u{039A}'), (0x4C, '\u{039B}'), (0x4D, '\u{039C}'), (0x4E, '\u{039D}'), (0x4F, '\u{039F}'),
+    (0x50, '\u{03A0}'), (0x51, '\u{0398}'), (0x52, '\u{03A1}'), (0x53, '\u{03A3}'), (0x54, '\u{03A4}'), (0x55, '\u{03A5}'),
+    (0x56, '\u{03C2}'), (0x57, '\u{2126}'), (0x58, '\u{039E}'), (0x59, '\u{03A8}'), (0x5A, '\u{0396}'), (0x5B, '['),
+    (0x5C, '\u{2234}'), (0x5D, ']'), (0x5E, '\u{22A5}'), (0x5F, '_'), (0x60, '\u{F8E5}'), (0x61, '\u{03B1}'),
+    (0x62, '\u{03B2}'), (0x63, '\u{03C7}'), (0x64, '\u{03B4}'), (0x65, '\u{03B5}'), (0x66, '\u{03C6}'), (0x67, '\u{03B3}'),
+    (0x68, '\u{03B7}'), (0x69, '\u{03B9}'), (0x6A, '\u{03D5}'), (0x6B, '\u{03BA}'), (0x6C, '\u{03BB}'), (0x6D, '\u{00B5}'),
+    (0x6E, '\u{03BD}'), (0x6F, '\u{03BF}'), (0x70, '\u{03C0}'), (0x71, '\u{03B8}'), (0x72, '\u{03C1}'), (0x73, '\u{03C3}'),
+    (0x74, '\u{03C4}'), (0x75, '\u{03C5}'), (0x76, '\u{03D6}'), (0x77, '\u{03C9}'), (0x78, '\u{03BE}'), (0x79, '\u{03C8}'),
+    (0x7A, '\u{03B6}'), (0x7B, '{'), (0x7C, '|'), (0x7D, '}'), (0x7E, '\u{223C}'), (0xA0, '\u{20AC}'),
+    (0xA1, '\u{03D2}'), (0xA2, '\u{2032}'), (0xA3, '\u{2264}'), (0xA4, '\u{2044}'), (0xA5, '\u{221E}'), (0xA6, '\u{0192}'),
+    (0xA7, '\u{2663}'), (0xA8, '\u{2666}'), (0xA9, '\u{2665}'), (0xAA, '\u{2660}'), (0xAB, '\u{2194}'), (0xAC, '\u{2190}'),
+    (0xAD, '\u{2191}'), (0xAE, '\u{2192}'), (0xAF, '\u{2193}'), (0xB0, '\u{00B0}'), (0xB1, '\u{00B1}'), (0xB2, '\u{2033}'),
+    (0xB3, '\u{2265}'), (0xB4, '\u{00D7}'), (0xB5, '\u{221D}'), (0xB6, '\u{2202}'), (0xB7, '\u{2022}'), (0xB8, '\u{00F7}'),
+    (0xB9, '\u{2260}'), (0xBA, '\u{2261}'), (0xBB, '\u{2248}'), (0xBC, '\u{2026}'), (0xBD, '\u{F8E6}'), (0xBE, '\u{F8E7}'),
+    (0xBF, '\u{21B5}'), (0xC0, '\u{2135}'), (0xC1, '\u{2111}'), (0xC2, '\u{211C}'), (0xC3, '\u{2118}'), (0xC4, '\u{2297}'),
+    (0xC5, '\u{2295}'), (0xC6, '\u{2205}'), (0xC7, '\u{2229}'), (0xC8, '\u{222A}'), (0xC9, '\u{2283}'), (0xCA, '\u{2287}'),
+    (0xCB, '\u{2284}'), (0xCC, '\u{2282}'), (0xCD, '\u{2286}'), (0xCE, '\u{2208}'), (0xCF, '\u{2209}'), (0xD0, '\u{2220}'),
+    (0xD1, '\u{2207}'), (0xD2, '\u{F6DA}'), (0xD3, '\u{F6D9}'), (0xD4, '\u{F6DB}'), (0xD5, '\u{220F}'), (0xD6, '\u{221A}'),
+    (0xD7, '\u{22C5}'), (0xD8, '\u{00AC}'), (0xD9, '\u{2227}'), (0xDA, '\u{2228}'), (0xDB, '\u{21D4}'), (0xDC, '\u{21D0}'),
+    (0xDD, '\u{21D1}'), (0xDE, '\u{21D2}'), (0xDF, '\u{21D3}'), (0xE0, '\u{25CA}'), (0xE1, '\u{2329}'), (0xE2, '\u{F8E8}'),
+    (0xE3, '\u{F8E9}'), (0xE4, '\u{F8EA}'), (0xE5, '\u{2211}'), (0xE6, '\u{F8EB}'), (0xE7, '\u{F8EC}'), (0xE8, '\u{F8ED}'),
+    (0xE9, '\u{F8EE}'), (0xEA, '\u{F8EF}'), (0xEB, '\u{F8F0}'), (0xEC, '\u{F8F1}'), (0xED, '\u{F8F2}'), (0xEE, '\u{F8F3}'),
+    (0xEF, '\u{F8F4}'), (0xF1, '\u{232A}'), (0xF2, '\u{222B}'), (0xF3, '\u{2320}'), (0xF4, '\u{F8F5}'), (0xF5, '\u{2321}'),
+    (0xF6, '\u{F8F6}'), (0xF7, '\u{F8F7}'), (0xF8, '\u{F8F8}'), (0xF9, '\u{F8F9}'), (0xFA, '\u{F8FA}'), (0xFB, '\u{F8FB}'),
+    (0xFC, '\u{F8FC}'), (0xFD, '\u{F8FD}'), (0xFE, '\u{F8FE}'),
+];
+
+#[rustfmt::skip]
+static ZAPFDINGBATS_ENCODING: &[(u8, char)] = &[
+    (0x20, ' '), (0x21, '\u{2701}'), (0x22, '\u{2702}'), (0x23, '\u{2703}'), (0x24, '\u{2704}'), (0x25, '\u{260E}'),
+    (0x26, '\u{2706}'), (0x27, '\u{2707}'), (0x28, '\u{2708}'), (0x29, '\u{2709}'), (0x2A, '\u{261B}'), (0x2B, '\u{261E}'),
+    (0x2C, '\u{270C}'), (0x2D, '\u{270D}'), (0x2E, '\u{270E}'), (0x2F, '\u{270F}'), (0x30, '\u{2710}'), (0x31, '\u{2711}'),
+    (0x32, '\u{2712}'), (0x33, '\u{2713}'), (0x34, '\u{2714}'), (0x35, '\u{2715}'), (0x36, '\u{2716}'), (0x37, '\u{2717}'),
+    (0x38, '\u{2718}'), (0x39, '\u{2719}'), (0x3A, '\u{271A}'), (0x3B, '\u{271B}'), (0x3C, '\u{271C}'), (0x3D, '\u{271D}'),
+    (0x3E, '\u{271E}'), (0x3F, '\u{271F}'), (0x40, '\u{2720}'), (0x41, '\u{2721}'), (0x42, '\u{2722}'), (0x43, '\u{2723}'),
+    (0x44, '\u{2724}'), (0x45, '\u{2725}'), (0x46, '\u{2726}'), (0x47, '\u{2727}'), (0x48, '\u{2605}'), (0x49, '\u{2729}'),
+    (0x4A, '\u{272A}'), (0x4B, '\u{272B}'), (0x4C, '\u{272C}'), (0x4D, '\u{272D}'), (0x4E, '\u{272E}'), (0x4F, '\u{272F}'),
+    (0x50, '\u{2730}'), (0x51, '\u{2731}'), (0x52, '\u{2732}'), (0x53, '\u{2733}'), (0x54, '\u{2734}'), (0x55, '\u{2735}'),
+    (0x56, '\u{2736}'), (0x57, '\u{2737}'), (0x58, '\u{2738}'), (0x59, '\u{2739}'), (0x5A, '\u{273A}'), (0x5B, '\u{273B}'),
+    (0x5C, '\u{273C}'), (0x5D, '\u{273D}'), (0x5E, '\u{273E}'), (0x5F, '\u{273F}'), (0x60, '\u{2740}'), (0x61, '\u{2741}'),
+    (0x62, '\u{2742}'), (0x63, '\u{2743}'), (0x64, '\u{2744}'), (0x65, '\u{2745}'), (0x66, '\u{2746}'), (0x67, '\u{2747}'),
+    (0x68, '\u{2748}'), (0x69, '\u{2749}'), (0x6A, '\u{274A}'), (0x6B, '\u{274B}'), (0x6C, '\u{25CF}'), (0x6D, '\u{274D}'),
+    (0x6E, '\u{25A0}'), (0x6F, '\u{274F}'), (0x70, '\u{2750}'), (0x71, '\u{2751}'), (0x72, '\u{2752}'), (0x73, '\u{25B2}'),
+    (0x74, '\u{25BC}'), (0x75, '\u{25C6}'), (0x76, '\u{2756}'), (0x77, '\u{25D7}'), (0x78, '\u{2758}'), (0x79, '\u{2759}'),
+    (0x7A, '\u{275A}'), (0x7B, '\u{275B}'), (0x7C, '\u{275C}'), (0x7D, '\u{275D}'), (0x7E, '\u{275E}'), (0x80, '\u{2768}'),
+    (0x81, '\u{2769}'), (0x82, '\u{276A}'), (0x83, '\u{276B}'), (0x84, '\u{276C}'), (0x85, '\u{276D}'), (0x86, '\u{276E}'),
+    (0x87, '\u{276F}'), (0x88, '\u{2770}'), (0x89, '\u{2771}'), (0x8A, '\u{2772}'), (0x8B, '\u{2773}'), (0x8C, '\u{2774}'),
+    (0x8D, '\u{2775}'), (0xA1, '\u{2761}'), (0xA2, '\u{2762}'), (0xA3, '\u{2763}'), (0xA4, '\u{2764}'), (0xA5, '\u{2765}'),
+    (0xA6, '\u{2766}'), (0xA7, '\u{2767}'), (0xA8, '\u{2663}'), (0xA9, '\u{2666}'), (0xAA, '\u{2665}'), (0xAB, '\u{2660}'),
+    (0xAC, '\u{2460}'), (0xAD, '\u{2461}'), (0xAE, '\u{2462}'), (0xAF, '\u{2463}'), (0xB0, '\u{2464}'), (0xB1, '\u{2465}'),
+    (0xB2, '\u{2466}'), (0xB3, '\u{2467}'), (0xB4, '\u{2468}'), (0xB5, '\u{2469}'), (0xB6, '\u{2776}'), (0xB7, '\u{2777}'),
+    (0xB8, '\u{2778}'), (0xB9, '\u{2779}'), (0xBA, '\u{277A}'), (0xBB, '\u{277B}'), (0xBC, '\u{277C}'), (0xBD, '\u{277D}'),
+    (0xBE, '\u{277E}'), (0xBF, '\u{277F}'), (0xC0, '\u{2780}'), (0xC1, '\u{2781}'), (0xC2, '\u{2782}'), (0xC3, '\u{2783}'),
+    (0xC4, '\u{2784}'), (0xC5, '\u{2785}'), (0xC6, '\u{2786}'), (0xC7, '\u{2787}'), (0xC8, '\u{2788}'), (0xC9, '\u{2789}'),
+    (0xCA, '\u{278A}'), (0xCB, '\u{278B}'), (0xCC, '\u{278C}'), (0xCD, '\u{278D}'), (0xCE, '\u{278E}'), (0xCF, '\u{278F}'),
+    (0xD0, '\u{2790}'), (0xD1, '\u{2791}'), (0xD2, '\u{2792}'), (0xD3, '\u{2793}'), (0xD4, '\u{2794}'), (0xD5, '\u{2192}'),
+    (0xD6, '\u{2194}'), (0xD7, '\u{2195}'), (0xD8, '\u{2798}'), (0xD9, '\u{2799}'), (0xDA, '\u{279A}'), (0xDB, '\u{279B}'),
+    (0xDC, '\u{279C}'), (0xDD, '\u{279D}'), (0xDE, '\u{279E}'), (0xDF, '\u{279F}'), (0xE0, '\u{27A0}'), (0xE1, '\u{27A1}'),
+    (0xE2, '\u{27A2}'), (0xE3, '\u{27A3}'), (0xE4, '\u{27A4}'), (0xE5, '\u{27A5}'), (0xE6, '\u{27A6}'), (0xE7, '\u{27A7}'),
+    (0xE8, '\u{27A8}'), (0xE9, '\u{27A9}'), (0xEA, '\u{27AA}'), (0xEB, '\u{27AB}'), (0xEC, '\u{27AC}'), (0xED, '\u{27AD}'),
+    (0xEE, '\u{27AE}'), (0xEF, '\u{27AF}'), (0xF1, '\u{27B1}'), (0xF2, '\u{27B2}'), (0xF3, '\u{27B3}'), (0xF4, '\u{27B4}'),
+    (0xF5, '\u{27B5}'), (0xF6, '\u{27B6}'), (0xF7, '\u{27B7}'), (0xF8, '\u{27B8}'), (0xF9, '\u{27B9}'), (0xFA, '\u{27BA}'),
+    (0xFB, '\u{27BB}'), (0xFC, '\u{27BC}'), (0xFD, '\u{27BD}'), (0xFE, '\u{27BE}'),
+];
+
+/// Every width table, for exhaustive testing.
+#[cfg(test)]
+static ALL_TABLES: &[(&str, &[(char, u16)])] = &[
+    ("COURIER", COURIER),
+    ("COURIER_BOLD", COURIER_BOLD),
+    ("COURIER_OBLIQUE", COURIER_OBLIQUE),
+    ("COURIER_BOLDOBLIQUE", COURIER_BOLDOBLIQUE),
+    ("HELVETICA", HELVETICA),
+    ("HELVETICA_BOLD", HELVETICA_BOLD),
+    ("HELVETICA_OBLIQUE", HELVETICA_OBLIQUE),
+    ("HELVETICA_BOLDOBLIQUE", HELVETICA_BOLDOBLIQUE),
+    ("TIMES_ROMAN", TIMES_ROMAN),
+    ("TIMES_BOLD", TIMES_BOLD),
+    ("TIMES_ITALIC", TIMES_ITALIC),
+    ("TIMES_BOLDITALIC", TIMES_BOLDITALIC),
+    ("SYMBOL", SYMBOL),
+    ("ZAPFDINGBATS", ZAPFDINGBATS),
+];
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn times_roman_ascii_widths() {
+        assert_eq!(base14_char_width("Times-Roman", ' '), Some(250));
+        assert_eq!(base14_char_width("Times-Roman", 'M'), Some(889));
+        assert_eq!(base14_char_width("Times-Roman", 'i'), Some(278));
+    }
+
+    #[test]
+    fn subset_prefix_and_aliases_normalize() {
+        assert_eq!(
+            base14_char_width("ABCDEF+Times-Bold", ' '),
+            base14_char_width("Times-Bold", ' ')
+        );
+        assert!(base14_char_width("ArialMT", 'a').is_some());
+        assert!(base14_char_width("TimesNewRomanPSMT", 'a').is_some());
+    }
+
+    #[test]
+    fn non_base14_returns_none() {
+        assert_eq!(base14_char_width("DejaVuSans", 'a'), None);
+        assert!(!is_base14_font("Garamond"));
+    }
+
+    #[test]
+    fn builtin_encoding_resolves_symbol_and_zapf_codes() {
+        // Symbol 0x61 renders alpha; Zapf 0x21 renders U+2701.
+        assert_eq!(builtin_encoding_char("Symbol", 0x61), Some('\u{03B1}'));
+        assert_eq!(builtin_encoding_char("Symbol", 0xA5), Some('\u{221E}'));
+        assert_eq!(
+            builtin_encoding_char("ZapfDingbats", 0x21),
+            Some('\u{2701}')
+        );
+        // Latin text fonts follow standard encodings — no builtin override.
+        assert_eq!(builtin_encoding_char("Times-Roman", 0x61), None);
+        // The resolved chars have real AFM widths.
+        let alpha_w = base14_char_width("Symbol", '\u{03B1}');
+        assert!(alpha_w.is_some() && alpha_w != Some(500));
+    }
+
+    #[test]
+    fn encoding_tables_are_sorted_for_binary_search() {
+        for table in [SYMBOL_ENCODING, ZAPFDINGBATS_ENCODING] {
+            assert!(table.windows(2).all(|w| w[0].0 < w[1].0));
+        }
+    }
+
+    #[test]
+    fn tables_are_sorted_for_binary_search() {
+        // Every table is queried by binary search, so all of them must be
+        // sorted — not just a sample.
+        for (name, table) in ALL_TABLES {
+            assert!(
+                table.windows(2).all(|w| w[0].0 < w[1].0),
+                "{name} is not sorted"
+            );
+        }
+    }
+}

src/extractor/content_stream.rs

@@ -14,9 +14,9 @@ use lopdf::{Document, Encoding, Object, ObjectId};
 use std::collections::HashMap;
 
 use super::fonts::{
-    build_font_encodings, build_font_widths, compute_string_width_ts, descriptor_style_flags,
-    extract_text_from_operand, get_font_file2_obj_num, get_operand_bytes, CMapDecisionCache,
-    FontStyleCache,
+    build_font_encodings, build_font_widths, build_type3_scales, compute_string_width_ts,
+    descriptor_style_flags, extract_text_from_operand, get_font_file2_obj_num, get_operand_bytes,
+    CMapDecisionCache, FontStyleCache,
 };
 use super::underline::UnderlineLine;
 use super::xobjects::{extract_form_xobject_text, get_page_xobjects, XObjectType};
@@ -41,6 +41,17 @@ fn strip_pdf_comments(data: &[u8]) -> Vec<u8> {
     while i < data.len() {
         let b = data[i];
         match b {
+            // Inside a string literal, a backslash escapes the next byte —
+            // `\(`, `\)`, and `\\` must not touch the nesting depth, or a
+            // later `%` glyph inside a string gets stripped as a comment,
+            // corrupting the stream.
+            b'\\' if in_string > 0 => {
+                result.push(b);
+                if let Some(&next) = data.get(i + 1) {
+                    result.push(next);
+                    i += 1;
+                }
+            }
             b'(' if !in_hex_string => {
                 in_string += 1;
                 result.push(b);
@@ -162,10 +173,11 @@ pub(crate) fn extract_page_text_items(
     let fonts = doc.get_page_fonts(page_id).unwrap_or_default();
 
     // Build font encoding maps from Differences arrays
-    let (font_encodings, has_gid_fonts) = build_font_encodings(doc, &fonts);
+    let (font_encodings, has_gid_fonts) = build_font_encodings(doc, &fonts, font_cmaps);
 
     // Build font width info for accurate text positioning
     let font_widths = build_font_widths(doc, &fonts);
+    let type3_scales = build_type3_scales(doc, &fonts);
 
     // Build maps of font resource names to their base font names and ToUnicode object refs
     let mut font_base_names: std::collections::HashMap<String, String> =
@@ -502,7 +514,8 @@ pub(crate) fn extract_page_text_items(
                     ) {
                         let combined =
                             multiply_matrices(&rise_adjusted(&text_matrix, text_rise), &ctm);
-                        let rendered_size = effective_font_size(current_font_size, &combined);
+                        let rendered_size = effective_font_size(current_font_size, &combined)
+                            * type3_scales.get(&current_font).copied().unwrap_or(1.0);
                         let (x, y) = (combined[4], combined[5]);
                         if combined[0].abs() >= combined[1].abs() {
                             rotation_votes.horizontal += 1;
@@ -673,7 +686,8 @@ pub(crate) fn extract_page_text_items(
                             } else {
                                 rotation_votes.rotated += 1;
                             }
-                            let rendered_size = effective_font_size(current_font_size, &combined);
+                            let rendered_size = effective_font_size(current_font_size, &combined)
+                                * type3_scales.get(&current_font).copied().unwrap_or(1.0);
                             let base_font = font_base_names
                                 .get(&current_font)
                                 .map(|s| s.as_str())
@@ -780,7 +794,8 @@ pub(crate) fn extract_page_text_items(
                             } else {
                                 rotation_votes.rotated += 1;
                             }
-                            let rendered_size = effective_font_size(current_font_size, &combined);
+                            let rendered_size = effective_font_size(current_font_size, &combined)
+                                * type3_scales.get(&current_font).copied().unwrap_or(1.0);
                             let (x, y) = (combined[4], combined[5]);
                             let width = w_ts_opt
                                 .map(|w_ts| {
@@ -932,7 +947,8 @@ pub(crate) fn extract_page_text_items(
                             } else {
                                 rotation_votes.rotated += 1;
                             }
-                            let rendered_size = effective_font_size(current_font_size, &combined);
+                            let rendered_size = effective_font_size(current_font_size, &combined)
+                                * type3_scales.get(&current_font).copied().unwrap_or(1.0);
                             let (x, y) = (combined[4], combined[5]);
                             // Width in device space from text matrix delta
                             let delta_ts = text_matrix[4] - start_tm[4];
@@ -1849,4 +1865,26 @@ BT 30 700 Tm <41> Tj ET";
             "ET should be preserved after comment stripping"
         );
     }
+
+    #[test]
+    fn test_strip_pdf_comments_escaped_parens() {
+        // An escaped `\)` must not close the string: the `%` after it is
+        // still string content, not a comment (subset fonts routinely map
+        // glyphs to `%` and to escaped parens in the same TJ array).
+        let input = b"[ (a\\)b) 1 (%) 1 (c) ] TJ\n";
+        let output = strip_pdf_comments(input);
+        assert_eq!(output, input.to_vec());
+
+        // Same for an escaped `\(` — must not open a phantom string that
+        // shields a real comment.
+        let input = b"(x\\(y) Tj % real comment\nET\n";
+        let output = strip_pdf_comments(input);
+        assert_eq!(output, b"(x\\(y) Tj  \nET\n");
+
+        // Escaped backslash before a real close-paren: `\\` ends the escape,
+        // the `)` does close the string, and the comment is stripped.
+        let input = b"(x\\\\) Tj % comment\nET\n";
+        let output = strip_pdf_comments(input);
+        assert_eq!(output, b"(x\\\\) Tj  \nET\n");
+    }
 }

src/extractor/fonts.rs

@@ -138,6 +138,93 @@ pub(crate) fn build_font_widths(
     widths
 }
 
+/// Visual-size scale factors for Type3 fonts, keyed by resource name.
+///
+/// A Type3 font's glyph space maps to text space through FontMatrix, so the
+/// visual height of its glyphs is `nominal_size × |matrix_y| × FontBBox
+/// height`. For a well-behaved font (matrix 0.001, bbox ≈ 1000 units) that
+/// factor is ≈ 1.0 and the nominal size is already right. TeX PK bitmap
+/// fonts (dvips → Distiller) instead use FontMatrix [1 0 0 -1 0 0] with
+/// nominal sizes like 0.12, which makes every downstream font-size heuristic
+/// (drop caps, sub/superscripts, small-font tables, line heights) see
+/// nonsense. Fonts without a usable FontBBox are omitted (treated as 1.0).
+pub(crate) fn build_type3_scales(
+    doc: &Document,
+    fonts: &std::collections::BTreeMap<Vec<u8>, &lopdf::Dictionary>,
+) -> HashMap<String, f32> {
+    let mut scales = HashMap::new();
+    for (font_name, font_dict) in fonts {
+        let is_type3 = font_dict
+            .get(b"Subtype")
+            .ok()
+            .and_then(|o| o.as_name().ok())
+            .is_some_and(|n| n == b"Type3");
+        if !is_type3 {
+            continue;
+        }
+        // Array elements may themselves be indirect references per PDF
+        // syntax — resolve before reading the numeric value.
+        let num = |o: &Object| {
+            let resolved = match o {
+                Object::Reference(r) => match doc.get_object(*r) {
+                    Ok(inner) => inner,
+                    Err(_) => return 0.0,
+                },
+                other => other,
+            };
+            match resolved {
+                Object::Integer(i) => *i as f32,
+                Object::Real(r) => *r,
+                _ => 0.0,
+            }
+        };
+        let Some(matrix) = font_dict
+            .get(b"FontMatrix")
+            .ok()
+            .and_then(|o| resolve_array(doc, o))
+        else {
+            continue;
+        };
+        let Some(bbox) = font_dict
+            .get(b"FontBBox")
+            .ok()
+            .and_then(|o| resolve_array(doc, o))
+        else {
+            continue;
+        };
+        if matrix.len() < 4 || bbox.len() < 4 {
+            continue;
+        }
+        let scale_y = (num(&matrix[2]).powi(2) + num(&matrix[3]).powi(2)).sqrt();
+        let bbox_h = (num(&bbox[3]) - num(&bbox[1])).abs();
+        let scale = bbox_h * scale_y;
+
+        // `scale` is the glyph box measured in text-space units. For a
+        // self-consistent font it lands near 1.0 — the FontMatrix is the
+        // reciprocal of the glyph-space em by construction — so the Tf
+        // operand is already the rendered size and must be left alone.
+        // A modest deviation is normal and must NOT trigger rescaling:
+        // FontBBox is the glyph bounding box, not the em box, so it is
+        // routinely somewhat smaller (descender..ascender ≈ 0.7) or larger
+        // (tall accents > 1.0).
+        //
+        // Only a wildly inconsistent font gets renormalized. dvips/PK
+        // bitmap fonts declare [1 0 0 -1 0 0] with glyphs spanning
+        // hundreds of units, giving scale ≈ 159 against a nominal size of
+        // 0.12pt — there the declared size carries no information. The
+        // band is deliberately wide so that only that class qualifies,
+        // while any matrix scale (including non-standard ones like 0.005
+        // with a full-em bbox, scale = 5.0) is judged on the product
+        // rather than on the matrix alone.
+        const CONSISTENT_LO: f32 = 0.25;
+        const CONSISTENT_HI: f32 = 4.0;
+        if scale.is_finite() && scale > 0.0 && !(CONSISTENT_LO..=CONSISTENT_HI).contains(&scale) {
+            scales.insert(String::from_utf8_lossy(font_name).to_string(), scale);
+        }
+    }
+    scales
+}
+
 /// Parse font widths from a font dictionary, dispatching by Subtype
 pub(crate) fn parse_font_widths(
     doc: &Document,
@@ -149,11 +236,71 @@ pub(crate) fn parse_font_widths(
 
     match subtype_name {
         b"Type0" => parse_type0_widths(doc, font_dict),
-        b"Type1" | b"TrueType" | b"MMType1" | b"Type3" => parse_simple_font_widths(doc, font_dict),
+        b"Type1" | b"TrueType" | b"MMType1" => parse_simple_font_widths(doc, font_dict)
+            .or_else(|| base14_fallback_widths(doc, font_dict)),
+        b"Type3" => parse_simple_font_widths(doc, font_dict),
         _ => None,
     }
 }
 
+/// Fallback metrics for non-embedded base-14 fonts whose dictionary omits
+/// `/FirstChar`/`/Widths` (legal per the PDF spec — the reader must supply
+/// standard-font metrics). Without this, every glyph advances 0 and all
+/// downstream gap-based logic (space synthesis, script detection, table
+/// columns) collapses — common in 1990s dvips/Distiller PDFs.
+///
+/// Widths are resolved per code through the font's Differences encoding when
+/// present, falling back to the same single-byte decode the text extractor
+/// uses (cp1252-style smart punctuation for 0x80..=0x9F, Latin-1 elsewhere) —
+/// so the width of a code always matches the char we extract for it.
+fn base14_fallback_widths(doc: &Document, font_dict: &lopdf::Dictionary) -> Option<FontWidthInfo> {
+    let base_font = font_dict
+        .get(b"BaseFont")
+        .ok()
+        .and_then(|o| o.as_name().ok())
+        .map(|n| String::from_utf8_lossy(n).to_string())?;
+    if !crate::extractor::base14::is_base14_font(&base_font) {
+        return None;
+    }
+
+    let enc_map = parse_font_encoding(doc, font_dict)
+        .map(|r| r.map)
+        .unwrap_or_default();
+
+    let mut widths = HashMap::new();
+    for code in 0u16..=255 {
+        // Resolution order: Differences override, then the font's BUILT-IN
+        // encoding (Symbol/ZapfDingbats glyphs live at positions unrelated
+        // to cp1252 — the renderer draws α for Symbol 0x61 no matter how
+        // the text decoder transliterates it, so the advance must be α's),
+        // then the cp1252-style fallback used by the text decoder.
+        let ch = enc_map
+            .get(&(code as u8))
+            .copied()
+            .or_else(|| crate::extractor::base14::builtin_encoding_char(&base_font, code as u8))
+            .unwrap_or_else(|| decode_single_byte_fallback_char(code as u8, true));
+        if let Some(w) = crate::extractor::base14::base14_char_width(&base_font, ch) {
+            widths.insert(code, w);
+        }
+    }
+    let space_width = widths.get(&32).copied().unwrap_or(250);
+
+    debug!(
+        "  base14 fallback widths for {} ({} codes mapped)",
+        base_font,
+        widths.len()
+    );
+
+    Some(FontWidthInfo {
+        widths,
+        default_width: 500,
+        space_width,
+        is_cid: false,
+        units_scale: 0.001,
+        wmode: 0,
+    })
+}
+
 /// Parse widths for simple fonts (Type1, TrueType, MMType1, Type3)
 /// Reads FirstChar, LastChar, and Widths array.
 /// For Type3 fonts, reads FontMatrix to determine the correct units_scale.
@@ -497,9 +644,13 @@ pub(crate) fn get_operand_bytes(obj: &Object) -> Option<&[u8]> {
 /// Build encoding maps for all fonts on a page.
 /// Returns `(encodings, has_gid_fonts)` where `has_gid_fonts` is true when
 /// any font uses raw glyph ID names (gidNNNNN) that can't be decoded.
+/// Gid names whose codes the font's own ToUnicode CMap maps are decodable
+/// and do not set the flag (LibreOffice subsets write /gidNNNN Differences
+/// names alongside a complete ToUnicode CMap).
 pub(crate) fn build_font_encodings(
     doc: &Document,
     fonts: &std::collections::BTreeMap<Vec<u8>, &lopdf::Dictionary>,
+    cmaps: &FontCMaps,
 ) -> (PageFontEncodings, bool) {
     let mut encodings = PageFontEncodings::new();
     let mut has_gid_fonts = false;
@@ -508,7 +659,9 @@ pub(crate) fn build_font_encodings(
         let resource_name = String::from_utf8_lossy(font_name).to_string();
 
         if let Some(result) = parse_font_encoding(doc, font_dict) {
-            if result.gid_glyph_count > 0 {
+            if !result.gid_codes.is_empty()
+                && !tounicode_maps_codes(font_dict, cmaps, &result.gid_codes)
+            {
                 has_gid_fonts = true;
             }
             if !result.map.is_empty() {
@@ -520,6 +673,34 @@ pub(crate) fn build_font_encodings(
     (encodings, has_gid_fonts)
 }
 
+/// True when the font's ToUnicode CMap maps the gid-named character codes,
+/// so the Differences entries still decode through the CMap.
+fn tounicode_maps_codes(font_dict: &lopdf::Dictionary, cmaps: &FontCMaps, codes: &[u8]) -> bool {
+    let Some(obj_ref) = font_dict
+        .get(b"ToUnicode")
+        .ok()
+        .and_then(|o| o.as_reference().ok())
+    else {
+        return false;
+    };
+    let Some(entry) = cmaps.get_by_obj(obj_ref.0) else {
+        return false;
+    };
+    // At least one gid code usably mapped means the CMap addresses these
+    // codes; remaining unmapped codes are subset leftovers (e.g. the
+    // component glyphs of an emoji ZWJ sequence mapped whole on its first
+    // code). A mapping is usable only when extraction would accept it —
+    // empty or U+FFFD results are rejected there as invalid. Fonts whose
+    // CMap ignores the gid codes entirely stay flagged, and the downstream
+    // garbage/encoding checks still catch partial damage.
+    codes.iter().any(|&code| {
+        entry
+            .primary
+            .lookup(code as u16)
+            .is_some_and(|s| !s.is_empty() && !s.contains('\u{FFFD}'))
+    })
+}
+
 /// Parse font encoding from a font dictionary
 pub(crate) fn parse_font_encoding(
     doc: &Document,
@@ -558,11 +739,10 @@ pub(crate) fn parse_font_encoding(
 /// Result of parsing an encoding dictionary's Differences array.
 pub(crate) struct EncodingResult {
     pub map: FontEncodingMap,
-    /// Number of glyph names matching the `gidNNNNN` pattern (raw glyph IDs).
-    /// These indicate a font with unresolvable encoding — the glyph IDs
-    /// reference the original font's glyph table, but without the original
-    /// font's cmap there is no way to map them to Unicode.
-    pub gid_glyph_count: u32,
+    /// Character codes whose glyph names match the `gidNNNNN` pattern (raw
+    /// glyph IDs). These reference the original font's glyph table and are
+    /// only decodable when the font's ToUnicode CMap maps the code.
+    pub gid_codes: Vec<u8>,
 }
 
 /// Parse an encoding dictionary with Differences array
@@ -588,7 +768,7 @@ pub(crate) fn parse_encoding_dictionary(
     let mut encoding_map = FontEncodingMap::new();
     let mut current_code: u8 = 0;
     let mut ligature_count = 0u32;
-    let mut gid_glyph_count = 0u32;
+    let mut gid_codes: Vec<u8> = Vec::new();
 
     for item in diff_array {
         match item {
@@ -614,7 +794,7 @@ pub(crate) fn parse_encoding_dictionary(
                     && glyph_name.len() >= 4
                     && glyph_name[3..].chars().all(|c| c.is_ascii_digit())
                 {
-                    gid_glyph_count += 1;
+                    gid_codes.push(current_code);
                 }
                 if let Some(ch) = mapped_char {
                     encoding_map.insert(current_code, ch);
@@ -638,16 +818,16 @@ pub(crate) fn parse_encoding_dictionary(
         );
     }
 
-    if gid_glyph_count > 0 {
+    if !gid_codes.is_empty() {
         debug!(
-            "  Differences: {} gid-encoded glyphs (unresolvable without original font)",
-            gid_glyph_count
+            "  Differences: {} gid-encoded glyphs (decodable only via ToUnicode)",
+            gid_codes.len()
         );
     }
 
     Some(EncodingResult {
         map: encoding_map,
-        gid_glyph_count,
+        gid_codes,
     })
 }
 
@@ -1187,10 +1367,37 @@ pub(crate) fn extract_text_from_operand(
     })();
     result.map(|text| {
         let text = clean_symbol_pua(text);
+        let text = remap_texcm_math_symbols(text, base_font_name);
         normalize_cp1252_controls(text, use_cp1252_fallback)
     })
 }
 
+/// Fix a known producer bug in "TeXCMMathsSymbols" subset fonts (IntechOpen
+/// and sibling academic pipelines): the Computer Modern symbol glyphs are
+/// misnamed after Latin lookalikes (equal → /onequarter, plus → /thorn, …)
+/// and the generated ToUnicode faithfully propagates the wrong names. The
+/// remap applies only to text decoded from that font, keyed on the glyphs'
+/// observed misnames.
+fn remap_texcm_math_symbols(text: String, base_font_name: Option<&str>) -> String {
+    let is_texcm = base_font_name.is_some_and(|n| {
+        let n = n.rsplit_once('+').map_or(n, |(_, s)| s);
+        n.eq_ignore_ascii_case("TeXCMMathsSymbols")
+    });
+    if !is_texcm {
+        return text;
+    }
+    text.chars()
+        .map(|c| match c {
+            '¼' => '=',
+            '½' => '-',
+            'þ' => '+',
+            'ð' => '(',
+            'Þ' => ')',
+            _ => c,
+        })
+        .collect()
+}
+
 fn decode_single_byte_fallback(bytes: &[u8], use_cp1252_fallback: bool) -> String {
     bytes
         .iter()
@@ -1409,6 +1616,107 @@ fn score_text(text: &str) -> i32 {
 
 #[cfg(test)]
 mod tests {
+
+    #[test]
+    fn type3_scale_resolves_indirect_matrix_and_bbox_numbers() {
+        use lopdf::{dictionary, Document, Object};
+        // FontMatrix/FontBBox elements may be indirect references per PDF
+        // syntax; the scale must use their resolved values, not zero.
+        let mut doc = Document::with_version("1.4");
+        let matrix_d = doc.add_object(Object::Real(-1.0));
+        let bbox_top = doc.add_object(Object::Integer(3));
+        let font_dict = dictionary! {
+            "Type" => "Font",
+            "Subtype" => "Type3",
+            "FontMatrix" => vec![
+                Object::Integer(1),
+                Object::Integer(0),
+                Object::Integer(0),
+                Object::Reference(matrix_d),
+                Object::Integer(0),
+                Object::Integer(0),
+            ],
+            "FontBBox" => vec![
+                Object::Integer(1),
+                Object::Integer(-156),
+                Object::Integer(37),
+                Object::Reference(bbox_top),
+            ],
+        };
+        let mut fonts = std::collections::BTreeMap::new();
+        fonts.insert(b"T2".to_vec(), &font_dict);
+        let scales = super::build_type3_scales(&doc, &fonts);
+        let scale = scales.get("T2").copied().unwrap_or(1.0);
+        // bbox height 159 x |matrix_y| 1.0
+        assert!(
+            (scale - 159.0).abs() < 0.5,
+            "scale should use resolved indirect values, got {scale}"
+        );
+    }
+
+    /// Build a one-font Type3 document and return its computed scale, if any.
+    #[cfg(test)]
+    fn type3_scale_for(matrix_y: f32, bbox_lo: i64, bbox_hi: i64) -> Option<f32> {
+        use lopdf::{dictionary, Document, Object};
+        let doc = Document::with_version("1.4");
+        let font_dict = dictionary! {
+            "Type" => "Font",
+            "Subtype" => "Type3",
+            "FontMatrix" => vec![
+                Object::Real(matrix_y), Object::Integer(0), Object::Integer(0),
+                Object::Real(matrix_y), Object::Integer(0), Object::Integer(0),
+            ],
+            "FontBBox" => vec![
+                Object::Integer(0), Object::Integer(bbox_lo),
+                Object::Integer(600), Object::Integer(bbox_hi),
+            ],
+        };
+        let mut fonts = std::collections::BTreeMap::new();
+        fonts.insert(b"T9".to_vec(), &font_dict);
+        super::build_type3_scales(&doc, &fonts).get("T9").copied()
+    }
+
+    #[test]
+    fn type3_scale_skips_self_consistent_fonts() {
+        // Conventional 1/1000 matrix with a descender..ascender bbox of 700
+        // units: scale 0.7. The Tf operand is already the rendered size, so
+        // renormalizing would report every size at 0.7x.
+        assert_eq!(type3_scale_for(0.001, -200, 500), None);
+        // Tall-accent bbox slightly over the em (1100 units, scale 1.1).
+        assert_eq!(type3_scale_for(0.001, -100, 1000), None);
+    }
+
+    #[test]
+    fn type3_scale_applies_to_inconsistent_fonts_at_any_matrix_scale() {
+        // Non-standard but valid matrix (0.005) with a full-em bbox:
+        // scale 5.0, so the declared size is off by 5x and must be fixed.
+        let s = type3_scale_for(0.005, 0, 1000).expect("0.005 matrix should rescale");
+        assert!((s - 5.0).abs() < 0.01, "got {s}");
+        // dvips/PK bitmap pattern: unit matrix, glyphs spanning ~159 units.
+        let s = type3_scale_for(1.0, -156, 3).expect("PK pattern should rescale");
+        assert!((s - 159.0).abs() < 0.5, "got {s}");
+    }
+
+    #[test]
+    fn type3_scale_ignores_degenerate_bbox() {
+        // [0 0 0 0] is legal and carries no size information.
+        assert_eq!(type3_scale_for(0.001, 0, 0), None);
+    }
+
+    #[test]
+    fn texcm_math_symbols_remap() {
+        assert_eq!(
+            super::remap_texcm_math_symbols("S ¼ kB þ 1".into(), Some("EEKVNO+TeXCMMathsSymbols")),
+            "S = kB + 1"
+        );
+        // Other fonts keep their genuine fractions/thorns.
+        assert_eq!(
+            super::remap_texcm_math_symbols("¼ cup þorn".into(), Some("Times-Roman")),
+            "¼ cup þorn"
+        );
+        assert_eq!(super::remap_texcm_math_symbols("¼".into(), None), "¼");
+    }
+
     use super::*;
     use lopdf::dictionary;
 
@@ -1896,4 +2204,113 @@ mod tests {
             false
         ));
     }
+
+    fn gid_font_doc(bfchar: Option<&str>) -> (Document, lopdf::ObjectId) {
+        use lopdf::Stream;
+        let mut doc = Document::with_version("1.4");
+        let cmap = format!(
+            "/CIDInit /ProcSet findresource begin
+12 dict begin
+begincmap
+1 begincodespacerange
+<00> <FF>
+endcodespacerange
+1 beginbfchar
+{}
+endbfchar
+endcmap
+CMapName currentdict /CMap defineresource pop
+end
+end",
+            bfchar.unwrap_or_default()
+        );
+        let tounicode_id = doc.add_object(Object::Stream(Stream::new(
+            dictionary! {},
+            cmap.into_bytes(),
+        )));
+        let enc_id = doc.add_object(dictionary! {
+            "Type" => "Encoding",
+            "Differences" => vec![
+                1.into(),
+                Object::Name(b"gid1283".to_vec()),
+                Object::Name(b"gid1464".to_vec()),
+            ],
+        });
+        let mut font = dictionary! {
+            "Type" => "Font",
+            "Subtype" => "TrueType",
+            "BaseFont" => "ABCDEF+OpenSymbol",
+            "Encoding" => Object::Reference(enc_id),
+        };
+        if bfchar.is_some() {
+            font.set("ToUnicode", Object::Reference(tounicode_id));
+        }
+        let font_id = doc.add_object(font);
+        let page_id = doc.add_object(dictionary! {
+            "Type" => "Page",
+            "Resources" => dictionary! {
+                "Font" => dictionary! { "F1" => Object::Reference(font_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 gid_flagged(bfchar: Option<&str>) -> bool {
+        let (doc, page_id) = gid_font_doc(bfchar);
+        let cmaps = FontCMaps::from_doc(&doc);
+        let fonts = doc.get_page_fonts(page_id).unwrap();
+        let (_, has_gid_fonts) = build_font_encodings(&doc, &fonts, &cmaps);
+        has_gid_fonts
+    }
+
+    #[test]
+    fn gid_differences_with_covering_tounicode_are_not_flagged() {
+        // LibreOffice subsets write /gidNNNN Differences names alongside a
+        // ToUnicode CMap that decodes those codes; the page must not be
+        // flagged as unresolvable (which would suppress the whole document's
+        // markdown when every page carries such a font).
+        assert!(!gid_flagged(Some("<01> <2022>\n<02> <25E6>")));
+    }
+
+    #[test]
+    fn gid_differences_with_partial_tounicode_are_not_flagged() {
+        // An emoji ZWJ sequence maps whole on its first code; the remaining
+        // component-glyph codes are subset leftovers, not damage.
+        assert!(!gid_flagged(Some(
+            "<01> <D83DDC68200DD83DDC69200DD83DDC67>"
+        )));
+    }
+
+    #[test]
+    fn gid_differences_without_tounicode_are_flagged() {
+        assert!(
+            gid_flagged(None),
+            "gid glyphs without ToUnicode are unresolvable"
+        );
+    }
+
+    #[test]
+    fn gid_differences_with_disjoint_tounicode_are_flagged() {
+        // A ToUnicode that never addresses the gid codes leaves them
+        // unresolvable.
+        assert!(gid_flagged(Some("<10> <0041>")));
+    }
+
+    #[test]
+    fn gid_differences_with_replacement_char_tounicode_are_flagged() {
+        // A mapping to U+FFFD is not usable — extraction rejects it as an
+        // invalid CMap result — so it must not clear the gid flag.
+        assert!(gid_flagged(Some("<01> <FFFD>\n<02> <FFFD>")));
+    }
 }

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();
 
@@ -146,32 +186,104 @@ 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,
     };
+    if fields.is_empty() {
+        return items;
+    }
+    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, field_ref, None, "", page_map, &mut items);
+            walk_form_fields(
+                doc,
+                field_ref,
+                None,
+                "",
+                page_map,
+                &annotation_pages,
+                &mut items,
+                &mut budget,
+                0,
+            );
         }
     }
 
     items
 }
 
+/// Map widget annotation objects back to the page whose `/Annots` array owns
+/// them. Some valid widgets omit `/P`, so the page tree is the only reliable
+/// ownership signal available for page-filtered extraction.
+fn annotation_page_map(
+    doc: &Document,
+    page_map: &HashMap<ObjectId, u32>,
+) -> HashMap<ObjectId, u32> {
+    let mut annotation_pages = HashMap::new();
+    for (&page_id, &page_num) in page_map {
+        let Some(annotations) = doc
+            .get_dictionary(page_id)
+            .ok()
+            .and_then(|page| page.get(b"Annots").ok())
+            .and_then(|annotations| resolve_array(doc, annotations))
+        else {
+            continue;
+        };
+        for annotation in annotations {
+            if let Ok(annotation_id) = annotation.as_reference() {
+                annotation_pages.insert(annotation_id, page_num);
+            }
+        }
+    }
+    annotation_pages
+}
+
 /// 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,
     parent_ft: Option<&[u8]>,
     parent_name: &str,
     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,
@@ -202,11 +314,31 @@ 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, kid_ref, ft, &full_name, page_map, items);
+                    walk_form_fields(
+                        doc,
+                        kid_ref,
+                        ft,
+                        &full_name,
+                        page_map,
+                        annotation_pages,
+                        items,
+                        budget,
+                        depth + 1,
+                    );
                 }
             }
             return;
@@ -299,6 +431,7 @@ pub(crate) fn walk_form_fields(
         .ok()
         .and_then(|o| o.as_reference().ok())
         .and_then(|p| page_map.get(&p).copied())
+        .or_else(|| annotation_pages.get(&field_id).copied())
         .unwrap_or(1);
 
     let text = if full_name.is_empty() {
@@ -324,3 +457,264 @@ pub(crate) fn walk_form_fields(
         mcid: None,
     });
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use lopdf::{dictionary, Object};
+
+    #[test]
+    fn widget_without_page_reference_uses_owning_page_annotation() {
+        let mut doc = Document::new();
+        let widget_id = doc.add_object(dictionary! {
+            "Type" => "Annot",
+            "Subtype" => "Widget",
+            "FT" => "Tx",
+            "T" => Object::string_literal("customer"),
+            "V" => Object::string_literal("Alice"),
+            "Rect" => vec![10.into(), 20.into(), 110.into(), 40.into()],
+        });
+        let page_one_id = doc.add_object(dictionary! {
+            "Type" => "Page",
+        });
+        let page_two_id = doc.add_object(dictionary! {
+            "Type" => "Page",
+            "Annots" => vec![Object::Reference(widget_id)],
+        });
+        let catalog_id = doc.add_object(dictionary! {
+            "Type" => "Catalog",
+            "AcroForm" => dictionary! {
+                "Fields" => vec![Object::Reference(widget_id)],
+            },
+        });
+        doc.trailer.set("Root", Object::Reference(catalog_id));
+
+        let page_map = HashMap::from([(page_one_id, 1), (page_two_id, 2)]);
+        let items = extract_form_fields(&doc, &page_map);
+
+        assert_eq!(items.len(), 1);
+        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/reading_order.rs

@@ -0,0 +1,591 @@
+//! Region-graph evidence for page reading order.
+//!
+//! Whole-page column histograms fail when images or spanning captions occupy
+//! only part of a page. This module turns image geometry and repeated row
+//! gutters into a small directed acyclic graph: content above a local column
+//! band, the left flow, the right flow, and content below it. The graph is
+//! deliberately evidence-gated; ordinary pages keep the established layout
+//! path.
+
+use crate::text_utils::{effective_width, is_cjk_char, is_rtl_text};
+use crate::types::TextItem;
+
+const MIN_IMAGE_WIDTH: f32 = 60.0;
+const MIN_IMAGE_HEIGHT: f32 = 40.0;
+const MIN_ROW_GUTTER: f32 = 8.0;
+const SPLIT_CLUSTER_TOLERANCE: f32 = 20.0;
+const MIN_ALIGNED_ROWS: usize = 4;
+
+pub(crate) type ImageRegion = (f32, f32, f32, f32);
+
+#[derive(Debug, Clone, Copy, PartialEq)]
+pub(crate) struct ColumnFlowBand {
+    pub(crate) split_x: f32,
+    pub(crate) y_bottom: f32,
+    pub(crate) y_top: f32,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub(crate) enum RegionKind {
+    FullWidth,
+    Column,
+}
+
+#[derive(Debug)]
+pub(crate) struct RegionNode {
+    pub(crate) kind: RegionKind,
+    pub(crate) items: Vec<TextItem>,
+}
+
+#[derive(Debug)]
+struct Row<'a> {
+    y: f32,
+    items: Vec<&'a TextItem>,
+}
+
+fn page_x_bounds(items: &[TextItem], images: &[ImageRegion]) -> Option<(f32, f32)> {
+    let text_min = items
+        .iter()
+        .map(|item| item.x)
+        .fold(f32::INFINITY, f32::min);
+    let text_max = items
+        .iter()
+        .map(|item| item.x + effective_width(item))
+        .fold(f32::NEG_INFINITY, f32::max);
+    let image_min = images
+        .iter()
+        .map(|region| region.0.min(region.2))
+        .fold(f32::INFINITY, f32::min);
+    let image_max = images
+        .iter()
+        .map(|region| region.0.max(region.2))
+        .fold(f32::NEG_INFINITY, f32::max);
+    let x_min = text_min.min(image_min);
+    let x_max = text_max.max(image_max);
+    (x_min.is_finite() && x_max.is_finite() && x_max > x_min).then_some((x_min, x_max))
+}
+
+fn group_rows(items: &[TextItem]) -> Vec<Row<'_>> {
+    const Y_TOLERANCE: f32 = 3.0;
+    let mut sorted: Vec<&TextItem> = items.iter().collect();
+    sorted.sort_by(|left, right| right.y.total_cmp(&left.y));
+    let mut rows: Vec<Row<'_>> = Vec::new();
+    for item in sorted {
+        if let Some(row) = rows
+            .last_mut()
+            .filter(|row| (row.y - item.y).abs() <= Y_TOLERANCE)
+        {
+            row.items.push(item);
+            row.y = row.items.iter().map(|member| member.y).sum::<f32>() / row.items.len() as f32;
+        } else {
+            rows.push(Row {
+                y: item.y,
+                items: vec![item],
+            });
+        }
+    }
+    for row in &mut rows {
+        row.items.sort_by(|left, right| left.x.total_cmp(&right.x));
+    }
+    rows
+}
+
+fn side_is_prose(items: &[&TextItem]) -> bool {
+    let text = items
+        .iter()
+        .map(|item| item.text.trim())
+        .collect::<Vec<_>>()
+        .join(" ");
+    let alphabetic_count = text
+        .chars()
+        .filter(|character| character.is_alphabetic())
+        .count();
+    let cjk_count = text
+        .chars()
+        .filter(|character| is_cjk_char(*character))
+        .count();
+    (text.split_whitespace().count() >= 3 || cjk_count >= 10) && alphabetic_count >= 10
+}
+
+fn aligned_row_split(row: &Row<'_>, x_min: f32, x_max: f32) -> Option<f32> {
+    if row.items.len() < 2 {
+        return None;
+    }
+    let page_width = x_max - x_min;
+    let center_low = x_min + page_width * 0.25;
+    let center_high = x_min + page_width * 0.75;
+    row.items
+        .windows(2)
+        .filter_map(|pair| {
+            let left_end = pair[0].x + effective_width(pair[0]);
+            let right_start = pair[1].x;
+            let gap = right_start - left_end;
+            let split_x = (left_end + right_start) / 2.0;
+            if gap < MIN_ROW_GUTTER || split_x < center_low || split_x > center_high {
+                return None;
+            }
+            let left: Vec<&TextItem> = row
+                .items
+                .iter()
+                .copied()
+                .filter(|item| item.x + effective_width(item) / 2.0 < split_x)
+                .collect();
+            let right: Vec<&TextItem> = row
+                .items
+                .iter()
+                .copied()
+                .filter(|item| item.x + effective_width(item) / 2.0 >= split_x)
+                .collect();
+            (side_is_prose(&left) && side_is_prose(&right)).then_some((split_x, gap))
+        })
+        .max_by(|left, right| left.1.total_cmp(&right.1))
+        .map(|candidate| candidate.0)
+}
+
+fn local_flow_below_full_width_image(
+    items: &[TextItem],
+    images: &[ImageRegion],
+    x_min: f32,
+    x_max: f32,
+) -> Option<ColumnFlowBand> {
+    let page_width = x_max - x_min;
+    let full_width_images: Vec<ImageRegion> = images
+        .iter()
+        .copied()
+        .filter(|&(x0, y0, x1, y1)| {
+            let width = (x1 - x0).abs();
+            let height = (y1 - y0).abs();
+            width >= page_width * 0.65 && height >= 60.0
+        })
+        .collect();
+    // A local column flow below an image is only unambiguous for a single,
+    // nearly square hero/figure. Wide report banners and full-page artwork
+    // frequently sit above unrelated page furniture whose aligned labels can
+    // mimic prose columns.
+    if full_width_images.len() != 1 {
+        return None;
+    }
+    let (image_x0, _, image_x1, _) = full_width_images[0];
+    let anchor_width = (image_x1 - image_x0).abs();
+    let anchor_height = (full_width_images[0].3 - full_width_images[0].1).abs();
+    if anchor_width < page_width * 0.85
+        || anchor_height < anchor_width * 0.85
+        || anchor_height > anchor_width * 1.2
+    {
+        return None;
+    }
+    let image_bottom = full_width_images
+        .iter()
+        .map(|&(_, y0, _, y1)| y0.min(y1))
+        .fold(f32::NEG_INFINITY, f32::max);
+    if !image_bottom.is_finite() {
+        return None;
+    }
+
+    let below: Vec<TextItem> = items
+        .iter()
+        .filter(|item| item.y < image_bottom && item.y >= image_bottom - 220.0)
+        .cloned()
+        .collect();
+    let candidates: Vec<(f32, f32)> = group_rows(&below)
+        .into_iter()
+        .filter_map(|row| aligned_row_split(&row, x_min, x_max).map(|split| (split, row.y)))
+        .collect();
+    if candidates.len() < MIN_ALIGNED_ROWS {
+        return None;
+    }
+
+    let mut clusters: Vec<Vec<(f32, f32)>> = Vec::new();
+    for candidate in candidates {
+        if let Some(cluster) = clusters.iter_mut().find(|cluster| {
+            let mean = cluster.iter().map(|entry| entry.0).sum::<f32>() / cluster.len() as f32;
+            (mean - candidate.0).abs() <= SPLIT_CLUSTER_TOLERANCE
+        }) {
+            cluster.push(candidate);
+        } else {
+            clusters.push(vec![candidate]);
+        }
+    }
+    let dominant = clusters.into_iter().max_by_key(Vec::len)?;
+    if dominant.len() < MIN_ALIGNED_ROWS {
+        return None;
+    }
+    let split_x = dominant.iter().map(|entry| entry.0).sum::<f32>() / dominant.len() as f32;
+    let y_top = dominant
+        .iter()
+        .map(|entry| entry.1)
+        .fold(f32::NEG_INFINITY, f32::max)
+        + 3.0;
+    let image_gap = image_bottom - y_top;
+    if !(60.0..=120.0).contains(&image_gap) {
+        return None;
+    }
+    let y_bottom = dominant
+        .iter()
+        .map(|entry| entry.1)
+        .fold(f32::INFINITY, f32::min)
+        - 3.0;
+    if y_top - y_bottom > 130.0 {
+        return None;
+    }
+    log::debug!(
+        "page {}: full-width image flow images={} aligned_rows={} split={:.1} page=[{:.1}..{:.1}] image_bottom={:.1} y=[{:.1}..{:.1}] full_width={:?}",
+        items.first().map_or(0, |item| item.page),
+        images.len(),
+        dominant.len(),
+        split_x,
+        x_min,
+        x_max,
+        image_bottom,
+        y_bottom,
+        y_top,
+        full_width_images
+    );
+    Some(ColumnFlowBand {
+        split_x,
+        y_bottom,
+        y_top,
+    })
+}
+
+fn paired_column_images(
+    items: &[TextItem],
+    images: &[ImageRegion],
+    split_x: f32,
+    x_min: f32,
+    x_max: f32,
+) -> Option<ColumnFlowBand> {
+    let page_width = x_max - x_min;
+    if split_x < x_min + page_width * 0.4 || split_x > x_min + page_width * 0.6 {
+        return None;
+    }
+    let qualifying: Vec<ImageRegion> = images
+        .iter()
+        .copied()
+        .filter(|&(x0, y0, x1, y1)| {
+            let image_left = x0.min(x1);
+            let image_right = x0.max(x1);
+            let confined_to_one_column = image_right <= split_x || image_left >= split_x;
+            confined_to_one_column
+                && (x1 - x0).abs() >= MIN_IMAGE_WIDTH
+                && (y1 - y0).abs() >= MIN_IMAGE_HEIGHT
+        })
+        .collect();
+    let wide_images: Vec<ImageRegion> = qualifying
+        .iter()
+        .copied()
+        .filter(|(x0, _, x1, _)| (x1 - x0).abs() >= page_width * 0.35)
+        .collect();
+    if qualifying.len() < 3 || wide_images.len() < 3 {
+        return None;
+    }
+    let has_left = qualifying
+        .iter()
+        .any(|&(x0, _, x1, _)| (x0 + x1) / 2.0 < split_x);
+    let has_right = qualifying
+        .iter()
+        .any(|&(x0, _, x1, _)| (x0 + x1) / 2.0 >= split_x);
+    if !has_left || !has_right {
+        return None;
+    }
+    // A meaningful image-backed column flow spans multiple vertical panels.
+    // Three same-row header/logo images can otherwise satisfy the image count
+    // and send an ordinary asymmetric page through sequential column order.
+    let image_y_min = wide_images
+        .iter()
+        .map(|region| region.1.min(region.3))
+        .fold(f32::INFINITY, f32::min);
+    let image_y_max = wide_images
+        .iter()
+        .map(|region| region.1.max(region.3))
+        .fold(f32::NEG_INFINITY, f32::max);
+    let has_vertical_stack = wide_images.iter().enumerate().any(|(index, left)| {
+        wide_images.iter().skip(index + 1).any(|right| {
+            let same_side =
+                ((left.0 + left.2) / 2.0 < split_x) == ((right.0 + right.2) / 2.0 < split_x);
+            let left_center = (left.1 + left.3) / 2.0;
+            let right_center = (right.1 + right.3) / 2.0;
+            let left_height = (left.3 - left.1).abs();
+            let right_height = (right.3 - right.1).abs();
+            let vertical_gap = if left.1.max(left.3) < right.1.min(right.3) {
+                right.1.min(right.3) - left.1.max(left.3)
+            } else if right.1.max(right.3) < left.1.min(left.3) {
+                left.1.min(left.3) - right.1.max(right.3)
+            } else {
+                0.0
+            };
+            same_side
+                && (left_center - right_center).abs() >= left_height.min(right_height) * 0.5
+                && vertical_gap <= left_height.max(right_height) * 0.5
+        })
+    });
+    if image_y_max - image_y_min < page_width * 0.45 || !has_vertical_stack {
+        return None;
+    }
+    let y_top = qualifying
+        .iter()
+        .map(|region| region.1.max(region.3))
+        .fold(f32::NEG_INFINITY, f32::max)
+        + 3.0;
+    // Only column-confined text proves the lower extent of the flow. A
+    // spanning heading or caption below the columns must become the trailing
+    // full-width node rather than stretching the column band to the page foot.
+    let y_bottom = items
+        .iter()
+        .filter(|item| {
+            let item_right = item.x + effective_width(item);
+            item.y <= y_top && (item_right <= split_x || item.x >= split_x)
+        })
+        .map(|item| item.y)
+        .fold(f32::INFINITY, f32::min)
+        - 3.0;
+    if !y_bottom.is_finite() {
+        return None;
+    }
+    let distinct_rows = |right: bool| {
+        let mut ys: Vec<f32> = items
+            .iter()
+            .filter(|item| {
+                item.y <= y_top && (item.x + effective_width(item) / 2.0 >= split_x) == right
+            })
+            .map(|item| item.y)
+            .collect();
+        ys.sort_by(|left, right| left.total_cmp(right));
+        ys.dedup_by(|left, right| (*left - *right).abs() <= 3.0);
+        ys.len()
+    };
+    let left_rows = distinct_rows(false);
+    let right_rows = distinct_rows(true);
+    let line_balance = left_rows.min(right_rows) as f32 / left_rows.max(right_rows).max(1) as f32;
+    (left_rows >= 5 && right_rows >= 5 && line_balance < 0.55).then(|| {
+        log::debug!(
+            "page {}: paired-image flow qualifying_images={} rows={}/{} split={:.1} page=[{:.1}..{:.1}] y=[{:.1}..{:.1}] images={:?}",
+            items.first().map_or(0, |item| item.page),
+            qualifying.len(),
+            left_rows,
+            right_rows,
+            split_x,
+            x_min,
+            x_max,
+            y_bottom,
+            y_top,
+            qualifying
+        );
+        ColumnFlowBand {
+            split_x,
+            y_bottom,
+            y_top,
+        }
+    })
+}
+
+pub(crate) fn infer_image_anchored_flow(
+    items: &[TextItem],
+    images: &[ImageRegion],
+    detected_split: Option<f32>,
+) -> Option<ColumnFlowBand> {
+    if items.is_empty() || images.is_empty() {
+        return None;
+    }
+    let (x_min, x_max) = page_x_bounds(items, images)?;
+    detected_split
+        .and_then(|split_x| paired_column_images(items, images, split_x, x_min, x_max))
+        .or_else(|| local_flow_below_full_width_image(items, images, x_min, x_max))
+}
+
+/// Partition a page into the topological order `above -> left -> right -> below`.
+/// These edges encode the reading-order DAG; empty nodes are omitted.
+pub(crate) fn build_region_graph(items: Vec<TextItem>, band: ColumnFlowBand) -> Vec<RegionNode> {
+    let mut above = Vec::new();
+    let mut left = Vec::new();
+    let mut right = Vec::new();
+    let mut below = Vec::new();
+    for item in items {
+        if item.y > band.y_top {
+            above.push(item);
+        } else if item.y < band.y_bottom {
+            below.push(item);
+        } else if item.x + effective_width(&item) / 2.0 < band.split_x {
+            left.push(item);
+        } else {
+            right.push(item);
+        }
+    }
+    let rtl = is_rtl_text(left.iter().chain(right.iter()).map(|item| &item.text));
+    let mut ordered = vec![(RegionKind::FullWidth, above)];
+    if rtl {
+        ordered.push((RegionKind::Column, right));
+        ordered.push((RegionKind::Column, left));
+    } else {
+        ordered.push((RegionKind::Column, left));
+        ordered.push((RegionKind::Column, right));
+    }
+    ordered.push((RegionKind::FullWidth, below));
+    ordered
+        .into_iter()
+        .filter_map(|(kind, items)| (!items.is_empty()).then_some(RegionNode { kind, items }))
+        .collect()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::types::ItemType;
+
+    fn item(text: &str, x: f32, y: f32, width: f32) -> TextItem {
+        TextItem {
+            text: text.into(),
+            x,
+            y,
+            width,
+            height: 11.0,
+            font: "F1".into(),
+            font_size: 11.0,
+            page: 1,
+            is_bold: false,
+            is_italic: false,
+            is_underline: false,
+            is_strikeout: false,
+            item_type: ItemType::Text,
+            mcid: None,
+        }
+    }
+
+    #[test]
+    fn full_width_image_anchors_local_two_column_flow() {
+        let mut items = vec![
+            item("A full width caption", 55.0, 230.0, 430.0),
+            item("A trailing full width heading", 55.0, 80.0, 430.0),
+        ];
+        for index in 0..5 {
+            let y = 170.0 - index as f32 * 14.0;
+            items.push(item("left column prose words", 55.0, y, 210.0));
+            items.push(item("right column prose words", 280.0, y, 210.0));
+        }
+        let images = vec![(55.0, 250.0, 490.0, 680.0)];
+        let band = infer_image_anchored_flow(&items, &images, None).unwrap();
+        assert!((band.split_x - 272.5).abs() < 2.0);
+        let graph = build_region_graph(items, band);
+        assert_eq!(graph.len(), 4);
+        assert_eq!(graph[0].kind, RegionKind::FullWidth);
+        assert_eq!(graph[1].kind, RegionKind::Column);
+        assert_eq!(graph[2].kind, RegionKind::Column);
+        assert_eq!(graph[3].kind, RegionKind::FullWidth);
+        assert_eq!(graph[3].items[0].text, "A trailing full width heading");
+    }
+
+    #[test]
+    fn full_width_image_anchors_cjk_column_flow() {
+        let mut items = Vec::new();
+        for index in 0..5 {
+            let y = 170.0 - index as f32 * 14.0;
+            items.push(item("左栏这是没有空格的正文内容", 55.0, y, 210.0));
+            items.push(item("右栏这是没有空格的正文内容", 280.0, y, 210.0));
+        }
+        let images = vec![(55.0, 250.0, 490.0, 680.0)];
+
+        assert!(infer_image_anchored_flow(&items, &images, None).is_some());
+    }
+
+    #[test]
+    fn paired_images_anchor_unbalanced_column_flows() {
+        let mut items = vec![
+            item("running header", 55.0, 700.0, 430.0),
+            item("trailing full width caption", 55.0, 300.0, 430.0),
+        ];
+        for index in 0..5 {
+            items.push(item(
+                "left prose words",
+                55.0,
+                500.0 - index as f32 * 14.0,
+                200.0,
+            ));
+            items.push(item(
+                "right prose words",
+                280.0,
+                520.0 - index as f32 * 14.0,
+                200.0,
+            ));
+        }
+        for index in 5..12 {
+            items.push(item(
+                "right continuation prose words",
+                280.0,
+                520.0 - index as f32 * 14.0,
+                200.0,
+            ));
+        }
+        let images = vec![
+            (55.0, 530.0, 255.0, 680.0),
+            (55.0, 380.0, 255.0, 530.0),
+            (280.0, 560.0, 490.0, 680.0),
+        ];
+        let band = infer_image_anchored_flow(&items, &images, Some(270.0)).unwrap();
+        let graph = build_region_graph(items, band);
+        assert_eq!(graph[0].kind, RegionKind::FullWidth);
+        assert_eq!(graph[1].kind, RegionKind::Column);
+        assert_eq!(graph[2].kind, RegionKind::Column);
+        assert_eq!(graph[3].kind, RegionKind::FullWidth);
+        assert_eq!(graph[3].items[0].text, "trailing full width caption");
+    }
+
+    #[test]
+    fn rtl_region_graph_reads_right_column_first() {
+        let items = vec![
+            item("A long English report header", 55.0, 250.0, 430.0),
+            item("نص العمود الأيسر", 55.0, 150.0, 180.0),
+            item("نص العمود الأيمن", 300.0, 150.0, 180.0),
+        ];
+        let graph = build_region_graph(
+            items,
+            ColumnFlowBand {
+                split_x: 270.0,
+                y_bottom: 100.0,
+                y_top: 200.0,
+            },
+        );
+
+        assert_eq!(graph.len(), 3);
+        assert_eq!(graph[0].kind, RegionKind::FullWidth);
+        assert!(graph[1].items[0].x > graph[2].items[0].x);
+    }
+
+    #[test]
+    fn paired_header_logos_do_not_anchor_page_columns() {
+        let mut items = Vec::new();
+        for index in 0..7 {
+            items.push(item(
+                "left prose words",
+                55.0,
+                700.0 - index as f32 * 14.0,
+                200.0,
+            ));
+        }
+        for index in 0..30 {
+            items.push(item(
+                "right prose words",
+                280.0,
+                700.0 - index as f32 * 14.0,
+                200.0,
+            ));
+        }
+        let images = vec![
+            (55.0, 720.0, 205.0, 770.0),
+            (60.0, 718.0, 210.0, 768.0),
+            (280.0, 720.0, 450.0, 770.0),
+        ];
+        assert!(infer_image_anchored_flow(&items, &images, Some(270.0)).is_none());
+    }
+
+    #[test]
+    fn wide_banner_does_not_anchor_local_columns() {
+        let mut items = Vec::new();
+        for index in 0..7 {
+            let y = 270.0 - index as f32 * 14.0;
+            items.push(item("left column prose words", 55.0, y, 210.0));
+            items.push(item("right column prose words", 280.0, y, 210.0));
+        }
+        let images = vec![(55.0, 310.0, 490.0, 550.0)];
+        assert!(infer_image_anchored_flow(&items, &images, None).is_none());
+    }
+}

src/extractor/underline.rs

@@ -44,6 +44,16 @@ const MIN_TABULAR_RULE_ITEMS: usize = 3;
 const MIN_TABULAR_RULE_GAPS: usize = 2;
 const TABULAR_RULE_GAP_EM: f32 = 2.0;
 
+/// Strikeout decorations are text-sized. Diagram connectors, signature
+/// lines, and chart rules often cross glyphs too, but extend well beyond the
+/// text they happen to intersect.
+const STRIKE_OWNER_PAD_EM: f32 = 0.75;
+const STRIKE_OWNER_MIN_PAD: f32 = 4.0;
+const STRIKE_ROW_Y_TOLERANCE_EM: f32 = 0.15;
+const STRIKE_ROW_Y_TOLERANCE_MIN: f32 = 5.0;
+const GRAPHIC_CONNECTION_EPS: f32 = 2.0;
+const GRAPHIC_CONNECTOR_MAX_THICKNESS: f32 = 4.0;
+
 #[derive(Clone)]
 pub(crate) struct UnderlineLine {
     pub(crate) x1: f32,
@@ -115,7 +125,13 @@ fn rules_from_graphics(rects: &[PdfRect], lines: &[UnderlineLine], page: u32) ->
     rules
 }
 
-fn discard_repeated_ruling_rules(rules: Vec<Rule>) -> Vec<Rule> {
+fn discard_repeated_ruling_rules(
+    rules: Vec<Rule>,
+    items: &[TextItem],
+    rects: &[PdfRect],
+    lines: &[UnderlineLine],
+    page: u32,
+) -> Vec<Rule> {
     if rules.len() < MIN_REPEATED_RULE_LEVELS {
         return rules;
     }
@@ -123,12 +139,142 @@ fn discard_repeated_ruling_rules(rules: Vec<Rule>) -> Vec<Rule> {
     rules
         .iter()
         .filter(|rule| {
-            !is_repeated_ruling_rule(rule, &rules) && !is_segmented_row_ruling_rule(rule, &rules)
+            // A rule snugly owned by one text line is an underline even when
+            // span-similar rules repeat down the page — documents that
+            // underline many full-width lines (dense CJK business docs) look
+            // exactly like table rulings to the repetition check, which used
+            // to discard every one of them. Table rulings fail snugness:
+            // row separators extend past their cells' text (or have no text
+            // on the baseline above), and multi-column matches are still
+            // culled by the tabular filter afterwards.
+            // Same-row segmented rules (column-header separators) are
+            // always rulings — each segment snugly owns its column label,
+            // so snugness must not override that check.
+            !is_segmented_row_ruling_rule(rule, &rules)
+                && ((has_snug_text_owner(rule, items)
+                    && !has_flanking_verticals(rule, rects, lines, page))
+                    || !is_repeated_ruling_rule(rule, &rules))
         })
         .cloned()
         .collect()
 }
 
+/// True when a single text item both matches the rule vertically (baseline
+/// window) and horizontally contains it: the rule may not extend past the
+/// item's span by more than ~0.75em on either side. Underlines are drawn to
+/// the width of the text they decorate; table/form rulings span cells or
+/// full table width and overshoot any single item.
+/// A rule flanked by vertical strokes at its ends is a table/box border
+/// row edge, not an underline — underlined text lines have no vertical
+/// rules rising from their ends. Checked against raw stroked lines: a
+/// near-vertical segment whose x sits at either end of the rule and whose
+/// y-range covers the rule's row.
+fn has_flanking_verticals(
+    rule: &Rule,
+    rects: &[PdfRect],
+    lines: &[UnderlineLine],
+    page: u32,
+) -> bool {
+    // A drawn rect that CONTAINS the rule vetoes rescue only with GRID
+    // EVIDENCE: another drawn rect abutting it vertically (cell rows tile).
+    // Height alone can't separate a table cell from a decorative callout
+    // panel — genuine underlines live inside isolated filled panels, and
+    // multiline table cells can be arbitrarily tall.
+    let norm = |r: &PdfRect| {
+        let (x_lo, x_hi) = if r.width >= 0.0 {
+            (r.x, r.x + r.width)
+        } else {
+            (r.x + r.width, r.x)
+        };
+        let (y_lo, y_hi) = if r.height >= 0.0 {
+            (r.y, r.y + r.height)
+        } else {
+            (r.y + r.height, r.y)
+        };
+        (x_lo, x_hi, y_lo, y_hi)
+    };
+    let page_rects: Vec<(f32, f32, f32, f32)> = rects
+        .iter()
+        .filter(|r| r.page == page && r.height.abs() > 6.0)
+        .map(norm)
+        .collect();
+    let rect_flank = page_rects.iter().any(|&(x_lo, x_hi, y_lo, y_hi)| {
+        let contains = x_lo <= rule.x1 + 2.0
+            && x_hi >= rule.x2 - 2.0
+            && y_lo <= rule.y + 2.0
+            && y_hi >= rule.y - 2.0;
+        if !contains {
+            return false;
+        }
+        // Grid evidence: a vertically abutting neighbor box with x-overlap.
+        page_rects.iter().any(|&(nx_lo, nx_hi, ny_lo, ny_hi)| {
+            let x_overlap = nx_hi.min(x_hi) - nx_lo.max(x_lo);
+            if x_overlap <= 10.0 {
+                return false;
+            }
+            (ny_lo - y_hi).abs() <= 3.0 || (y_lo - ny_hi).abs() <= 3.0
+        })
+    });
+    if rect_flank {
+        return true;
+    }
+    lines.iter().any(|l| {
+        if l.page != page || (l.x1 - l.x2).abs() > 2.0 {
+            return false;
+        }
+        let x = (l.x1 + l.x2) / 2.0;
+        let near_end = (x - rule.x1).abs() <= 6.0 || (x - rule.x2).abs() <= 6.0;
+        if !near_end {
+            return false;
+        }
+        let (y_lo, y_hi) = if l.y1 <= l.y2 {
+            (l.y1, l.y2)
+        } else {
+            (l.y2, l.y1)
+        };
+        y_lo <= rule.y + 2.0 && y_hi >= rule.y - 2.0
+    })
+}
+
+fn has_snug_text_owner(rule: &Rule, items: &[TextItem]) -> bool {
+    // Underlines are drawn to the width of the text they decorate, but the
+    // text may be split into several runs (CJK lines mix scripts and font
+    // switches) — so ownership is judged against the UNION of the runs on
+    // the rule's baseline row. Table/form rulings overshoot their row's
+    // text (row separators span cell padding and empty columns), so they
+    // fail either containment or coverage.
+    let matched: Vec<&TextItem> = items
+        .iter()
+        .filter(|item| is_underline_candidate(item) && rule_matches_item(rule, item))
+        .collect();
+    if matched.is_empty() {
+        return false;
+    }
+    let x1 = matched.iter().map(|i| i.x).fold(f32::INFINITY, f32::min);
+    let x2 = matched
+        .iter()
+        .map(|i| i.x + i.width)
+        .fold(f32::NEG_INFINITY, f32::max);
+    let max_fs = matched.iter().map(|i| i.font_size).fold(0.0, f32::max);
+    let pad = (max_fs * 0.75).max(4.0);
+    if rule.x1 < x1 - pad || rule.x2 > x2 + pad {
+        return false;
+    }
+    let covered: f32 = matched.iter().map(|i| i.width).sum();
+    if covered < rule.width() * 0.6 {
+        return false;
+    }
+    // A table row also unions to the rule's span — but its cells sit apart.
+    // An underlined text line is contiguous runs with word-sized gaps; any
+    // column-sized hole between matched runs means this is a row ruling.
+    let mut sorted = matched;
+    sorted.sort_by(|a, b| a.x.total_cmp(&b.x));
+    sorted.windows(2).all(|pair| {
+        let gap = pair[1].x - (pair[0].x + pair[0].width);
+        gap <= (max_fs * 2.0).max(12.0)
+    })
+}
+
 fn is_repeated_ruling_rule(rule: &Rule, rules: &[Rule]) -> bool {
     let mut y_levels: Vec<f32> = rules
         .iter()
@@ -215,9 +361,11 @@ fn is_underline_candidate(item: &TextItem) -> bool {
 
 fn rule_matches_item(rule: &Rule, item: &TextItem) -> bool {
     // Vertical window: underlines sit at or slightly below the baseline.
-    // Fonts draw them at roughly 5-15% of the em below; allow up to 35%
-    // (min 3pt) below and 1pt above for rounding.
-    let below = (item.font_size * 0.35).max(3.0);
+    // Latin fonts draw them at roughly 5-15% of the em below; CJK layouts
+    // put them under the full em box, measured up to ~0.67em below the
+    // baseline (text_dense__underline). Allow 0.72em (min 3pt) below and
+    // 1pt above for rounding.
+    let below = (item.font_size * 0.72).max(3.0);
     let y_min = item.y - below;
     let y_max = item.y + 1.0;
     if rule.y < y_min || rule.y > y_max {
@@ -249,6 +397,206 @@ fn rule_strikes_item(rule: &Rule, item: &TextItem) -> bool {
     overlap >= min_overlap
 }
 
+fn is_bare_list_marker(text: &str) -> bool {
+    matches!(
+        text.trim(),
+        "•" | "◦" | "▪" | "▫" | "‣" | "⁃" | "●" | "○" | "■" | "□" | "-" | "*"
+    )
+}
+
+fn same_strike_row(left: &TextItem, right: &TextItem) -> bool {
+    let font_size = left.font_size.max(right.font_size);
+    let tolerance = (font_size * STRIKE_ROW_Y_TOLERANCE_EM).max(STRIKE_ROW_Y_TOLERANCE_MIN);
+    (left.y - right.y).abs() <= tolerance
+}
+
+fn is_inline_script(rule: &Rule, candidate: &TextItem, parent: &TextItem) -> bool {
+    if !is_underline_candidate(candidate)
+        || is_bare_list_marker(&candidate.text)
+        || candidate.font_size <= 0.0
+        || candidate.font_size >= parent.font_size * 0.75
+        || candidate.text.len() > 4
+        || !candidate.text.chars().all(|c| c.is_ascii_digit())
+        || (candidate.y - parent.y).abs() > 5.0
+    {
+        return false;
+    }
+
+    let parent_ends_with_letter = parent.text.chars().last().is_some_and(char::is_alphabetic);
+    if !parent_ends_with_letter {
+        return false;
+    }
+
+    let parent_right = parent.x + parent.width;
+    let gap = candidate.x - parent_right;
+    if gap >= parent.font_size * 0.2 || gap <= -parent.font_size * 0.3 {
+        return false;
+    }
+
+    let overlap = rule.x2.min(candidate.x + candidate.width) - rule.x1.max(candidate.x);
+    overlap >= candidate.width * MIN_X_OVERLAP
+}
+
+/// Return the items owned by a snug mid-glyph rule.
+///
+/// Real strikeout decorations track the width of the deleted text, including
+/// runs split by font/style changes and adjacent numeric super/subscripts.
+/// Non-text graphics can cross the same vertical window, but arrow shafts,
+/// signature lines, fraction bars, and chart rules extend materially beyond
+/// the intersected glyphs. Requiring the rule to stay within a small em-sized
+/// pad of a contiguous matched row separates those cases without relying on
+/// document-specific fonts or coordinates.
+///
+/// Ownership is computed once per rule. This keeps the strikeout pass at the
+/// same rule-by-item scale as underline detection instead of rescanning the
+/// whole page for every matching item.
+fn snug_strike_owner_indices(rule: &Rule, items: &[TextItem]) -> Vec<usize> {
+    let mut struck_indices: Vec<usize> = items
+        .iter()
+        .enumerate()
+        .filter_map(|(index, candidate)| {
+            (is_underline_candidate(candidate)
+                && !is_bare_list_marker(&candidate.text)
+                && rule_strikes_item(rule, candidate))
+            .then_some(index)
+        })
+        .collect();
+    if struck_indices.is_empty() {
+        return Vec::new();
+    }
+    struck_indices.sort_by(|&left, &right| items[left].y.total_cmp(&items[right].y));
+
+    let mut rows: Vec<Vec<usize>> = Vec::new();
+    for index in struck_indices {
+        if let Some(row) = rows
+            .last_mut()
+            .filter(|row| same_strike_row(&items[row[0]], &items[index]))
+        {
+            row.push(index);
+        } else {
+            rows.push(vec![index]);
+        }
+    }
+
+    let mut owned_indices = Vec::new();
+    for mut row in rows {
+        row.sort_by(|&left, &right| items[left].x.total_cmp(&items[right].x));
+
+        // Underline detection runs before the extractor's script-merging
+        // pass. Include the same tightly adjacent numeric script shape here
+        // when the rule spans it, so the owner width and semantic mark both
+        // survive that later merge.
+        let scripts: Vec<usize> = items
+            .iter()
+            .enumerate()
+            .filter_map(|(index, candidate)| {
+                let parent_pos =
+                    row.partition_point(|&row_index| items[row_index].x <= candidate.x);
+                let parent_index = parent_pos.checked_sub(1).map(|pos| row[pos])?;
+                is_inline_script(rule, candidate, &items[parent_index]).then_some(index)
+            })
+            .collect();
+        row.extend(scripts);
+        row.sort_by(|&left, &right| items[left].x.total_cmp(&items[right].x));
+        row.dedup();
+
+        let x1 = row
+            .iter()
+            .map(|&index| items[index].x)
+            .fold(f32::INFINITY, f32::min);
+        let x2 = row
+            .iter()
+            .map(|&index| items[index].x + items[index].width)
+            .fold(f32::NEG_INFINITY, f32::max);
+        let max_font_size = row
+            .iter()
+            .map(|&index| items[index].font_size)
+            .fold(0.0, f32::max);
+        let pad = (max_font_size * STRIKE_OWNER_PAD_EM).max(STRIKE_OWNER_MIN_PAD);
+
+        if rule.x1 < x1 - pad || rule.x2 > x2 + pad {
+            continue;
+        }
+
+        let contiguous = row.windows(2).all(|pair| {
+            let gap = items[pair[1]].x - (items[pair[0]].x + items[pair[0]].width);
+            gap <= (max_font_size * 2.0).max(12.0)
+        });
+        if contiguous {
+            owned_indices.extend(row);
+        }
+    }
+
+    owned_indices.sort_unstable();
+    owned_indices.dedup();
+    owned_indices
+}
+
+/// Diagram and table rules participate in larger path geometry. A vertical
+/// or diagonal segment meeting the candidate rule is strong evidence that
+/// the horizontal segment is a connector, border, arrow, or symbol rather
+/// than an isolated text decoration.
+fn has_connected_nonhorizontal_segment(rule: &Rule, lines: &[UnderlineLine], page: u32) -> bool {
+    lines.iter().any(|line| {
+        if line.page != page {
+            return false;
+        }
+
+        let dx = line.x2 - line.x1;
+        let dy = line.y2 - line.y1;
+        if dy.abs() <= MAX_RULE_THICKNESS {
+            return false;
+        }
+
+        let y_min = line.y1.min(line.y2) - GRAPHIC_CONNECTION_EPS;
+        let y_max = line.y1.max(line.y2) + GRAPHIC_CONNECTION_EPS;
+        if rule.y < y_min || rule.y > y_max {
+            return false;
+        }
+
+        let t = (rule.y - line.y1) / dy;
+        if !(-0.05..=1.05).contains(&t) {
+            return false;
+        }
+        let intersection_x = line.x1 + t * dx;
+        intersection_x >= rule.x1 - GRAPHIC_CONNECTION_EPS
+            && intersection_x <= rule.x2 + GRAPHIC_CONNECTION_EPS
+    })
+}
+
+/// Filled diagrams often build connectors from intersecting thin rectangles
+/// instead of stroked path segments. Treat only narrow, vertically elongated
+/// rectangles as connector geometry; broad fills can legitimately sit behind
+/// struck text and must not veto its decoration.
+fn has_connected_nonhorizontal_rect(rule: &Rule, rects: &[PdfRect], page: u32) -> bool {
+    rects.iter().any(|rect| {
+        if rect.page != page {
+            return false;
+        }
+
+        let (x1, x2) = if rect.width >= 0.0 {
+            (rect.x, rect.x + rect.width)
+        } else {
+            (rect.x + rect.width, rect.x)
+        };
+        let (y1, y2) = if rect.height >= 0.0 {
+            (rect.y, rect.y + rect.height)
+        } else {
+            (rect.y + rect.height, rect.y)
+        };
+        let width = x2 - x1;
+        let height = y2 - y1;
+
+        width > 0.0
+            && width <= GRAPHIC_CONNECTOR_MAX_THICKNESS
+            && height > width * 2.0
+            && rule.y >= y1 - GRAPHIC_CONNECTION_EPS
+            && rule.y <= y2 + GRAPHIC_CONNECTION_EPS
+            && x2 >= rule.x1 - GRAPHIC_CONNECTION_EPS
+            && x1 <= rule.x2 + GRAPHIC_CONNECTION_EPS
+    })
+}
+
 /// Mark `is_underline` on text items that have a horizontal rule just
 /// below their baseline, and `is_strikeout` on items whose glyphs a rule
 /// crosses at mid x-height. `items`, `rects`, and `lines` are a single
@@ -260,30 +608,85 @@ pub(crate) fn mark_underlined_items(
     lines: &[UnderlineLine],
     page: u32,
 ) {
-    let rules = discard_repeated_ruling_rules(rules_from_graphics(rects, lines, page));
+    let rules = discard_repeated_ruling_rules(
+        rules_from_graphics(rects, lines, page),
+        items,
+        rects,
+        lines,
+        page,
+    );
     if rules.is_empty() {
         return;
     }
     let tabular_rules = tabular_row_separator_rule_indices(&rules, items);
 
-    for item in items.iter_mut() {
+    // Math fraction bars are short horizontal lines with the numerator just
+    // above AND the denominator just below — underline geometry from above,
+    // but no underline has text hanging directly beneath it at fraction
+    // distance. Only narrow rules qualify: real underlines under short
+    // labels have their next text line a full line-pitch away.
+    let fraction_rules: HashSet<usize> = rules
+        .iter()
+        .enumerate()
+        .filter(|(_, rule)| {
+            rule.width() <= 60.0
+                && items.iter().any(|item| {
+                    if !is_underline_candidate(item) {
+                        return false;
+                    }
+                    // A denominator HUGS the bar (fraction typesetting
+                    // leaves ~0.1-0.2em) and is bar-sized. Both bounds
+                    // matter: a short last-line of a paragraph at normal
+                    // leading sits further below, and a full next text
+                    // line is far wider than the rule.
+                    let dy = rule.y - (item.y + item.height);
+                    let overlap = rule.x2.min(item.x + item.width) - rule.x1.max(item.x);
+                    dy > 0.0
+                        && dy <= item.font_size * 0.3
+                        && overlap > rule.width() * 0.5
+                        && item.width <= rule.width() * 1.5
+                })
+        })
+        .map(|(i, _)| i)
+        .collect();
+
+    let mut strikeout_items = vec![false; items.len()];
+    for (rule_idx, rule) in rules.iter().enumerate() {
+        if tabular_rules.contains(&rule_idx)
+            || has_connected_nonhorizontal_segment(rule, lines, page)
+            || has_connected_nonhorizontal_rect(rule, rects, page)
+        {
+            continue;
+        }
+        for item_idx in snug_strike_owner_indices(rule, items) {
+            strikeout_items[item_idx] = true;
+        }
+    }
+
+    let underlined_items: HashSet<usize> = items
+        .iter()
+        .enumerate()
+        .filter(|(_, item)| {
+            is_underline_candidate(item)
+                && rules.iter().enumerate().any(|(rule_idx, rule)| {
+                    !tabular_rules.contains(&rule_idx)
+                        && !fraction_rules.contains(&rule_idx)
+                        && rule_matches_item(rule, item)
+                })
+        })
+        .map(|(item_idx, _)| item_idx)
+        .collect();
+
+    for (item_idx, item) in items.iter_mut().enumerate() {
         if !is_underline_candidate(item) {
             continue;
         }
 
-        for (rule_idx, rule) in rules.iter().enumerate() {
-            if tabular_rules.contains(&rule_idx) {
-                continue;
-            }
-            if rule_matches_item(rule, item) {
-                item.is_underline = true;
-            }
-            if rule_strikes_item(rule, item) {
-                item.is_strikeout = true;
-            }
-            if item.is_underline && item.is_strikeout {
-                break;
-            }
+        if strikeout_items[item_idx] {
+            item.is_strikeout = true;
+        }
+        if underlined_items.contains(&item_idx) {
+            item.is_underline = true;
         }
     }
 }
@@ -323,6 +726,16 @@ mod tests {
         }
     }
 
+    fn cell_rect(x: f32, y: f32, width: f32, height: f32) -> PdfRect {
+        PdfRect {
+            x,
+            y,
+            width,
+            height,
+            page: 1,
+        }
+    }
+
     fn thin_rect(x: f32, y: f32, width: f32) -> PdfRect {
         PdfRect {
             x,
@@ -393,6 +806,148 @@ mod tests {
         assert!(!items[0].is_underline);
     }
 
+    #[test]
+    fn long_connector_crossing_text_is_not_a_strikeout() {
+        let mut items = vec![item("diagram label", 160.0, 500.0, 60.0, 10.0)];
+        let lines = vec![hline(100.0, 280.0, 503.0)];
+
+        mark_underlined_items(&mut items, &[], &lines, 1);
+
+        assert!(!items[0].is_strikeout);
+    }
+
+    #[test]
+    fn chart_rule_ending_inside_short_label_is_not_a_strikeout() {
+        let mut items = vec![item("T 18", 300.0, 500.0, 20.0, 10.0)];
+        let lines = vec![hline(100.0, 315.0, 503.0)];
+
+        mark_underlined_items(&mut items, &[], &lines, 1);
+
+        assert!(!items[0].is_strikeout);
+    }
+
+    #[test]
+    fn connected_diagram_segment_is_not_a_strikeout() {
+        let mut items = vec![item("V8", 100.0, 500.0, 12.0, 10.0)];
+        let lines = vec![
+            hline(99.0, 113.0, 503.0),
+            UnderlineLine {
+                x1: 106.0,
+                y1: 496.0,
+                x2: 109.0,
+                y2: 510.0,
+                stroke_width: 1.0,
+                page: 1,
+            },
+        ];
+
+        mark_underlined_items(&mut items, &[], &lines, 1);
+
+        assert!(!items[0].is_strikeout);
+    }
+
+    #[test]
+    fn connected_filled_rect_is_not_a_strikeout() {
+        let mut items = vec![item("V8", 100.0, 500.0, 12.0, 10.0)];
+        let rects = vec![
+            thin_rect(99.0, 502.6, 14.0),
+            PdfRect {
+                x: 106.0,
+                y: 496.0,
+                width: 2.0,
+                height: 14.0,
+                page: 1,
+            },
+        ];
+
+        mark_underlined_items(&mut items, &rects, &[], 1);
+
+        assert!(!items[0].is_strikeout);
+    }
+
+    #[test]
+    fn broad_fill_behind_text_does_not_block_strikeout() {
+        let mut items = vec![item("deleted", 100.0, 500.0, 40.0, 10.0)];
+        let rects = vec![
+            thin_rect(99.0, 502.6, 42.0),
+            PdfRect {
+                x: 90.0,
+                y: 490.0,
+                width: 100.0,
+                height: 20.0,
+                page: 1,
+            },
+        ];
+
+        mark_underlined_items(&mut items, &rects, &[], 1);
+
+        assert!(items[0].is_strikeout);
+    }
+
+    #[test]
+    fn list_bullet_is_not_a_strikeout() {
+        for marker in ["•", "-", "*"] {
+            let mut items = vec![item(marker, 100.0, 500.0, 6.0, 10.0)];
+            let lines = vec![hline(99.0, 107.0, 503.0)];
+
+            mark_underlined_items(&mut items, &[], &lines, 1);
+
+            assert!(!items[0].is_strikeout, "marker {marker:?}");
+        }
+    }
+
+    #[test]
+    fn snug_rule_marks_adjacent_split_runs_as_strikeout() {
+        let mut items = vec![
+            item("deleted", 100.0, 500.0, 40.0, 10.0),
+            item("text", 142.0, 500.0, 25.0, 10.0),
+        ];
+        let lines = vec![hline(99.0, 168.0, 503.0)];
+
+        mark_underlined_items(&mut items, &[], &lines, 1);
+
+        assert!(items.iter().all(|item| item.is_strikeout));
+    }
+
+    #[test]
+    fn snug_rule_groups_split_runs_with_baseline_drift() {
+        let mut items = vec![
+            item("deleted", 100.0, 500.0, 40.0, 10.0),
+            item("text", 142.0, 498.0, 25.0, 10.0),
+        ];
+        let lines = vec![hline(99.0, 168.0, 503.0)];
+
+        mark_underlined_items(&mut items, &[], &lines, 1);
+
+        assert!(items.iter().all(|item| item.is_strikeout));
+    }
+
+    #[test]
+    fn snug_rule_keeps_inline_superscript_in_strike_owner() {
+        let mut items = vec![
+            item("deleted", 100.0, 500.0, 40.0, 10.0),
+            item("2", 140.5, 503.0, 4.0, 6.0),
+        ];
+        let lines = vec![hline(99.0, 145.0, 503.0)];
+
+        mark_underlined_items(&mut items, &[], &lines, 1);
+
+        assert!(items.iter().all(|item| item.is_strikeout));
+    }
+
+    #[test]
+    fn snug_rule_keeps_inline_subscript_in_strike_owner() {
+        let mut items = vec![
+            item("deleted", 100.0, 500.0, 40.0, 10.0),
+            item("2", 140.5, 497.0, 4.0, 6.0),
+        ];
+        let lines = vec![hline(99.0, 145.0, 503.0)];
+
+        mark_underlined_items(&mut items, &[], &lines, 1);
+
+        assert!(items.iter().all(|item| item.is_strikeout));
+    }
+
     #[test]
     fn baseline_rule_marks_underline_not_strikeout() {
         let mut items = vec![item("underlined", 100.0, 500.0, 60.0, 10.0)];
@@ -543,6 +1098,100 @@ mod tests {
         assert!(items.iter().all(|item| !item.is_underline));
     }
 
+    #[test]
+    fn repeated_snug_underlines_survive_ruling_filter() {
+        // Dense docs underline many full-width lines: span-similar rules at
+        // 3+ y-levels used to be discarded wholesale as table rulings.
+        // Each rule here snugly matches one text line, so all must mark.
+        let mut items = vec![
+            item("first underlined line of text", 50.0, 700.0, 300.0, 11.0),
+            item("second underlined line here", 50.0, 650.0, 300.0, 11.0),
+            item("third underlined line as well", 50.0, 600.0, 300.0, 11.0),
+        ];
+        let lines = vec![
+            hline(50.0, 350.0, 697.0),
+            hline(50.0, 350.0, 647.0),
+            hline(50.0, 350.0, 597.0),
+        ];
+
+        mark_underlined_items(&mut items, &[], &lines, 1);
+
+        assert!(items.iter().all(|item| item.is_underline));
+    }
+
+    #[test]
+    fn snug_rescue_spans_split_runs_on_one_line() {
+        // A single underlined line is often split into several runs (script
+        // or font switches). The union of touching runs owns the rule.
+        let mut items = vec![
+            item("run one", 50.0, 700.0, 100.0, 11.0),
+            item("run two", 150.5, 700.0, 100.0, 11.0),
+            item("run three", 251.0, 700.0, 99.0, 11.0),
+            item("other a", 50.0, 650.0, 300.0, 11.0),
+            item("other b", 50.0, 600.0, 300.0, 11.0),
+        ];
+        let lines = vec![
+            hline(50.0, 350.0, 697.0),
+            hline(50.0, 350.0, 647.0),
+            hline(50.0, 350.0, 597.0),
+        ];
+
+        mark_underlined_items(&mut items, &[], &lines, 1);
+
+        assert!(items[0].is_underline && items[1].is_underline && items[2].is_underline);
+    }
+
+    #[test]
+    fn snug_rescue_denied_for_row_with_cell_gaps() {
+        // A full-width rule whose baseline row is several items separated by
+        // column-sized gaps is a table row separator, not an underline —
+        // even when span-similar rules repeat down the page.
+        let mut items = vec![
+            item("cell a", 50.0, 700.0, 60.0, 11.0),
+            item("cell b", 190.0, 700.0, 60.0, 11.0),
+            item("cell c", 330.0, 700.0, 70.0, 11.0),
+            item("cell d", 50.0, 650.0, 60.0, 11.0),
+            item("cell e", 190.0, 650.0, 60.0, 11.0),
+            item("cell f", 330.0, 650.0, 70.0, 11.0),
+        ];
+        let lines = vec![
+            hline(50.0, 400.0, 697.0),
+            hline(50.0, 400.0, 647.0),
+            hline(50.0, 400.0, 597.0),
+        ];
+
+        mark_underlined_items(&mut items, &[], &lines, 1);
+
+        assert!(items.iter().all(|item| !item.is_underline));
+    }
+
+    #[test]
+    fn snug_rescue_denied_inside_cell_box() {
+        // A rule snugly under one text line but enclosed by a drawn cell
+        // box that TILES with vertical neighbors (grid evidence) is a row
+        // ruling of a rect-grid table. Isolated boxes (callout panels) do
+        // not veto — see repeated_snug_underlines_survive_ruling_filter.
+        let mut items = vec![
+            item("one wide cell row", 50.0, 700.0, 300.0, 11.0),
+            item("second wide cell", 50.0, 650.0, 300.0, 11.0),
+            item("third wide cell", 50.0, 600.0, 300.0, 11.0),
+        ];
+        let lines = vec![
+            hline(50.0, 350.0, 697.0),
+            hline(50.0, 350.0, 647.0),
+            hline(50.0, 350.0, 597.0),
+        ];
+        let boxes = vec![
+            cell_rect(45.0, 690.0, 320.0, 50.0),
+            cell_rect(45.0, 640.0, 320.0, 50.0),
+            cell_rect(45.0, 590.0, 320.0, 50.0),
+        ];
+
+        mark_underlined_items(&mut items, &boxes, &lines, 1);
+
+        assert!(items.iter().all(|item| !item.is_underline));
+    }
+
     #[test]
     fn same_row_spaced_rule_segments_do_not_mark_column_labels() {
         let mut items = vec![

src/extractor/xobjects.rs

@@ -8,8 +8,9 @@ use lopdf::{Document, Encoding, Object, ObjectId};
 use std::collections::HashMap;
 
 use super::fonts::{
-    build_font_encodings, build_font_widths, compute_string_width_ts, extract_text_from_operand,
-    get_font_file2_obj_num, get_operand_bytes, CMapDecisionCache, FontStyleCache,
+    build_font_encodings, build_font_widths, build_type3_scales, compute_string_width_ts,
+    extract_text_from_operand, get_font_file2_obj_num, get_operand_bytes, CMapDecisionCache,
+    FontStyleCache,
 };
 use super::{get_number, image_bbox_from_ctm, multiply_matrices};
 
@@ -162,10 +163,11 @@ fn extract_form_xobject_text_inner(
 
     // Get fonts from the Form's Resources
     let form_fonts = get_form_fonts(doc, &stream.dict);
-    let (font_encodings, _has_gid_fonts) = build_font_encodings(doc, &form_fonts);
+    let (font_encodings, _has_gid_fonts) = build_font_encodings(doc, &form_fonts, font_cmaps);
 
     // Build font width info for the form
     let font_widths = build_font_widths(doc, &form_fonts);
+    let type3_scales = build_type3_scales(doc, &form_fonts);
 
     // Build font base names and ToUnicode refs for the form
     let mut font_base_names: HashMap<String, String> = HashMap::new();
@@ -413,7 +415,8 @@ fn extract_form_xobject_text_inner(
                         &font_widths,
                     ) {
                         let combined = multiply_matrices(&text_matrix, &ctm);
-                        let rendered_size = effective_font_size(current_font_size, &combined);
+                        let rendered_size = effective_font_size(current_font_size, &combined)
+                            * 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]) {
@@ -572,7 +575,8 @@ fn extract_form_xobject_text_inner(
                         }
                         if !sub_items.is_empty() {
                             let combined = multiply_matrices(&text_matrix, &ctm);
-                            let rendered_size = effective_font_size(current_font_size, &combined);
+                            let rendered_size = effective_font_size(current_font_size, &combined)
+                                * type3_scales.get(&current_font).copied().unwrap_or(1.0);
                             let base_font = font_base_names
                                 .get(&current_font)
                                 .map(|s| s.as_str())

src/lib.rs

@@ -50,10 +50,11 @@ pub use detector::{
 };
 pub use extractor::{
     extract_text, extract_text_with_positions, extract_text_with_positions_mem,
-    extract_text_with_positions_pages,
+    extract_text_with_positions_pages, extract_text_with_positions_pages_with_password,
 };
 pub use markdown::{
-    to_markdown, to_markdown_from_items, to_markdown_from_items_with_rects, MarkdownOptions,
+    to_markdown, to_markdown_from_items, to_markdown_from_items_with_rects,
+    to_markdown_from_items_with_rects_and_page_count, MarkdownOptions, MarkdownProfile,
 };
 pub use process_mode::ProcessMode;
 pub use types::{LayoutComplexity, PdfLine, PdfRect, TextItem};
@@ -67,6 +68,40 @@ use text_quality::{
 };
 use tounicode::FontCMaps;
 
+#[cfg(not(target_arch = "wasm32"))]
+struct ProcessingTimer(std::time::Instant);
+
+#[cfg(target_arch = "wasm32")]
+struct ProcessingTimer;
+
+impl ProcessingTimer {
+    fn start() -> Self {
+        #[cfg(not(target_arch = "wasm32"))]
+        {
+            Self(std::time::Instant::now())
+        }
+
+        #[cfg(target_arch = "wasm32")]
+        {
+            Self
+        }
+    }
+
+    fn elapsed_ms(&self) -> u64 {
+        #[cfg(not(target_arch = "wasm32"))]
+        {
+            self.0.elapsed().as_millis() as u64
+        }
+
+        #[cfg(target_arch = "wasm32")]
+        {
+            // The wasm32-unknown-unknown standard library has no clock.
+            // Browser bindings measure with JavaScript's host clock.
+            0
+        }
+    }
+}
+
 /// OCR reason emitted when the extracted text layer appears garbled due to
 /// broken font decoding or mojibake.
 pub const OCR_REASON_SUSPECTED_GARBLED_TEXT: &str = "suspected_garbled_text";
@@ -249,7 +284,7 @@ pub fn process_pdf_with_options<P: AsRef<Path>>(
     path: P,
     options: PdfOptions,
 ) -> Result<PdfProcessResult, PdfError> {
-    let start = std::time::Instant::now();
+    let start = ProcessingTimer::start();
     validate_pdf_file(&path)?;
 
     // Load the document once — shared by detection AND extraction.
@@ -276,7 +311,7 @@ pub fn process_pdf_mem_with_options(
     buffer: &[u8],
     options: PdfOptions,
 ) -> Result<PdfProcessResult, PdfError> {
-    let start = std::time::Instant::now();
+    let start = ProcessingTimer::start();
     validate_pdf_bytes(buffer)?;
 
     let (doc, page_count) =
@@ -427,16 +462,46 @@ pub fn extract_pages_markdown_mem(
     let (doc, page_count) = load_document_from_mem(buffer)?;
     let font_cmaps = FontCMaps::from_doc(&doc);
 
-    // Extract ALL pages to get accurate, document-wide font stats.
+    // Extract ALL pages to get accurate, document-wide font stats. A malformed
+    // unselected page cannot make a valid requested page fail, but errors on a
+    // requested page retain the normal extraction semantics.
+    let required_pages: Option<HashSet<u32>> = pages.map(|pages| {
+        pages
+            .iter()
+            .filter_map(|page| page.checked_add(1))
+            .collect()
+    });
     let ((all_items, all_rects, all_lines), page_thresholds, gid_pages) =
-        extractor::extract_positioned_text_from_doc(&doc, &font_cmaps, None)?;
+        if let Some(required_pages) = required_pages.as_ref() {
+            extractor::extract_positioned_text_for_document_analysis(
+                &doc,
+                &font_cmaps,
+                required_pages,
+            )?
+        } else {
+            extractor::extract_positioned_text_from_doc(&doc, &font_cmaps, None)?
+        };
     let text_quality = analyze_text_quality(&all_items);
 
-    // Compute layout complexity from full document (near-zero cost).
-    let complexity = compute_layout_complexity(&all_items, &all_rects, &all_lines);
+    // Resolve page numbers with full-document context before partitioning.
+    // Per-page Markdown receives the original items plus these decisions so
+    // table detection can retain legitimate numeric cells.
+    let (filtered_items, removed_page_number_pages, page_number_removal_mask) =
+        extractor::filter_markdown_page_numbers_with_removed_pages(all_items.clone(), page_count);
+
+    // Tables need the original numeric cells; columns use folio-cleaned
+    // evidence so removed page numbers cannot create false layout metadata.
+    let chart_regions = markdown::chart_regions_by_page(&all_items, &all_rects, &all_lines);
+    let complexity = compute_layout_complexity_with_chart_regions(
+        &all_items,
+        &filtered_items,
+        &all_rects,
+        &all_lines,
+        &chart_regions,
+    );
 
     // Compute font stats from full document (cross-page consistency).
-    let font_stats = markdown::analysis::calculate_font_stats_from_items(&all_items);
+    let font_stats = markdown::analysis::calculate_font_stats_from_items(&filtered_items);
 
     // When caller doesn't specify pages, return every page in document order.
     let all_pages: Vec<u32>;
@@ -451,6 +516,7 @@ pub fn extract_pages_markdown_mem(
     let mut results = Vec::with_capacity(pages_slice.len());
     let mut pages_needing_ocr = Vec::new();
     let mut ocr_reasons_by_page = BTreeMap::new();
+    let lopdf_pages = doc.get_pages();
 
     for &page_0idx in pages_slice {
         // Out-of-range pages → empty + needs_ocr
@@ -467,12 +533,13 @@ pub fn extract_pages_markdown_mem(
 
         let page_1idx = page_0idx + 1;
 
-        // Filter items/rects for this page only
-        let page_items: Vec<TextItem> = all_items
+        // Partition items, removal decisions, and rects for this page only.
+        let (page_items, page_number_removal_mask): (Vec<TextItem>, Vec<bool>) = all_items
             .iter()
-            .filter(|i| i.page == page_1idx)
-            .cloned()
-            .collect();
+            .zip(&page_number_removal_mask)
+            .filter(|(item, _)| item.page == page_1idx)
+            .map(|(item, remove)| (item.clone(), *remove))
+            .unzip();
 
         let page_rects: Vec<PdfRect> = all_rects
             .iter()
@@ -483,6 +550,25 @@ pub fn extract_pages_markdown_mem(
         let has_gid = gid_pages.contains(&page_1idx);
         let has_text_quality_issue = text_quality.pages_needing_ocr.contains(&page_1idx);
 
+        // A page can extract cleanly (no decoding issues, non-empty text)
+        // while still being fundamentally a scan: a full-page raster with
+        // a little genuine native text drawn over it (a header, a stamp, a
+        // cover-sheet annotation). Text-quality signals alone can't see
+        // that — consult the same "large background image" signal
+        // classify_pdf/detect_pdf_type already uses, so the two APIs can't
+        // silently disagree on whether a page needs OCR. See #227.
+        // Also covers vector-outlined text (glyphs drawn as paths, not
+        // shown via a text-showing operator): a hybrid page with real
+        // embedded-font body text elsewhere would otherwise still extract
+        // non-empty, non-garbled markdown and miss OCR routing entirely.
+        // detect_from_document's Mixed-type per-page routing always sends
+        // these pages to OCR; mirror that here too. Both signals share one
+        // analyze_page_content pass — see page_ocr_signals's doc comment.
+        let (has_template_image, has_vector_text) = lopdf_pages
+            .get(&page_1idx)
+            .map(|&page_id| detector::page_ocr_signals(&doc, page_id))
+            .unwrap_or((false, false));
+
         // Build markdown with document-wide font stats
         let options = MarkdownOptions {
             base_font_size: Some(font_stats.most_common_size),
@@ -499,9 +585,15 @@ pub fn extract_pages_markdown_mem(
                 options,
                 &page_rects,
                 &[],
-                &page_thresholds,
-                None,
-                &[],
+                markdown::MarkdownDocumentContext {
+                    page_thresholds: &page_thresholds,
+                    struct_roles: None,
+                    struct_tables: &[],
+                    page_count,
+                    prefiltered_page_number_pages: Some(&removed_page_number_pages),
+                    prefiltered_page_number_mask: Some(&page_number_removal_mask),
+                    precomputed_chart_regions: Some(&chart_regions),
+                },
             )
         };
 
@@ -514,10 +606,20 @@ pub fn extract_pages_markdown_mem(
                 OCR_REASON_SUSPECTED_GARBLED_TEXT,
             );
         }
+        if has_template_image {
+            add_ocr_reason(&mut ocr_reasons_by_page, page_1idx, OCR_REASON_SCANNED);
+        }
+        if has_vector_text {
+            add_ocr_reason(&mut ocr_reasons_by_page, page_1idx, OCR_REASON_VECTOR_TEXT);
+        }
         let ocr_reason = page_ocr_reason(&ocr_reasons_by_page, page_1idx);
 
-        let needs_ocr =
-            ocr_reason.is_some() || md.trim().is_empty() || has_gid || is_garbage_text(&md);
+        let needs_ocr = ocr_reason.is_some()
+            || md.trim().is_empty()
+            || has_gid
+            || is_garbage_text(&md)
+            || has_template_image
+            || has_vector_text;
 
         if needs_ocr {
             pages_needing_ocr.push(page_1idx);
@@ -673,18 +775,51 @@ pub fn extract_text_in_regions_mem(
 
         let mut page_results = Vec::with_capacity(regions.len());
 
-        for rect in regions {
-            let [rx1, ry1, rx2, ry2] = *rect;
+        // Exclusive item->region assignment: overlapping layout regions used
+        // to extract shared items into EVERY region they touched (the
+        // 1.5pt inclusion margin makes borders generous), duplicating whole
+        // lines in the final markdown on 21% of bench docs — and downstream
+        // duplicate-handling sometimes dropped the variant holding a
+        // sentence tail, turning duplication into content LOSS. Each item
+        // now belongs to the single region with the largest overlap area;
+        // items are partitioned, never suppressed, so no content can vanish.
+        let all_bounds: Vec<RegionBounds> = regions
+            .iter()
+            .map(|rect| {
+                let [rx1, ry1, rx2, ry2] = *rect;
+                region_bounds(rx1, ry1, rx2, ry2, page_h, coords)
+            })
+            .collect();
+        // Single pass over items: assign each to the best-overlap region and
+        // bucket the clone directly (review: avoid a second O(items x
+        // regions) traversal). `had_candidates` marks regions that touched
+        // at least one item even if every one was assigned elsewhere.
+        let mut region_items: Vec<Vec<TextItem>> = vec![Vec::new(); regions.len()];
+        let mut had_candidates: Vec<bool> = vec![false; regions.len()];
+        if let Some(items) = items {
+            for item in items {
+                let mut best: Option<usize> = None;
+                let mut best_area = 0.0_f32;
+                for (ri, b) in all_bounds.iter().enumerate() {
+                    if !region_overlaps_item(item, *b) {
+                        continue;
+                    }
+                    had_candidates[ri] = true;
+                    let area = region_item_overlap_area(item, *b);
+                    if area > best_area {
+                        best_area = area;
+                        best = Some(ri);
+                    }
+                }
+                if let Some(ri) = best {
+                    region_items[ri].push(item.clone());
+                }
+            }
+        }
 
-            let bounds = region_bounds(rx1, ry1, rx2, ry2, page_h, coords);
-            let matched: Vec<TextItem> = match items {
-                Some(items) => items
-                    .iter()
-                    .filter(|item| region_overlaps_item(item, bounds))
-                    .cloned()
-                    .collect(),
-                None => Vec::new(),
-            };
+        for (region_idx, _rect) in regions.iter().enumerate() {
+            let matched: Vec<TextItem> = std::mem::take(&mut region_items[region_idx]);
+            let assigned_count = matched.len();
             let has_text_quality_issue = region_items_have_decoding_issue(&matched);
             let text = collect_text_from_matched_items(matched, adaptive_threshold);
             let has_cid_issue = is_cid_garbage(&text);
@@ -698,8 +833,21 @@ pub fn extract_text_in_regions_mem(
             // Check per-region text quality instead of blanket page-level
             // GID rejection. A GID font in a logo elsewhere on the page
             // shouldn't force GPU OCR for clean text regions.
-            let needs_ocr =
-                ocr_reason.is_some() || text.trim().is_empty() || is_garbage_text(&text);
+            // A region whose ONLY overlapping items were assigned to a
+            // better-overlapping neighbor must not fall back to OCR: the
+            // pixels it would re-read belong to that neighbor, and OCR
+            // would reintroduce the duplication exclusivity removed.
+            // Before exclusive assignment these regions were non-empty
+            // native (no OCR), so this preserves the old OCR load too.
+            // Requires ZERO items assigned HERE: a region whose own
+            // assigned items materialize to empty text (whitespace-only,
+            // collector-filtered) keeps its OCR fallback.
+            let lost_to_neighbor = text.trim().is_empty()
+                && ocr_reason.is_none()
+                && assigned_count == 0
+                && had_candidates[region_idx];
+            let needs_ocr = !lost_to_neighbor
+                && (ocr_reason.is_some() || text.trim().is_empty() || is_garbage_text(&text));
 
             page_results.push(RegionText {
                 text,
@@ -980,7 +1128,12 @@ pub fn extract_tables_in_regions_mem(
                     candidates.push(candidate);
                 }
             }
-            let detected = tables::detect_tables(&matched, base_font_size, false);
+            let detected = tables::detect_tables_with_page_width(
+                &matched,
+                base_font_size,
+                false,
+                items.map_or(1.0, |items| tables::content_width(items)),
+            );
             if let Some(candidate) = detected
                 .iter()
                 .find_map(|t| evaluate(TableCandidateSource::Heuristic, t))
@@ -1127,7 +1280,7 @@ pub fn detect_vector_grid_in_region_mem(
     }
 
     let line_tables =
-        tables::detect_tables_from_lines(&items_in_region, &lines_in_region, page_1idx);
+        tables::detect_vector_grid_tables_from_lines(&items_in_region, &lines_in_region, page_1idx);
     for table in line_tables {
         if let Some(result) = vector_grid_result_from_table(
             &table,
@@ -1480,16 +1633,10 @@ mod vector_grid_tests {
     /// strip while body rows are drawn with `m`/`l` operators, so the rect
     /// cluster has only 2 Y-edges and `try_build_grid` rejects.
     ///
-    /// IGNORED: lifting this shape required the exact-duplicate early-dedup
-    /// (PR #76 first iteration), which had broad collateral damage on
-    /// SEC 10-K TOCs and similar docs that draw rule-rects above + below
-    /// section dividers (production diff: 0001104659-25-093871 lost its
-    /// TOC structure, perf-graph data table, and qualifications matrix).
-    /// Re-enable once a more surgical lift exists in `try_build_grid` or
-    /// `snap_edges` that handles cell-border + inner-fill + text-bg rect
-    /// triplets without page-wide dedup.
+    /// The page repeats a full-page background many times. Those fills must be
+    /// removed from clustering before chart/table evidence is evaluated, or
+    /// they swamp the real cell rectangles and make this table look chart-like.
     #[test]
-    #[ignore]
     fn greencomp_competence_two_cols() {
         let tables = detect_rect_tables_in_fixture("tests/fixtures/greencomp_competence.pdf");
         assert!(
@@ -3215,8 +3362,26 @@ fn region_bounds(
     }
 }
 
+/// Inclusion margin shared by the region/item overlap predicates and the
+/// exclusive-assignment area score — these MUST stay in sync: an item that
+/// passes the boolean guard must always have positive overlap area.
+const REGION_MARGIN: f32 = 1.5;
+
+/// Overlap area between an item and region bounds (same margin as the
+/// boolean test) — the exclusive-assignment score.
+fn region_item_overlap_area(item: &TextItem, bounds: RegionBounds) -> f32 {
+    let item_x_max = item.x + text_utils::effective_width(item);
+    let item_y_max = item.y + item.height;
+    let x_overlap = (item_x_max.min(bounds.x_max + REGION_MARGIN)
+        - item.x.max(bounds.x_min - REGION_MARGIN))
+    .max(0.0);
+    let y_overlap = (item_y_max.min(bounds.y_max + REGION_MARGIN)
+        - item.y.max(bounds.y_min - REGION_MARGIN))
+    .max(0.0);
+    x_overlap * y_overlap
+}
+
 fn region_overlaps_item(item: &TextItem, bounds: RegionBounds) -> bool {
-    const REGION_MARGIN: f32 = 1.5;
     let item_x_min = item.x;
     let item_x_max = item.x + text_utils::effective_width(item);
     let item_y_min = item.y;
@@ -3232,7 +3397,6 @@ fn region_overlaps_item(item: &TextItem, bounds: RegionBounds) -> bool {
 }
 
 fn region_overlaps_rect(rect: &PdfRect, bounds: RegionBounds) -> bool {
-    const REGION_MARGIN: f32 = 1.5;
     let (x_min, y_min, x_max, y_max) = normalized_rect_edges(rect);
     ranges_overlap(
         x_min,
@@ -3248,7 +3412,6 @@ fn region_overlaps_rect(rect: &PdfRect, bounds: RegionBounds) -> bool {
 }
 
 fn region_overlaps_line(line: &PdfLine, bounds: RegionBounds) -> bool {
-    const REGION_MARGIN: f32 = 1.5;
     let x_min = line.x1.min(line.x2);
     let x_max = line.x1.max(line.x2);
     let y_min = line.y1.min(line.y2);
@@ -3390,6 +3553,7 @@ fn repair_pdf_container_candidates(buf: &[u8]) -> Vec<Vec<u8>> {
     let mut candidates = Vec::new();
 
     add_repair_candidate(&mut candidates, append_missing_eof_marker(buf), buf);
+    add_repair_candidate(&mut candidates, recover_startxref_pointer(buf), buf);
 
     let stripped = strip_leading_pdf_container_bytes(buf);
     if let Some(stripped_buf) = stripped.as_deref() {
@@ -3399,11 +3563,112 @@ fn repair_pdf_container_candidates(buf: &[u8]) -> Vec<Vec<u8>> {
             append_missing_eof_marker(stripped_buf),
             buf,
         );
+        add_repair_candidate(
+            &mut candidates,
+            recover_startxref_pointer(stripped_buf),
+            buf,
+        );
     }
 
     candidates
 }
 
+/// Some PDF writers emit a `startxref` pointer that doesn't actually point
+/// at the cross-reference table — a single corrupted byte in the offset is
+/// enough. lopdf trusts that pointer outright and fails to load rather than
+/// searching for the real table, unlike pypdf/pdfium which both recover by
+/// locating it directly. This finds the real (classic, non-stream) `xref`
+/// table by scanning for the keyword — validating that a plausible
+/// subsection header follows, not just any standalone "xref" token, since
+/// this crate processes untrusted input and a coincidental match inside
+/// unrelated stream/string content must not get "repaired" against a bogus
+/// offset (lopdf would then load successfully against garbage instead of
+/// returning a clean error) — and appends a corrected trailing
+/// `startxref`/`%%EOF` block. lopdf's own `get_xref_start` always uses the
+/// *last* `%%EOF` in the final 512 bytes of the buffer, so ours
+/// transparently supersedes the broken one without needing to touch
+/// anything already in the file.
+///
+/// Doesn't cover cross-reference *streams* (`N 0 obj << /Type /XRef ...`,
+/// used by some PDF 1.5+ writers instead of a classic table) — recovering
+/// those needs the containing object's number, not just a byte offset.
+fn recover_startxref_pointer(buf: &[u8]) -> Option<Vec<u8>> {
+    let xref_pos = find_last_valid_xref_table_start(buf)?;
+
+    let mut repaired = Vec::with_capacity(buf.len() + 32);
+    repaired.extend_from_slice(buf);
+    if !repaired.ends_with(b"\n") {
+        repaired.push(b'\n');
+    }
+    repaired.extend_from_slice(format!("startxref\n{xref_pos}\n%%EOF\n").as_bytes());
+    Some(repaired)
+}
+
+/// Finds the last standalone `xref` token in `buf` that is immediately
+/// followed by a plausible classic cross-reference subsection header
+/// (`<start-id> <count>`, e.g. "0 6") — the shape every real classic xref
+/// table starts with. A single reverse byte scan: O(n) even on a
+/// pathological buffer with many non-matching or non-standalone "xref"
+/// occurrences, unlike repeatedly re-searching a shrinking prefix.
+fn find_last_valid_xref_table_start(buf: &[u8]) -> Option<usize> {
+    const KEYWORD: &[u8] = b"xref";
+    if buf.len() < KEYWORD.len() {
+        return None;
+    }
+    let mut pos = buf.len() - KEYWORD.len();
+    loop {
+        if &buf[pos..pos + KEYWORD.len()] == KEYWORD {
+            let before_ok = pos == 0 || buf[pos - 1].is_ascii_whitespace();
+            let after_ok = buf
+                .get(pos + KEYWORD.len())
+                .is_none_or(|c| c.is_ascii_whitespace());
+            if before_ok && after_ok && looks_like_xref_subsection_header(buf, pos + KEYWORD.len())
+            {
+                return Some(pos);
+            }
+        }
+        if pos == 0 {
+            return None;
+        }
+        pos -= 1;
+    }
+}
+
+/// Checks that `buf[pos..]` starts (after whitespace) with two
+/// whitespace-separated runs of ASCII digits — `<start-id> <count>`, the
+/// first subsection header of a classic PDF cross-reference table.
+fn looks_like_xref_subsection_header(buf: &[u8], pos: usize) -> bool {
+    fn skip_ws(buf: &[u8], mut pos: usize) -> usize {
+        while buf.get(pos).is_some_and(u8::is_ascii_whitespace) {
+            pos += 1;
+        }
+        pos
+    }
+    fn skip_digits(buf: &[u8], mut pos: usize) -> usize {
+        while buf.get(pos).is_some_and(u8::is_ascii_digit) {
+            pos += 1;
+        }
+        pos
+    }
+
+    let pos = skip_ws(buf, pos);
+    let after_first_digits = skip_digits(buf, pos);
+    if after_first_digits == pos {
+        return false; // no start-id
+    }
+    let sep = skip_ws(buf, after_first_digits);
+    if sep == after_first_digits {
+        return false; // start-id and count must be whitespace-separated
+    }
+    let after_count = skip_digits(buf, sep);
+    if after_count == sep {
+        return false; // no count
+    }
+    // The count run must end at whitespace/buffer-end, not run into trailing
+    // garbage (e.g. a coincidental "xref\n0 6garbage" in stream content).
+    buf.get(after_count).is_none_or(u8::is_ascii_whitespace)
+}
+
 fn add_repair_candidate(
     candidates: &mut Vec<Vec<u8>>,
     candidate: Option<Vec<u8>>,
@@ -3469,7 +3734,7 @@ fn process_document(
     doc: Document,
     page_count: u32,
     options: PdfOptions,
-    start: std::time::Instant,
+    start: ProcessingTimer,
 ) -> Result<PdfProcessResult, PdfError> {
     // Step 1 — Detection (cheap: scans content streams for text operators)
     let detection = detector::detect_from_document(&doc, page_count, &options.detection)?;
@@ -3485,7 +3750,7 @@ fn process_document(
             pdf_type,
             markdown: None,
             page_count,
-            processing_time_ms: start.elapsed().as_millis() as u64,
+            processing_time_ms: start.elapsed_ms(),
             pages_needing_ocr,
             ocr_reasons_by_page: page_ocr_reasons_vec(detection_ocr_reasons),
             title,
@@ -3501,7 +3766,7 @@ fn process_document(
             pdf_type,
             markdown: None,
             page_count,
-            processing_time_ms: start.elapsed().as_millis() as u64,
+            processing_time_ms: start.elapsed_ms(),
             pages_needing_ocr,
             ocr_reasons_by_page: page_ocr_reasons_vec(detection_ocr_reasons),
             title,
@@ -3514,7 +3779,10 @@ fn process_document(
     // Step 2 — Extraction (reuses the already-loaded document)
     let extracted = {
         let font_cmaps = FontCMaps::from_doc(&doc);
-        let result = extractor::extract_positioned_text_from_doc(
+        // Most page-filtered requests extract only the selected pages. Gather
+        // other pages only when a selected contextual folio needs cross-page
+        // evidence; failures on those context-only pages are non-fatal.
+        let result = extractor::extract_positioned_text_with_folio_context(
             &doc,
             &font_cmaps,
             options.page_filter.as_ref(),
@@ -3525,9 +3793,19 @@ fn process_document(
         // This unlocks OCR text layers behind scanned images.
         if pdf_type == PdfType::Mixed {
             if let Ok((ref items, _, _)) = result.as_ref().map(|(e, _, _)| e) {
-                let sample: String = items.iter().take(200).map(|i| i.text.as_str()).collect();
+                let sample: String = items
+                    .iter()
+                    .filter(|item| {
+                        options
+                            .page_filter
+                            .as_ref()
+                            .is_none_or(|filter| filter.contains(&item.page))
+                    })
+                    .take(200)
+                    .map(|item| item.text.as_str())
+                    .collect();
                 if is_garbage_text(&sample) || sample.trim().is_empty() {
-                    extractor::extract_positioned_text_include_invisible(
+                    extractor::extract_positioned_text_include_invisible_with_folio_context(
                         &doc,
                         &font_cmaps,
                         options.page_filter.as_ref(),
@@ -3537,7 +3815,7 @@ fn process_document(
                 }
             } else {
                 // Normal extraction failed — try invisible as fallback
-                extractor::extract_positioned_text_include_invisible(
+                extractor::extract_positioned_text_include_invisible_with_folio_context(
                     &doc,
                     &font_cmaps,
                     options.page_filter.as_ref(),
@@ -3601,6 +3879,13 @@ fn process_document(
                     let mut garbage_pages: std::collections::HashSet<u32> =
                         std::collections::HashSet::new();
                     for &pg in &ocr_set {
+                        if options
+                            .page_filter
+                            .as_ref()
+                            .is_some_and(|filter| !filter.contains(&pg))
+                        {
+                            continue;
+                        }
                         let page_text: String = items
                             .iter()
                             .filter(|i| i.page == pg)
@@ -3640,9 +3925,45 @@ fn process_document(
                     }
                 };
 
+            let selected_page = |page: u32| {
+                options
+                    .page_filter
+                    .as_ref()
+                    .is_none_or(|filter| filter.contains(&page))
+            };
+            let rects: Vec<_> = rects
+                .into_iter()
+                .filter(|rect| selected_page(rect.page))
+                .collect();
+            let lines: Vec<_> = lines
+                .into_iter()
+                .filter(|line| selected_page(line.page))
+                .collect();
+            let gid_encoded_pages: HashSet<_> = gid_encoded_pages
+                .into_iter()
+                .filter(|page| selected_page(*page))
+                .collect();
+            let FolioFilteredItems {
+                items,
+                layout_items,
+                removal_mask,
+                removed_pages,
+            } = select_items_with_document_folio_context(
+                items,
+                page_count,
+                options.page_filter.as_ref(),
+            );
+
             let text_quality = analyze_text_quality(&items);
             merge_ocr_reasons(&mut ocr_reasons_by_page, text_quality.reasons_by_page);
-            let layout = compute_layout_complexity(&items, &rects, &lines);
+            let chart_regions = markdown::chart_regions_by_page(&items, &rects, &lines);
+            let layout = compute_layout_complexity_with_chart_regions(
+                &items,
+                &layout_items,
+                &rects,
+                &lines,
+                &chart_regions,
+            );
 
             let md = if options.mode == ProcessMode::Analyze {
                 None
@@ -3652,9 +3973,15 @@ fn process_document(
                     options.markdown,
                     &rects,
                     &lines,
-                    &page_thresholds,
-                    struct_roles.as_ref(),
-                    &struct_tables,
+                    markdown::MarkdownDocumentContext {
+                        page_thresholds: &page_thresholds,
+                        struct_roles: struct_roles.as_ref(),
+                        struct_tables: &struct_tables,
+                        page_count,
+                        prefiltered_page_number_pages: Some(&removed_pages),
+                        prefiltered_page_number_mask: Some(removal_mask.as_slice()),
+                        precomputed_chart_regions: Some(&chart_regions),
+                    },
                 ))
             };
 
@@ -3767,7 +4094,7 @@ fn process_document(
         pdf_type,
         markdown,
         page_count,
-        processing_time_ms: start.elapsed().as_millis() as u64,
+        processing_time_ms: start.elapsed_ms(),
         pages_needing_ocr,
         ocr_reasons_by_page: {
             // Detector reasons (scanned / no_text / vector_text / garbled) merged
@@ -5413,11 +5740,70 @@ mod looks_like_partial_table_tests {
     }
 }
 
+struct FolioFilteredItems {
+    items: Vec<types::TextItem>,
+    layout_items: Vec<types::TextItem>,
+    removal_mask: Vec<bool>,
+    removed_pages: HashSet<u32>,
+}
+
+/// Resolve folios with complete document context, then select the caller's
+/// requested pages without losing those decisions.
+fn select_items_with_document_folio_context(
+    all_items: Vec<types::TextItem>,
+    page_count: u32,
+    page_filter: Option<&HashSet<u32>>,
+) -> FolioFilteredItems {
+    let (all_layout_items, all_removed_pages, all_removal_mask) =
+        extractor::filter_markdown_page_numbers_with_removed_pages(all_items.clone(), page_count);
+    let selected_page = |page: u32| page_filter.is_none_or(|filter| filter.contains(&page));
+
+    let (items, removal_mask) = all_items
+        .into_iter()
+        .zip(all_removal_mask)
+        .filter(|(item, _)| selected_page(item.page))
+        .unzip();
+    let layout_items = all_layout_items
+        .into_iter()
+        .filter(|item| selected_page(item.page))
+        .collect();
+    let removed_pages = all_removed_pages
+        .into_iter()
+        .filter(|page| selected_page(*page))
+        .collect();
+
+    FolioFilteredItems {
+        items,
+        layout_items,
+        removal_mask,
+        removed_pages,
+    }
+}
+
 /// Analyse extracted items and rects for layout complexity.
+#[cfg(test)]
 fn compute_layout_complexity(
     items: &[types::TextItem],
+    column_items: &[types::TextItem],
     rects: &[types::PdfRect],
     lines: &[types::PdfLine],
+) -> LayoutComplexity {
+    let page_chart_regions = markdown::chart_regions_by_page(items, rects, lines);
+    compute_layout_complexity_with_chart_regions(
+        items,
+        column_items,
+        rects,
+        lines,
+        &page_chart_regions,
+    )
+}
+
+fn compute_layout_complexity_with_chart_regions(
+    items: &[types::TextItem],
+    column_items: &[types::TextItem],
+    rects: &[types::PdfRect],
+    lines: &[types::PdfLine],
+    page_chart_regions: &markdown::PageChartRegions,
 ) -> LayoutComplexity {
     use markdown::analysis::calculate_font_stats_from_items;
 
@@ -5437,7 +5823,12 @@ fn compute_layout_complexity(
 
         // Check for side-by-side layout
         let owned_items: Vec<types::TextItem> = page_items.iter().map(|i| (*i).clone()).collect();
+        let page_content_width = tables::content_width(&owned_items);
         let bands = markdown::split_side_by_side(&owned_items);
+        let chart_regions = page_chart_regions
+            .get(&page)
+            .map(Vec::as_slice)
+            .unwrap_or_default();
 
         let band_ranges: Vec<(f32, f32)> = if bands.is_empty() {
             // Single region — use sentinel range that includes everything
@@ -5452,7 +5843,8 @@ fn compute_layout_complexity(
             let band_items: Vec<types::TextItem> = owned_items
                 .iter()
                 .filter(|item| {
-                    x_lo == f32::MIN || (item.x >= x_lo - margin && item.x < x_hi + margin)
+                    (x_lo == f32::MIN || (item.x >= x_lo - margin && item.x < x_hi + margin))
+                        && !markdown::item_is_in_chart_region(item, chart_regions)
                 })
                 .cloned()
                 .collect();
@@ -5487,7 +5879,12 @@ fn compute_layout_complexity(
                 break;
             }
             // Heuristic fallback for borderless tables
-            let heuristic_tables = tables::detect_tables(&band_items, base_size, false);
+            let heuristic_tables = tables::detect_tables_with_page_width(
+                &band_items,
+                base_size,
+                false,
+                page_content_width,
+            );
             if has_data_table(&heuristic_tables) {
                 found_table = true;
                 break;
@@ -5499,8 +5896,20 @@ fn compute_layout_complexity(
     }
 
     let mut pages_with_columns: Vec<u32> = Vec::new();
-    for page in seen_pages {
-        let cols = extractor::detect_columns(items, page, pages_with_tables.contains(&page));
+    for &page in &seen_pages {
+        let chart_regions = page_chart_regions
+            .get(&page)
+            .map(Vec::as_slice)
+            .unwrap_or_default();
+        let page_column_items: Vec<types::TextItem> = column_items
+            .iter()
+            .filter(|item| {
+                item.page == page && !markdown::item_is_in_chart_region(item, chart_regions)
+            })
+            .cloned()
+            .collect();
+        let cols =
+            extractor::detect_columns(&page_column_items, page, pages_with_tables.contains(&page));
         if cols.len() >= 2 {
             pages_with_columns.push(page);
         }
@@ -5712,6 +6121,128 @@ mod tests {
         }
     }
 
+    #[test]
+    fn removed_sparse_folios_leave_no_layout_evidence() {
+        let items = vec![
+            test_item("1", 25.0, 20.0, 12.0, 10.0),
+            test_item("2", 520.0, 60.0, 12.0, 10.0),
+        ];
+        let (filtered, _, _) =
+            extractor::filter_markdown_page_numbers_with_removed_pages(items.clone(), 1);
+        assert!(filtered.is_empty());
+
+        let filtered = compute_layout_complexity(&items, &filtered, &[], &[]);
+
+        assert!(!filtered.is_complex);
+        assert!(filtered.pages_with_tables.is_empty());
+        assert!(filtered.pages_with_columns.is_empty());
+    }
+
+    #[test]
+    fn dense_chart_panel_is_not_reported_as_a_table() {
+        let mut items: Vec<TextItem> = (0..8)
+            .flat_map(|row| {
+                (0..6).map(move |column| {
+                    test_item(
+                        &format!("{}", row * 10 + column),
+                        105.0 + column as f32 * 35.0,
+                        525.0 - row as f32 * 15.0,
+                        24.0,
+                        10.0,
+                    )
+                })
+            })
+            .collect();
+        for row in 0..6 {
+            items.push(test_item(
+                "Left column prose continues here",
+                80.0,
+                320.0 - row as f32 * 15.0,
+                160.0,
+                10.0,
+            ));
+            items.push(test_item(
+                "Right column prose continues here",
+                300.0,
+                320.0 - row as f32 * 15.0,
+                160.0,
+                10.0,
+            ));
+        }
+        let mut lines: Vec<PdfLine> = (0..30)
+            .map(|column| PdfLine {
+                x1: 100.0 + column as f32 * 8.0,
+                y1: 400.0,
+                x2: 100.0 + column as f32 * 8.0,
+                y2: 550.0,
+                page: 1,
+            })
+            .collect();
+        lines.extend((0..6).map(|row| PdfLine {
+            x1: 100.0,
+            y1: 400.0 + row as f32 * 30.0,
+            x2: 332.0,
+            y2: 400.0 + row as f32 * 30.0,
+            page: 1,
+        }));
+        let rects = vec![PdfRect {
+            x: 80.0,
+            y: 350.0,
+            width: 280.0,
+            height: 240.0,
+            page: 1,
+        }];
+
+        let complexity = compute_layout_complexity(&items, &items, &rects, &lines);
+
+        assert!(complexity.pages_with_tables.is_empty());
+    }
+
+    #[test]
+    fn page_selection_keeps_document_wide_folio_layout_decisions() {
+        let mut items = Vec::new();
+        for page in 1..=4 {
+            for row in 0..8 {
+                let y = 20.0 + row as f32 * 8.0;
+                let mut folio = test_item(&(row * 10 + page).to_string(), 25.0, y, 12.0, 10.0);
+                folio.page = page;
+                let mut footer =
+                    test_item(&format!("Footer row {row} summary"), 43.0, y, 470.0, 10.0);
+                footer.page = page;
+                let mut body = test_item(&format!("Body{row}"), 530.0, y, 55.0, 10.0);
+                body.page = page;
+                items.extend([folio, footer, body]);
+            }
+        }
+
+        let page_one_items: Vec<_> = items
+            .iter()
+            .filter(|item| item.page == 1)
+            .cloned()
+            .collect();
+        let (page_local_layout, _, _) =
+            extractor::filter_markdown_page_numbers_with_removed_pages(page_one_items.clone(), 4);
+        let page_local = compute_layout_complexity(&page_one_items, &page_local_layout, &[], &[]);
+        assert!(
+            page_local.pages_with_columns.contains(&1),
+            "fixture must reproduce page-local folio column evidence"
+        );
+
+        let selected =
+            select_items_with_document_folio_context(items, 4, Some(&HashSet::from([1])));
+        assert_eq!(
+            selected
+                .removal_mask
+                .iter()
+                .filter(|remove| **remove)
+                .count(),
+            8
+        );
+        let document_wide =
+            compute_layout_complexity(&selected.items, &selected.layout_items, &[], &[]);
+        assert!(!document_wide.pages_with_columns.contains(&1));
+    }
+
     #[test]
     fn test_detect_encoding_issues_fffd() {
         assert!(detect_encoding_issues(
@@ -6670,4 +7201,65 @@ mod tests {
         // Pre-filled cell was not touched.
         assert_eq!(cells[1].text, "Pre-filled");
     }
+
+    // -- recover_startxref_pointer / find_last_valid_xref_table_start ------
+    //
+    // Direct unit tests on the byte-level scan, addressing review feedback
+    // on #230: a coincidental standalone "xref" token that isn't actually
+    // followed by a subsection header (start-id + count) must not be
+    // treated as a real table — accepting it would let lopdf "succeed"
+    // against a bogus offset and silently return garbled/empty content
+    // instead of a clean error.
+
+    #[test]
+    fn find_xref_rejects_standalone_token_without_subsection_header() {
+        // "xref" appears as a real standalone word, but nothing that looks
+        // like "<start-id> <count>" follows it.
+        let buf = b"Please refer to the xref appendix for details.";
+        assert_eq!(find_last_valid_xref_table_start(buf), None);
+    }
+
+    #[test]
+    fn find_xref_accepts_real_classic_table_header() {
+        let buf = b"garbage\nxref\n0 6\n0000000000 65535 f \n%%EOF";
+        let pos = find_last_valid_xref_table_start(buf).expect("should find the real table");
+        assert_eq!(&buf[pos..pos + 4], b"xref");
+        assert_eq!(&buf[pos..], b"xref\n0 6\n0000000000 65535 f \n%%EOF");
+    }
+
+    #[test]
+    fn find_xref_skips_coincidental_match_and_finds_real_table_before_it() {
+        // A coincidental "xref" (no subsection header) appears *after* the
+        // real table in the buffer — the scan must not stop at the first
+        // (rightmost) standalone token it finds; it must keep looking
+        // backward until one actually validates.
+        let buf = b"xref\n0 3\n0000000000 65535 f \ntrailer\nsee the xref\n";
+        let pos = find_last_valid_xref_table_start(buf).expect("should find the real table");
+        assert_eq!(pos, 0);
+    }
+
+    #[test]
+    fn find_xref_rejects_substring_of_startxref() {
+        // "xref" is a substring of "startxref" but isn't a standalone
+        // token there (not preceded by whitespace) — must not match, even
+        // though a number immediately follows it.
+        let buf = b"startxref\n1234\n%%EOF";
+        assert_eq!(find_last_valid_xref_table_start(buf), None);
+    }
+
+    #[test]
+    fn find_xref_rejects_count_run_with_trailing_garbage() {
+        // "xref\n0 6garbage" has the right shape (digits, whitespace,
+        // digits) but the count run doesn't end at whitespace/EOF — it
+        // runs straight into non-digit garbage, so this must not be
+        // accepted as a real subsection header.
+        let buf = b"xref\n0 6garbage\n%%EOF";
+        assert_eq!(find_last_valid_xref_table_start(buf), None);
+    }
+
+    #[test]
+    fn recover_startxref_pointer_returns_none_without_a_valid_table() {
+        let buf = b"Please refer to the xref appendix for details.";
+        assert!(recover_startxref_pointer(buf).is_none());
+    }
 }

src/markdown/analysis.rs

@@ -171,6 +171,74 @@ pub(crate) fn is_toc_marker_heading(text: &str) -> bool {
 /// equation and absent from name-plus-number headings. A bare trailing colon
 /// is NOT a fragment signal either: real headings frequently end with colons
 /// ("Procedure:", "Steps for Using the Microscope:").
+/// True when the line opens with a section number ("3.", "2.1.4", "IV)").
+///
+/// Mirrors the acceptance of `heading::parse_numbering` rather than the
+/// stricter `convert::starts_with_section_number`, which deliberately
+/// requires two components because it bypasses isolation checks. Here a
+/// single "1." counts: numbering is independent evidence of a heading, and
+/// `heading.rs` applies its numbered-prefix allowance *after* consulting
+/// `is_heading_fragment`, so without this exemption a numbered
+/// sentence-case heading would be vetoed before that allowance can run.
+fn starts_with_numbering_prefix(t: &str) -> bool {
+    let Some(first) = t.split_whitespace().next() else {
+        return false;
+    };
+    let has_delimiter = first.ends_with(['.', ')', ':']);
+    let token = first.trim_end_matches(['.', ')', ':']);
+    if token.is_empty() {
+        return false;
+    }
+    let parts: Vec<&str> = token.split('.').collect();
+    let decimal = parts
+        .iter()
+        .all(|p| !p.is_empty() && p.len() <= 3 && p.chars().all(|c| c.is_ascii_digit()));
+    if decimal {
+        // "1." / "2.1." carry a delimiter; "2.3 Title" is written without
+        // one, so a multi-component number is accepted bare. A bare single
+        // number ("3 apples") is not — that is ordinary prose.
+        return has_delimiter || parts.len() >= 2;
+    }
+    // Roman numerals go through the heading parser's own grammar so the two
+    // agree: uppercase I/V/X/L/C only, at most 8 characters. A looser rule
+    // here would exempt markers the parser rejects — "iv)" or "d)" from an
+    // alphabetical list — letting an ordinary list item bypass the veto and
+    // reach heading promotion.
+    //
+    // A delimiter is also required: a bare leading "I" is the pronoun far
+    // more often than a section number.
+    has_delimiter && crate::markdown::heading::roman_value(token).is_some()
+}
+
+/// True when the line reads as a title rather than a sentence: every
+/// content word (ignoring minor words) starts uppercase. Used to spare real
+/// headings from the dangling-verb veto — "Bond Yields" is a section title,
+/// "the method yields" is a stranded clause, and only the casing tells them
+/// apart.
+fn looks_title_case(t: &str) -> bool {
+    const MINOR: &[&str] = &[
+        "a", "an", "the", "of", "and", "or", "for", "to", "in", "on", "at", "by", "with", "from",
+        "as", "is", "are", "that", "than", "into",
+    ];
+    let mut content = 0usize;
+    let mut capitalized = 0usize;
+    for w in t.split_whitespace() {
+        let cleaned: String = w.chars().filter(|c| c.is_alphabetic()).collect();
+        if cleaned.is_empty() {
+            continue;
+        }
+        if MINOR.contains(&cleaned.to_lowercase().as_str()) {
+            continue;
+        }
+        content += 1;
+        if cleaned.chars().next().is_some_and(char::is_uppercase) {
+            capitalized += 1;
+        }
+    }
+    // A single content word ("Yields") is a title by default.
+    content == 0 || capitalized == content
+}
+
 pub(crate) fn is_heading_fragment(text: &str) -> bool {
     let t = text.trim_end();
 
@@ -244,9 +312,133 @@ pub(crate) fn is_heading_fragment(text: &str) -> bool {
     if t.ends_with(':') && t.split_whitespace().any(is_equation_number) {
         return true;
     }
+
+    // Dangling clause: a stranded sentence lead-in ends on a relational
+    // verb with no terminal punctuation — "Note that the exact error equals"
+    // left ahead of its formula when a phantom table dissolved.
+    //
+    // Gated on the line reading as prose rather than a title. Case is the
+    // discriminator the trailing word alone cannot provide: a heading is
+    // title case ("Bond Yields", "The Method Yields") while a stranded
+    // lead-in is sentence case ("the method yields"). Without this gate the
+    // veto eats real headings — "Bond Yields", "Crop Yields" and any wrapped
+    // title-case heading the preprocessor failed to merge.
+    if !t.ends_with(['.', '!', '?', ':', ';', ')', ']'])
+        && !looks_title_case(t)
+        && !starts_with_numbering_prefix(t)
+    {
+        if let Some(last) = t.split_whitespace().next_back() {
+            let word: String = last
+                .trim_matches(|c: char| !c.is_alphanumeric())
+                .to_lowercase();
+            // Relational verbs only, and only those with no common noun
+            // sense. "yields" was dropped for exactly that reason: "Bond
+            // Yields" is a real section title. Function words, copulas and
+            // auxiliaries were measured and rejected outright — a heading
+            // that wraps across lines ends on those, and suppressing them
+            // destroyed real IRS Publication 17 headings.
+            const DANGLING_TAIL: &[&str] =
+                &["equals", "denotes", "implies", "satisfies", "signifies"];
+            if DANGLING_TAIL.contains(&word.as_str()) {
+                return true;
+            }
+        }
+    }
     false
 }
 
+#[cfg(test)]
+mod fragment_heading_tests {
+    use super::is_heading_fragment;
+
+    #[test]
+    fn dangling_tail_marks_stranded_clause() {
+        // opendataloader 01030000000144: left behind when a phantom table
+        // dissolved, ahead of its formula on the next line.
+        assert!(is_heading_fragment("Note that the exact error equals"));
+        assert!(is_heading_fragment("The remainder term satisfies"));
+        assert!(is_heading_fragment("we conclude that the sum equals"));
+    }
+
+    #[test]
+    fn real_headings_survive() {
+        assert!(!is_heading_fragment("Introduction"));
+        assert!(!is_heading_fragment("Error Analysis"));
+        assert!(!is_heading_fragment("Materials and Methods"));
+        assert!(!is_heading_fragment("Results"));
+        assert!(!is_heading_fragment("3.2 Richardson Extrapolation"));
+        assert!(!is_heading_fragment("Discussion and Conclusions"));
+        // Terminal punctuation means the clause is complete.
+        assert!(!is_heading_fragment("What is a Derivative?"));
+        assert!(!is_heading_fragment("Procedure:"));
+        assert!(!is_heading_fragment("Note that this is important."));
+    }
+
+    #[test]
+    fn title_case_headings_ending_in_a_verb_survive() {
+        // "yields" is also a plural noun; these are real section titles.
+        assert!(!is_heading_fragment("Bond Yields"));
+        assert!(!is_heading_fragment("Crop Yields"));
+        assert!(!is_heading_fragment("Dividend Yields"));
+        assert!(!is_heading_fragment("Yields"));
+        // A wrapped title-case heading whose first line ends on a listed
+        // verb must survive even if the preprocessor failed to merge it.
+        assert!(!is_heading_fragment("The Theorem Implies"));
+        assert!(!is_heading_fragment("What This Denotes"));
+    }
+
+    #[test]
+    fn numbered_sentence_case_headings_survive() {
+        // heading.rs consults is_heading_fragment BEFORE applying its
+        // numbered-prefix allowance, so the veto must not pre-empt it.
+        assert!(!is_heading_fragment("1. What the model implies"));
+        assert!(!is_heading_fragment("2.3 How the estimator satisfies"));
+        assert!(!is_heading_fragment("IV) What this denotes"));
+        // Without numbering the same wording is still a stranded clause.
+        assert!(is_heading_fragment("What the model implies"));
+        // A bare leading number or pronoun is prose, not numbering.
+        assert!(is_heading_fragment("3 apples and what that implies"));
+        assert!(is_heading_fragment("I think the model implies"));
+        // Markers heading::parse_numbering rejects must not be exempted
+        // either, or an ordinary list item bypasses the veto: lowercase
+        // roman, alphabetical markers, and over-long tokens.
+        assert!(is_heading_fragment("iv) the estimator satisfies"));
+        assert!(is_heading_fragment("d) the value implies"));
+        // Unsupported character (M is outside the parser's I/V/X/L/C set).
+        assert!(is_heading_fragment("MMMM. the value implies"));
+        // Over-long token: nine valid characters, so this exercises the
+        // 8-character bound rather than the character set.
+        assert!(is_heading_fragment("IIIIIIIII. the value implies"));
+        // Eight is still within the bound and stays exempt.
+        assert!(!is_heading_fragment("IIIIIIII. What this implies"));
+        // Uppercase roman within the parser's grammar is still exempt.
+        assert!(!is_heading_fragment("IV. What this denotes"));
+        assert!(!is_heading_fragment("XII) What this implies"));
+    }
+
+    #[test]
+    fn wrapped_headings_are_not_fragments() {
+        // A heading that wraps across lines ends on a function word. These
+        // are real headings from IRS Publication 17 and must survive.
+        assert!(!is_heading_fragment("Casualty and"));
+        assert!(!is_heading_fragment("Rule 10. You Must Be at"));
+        assert!(!is_heading_fragment("Higher Standard Deduction for"));
+        assert!(!is_heading_fragment("Qualifying Child of"));
+        assert!(!is_heading_fragment("When Can I Withdraw or"));
+        // Copulas and auxiliaries also end real wrapped headings.
+        assert!(!is_heading_fragment("Rule 15. Your AGI Must Be"));
+        assert!(!is_heading_fragment("What Medical Expenses Are"));
+        assert!(!is_heading_fragment("Rule 13. You Must Have"));
+        assert!(!is_heading_fragment("When Can a Roth IRA Be"));
+    }
+
+    #[test]
+    fn dangling_check_is_case_insensitive() {
+        // All-caps is not sentence case, so the veto must not fire there.
+        assert!(!is_heading_fragment("THE REMAINDER EQUALS"));
+    }
+}
+
 /// Compute the Y-gap threshold for paragraph break detection.
 ///
 /// Instead of using a fixed multiple of base_size (which fails for double-spaced
@@ -374,6 +566,15 @@ pub(crate) fn compute_heading_tiers(lines: &[TextLine], base_size: f32) -> Vec<f
     for line in lines {
         if let Some(first) = line.items.first() {
             if first.font_size / base_size >= 1.2 {
+                // Digit-only lines (page numbers, issue numbers) must not
+                // define heading tiers: a large bold folio claims tier 0 and
+                // blocks the bold-size fallback for the document's real
+                // same-size headings.
+                let text = line.text();
+                let t = text.trim();
+                if !t.is_empty() && t.chars().all(|c| !c.is_alphabetic()) {
+                    continue;
+                }
                 heading_sizes.push(first.font_size);
             }
         }
@@ -391,11 +592,45 @@ pub(crate) fn compute_heading_tiers(lines: &[TextLine], base_size: f32) -> Vec<f
         }
     }
 
+    // Books often set section headings barely above body size (e.g. 11pt
+    // bold over 10pt text). When nothing clears the 1.2x ratio gate, fall
+    // back to bold lines modestly larger than body so those documents still
+    // get an H1 instead of every bold heading defaulting to H2.
+    if tiers.is_empty() {
+        let mut bold_sizes: Vec<f32> = lines
+            .iter()
+            .filter(|line| {
+                let text = line.text();
+                let t = text.trim();
+                !t.is_empty() && t.chars().any(|c| c.is_alphabetic())
+            })
+            .filter_map(|line| line.items.first())
+            .filter(|it| it.is_bold && it.font_size / base_size >= 1.05)
+            .map(|it| it.font_size)
+            .collect();
+        bold_sizes.sort_by(|a, b| b.total_cmp(a));
+        for size in bold_sizes {
+            if !tiers.iter().any(|&t| (t - size).abs() < 0.5) {
+                tiers.push(size);
+            }
+        }
+    }
+
     // Cap at 4 tiers
     tiers.truncate(4);
     tiers
 }
 
+/// Boldness of a line judged by character mass, so a heading with an
+/// unbold section-number prefix ("4. " + bold title) still counts as bold.
+pub(crate) fn line_is_mostly_bold(line: &TextLine) -> bool {
+    let (bold, total) = line.items.iter().fold((0usize, 0usize), |(b, t), it| {
+        let n = it.text.trim().chars().count();
+        (b + if it.is_bold { n } else { 0 }, t + n)
+    });
+    total > 0 && bold * 2 >= total
+}
+
 /// Detect header level from font size using document-specific heading tiers.
 /// When tiers are available, maps tier 0→H1, tier 1→H2, etc.
 /// Falls back to ratio-based thresholds when no tiers exist.
@@ -403,9 +638,21 @@ pub(crate) fn detect_header_level(
     font_size: f32,
     base_size: f32,
     heading_tiers: &[f32],
+    is_bold: bool,
 ) -> Option<usize> {
     let ratio = font_size / base_size;
 
+    // Tier matches are trusted below the 1.2x gate (down to 1.05x) only for
+    // bold lines: sub-gate tiers come from the bold fallback, and honoring
+    // them for non-bold text at the same size would promote captions.
+    if (1.05..1.2).contains(&ratio) && is_bold && !heading_tiers.is_empty() {
+        for (i, &tier_size) in heading_tiers.iter().enumerate() {
+            if (font_size - tier_size).abs() < 0.5 {
+                return Some(i + 1); // tier 0 → H1, tier 1 → H2, etc.
+            }
+        }
+    }
+
     if ratio < 1.2 {
         return None; // Regular text
     }
@@ -442,6 +689,75 @@ pub(crate) fn detect_header_level(
 mod tests {
     use super::*;
 
+    fn line_of(text: &str, font_size: f32, bold: bool, y: f32) -> crate::types::TextLine {
+        let item = crate::types::TextItem {
+            text: text.into(),
+            x: 72.0,
+            y,
+            width: text.len() as f32 * font_size * 0.5,
+            height: font_size,
+            font: "Test".into(),
+            font_size,
+            page: 1,
+            is_bold: bold,
+            is_italic: false,
+            is_underline: false,
+            is_strikeout: false,
+            item_type: crate::types::ItemType::Text,
+            mcid: None,
+        };
+        crate::types::TextLine {
+            items: vec![item],
+            y,
+            page: 1,
+            adaptive_threshold: 0.10,
+        }
+    }
+
+    #[test]
+    fn digit_only_lines_do_not_define_tiers() {
+        // A 14pt bold page number must not claim tier 0 — that both demotes
+        // every real heading a level and blocks the bold-size fallback.
+        let lines = vec![
+            line_of("76", 14.0, true, 760.0),
+            line_of("Replace", 11.0, true, 700.0),
+            line_of("body text at eleven points", 11.0, false, 680.0),
+        ];
+        let tiers = compute_heading_tiers(&lines, 11.0);
+        assert!(tiers.is_empty(), "page number claimed a tier: {tiers:?}");
+    }
+
+    #[test]
+    fn bold_fallback_tiers_when_nothing_clears_ratio_gate() {
+        // 10pt body, 11pt bold section headings (book-style): no size clears
+        // 1.2x, so bold sizes modestly above body form the tiers.
+        let lines = vec![
+            line_of("4. Entropy", 11.0, true, 700.0),
+            line_of("body text about entropy", 10.0, false, 680.0),
+            line_of("5. The dynamics", 11.0, true, 500.0),
+        ];
+        let tiers = compute_heading_tiers(&lines, 10.0);
+        assert_eq!(tiers, vec![11.0]);
+        assert_eq!(detect_header_level(11.0, 10.0, &tiers, true), Some(1));
+        // Non-bold text at the fallback size must not become a heading.
+        assert_eq!(detect_header_level(11.0, 10.0, &tiers, false), None);
+        // Non-tier body text stays regular.
+        assert_eq!(detect_header_level(10.0, 10.0, &tiers, true), None);
+    }
+
+    #[test]
+    fn bold_fallback_skipped_when_real_tiers_exist() {
+        let lines = vec![
+            line_of("Chapter One", 18.0, false, 700.0),
+            line_of("bold label", 11.0, true, 600.0),
+            line_of("body", 10.0, false, 580.0),
+        ];
+        let tiers = compute_heading_tiers(&lines, 10.0);
+        assert_eq!(tiers, vec![18.0]);
+        // The 11pt bold label does not match any tier and stays non-heading.
+        assert_eq!(detect_header_level(11.0, 10.0, &tiers, true), None);
+    }
+
     #[test]
     fn toc_entry_with_single_dot_group() {
         assert!(is_toc_entry_line("Measurement Lab worksheet ... 3"));

src/markdown/convert.rs

@@ -1,5 +1,6 @@
 //! Core line-to-markdown conversion loop with table/image interleaving.
 
+use std::cmp::Ordering;
 use std::collections::{HashMap, HashSet};
 
 use crate::structure_tree::StructRole;
@@ -13,9 +14,149 @@ use super::analysis::{
 use super::classify::{
     format_list_item, is_caption_line, is_list_item, is_monospace_font, starts_with_bullet_marker,
 };
+use super::heading::classify_heading_sequences;
 use super::postprocess::clean_markdown;
 use super::preprocess::{merge_drop_caps, merge_heading_lines};
-use super::MarkdownOptions;
+use super::{item_is_in_chart_region, MarkdownOptions, CHART_SEPARATOR_PAD};
+
+/// Logical stream geometry for a page where one full-width chart separates
+/// two prose columns. Positioned non-text blocks use this same ordering so a
+/// right-column table or image cannot jump ahead of left-column prose.
+#[derive(Debug, Clone, Copy)]
+pub(super) struct ChartProseOrder {
+    split_x: f32,
+    chart_region: (f32, f32, f32, f32),
+}
+
+impl ChartProseOrder {
+    pub(super) fn new(split_x: f32, chart_region: (f32, f32, f32, f32)) -> Self {
+        Self {
+            split_x,
+            chart_region,
+        }
+    }
+}
+
+/// Markdown block with its physical position and optional logical chart-page
+/// stream. Tables and images share this representation because both are
+/// removed before text-line grouping and reinserted during conversion.
+#[derive(Debug, Clone)]
+pub(super) struct PositionedMarkdown {
+    y: f32,
+    x: f32,
+    markdown: String,
+    chart_order: Option<ChartProseOrder>,
+}
+
+impl PositionedMarkdown {
+    pub(super) fn new(
+        y: f32,
+        x: f32,
+        markdown: String,
+        chart_order: Option<ChartProseOrder>,
+    ) -> Self {
+        Self {
+            y,
+            x,
+            markdown,
+            chart_order,
+        }
+    }
+}
+
+fn chart_stream_position(
+    y: f32,
+    x: f32,
+    claimed_by_chart: bool,
+    order: ChartProseOrder,
+) -> (u8, u8) {
+    let (_, y0, _, y1) = order.chart_region;
+    let low = y0.min(y1) - CHART_SEPARATOR_PAD;
+    let high = y0.max(y1) + CHART_SEPARATOR_PAD;
+    let in_chart_zone = claimed_by_chart || (y >= low && y <= high);
+    let zone = if in_chart_zone {
+        1
+    } else if y > high {
+        0
+    } else {
+        2
+    };
+    let column = if in_chart_zone || x < order.split_x {
+        0
+    } else {
+        1
+    };
+    (zone, column)
+}
+
+fn positioned_block_precedes_line(block: &PositionedMarkdown, line: &TextLine) -> bool {
+    let Some(order) = block.chart_order else {
+        return block.y > line.y;
+    };
+    let line_x = line.items.first().map(|item| item.x).unwrap_or(0.0);
+    let line_claimed_by_chart = line
+        .items
+        .iter()
+        .any(|item| item_is_in_chart_region(item, &[order.chart_region]));
+    let block_position = chart_stream_position(block.y, block.x, false, order);
+    let line_position = chart_stream_position(line.y, line_x, line_claimed_by_chart, order);
+    block_position < line_position || (block_position == line_position && block.y > line.y)
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
+enum PositionedBlockKind {
+    Table,
+    Image,
+}
+
+type PositionedBlockRef<'a> = (PositionedBlockKind, usize, &'a PositionedMarkdown);
+
+fn compare_positioned_blocks(
+    (a_kind, a_idx, a): &PositionedBlockRef<'_>,
+    (b_kind, b_idx, b): &PositionedBlockRef<'_>,
+) -> Ordering {
+    if let (Some(a_order), Some(b_order)) = (a.chart_order, b.chart_order) {
+        let a_position = chart_stream_position(a.y, a.x, false, a_order);
+        let b_position = chart_stream_position(b.y, b.x, false, b_order);
+        return a_position
+            .cmp(&b_position)
+            .then_with(|| b.y.total_cmp(&a.y))
+            .then_with(|| a.x.total_cmp(&b.x))
+            .then_with(|| a_kind.cmp(b_kind))
+            .then_with(|| a_idx.cmp(b_idx));
+    }
+
+    // Preserve the legacy ordering for ordinary pages: tables in detection
+    // order, followed by images in input order. Chart pages give every block
+    // a chart order and use the logical stream comparison above.
+    a_kind.cmp(b_kind).then_with(|| a_idx.cmp(b_idx))
+}
+
+fn positioned_blocks_for_page<'a>(
+    page: u32,
+    page_tables: &'a HashMap<u32, Vec<PositionedMarkdown>>,
+    page_images: &'a HashMap<u32, Vec<PositionedMarkdown>>,
+) -> Vec<PositionedBlockRef<'a>> {
+    let mut blocks = Vec::new();
+    if let Some(tables) = page_tables.get(&page) {
+        blocks.extend(
+            tables
+                .iter()
+                .enumerate()
+                .map(|(idx, table)| (PositionedBlockKind::Table, idx, table)),
+        );
+    }
+    if let Some(images) = page_images.get(&page) {
+        blocks.extend(
+            images
+                .iter()
+                .enumerate()
+                .map(|(idx, image)| (PositionedBlockKind::Image, idx, image)),
+        );
+    }
+    blocks.sort_by(compare_positioned_blocks);
+    blocks
+}
 
 /// Pre-scan struct heading tags to find levels that are overused — i.e., tagged on
 /// so many lines that they clearly represent body text, not real headings.
@@ -160,6 +301,112 @@ fn find_isolated_lines(lines: &[TextLine], base_size: f32, para_threshold: f32)
 /// wrapped visual line as "standalone" once the first line is misclassified,
 /// producing a stack of `##` headings.  Multi-line body-size bold runs with a
 /// paragraph-sized word count should stay paragraph text.
+/// Merge 2-3 consecutive all-bold body-size lines into one line when the
+/// group is isolated (paragraph break before and after) and short enough to
+/// be a heading. Longer/wordier bold runs are wrapped bold paragraphs and
+/// are left for `find_wrapped_bold_paragraph_lines` to suppress.
+/// "9.5. ", "12.3.1. " — section-numbered heading prefix followed by a word.
+fn starts_with_section_number(t: &str) -> bool {
+    let t = t.trim_start();
+    let mut rest = t;
+    let mut groups = 0;
+    loop {
+        let digits = rest.chars().take_while(|c| c.is_ascii_digit()).count();
+        if digits == 0 || digits > 3 {
+            break;
+        }
+        groups += 1;
+        rest = &rest[digits..];
+        if let Some(r) = rest.strip_prefix('.') {
+            rest = r;
+        } else {
+            break;
+        }
+    }
+    // Two components minimum ("9.5. "): a single "1. " is an ordered list
+    // item, and this prefix bypasses the isolation checks entirely.
+    groups >= 2
+        && rest.starts_with(char::is_whitespace)
+        && rest.trim_start().starts_with(|c: char| c.is_alphabetic())
+}
+
+fn merge_wrapped_bold_heading_groups(
+    lines: Vec<TextLine>,
+    base_size: f32,
+    para_threshold: f32,
+) -> Vec<TextLine> {
+    let mut out: Vec<TextLine> = Vec::with_capacity(lines.len());
+    let mut i = 0usize;
+    while i < lines.len() {
+        if !is_body_size_all_bold_line(&lines[i], base_size) {
+            out.push(lines[i].clone());
+            i += 1;
+            continue;
+        }
+        let start = i;
+        let mut end = i;
+        let mut word_count = lines[i].text().split_whitespace().count();
+        while end + 1 < lines.len()
+            && is_body_size_all_bold_line(&lines[end + 1], base_size)
+            && is_wrapped_same_style_line(&lines[end], &lines[end + 1], para_threshold)
+        {
+            end += 1;
+            word_count += lines[end].text().split_whitespace().count();
+        }
+        let line_count = end - start + 1;
+        // Column-local isolation: on interleaved multi-column pages the
+        // vector neighbors may be the other column's lines, so judge the
+        // break by x-overlapping lines only.
+        let gx0 = lines[start..=end]
+            .iter()
+            .flat_map(|l| l.items.iter().map(|i| i.x))
+            .fold(f32::INFINITY, f32::min);
+        let gx1 = lines[start..=end]
+            .iter()
+            .flat_map(|l| l.items.iter().map(|i| i.x + i.width))
+            .fold(f32::NEG_INFINITY, f32::max);
+        let overlaps_x = |l: &TextLine| {
+            let lx0 = l.items.iter().map(|i| i.x).fold(f32::INFINITY, f32::min);
+            let lx1 = l
+                .items
+                .iter()
+                .map(|i| i.x + i.width)
+                .fold(f32::NEG_INFINITY, f32::max);
+            lx0 < gx1 && lx1 > gx0
+        };
+        let page = lines[start].page;
+        let break_before = !lines.iter().any(|l| {
+            l.page == page
+                && l.y > lines[start].y
+                && l.y - lines[start].y <= para_threshold
+                && overlaps_x(l)
+        });
+        let break_after = !lines.iter().any(|l| {
+            l.page == page
+                && l.y < lines[end].y
+                && lines[end].y - l.y <= para_threshold
+                && overlaps_x(l)
+        });
+        let numbered = starts_with_section_number(&lines[start].text());
+        if (2..=3).contains(&line_count)
+            && word_count <= 15
+            && ((break_before && break_after) || numbered)
+        {
+            let mut merged = lines[start].clone();
+            for l in &lines[start + 1..=end] {
+                merged.items.extend(l.items.iter().cloned());
+            }
+            out.push(merged);
+        } else {
+            for l in &lines[start..=end] {
+                out.push(l.clone());
+            }
+        }
+        i = end + 1;
+    }
+    out
+}
+
 fn find_wrapped_bold_paragraph_lines(
     lines: &[TextLine],
     base_size: f32,
@@ -277,7 +524,7 @@ fn struct_role_heading_level(role: &StructRole) -> Option<usize> {
 /// We strip their header+separator rows and append their data rows to the first page's
 /// table, then remove them from later pages.
 pub(super) fn merge_continuation_tables(
-    page_tables: &mut std::collections::HashMap<u32, Vec<(f32, String)>>,
+    page_tables: &mut std::collections::HashMap<u32, Vec<PositionedMarkdown>>,
     table_only_pages: &HashSet<u32>,
 ) {
     let mut sorted_pages: Vec<u32> = page_tables.keys().copied().collect();
@@ -305,7 +552,7 @@ pub(super) fn merge_continuation_tables(
             continue;
         }
 
-        let first_col_count = count_table_columns(&first_tables[0].1);
+        let first_col_count = count_table_columns(&first_tables[0].markdown);
         if first_col_count == 0 {
             i += 1;
             continue;
@@ -336,7 +583,7 @@ pub(super) fn merge_continuation_tables(
                 _ => break,
             };
 
-            let next_col_count = count_table_columns(&next_tables[0].1);
+            let next_col_count = count_table_columns(&next_tables[0].markdown);
             if next_col_count != first_col_count {
                 break;
             }
@@ -350,7 +597,7 @@ pub(super) fn merge_continuation_tables(
             let mut extra_rows = String::new();
             for &cont_page in &continuation_pages {
                 if let Some(tables) = page_tables.get(&cont_page) {
-                    let table_md = &tables[0].1;
+                    let table_md = &tables[0].markdown;
                     // Skip header row (line 1) and separator row (line 2), keep the rest
                     for (line_idx, line) in table_md.lines().enumerate() {
                         if line_idx >= 2 {
@@ -363,7 +610,7 @@ pub(super) fn merge_continuation_tables(
 
             // Append continuation rows to the first page's table
             if let Some(tables) = page_tables.get_mut(&first_page) {
-                tables[0].1.push_str(&extra_rows);
+                tables[0].markdown.push_str(&extra_rows);
             }
 
             // Remove continuation pages from the map
@@ -395,37 +642,35 @@ fn count_table_columns(table_md: &str) -> usize {
 /// Flush any remaining tables and images for a given page
 fn flush_page_tables_and_images(
     page: u32,
-    page_tables: &std::collections::HashMap<u32, Vec<(f32, String)>>,
-    page_images: &std::collections::HashMap<u32, Vec<(f32, String)>>,
+    page_blocks: &HashMap<u32, Vec<PositionedBlockRef<'_>>>,
     inserted_tables: &mut HashSet<(u32, usize)>,
     inserted_images: &mut HashSet<(u32, usize)>,
     output: &mut String,
     in_paragraph: &mut bool,
 ) {
-    if let Some(tables) = page_tables.get(&page) {
-        for (idx, (_, table_md)) in tables.iter().enumerate() {
-            if !inserted_tables.contains(&(page, idx)) {
-                if *in_paragraph {
-                    output.push_str("\n\n");
-                    *in_paragraph = false;
-                }
-                output.push('\n');
-                output.push_str(table_md);
-                output.push('\n');
+    let Some(blocks) = page_blocks.get(&page) else {
+        return;
+    };
+    for &(kind, idx, block) in blocks {
+        let already_inserted = match kind {
+            PositionedBlockKind::Table => inserted_tables.contains(&(page, idx)),
+            PositionedBlockKind::Image => inserted_images.contains(&(page, idx)),
+        };
+        if already_inserted {
+            continue;
+        }
+        if *in_paragraph {
+            output.push_str("\n\n");
+            *in_paragraph = false;
+        }
+        output.push('\n');
+        output.push_str(&block.markdown);
+        output.push('\n');
+        match kind {
+            PositionedBlockKind::Table => {
                 inserted_tables.insert((page, idx));
             }
-        }
-    }
-    if let Some(images) = page_images.get(&page) {
-        for (idx, (_, image_md)) in images.iter().enumerate() {
-            if !inserted_images.contains(&(page, idx)) {
-                if *in_paragraph {
-                    output.push_str("\n\n");
-                    *in_paragraph = false;
-                }
-                output.push('\n');
-                output.push_str(image_md);
-                output.push('\n');
+            PositionedBlockKind::Image => {
                 inserted_images.insert((page, idx));
             }
         }
@@ -436,8 +681,9 @@ fn flush_page_tables_and_images(
 pub(super) fn to_markdown_from_lines_with_tables_and_images(
     lines: Vec<TextLine>,
     options: MarkdownOptions,
-    page_tables: std::collections::HashMap<u32, Vec<(f32, String)>>,
-    page_images: std::collections::HashMap<u32, Vec<(f32, String)>>,
+    page_tables: std::collections::HashMap<u32, Vec<PositionedMarkdown>>,
+    page_images: std::collections::HashMap<u32, Vec<PositionedMarkdown>>,
+    page_chart_regions: &std::collections::HashMap<u32, Vec<(f32, f32, f32, f32)>>,
     band_split_pages: &HashSet<u32>,
     struct_roles: Option<
         &std::collections::HashMap<u32, std::collections::HashMap<i64, StructRole>>,
@@ -468,6 +714,17 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
     // threshold and cause every line to be treated as a paragraph break.
     let para_threshold = compute_paragraph_threshold(&lines, base_size);
 
+    // Merge wrapped bold headings: a 2-3 line group of consecutive all-bold
+    // body-size lines that is isolated as a group (paragraph break before
+    // and after) is one heading that wrapped. Left split, the internal line
+    // gap breaks each line's isolation and neither classifies as a heading —
+    // the whole group then merges into the following body paragraph.
+    let lines = if std::env::var("PI_NO_MERGE").is_ok() {
+        lines
+    } else {
+        merge_wrapped_bold_heading_groups(lines, base_size, para_threshold)
+    };
+
     // Pre-scan: identify isolated lines (paragraph break before AND after).
     // These are heading candidates even without bold/large font — common in
     // academic papers where section titles like "Acknowledgements" sit alone
@@ -477,6 +734,33 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
     let wrapped_bold_paragraph_lines =
         find_wrapped_bold_paragraph_lines(&lines, base_size, para_threshold);
 
+    let mut sequence_excluded_lines = wrapped_bold_paragraph_lines.clone();
+    for (line_idx, line) in lines.iter().enumerate() {
+        if page_chart_regions.get(&line.page).is_some_and(|regions| {
+            line.items
+                .iter()
+                .any(|item| item_is_in_chart_region(item, regions))
+        }) {
+            sequence_excluded_lines.insert(line_idx);
+        }
+    }
+    if let Some(roles) = struct_roles {
+        for (line_idx, line) in lines.iter().enumerate() {
+            if resolve_line_struct_role(line, roles)
+                .is_some_and(|role| role.is_non_heading_content())
+            {
+                sequence_excluded_lines.insert(line_idx);
+            }
+        }
+    }
+    let sequence_heading_levels = classify_heading_sequences(
+        &lines,
+        base_size,
+        &heading_tiers,
+        &isolated_lines,
+        &sequence_excluded_lines,
+    );
+
     // Detect struct heading levels that are overused (body text mistagged as headings)
     let overused_heading_levels = detect_overused_struct_heading_levels(&lines, struct_roles);
 
@@ -502,6 +786,18 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
         .collect();
     all_content_pages.sort();
     all_content_pages.dedup();
+    // Build the unified table/image order once per page. This is only a
+    // meaningful sort on chart/prose pages; ordinary pages retain their
+    // legacy table-then-image order without repeating work for every line.
+    let page_blocks: HashMap<u32, Vec<PositionedBlockRef<'_>>> = all_content_pages
+        .iter()
+        .map(|&page| {
+            (
+                page,
+                positioned_blocks_for_page(page, &page_tables, &page_images),
+            )
+        })
+        .collect();
 
     for (line_idx, line) in lines.iter().enumerate() {
         // Page break
@@ -514,8 +810,7 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
                 }
                 flush_page_tables_and_images(
                     current_page,
-                    &page_tables,
-                    &page_images,
+                    &page_blocks,
                     &mut inserted_tables,
                     &mut inserted_images,
                     &mut output,
@@ -539,8 +834,7 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
                 }
                 flush_page_tables_and_images(
                     p,
-                    &page_tables,
-                    &page_images,
+                    &page_blocks,
                     &mut inserted_tables,
                     &mut inserted_images,
                     &mut output,
@@ -563,38 +857,32 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
             }
         }
 
-        // Check if we should insert a table before this line
-        if let Some(tables) = page_tables.get(&current_page) {
-            for (idx, (table_y, table_md)) in tables.iter().enumerate() {
-                // Insert table when we pass its Y position
-                if *table_y > line.y && !inserted_tables.contains(&(current_page, idx)) {
+        // Insert tables and images through one ordered stream. Chart/prose
+        // pages sort by zone, column, and physical Y; ordinary pages retain
+        // the legacy table-then-image input order.
+        if let Some(blocks) = page_blocks.get(&current_page) {
+            for &(kind, idx, block) in blocks {
+                let already_inserted = match kind {
+                    PositionedBlockKind::Table => inserted_tables.contains(&(current_page, idx)),
+                    PositionedBlockKind::Image => inserted_images.contains(&(current_page, idx)),
+                };
+                if positioned_block_precedes_line(block, line) && !already_inserted {
                     if in_paragraph {
                         output.push_str("\n\n");
                         in_paragraph = false;
                         paragraph_in_wrapped_bold_run = false;
                     }
                     output.push('\n');
-                    output.push_str(table_md);
+                    output.push_str(&block.markdown);
                     output.push('\n');
-                    inserted_tables.insert((current_page, idx));
-                }
-            }
-        }
-
-        // Check if we should insert an image before this line
-        if let Some(images) = page_images.get(&current_page) {
-            for (idx, (image_y, image_md)) in images.iter().enumerate() {
-                // Insert image when we pass its Y position
-                if *image_y > line.y && !inserted_images.contains(&(current_page, idx)) {
-                    if in_paragraph {
-                        output.push_str("\n\n");
-                        in_paragraph = false;
-                        paragraph_in_wrapped_bold_run = false;
+                    match kind {
+                        PositionedBlockKind::Table => {
+                            inserted_tables.insert((current_page, idx));
+                        }
+                        PositionedBlockKind::Image => {
+                            inserted_images.insert((current_page, idx));
+                        }
                     }
-                    output.push('\n');
-                    output.push_str(image_md);
-                    output.push('\n');
-                    inserted_images.insert((current_page, idx));
                 }
             }
         }
@@ -721,7 +1009,13 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
             && toc_suppress_page != Some(line.page)
         {
             let line_font_size = line.items.first().map(|i| i.font_size).unwrap_or(base_size);
-            detect_header_level(line_font_size, base_size, &heading_tiers).or_else(|| {
+            detect_header_level(
+                line_font_size,
+                base_size,
+                &heading_tiers,
+                crate::markdown::analysis::line_is_mostly_bold(line),
+            )
+            .or_else(|| {
                 // Rarity-based heading detection (inspired by opendataloader).
                 // Heading probability scoring with lookahead context.
                 // Score = rarity * 0.5 + bold * 0.3 + standalone * 0.2
@@ -754,16 +1048,23 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
                 // paragraph continuity and minor font-size variation
                 // inflates rarity scores.
                 let has_strong_signal = all_bold || isolated || (rarity >= 0.97 && word_count <= 8);
-                // Single-word headings ("IMPLEMENTATION", "CONTENTS") are common;
-                // accept them only with the strongest signal combination.
-                let enough_words =
-                    word_count >= 2 || (all_bold && isolated && plain_trimmed.len() >= 4);
-                if score >= 0.5 && standalone && enough_words && has_strong_signal {
+                // Single-word headings ("IMPLEMENTATION", "CONTENTS",
+                // "Replace") are common. All-bold single words qualify when
+                // standalone (paragraph break before / page top) — headings
+                // hug their section's first paragraph, so requiring a break
+                // after as well missed most of them. Mixed bold lead-ins
+                // ("Note: ...") are excluded by all_bold.
+                let enough_words = word_count >= 2 || (all_bold && plain_trimmed.len() >= 4);
+                let numbered_bold = all_bold && starts_with_section_number(plain_trimmed);
+                if numbered_bold
+                    || (score >= 0.5 && standalone && enough_words && has_strong_signal)
+                {
                     Some(bold_heading_level(&heading_tiers))
                 } else {
                     None
                 }
             })
+            .or_else(|| sequence_heading_levels.get(&line_idx).copied())
         } else {
             None
         };
@@ -915,8 +1216,7 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
     // (handles table-only pages after the last text line, and trailing image-only pages)
     flush_page_tables_and_images(
         current_page,
-        &page_tables,
-        &page_images,
+        &page_blocks,
         &mut inserted_tables,
         &mut inserted_images,
         &mut output,
@@ -928,8 +1228,7 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
         }
         flush_page_tables_and_images(
             p,
-            &page_tables,
-            &page_images,
+            &page_blocks,
             &mut inserted_tables,
             &mut inserted_images,
             &mut output,
@@ -973,6 +1272,13 @@ pub fn to_markdown_from_lines(lines: Vec<TextLine>, options: MarkdownOptions) ->
     let isolated_lines = find_isolated_lines(&lines, base_size, para_threshold);
     let wrapped_bold_paragraph_lines =
         find_wrapped_bold_paragraph_lines(&lines, base_size, para_threshold);
+    let sequence_heading_levels = classify_heading_sequences(
+        &lines,
+        base_size,
+        &heading_tiers,
+        &isolated_lines,
+        &wrapped_bold_paragraph_lines,
+    );
 
     let mut output = String::new();
     let mut current_page = 0u32;
@@ -1067,33 +1373,39 @@ pub fn to_markdown_from_lines(lines: Vec<TextLine>, options: MarkdownOptions) ->
             && !(options.detect_code && line.items.iter().any(|i| is_monospace_font(&i.font)))
         {
             let line_font_size = line.items.first().map(|i| i.font_size).unwrap_or(base_size);
-            if let Some(header_level) =
-                detect_header_level(line_font_size, base_size, &heading_tiers).or_else(|| {
-                    if line_font_size < base_size * 0.95 {
-                        return None;
-                    }
-                    let word_count = plain_trimmed.split_whitespace().count();
-                    if !(1..=15).contains(&word_count) {
-                        return None;
-                    }
-                    if wrapped_bold_paragraph_lines.contains(&line_idx) {
-                        return None;
-                    }
-                    let rarity = font_size_rarity(line_font_size, &font_stats);
-                    let all_bold = !line.items.is_empty() && line.items.iter().all(|i| i.is_bold);
-                    let standalone = !in_paragraph;
-                    let isolated = isolated_lines.contains(&line_idx);
-                    let score = rarity * 0.5
-                        + if all_bold { 0.3 } else { 0.0 }
-                        + if standalone { 0.2 } else { 0.0 }
-                        + if isolated { 0.3 } else { 0.0 };
-                    let enough_words =
-                        word_count >= 2 || (all_bold && isolated && plain_trimmed.len() >= 4);
-                    if score >= 0.5 && standalone && enough_words {
-                        return Some(bold_heading_level(&heading_tiers));
-                    }
-                    None
-                })
+            if let Some(header_level) = detect_header_level(
+                line_font_size,
+                base_size,
+                &heading_tiers,
+                crate::markdown::analysis::line_is_mostly_bold(line),
+            )
+            .or_else(|| {
+                if line_font_size < base_size * 0.95 {
+                    return None;
+                }
+                let word_count = plain_trimmed.split_whitespace().count();
+                if !(1..=15).contains(&word_count) {
+                    return None;
+                }
+                if wrapped_bold_paragraph_lines.contains(&line_idx) {
+                    return None;
+                }
+                let rarity = font_size_rarity(line_font_size, &font_stats);
+                let all_bold = !line.items.is_empty() && line.items.iter().all(|i| i.is_bold);
+                let standalone = !in_paragraph;
+                let isolated = isolated_lines.contains(&line_idx);
+                let score = rarity * 0.5
+                    + if all_bold { 0.3 } else { 0.0 }
+                    + if standalone { 0.2 } else { 0.0 }
+                    + if isolated { 0.3 } else { 0.0 };
+                let enough_words =
+                    word_count >= 2 || (all_bold && isolated && plain_trimmed.len() >= 4);
+                if score >= 0.5 && standalone && enough_words {
+                    return Some(bold_heading_level(&heading_tiers));
+                }
+                None
+            })
+            .or_else(|| sequence_heading_levels.get(&line_idx).copied())
             {
                 if in_paragraph {
                     output.push_str("\n\n");
@@ -1204,6 +1516,20 @@ pub fn to_markdown_from_lines(lines: Vec<TextLine>, options: MarkdownOptions) ->
 
 #[cfg(test)]
 mod tests {
+
+    #[test]
+    fn section_number_prefix_detection() {
+        assert!(starts_with_section_number(
+            "9.5. Adapting to the New Normal"
+        ));
+        assert!(starts_with_section_number("12.3.1. Deep subsection"));
+        assert!(starts_with_section_number("2.1 Systems thinking"));
+        assert!(!starts_with_section_number("1. First item in a list"));
+        assert!(!starts_with_section_number("24% in October 2020."));
+        assert!(!starts_with_section_number("2020 was a hard year"));
+        assert!(!starts_with_section_number("Introduction"));
+    }
+
     use super::*;
     use crate::structure_tree::StructRole;
     use crate::types::TextItem;
@@ -1245,6 +1571,91 @@ mod tests {
         make_line(vec![item])
     }
 
+    #[test]
+    fn chart_page_blocks_follow_zone_and_column_stream() {
+        let line = |text: &str, x: f32, y: f32| {
+            let mut item = make_item(text, 1, None);
+            item.x = x;
+            item.y = y;
+            make_line(vec![item])
+        };
+        // Logical newspaper order for the prose zone: all left-column lines,
+        // then all right-column lines, even though their physical Y values
+        // jump back upward at the column switch.
+        let lines = vec![
+            line("Left column upper prose.", 90.0, 700.0),
+            line("Left column lower prose.", 90.0, 500.0),
+            line("Right column upper prose.", 340.0, 700.0),
+            line("Right column lower prose.", 340.0, 500.0),
+        ];
+        let order = ChartProseOrder::new(280.0, (100.0, 300.0, 500.0, 400.0));
+        let mut tables = HashMap::new();
+        tables.insert(
+            1,
+            vec![
+                // Detection order is deliberately right before left.
+                PositionedMarkdown::new(
+                    600.0,
+                    340.0,
+                    "| Right metric | Value |\n|---|---|\n| A | 1 |\n".into(),
+                    Some(order),
+                ),
+                PositionedMarkdown::new(
+                    550.0,
+                    90.0,
+                    "| Left metric | Value |\n|---|---|\n| B | 2 |\n".into(),
+                    Some(order),
+                ),
+            ],
+        );
+        let mut images = HashMap::new();
+        images.insert(
+            1,
+            vec![
+                PositionedMarkdown::new(
+                    575.0,
+                    90.0,
+                    "![Left column figure](left-image)\n".into(),
+                    Some(order),
+                ),
+                PositionedMarkdown::new(
+                    575.0,
+                    340.0,
+                    "![Right column figure](right-image)\n".into(),
+                    Some(order),
+                ),
+            ],
+        );
+
+        let md = to_markdown_from_lines_with_tables_and_images(
+            lines,
+            MarkdownOptions::default(),
+            tables,
+            images,
+            &HashMap::new(),
+            &HashSet::from([1]),
+            None,
+        );
+        let positions = [
+            "Left column upper prose.",
+            "![Left column figure](left-image)",
+            "| Left metric | Value |",
+            "Left column lower prose.",
+            "Right column upper prose.",
+            "| Right metric | Value |",
+            "![Right column figure](right-image)",
+            "Right column lower prose.",
+        ]
+        .map(|needle| {
+            md.find(needle)
+                .unwrap_or_else(|| panic!("missing {needle:?} in {md}"))
+        });
+        assert!(
+            positions.windows(2).all(|pair| pair[0] < pair[1]),
+            "blocks must follow the logical chart-page stream: {md}"
+        );
+    }
+
     #[test]
     fn isolated_lines_kept_on_sparse_pages() {
         // A ToC page with a lone title and one entry far below: the density
@@ -1297,6 +1708,7 @@ mod tests {
             MarkdownOptions::default(),
             HashMap::new(),
             HashMap::new(),
+            &HashMap::new(),
             &std::collections::HashSet::new(),
             Some(&roles),
         );
@@ -1325,6 +1737,7 @@ mod tests {
             MarkdownOptions::default(),
             HashMap::new(),
             HashMap::new(),
+            &HashMap::new(),
             &std::collections::HashSet::new(),
             Some(&roles),
         );
@@ -1366,6 +1779,7 @@ mod tests {
             MarkdownOptions::default(),
             HashMap::new(),
             HashMap::new(),
+            &HashMap::new(),
             &std::collections::HashSet::new(),
             Some(&roles),
         );
@@ -1402,6 +1816,7 @@ mod tests {
             MarkdownOptions::default(),
             HashMap::new(),
             HashMap::new(),
+            &HashMap::new(),
             &std::collections::HashSet::new(),
             Some(&roles),
         );
@@ -1437,6 +1852,7 @@ mod tests {
             MarkdownOptions::default(),
             HashMap::new(),
             HashMap::new(),
+            &HashMap::new(),
             &std::collections::HashSet::new(),
             Some(&roles),
         );
@@ -1466,6 +1882,7 @@ mod tests {
             MarkdownOptions::default(),
             HashMap::new(),
             HashMap::new(),
+            &HashMap::new(),
             &std::collections::HashSet::new(),
             None,
         );
@@ -1507,6 +1924,7 @@ mod tests {
             MarkdownOptions::default(),
             HashMap::new(),
             HashMap::new(),
+            &HashMap::new(),
             &std::collections::HashSet::new(),
             Some(&roles),
         );
@@ -1555,6 +1973,7 @@ mod tests {
             MarkdownOptions::default(),
             HashMap::new(),
             HashMap::new(),
+            &HashMap::new(),
             &std::collections::HashSet::new(),
             None,
         );
@@ -1656,6 +2075,7 @@ mod tests {
             MarkdownOptions::default(),
             HashMap::new(),
             HashMap::new(),
+            &HashMap::new(),
             &std::collections::HashSet::new(),
             None,
         );
@@ -1703,6 +2123,7 @@ mod tests {
             MarkdownOptions::default(),
             HashMap::new(),
             HashMap::new(),
+            &HashMap::new(),
             &std::collections::HashSet::new(),
             Some(&roles),
         );
@@ -1842,6 +2263,7 @@ mod tests {
             MarkdownOptions::default(),
             HashMap::new(),
             HashMap::new(),
+            &HashMap::new(),
             &std::collections::HashSet::new(),
             None,
         );

src/markdown/postprocess.rs

@@ -2,12 +2,16 @@
 
 use regex::Regex;
 
-use super::MarkdownOptions;
+use super::{MarkdownOptions, MarkdownProfile};
+use crate::text_utils::is_page_number_line;
 
 /// Clean up markdown output with post-processing
 pub(crate) fn clean_markdown(mut text: String, options: &MarkdownOptions) -> String {
-    // Collapse dot leaders (e.g. TOC entries: "Introduction...............................1")
-    text = collapse_dot_leaders(&text);
+    if options.profile == MarkdownProfile::Compact {
+        // Dot-leader collapse saves tokens but changes source text, so it is
+        // reserved for the explicit compact profile.
+        text = collapse_dot_leaders(&text);
+    }
 
     // Fix hyphenation first (before other processing)
     if options.fix_hyphenation {
@@ -142,7 +146,7 @@ fn fix_hyphenation(text: &str) -> String {
     result
 }
 
-/// Remove standalone page numbers (lines that are just 1-4 digit numbers)
+/// Remove isolated page-number expressions from Markdown.
 fn remove_page_numbers(text: &str) -> String {
     let mut result = Vec::new();
     let lines: Vec<&str> = text.lines().collect();
@@ -180,69 +184,6 @@ fn remove_page_numbers(text: &str) -> String {
     result.join("\n")
 }
 
-/// Check if a line looks like a page number
-fn is_page_number_line(trimmed: &str) -> bool {
-    // Empty lines are not page numbers
-    if trimmed.is_empty() {
-        return false;
-    }
-
-    // Pattern 1: Just a number (1-4 digits)
-    if trimmed.len() <= 4 && trimmed.chars().all(|c| c.is_ascii_digit()) {
-        return true;
-    }
-
-    // Pattern 2: "Page X of Y" or "Page X" or "Page   of" (placeholder)
-    let lower = trimmed.to_lowercase();
-    if let Some(rest) = lower.strip_prefix("page") {
-        let rest = rest.trim();
-        // "Page   of" (empty page numbers)
-        if rest == "of" || rest.starts_with("of ") {
-            return true;
-        }
-        // "Page X" or "Page X of Y"
-        if rest
-            .chars()
-            .next()
-            .map(|c| c.is_ascii_digit())
-            .unwrap_or(false)
-        {
-            return true;
-        }
-        // Just "Page" followed by whitespace and maybe "of"
-        if rest.is_empty()
-            || rest
-                .split_whitespace()
-                .all(|w| w == "of" || w.chars().all(|c| c.is_ascii_digit()))
-        {
-            return true;
-        }
-    }
-
-    // Pattern 3: "X of Y" where X and Y are numbers
-    if let Some(of_idx) = trimmed.find(" of ") {
-        let before = trimmed[..of_idx].trim();
-        let after = trimmed[of_idx + 4..].trim();
-        if before.chars().all(|c| c.is_ascii_digit())
-            && after.chars().all(|c| c.is_ascii_digit())
-            && !before.is_empty()
-            && !after.is_empty()
-        {
-            return true;
-        }
-    }
-
-    // Pattern 4: "- X -" centered page number
-    if trimmed.len() >= 3 && trimmed.starts_with('-') && trimmed.ends_with('-') {
-        let inner = trimmed[1..trimmed.len() - 1].trim();
-        if inner.chars().all(|c| c.is_ascii_digit()) && !inner.is_empty() {
-            return true;
-        }
-    }
-
-    false
-}
-
 /// Convert URLs to markdown links
 fn format_urls(text: &str) -> String {
     use once_cell::sync::Lazy;
@@ -355,6 +296,23 @@ fn format_urls(text: &str) -> String {
 mod tests {
     use super::*;
 
+    #[test]
+    fn fidelity_profile_preserves_dot_leaders() {
+        let input = "Introduction............................1".to_string();
+        let result = clean_markdown(input.clone(), &MarkdownOptions::default());
+        assert_eq!(result, format!("{input}\n"));
+    }
+
+    #[test]
+    fn compact_profile_collapses_dot_leaders() {
+        let input = "Introduction............................1".to_string();
+        let options = MarkdownOptions {
+            profile: MarkdownProfile::Compact,
+            ..MarkdownOptions::default()
+        };
+        assert_eq!(clean_markdown(input, &options), "Introduction ... 1\n");
+    }
+
     // --- collapse_dot_leaders ---
 
     #[test]
@@ -469,12 +427,14 @@ mod tests {
     fn test_is_page_number_page_x() {
         assert!(is_page_number_line("Page 5"));
         assert!(is_page_number_line("page 12"));
+        assert!(is_page_number_line("Page123"));
     }
 
     #[test]
     fn test_is_page_number_page_x_of_y() {
         assert!(is_page_number_line("Page 3 of 10"));
         assert!(is_page_number_line("page 1 of 5"));
+        assert!(is_page_number_line("Page 3 of 10 Report header"));
     }
 
     #[test]
@@ -504,6 +464,13 @@ mod tests {
         assert!(!is_page_number_line("Hello World"));
         assert!(!is_page_number_line("Chapter 1"));
         assert!(!is_page_number_line("Total: 500"));
+        assert!(!is_page_number_line("PAGE0-PARA2-END-MARKER-0"));
+    }
+
+    #[test]
+    fn test_is_page_number_labeled_running_header() {
+        assert!(is_page_number_line("Page 42 Chapter 5"));
+        assert!(is_page_number_line("Page 42 explains the result"));
     }
 
     // --- remove_page_numbers ---
@@ -531,6 +498,24 @@ mod tests {
         assert!(result.contains("42"));
     }
 
+    #[test]
+    fn test_remove_page_numbers_labeled_header_with_content() {
+        let input = "Content\n\nPage 42 explains the result\n---\nEnd";
+        let result = remove_page_numbers(input);
+
+        assert!(!result.contains("Page 42 explains the result"));
+        assert!(result.contains("Content"));
+        assert!(result.contains("End"));
+    }
+
+    #[test]
+    fn test_remove_page_numbers_preserves_page_prefixed_content() {
+        let input = "PAGE0-PARA2-START substantive report text PAGE0-PARA2-END-MARKER-0";
+        let result = remove_page_numbers(input);
+
+        assert_eq!(result, input);
+    }
+
     #[test]
     fn test_remove_page_numbers_multiple_patterns() {
         let input = "\n1\n\nContent\n\n2\n\n---\nMore\n\n3\n";

src/markdown/preprocess.rs

@@ -42,7 +42,12 @@ fn effective_heading_level(
 
     // Fall back to font-size heuristic
     let font = line.items.first().map(|i| i.font_size).unwrap_or(base_size);
-    detect_header_level(font, base_size, heading_tiers)
+    detect_header_level(
+        font,
+        base_size,
+        heading_tiers,
+        crate::markdown::analysis::line_is_mostly_bold(line),
+    )
 }
 
 /// Merge consecutive heading lines at the same level into a single line.

src/structure_tree.rs

@@ -9,7 +9,7 @@
 use log::debug;
 use lopdf::{Document, Object, ObjectId};
 use std::borrow::Cow;
-use std::collections::HashMap;
+use std::collections::{HashMap, HashSet};
 
 // ─── Standard structure types ────────────────────────────────────────
 
@@ -252,10 +252,28 @@ impl StructTree {
         let role_map = parse_role_map(doc, struct_root);
         debug!("structure tree: {} role map entries", role_map.len());
 
+        // Seed the cycle guard with the struct-root's own object id so a `/K`
+        // that points back at the root is treated as a cycle, and bound total
+        // node materialization with a global budget.
+        let mut walk = StructWalk::new();
+        if let Ok(root_id) = struct_root_obj.as_reference() {
+            walk.active.insert(root_id);
+        }
+
         // Parse child elements from /K
-        let children = parse_kids(doc, struct_root, &role_map, None, 0);
+        let children = parse_kids(doc, struct_root, &role_map, None, 0, &mut walk);
         debug!("structure tree: {} top-level elements", children.len());
 
+        if walk.truncated {
+            log::warn!(
+                "structure tree parsing was truncated (node budget of \
+                 {MAX_STRUCT_NODES} or traversal budget of {MAX_STRUCT_WORK} \
+                 reached, a `/K` reference cycle, or the max nesting depth of \
+                 {MAX_DEPTH}); tagged roles/tables may be incomplete (likely a \
+                 very large or malformed tagged PDF)"
+            );
+        }
+
         if children.is_empty() {
             return None;
         }
@@ -479,6 +497,123 @@ fn parse_role_map(doc: &Document, struct_root: &lopdf::Dictionary) -> HashMap<St
 /// malformed PDFs).
 const MAX_DEPTH: usize = 64;
 
+/// Global cap on the number of structure-tree nodes materialized in a single
+/// parse. Real tagged trees are far smaller; a crafted PDF can alias one struct
+/// element into its own `/K` (e.g. `/K [n 0 R n 0 R]`) so the tree branches
+/// exponentially (2^depth) before the depth cap is reached, exhausting memory.
+/// This budget bounds total work and allocation regardless of tree shape.
+const MAX_STRUCT_NODES: usize = 500_000;
+
+/// Cap on the number of `/K` items *examined* during a single parse, regardless
+/// of whether they materialize anything. Bounds CPU for crafted wide `/K` arrays
+/// of non-materializing entries (unsupported value types, `/OBJR` dicts, cycle
+/// back-edges) that would otherwise be scanned in full without ever touching the
+/// node budget. Kept well above the node budget so it never truncates content
+/// that already fits within `MAX_STRUCT_NODES`.
+const MAX_STRUCT_WORK: usize = 2_000_000;
+
+/// Traversal state shared across the recursive structure-tree parse.
+///
+/// `budget` is a global allowance charged once per materialized item — each
+/// struct-element node and each marked-content reference — so total work is
+/// bounded even for aliased/DAG-shaped `/K` graphs of distinct objects or a
+/// single element with a very wide `/K` array. `active` holds the object IDs
+/// currently on the depth-first path so a struct element that references itself
+/// (or an ancestor) is not expanded into an unbounded/exponential subtree.
+/// `budget` bounds *materialization* (nodes + content refs). `work` separately
+/// bounds *traversal* — every `/K` item examined is charged against it, even
+/// ones that materialize nothing (unsupported values, `/OBJR`, cycle back-edges)
+/// — so a wide malformed array cannot force an unbounded scan, and those skipped
+/// items don't drain the materialization budget and truncate real content.
+/// `truncated` records whether any parse work was skipped — the budget was
+/// exhausted, a `/K` reference cycle was broken, or the depth cap was hit — so
+/// the caller can log it once rather than per skipped item. `stalled` is set
+/// when an atomic multi-unit reservation could not fit in the remaining budget;
+/// it makes [`exhausted`](Self::exhausted) report done so a wide `/K` array is
+/// not scanned to the end once no further leaf can be materialized.
+struct StructWalk {
+    budget: usize,
+    work: usize,
+    active: HashSet<ObjectId>,
+    truncated: bool,
+    stalled: bool,
+}
+
+impl StructWalk {
+    fn new() -> Self {
+        Self {
+            budget: MAX_STRUCT_NODES,
+            work: MAX_STRUCT_WORK,
+            active: HashSet::new(),
+            truncated: false,
+            stalled: false,
+        }
+    }
+
+    /// Charge one unit of traversal work for an examined `/K` item, whether or
+    /// not it materializes anything. Returns `false` (flagging truncation) once
+    /// the traversal budget is spent, so an enclosing loop stops instead of
+    /// scanning the rest of a wide array of non-materializing entries.
+    fn spend_work(&mut self) -> bool {
+        if self.work == 0 {
+            self.truncated = true;
+            return false;
+        }
+        self.work -= 1;
+        true
+    }
+
+    /// Record that some parse work was skipped for a non-budget reason (a `/K`
+    /// reference cycle or the depth cap), so the one-shot truncation warning
+    /// also covers malformed/over-deep trees, not just budget exhaustion.
+    fn note_skipped(&mut self) {
+        self.truncated = true;
+    }
+
+    /// Charge one unit against the budget for a materialized item (a struct
+    /// element node or a marked-content reference). Returns `false` — without
+    /// underflowing — once the budget is exhausted, so callers skip the item.
+    fn charge(&mut self) -> bool {
+        if self.budget == 0 {
+            self.truncated = true;
+            return false;
+        }
+        self.budget -= 1;
+        true
+    }
+
+    /// Atomically charge `n` units for a single item that materializes several
+    /// budget-counted parts at once (a leaf wrapper node *plus* its content
+    /// reference). Charges nothing when fewer than `n` units remain — so a
+    /// partial reservation never wastes capacity — and marks the walk `stalled`
+    /// so the enclosing loop stops instead of scanning the rest of a wide `/K`
+    /// array that can no longer fit any leaf.
+    fn charge_n(&mut self, n: usize) -> bool {
+        if self.budget < n {
+            self.truncated = true;
+            self.stalled = true;
+            return false;
+        }
+        self.budget -= n;
+        true
+    }
+
+    /// Whether traversal should stop: the budget is spent, or a multi-unit
+    /// reservation could not fit (`stalled`) so no further leaf will materialize.
+    /// Use this at the guards that break/return to skip remaining items; it
+    /// records that truncation occurred (a guard only fires while an item is
+    /// still pending), so callers that drop work without going through
+    /// [`charge`](Self::charge) still flag the truncation for logging.
+    fn exhausted(&mut self) -> bool {
+        if self.budget == 0 || self.stalled {
+            self.truncated = true;
+            true
+        } else {
+            false
+        }
+    }
+}
+
 /// Parse child elements from a `/K` entry.
 fn parse_kids(
     doc: &Document,
@@ -486,8 +621,13 @@ fn parse_kids(
     role_map: &HashMap<String, String>,
     inherited_page: Option<ObjectId>,
     depth: usize,
+    walk: &mut StructWalk,
 ) -> Vec<StructElement> {
     if depth >= MAX_DEPTH {
+        walk.note_skipped();
+        return Vec::new();
+    }
+    if walk.exhausted() {
         return Vec::new();
     }
 
@@ -498,22 +638,59 @@ fn parse_kids(
     // /Pg on this element (inherited by children)
     let page_id = get_page_ref(doc, dict).or(inherited_page);
 
+    let mut children = Vec::new();
     match k_obj {
         Object::Array(arr) => {
-            let mut children = Vec::new();
             for item in arr {
-                let resolved = resolve_obj(doc, item);
-                parse_kid(doc, resolved, role_map, page_id, depth, &mut children);
+                if walk.exhausted() || !walk.spend_work() {
+                    break;
+                }
+                process_kid_item(doc, item, role_map, page_id, depth, &mut children, walk);
             }
-            children
         }
         other => {
-            let resolved = resolve_obj(doc, other);
-            let mut children = Vec::new();
-            parse_kid(doc, resolved, role_map, page_id, depth, &mut children);
-            children
+            process_kid_item(doc, other, role_map, page_id, depth, &mut children, walk);
         }
     }
+    children
+}
+
+/// Resolve one `/K` array item (following at most one level of indirection),
+/// guarding against reference cycles and the global node budget, then dispatch
+/// it via [`parse_kid`].
+fn process_kid_item(
+    doc: &Document,
+    item: &Object,
+    role_map: &HashMap<String, String>,
+    inherited_page: Option<ObjectId>,
+    depth: usize,
+    out: &mut Vec<StructElement>,
+    walk: &mut StructWalk,
+) {
+    if walk.exhausted() {
+        return;
+    }
+    if depth >= MAX_DEPTH {
+        walk.note_skipped();
+        return;
+    }
+    // If this child is an indirect reference, track its id on the active path so
+    // a self/ancestor reference is not expanded into an exponential subtree.
+    let ref_id = match item {
+        Object::Reference(id) => Some(*id),
+        _ => None,
+    };
+    if let Some(id) = ref_id {
+        if !walk.active.insert(id) {
+            walk.note_skipped();
+            return; // cycle: this object is already on the current path
+        }
+    }
+    let resolved = resolve_obj(doc, item);
+    parse_kid(doc, resolved, role_map, inherited_page, depth, out, walk);
+    if let Some(id) = ref_id {
+        walk.active.remove(&id);
+    }
 }
 
 /// Parse a single child (either a struct element dict or an MCID integer).
@@ -524,10 +701,16 @@ fn parse_kid(
     inherited_page: Option<ObjectId>,
     depth: usize,
     out: &mut Vec<StructElement>,
+    walk: &mut StructWalk,
 ) {
     match obj {
         // Direct MCID integer — create a leaf wrapper
         Object::Integer(mcid) => {
+            // A wrapper node plus its content reference — two items — reserved
+            // atomically so we never consume one unit without emitting both.
+            if !walk.charge_n(2) {
+                return;
+            }
             // This is a bare MCID at the struct-element level.
             // We attach it to the parent element, so we create a wrapper struct element.
             // Actually, bare MCIDs inside /K are content refs for the parent,
@@ -546,11 +729,11 @@ fn parse_kid(
             });
         }
         Object::Dictionary(d) => {
-            parse_struct_element_dict(doc, d, role_map, inherited_page, depth, out);
+            parse_struct_element_dict(doc, d, role_map, inherited_page, depth, out, walk);
         }
         Object::Stream(s) => {
             // Some PDFs wrap struct elements in streams (rare)
-            parse_struct_element_dict(doc, &s.dict, role_map, inherited_page, depth, out);
+            parse_struct_element_dict(doc, &s.dict, role_map, inherited_page, depth, out, walk);
         }
         _ => {}
     }
@@ -565,13 +748,23 @@ fn parse_struct_element_dict(
     inherited_page: Option<ObjectId>,
     depth: usize,
     out: &mut Vec<StructElement>,
+    walk: &mut StructWalk,
 ) {
     if depth >= MAX_DEPTH {
+        walk.note_skipped();
         return;
     }
-    // Check if this is a marked-content reference dict (has /Type /MCR)
+
+    // A marked-content reference dict materializes a wrapper node + one content
+    // reference (two items). Reserve both atomically *before* the node charge so
+    // we never consume a unit without emitting the reference — which would also
+    // deny that unit to a later element that would have fit. This matches the
+    // bare-MCID path.
     if is_mcr_dict(dict) {
         if let Ok(Object::Integer(mcid)) = dict.get(b"MCID") {
+            if !walk.charge_n(2) {
+                return;
+            }
             let page_id = get_page_ref(doc, dict).or(inherited_page);
             out.push(StructElement {
                 role: StructRole::Span,
@@ -588,12 +781,15 @@ fn parse_struct_element_dict(
         return;
     }
 
-    // Check if this is an object reference dict (has /Type /OBJR) — skip these
+    // Skip object-reference dicts (`/Type /OBJR`) — they materialize no node, so
+    // recognize and return *before* charging the budget (otherwise a document
+    // full of OBJRs would drain the shared budget and truncate real content).
     if is_objr_dict(dict) {
         return;
     }
 
-    // It's a struct element — parse its /S (structure type)
+    // It's a struct element — parse its /S (structure type). A dict without a
+    // valid /S also materializes nothing, so validate before charging.
     let role_name = match dict.get(b"S") {
         Ok(s_obj) => {
             let resolved = resolve_obj(doc, s_obj);
@@ -605,6 +801,12 @@ fn parse_struct_element_dict(
         Err(_) => return,
     };
 
+    // Charge the node only now that we know it will materialize (bounds
+    // aliased/DAG-shaped `/K` graphs the per-path cycle guard alone cannot stop).
+    if !walk.charge() {
+        return;
+    }
+
     let role = StructRole::from_name_with_role_map(&role_name, role_map);
     let page_id = get_page_ref(doc, dict).or(inherited_page);
 
@@ -621,51 +823,73 @@ fn parse_struct_element_dict(
         let k_resolved = resolve_obj(doc, k_obj);
         match k_resolved {
             Object::Integer(mcid) => {
-                content_refs.push(MarkedContentRef {
-                    mcid: *mcid,
-                    page_id,
-                });
+                if walk.charge() {
+                    content_refs.push(MarkedContentRef {
+                        mcid: *mcid,
+                        page_id,
+                    });
+                }
             }
             Object::Array(arr) => {
                 for item in arr {
+                    if walk.exhausted() || !walk.spend_work() {
+                        break;
+                    }
+                    // Only content-ref items (bare MCIDs / MCR dicts) are charged
+                    // here — those are the unbounded allocations. Structural
+                    // children are charged once at their own node entry in the
+                    // recursive call, so charging them here too would double-count
+                    // and drain the budget ~2× faster than the per-node semantics.
+                    let ref_id = match item {
+                        Object::Reference(id) => Some(*id),
+                        _ => None,
+                    };
                     let resolved = resolve_obj(doc, item);
                     match resolved {
                         Object::Integer(mcid) => {
-                            content_refs.push(MarkedContentRef {
-                                mcid: *mcid,
-                                page_id,
-                            });
+                            if walk.charge() {
+                                content_refs.push(MarkedContentRef {
+                                    mcid: *mcid,
+                                    page_id,
+                                });
+                            }
                         }
                         Object::Dictionary(d) => {
                             if is_mcr_dict(d) {
                                 if let Ok(Object::Integer(mcid)) = d.get(b"MCID") {
-                                    let pg = get_page_ref(doc, d).or(page_id);
-                                    content_refs.push(MarkedContentRef {
-                                        mcid: *mcid,
-                                        page_id: pg,
-                                    });
+                                    if walk.charge() {
+                                        let pg = get_page_ref(doc, d).or(page_id);
+                                        content_refs.push(MarkedContentRef {
+                                            mcid: *mcid,
+                                            page_id: pg,
+                                        });
+                                    }
                                 }
                             } else if is_objr_dict(d) {
                                 // Skip object references
                             } else {
-                                parse_struct_element_dict(
+                                recurse_struct_child(
                                     doc,
+                                    ref_id,
                                     d,
                                     role_map,
                                     page_id,
-                                    depth + 1,
+                                    depth,
                                     &mut children,
+                                    walk,
                                 );
                             }
                         }
                         Object::Stream(s) => {
-                            parse_struct_element_dict(
+                            recurse_struct_child(
                                 doc,
+                                ref_id,
                                 &s.dict,
                                 role_map,
                                 page_id,
-                                depth + 1,
+                                depth,
                                 &mut children,
+                                walk,
                             );
                         }
                         _ => {}
@@ -675,14 +899,29 @@ fn parse_struct_element_dict(
             Object::Dictionary(d) => {
                 if is_mcr_dict(d) {
                     if let Ok(Object::Integer(mcid)) = d.get(b"MCID") {
-                        let pg = get_page_ref(doc, d).or(page_id);
-                        content_refs.push(MarkedContentRef {
-                            mcid: *mcid,
-                            page_id: pg,
-                        });
+                        if walk.charge() {
+                            let pg = get_page_ref(doc, d).or(page_id);
+                            content_refs.push(MarkedContentRef {
+                                mcid: *mcid,
+                                page_id: pg,
+                            });
+                        }
                     }
                 } else {
-                    parse_struct_element_dict(doc, d, role_map, page_id, depth + 1, &mut children);
+                    let ref_id = match k_obj {
+                        Object::Reference(id) => Some(*id),
+                        _ => None,
+                    };
+                    recurse_struct_child(
+                        doc,
+                        ref_id,
+                        d,
+                        role_map,
+                        page_id,
+                        depth,
+                        &mut children,
+                        walk,
+                    );
                 }
             }
             _ => {}
@@ -699,6 +938,37 @@ fn parse_struct_element_dict(
     });
 }
 
+/// Recurse into a child struct-element dictionary, guarding against reference
+/// cycles (via the active-path object-id set) and the global node budget.
+///
+/// `ref_id` is the object id of the child when it was reached through an
+/// indirect reference (`None` for an inline dictionary, which cannot alias).
+#[allow(clippy::too_many_arguments)]
+fn recurse_struct_child(
+    doc: &Document,
+    ref_id: Option<ObjectId>,
+    dict: &lopdf::Dictionary,
+    role_map: &HashMap<String, String>,
+    inherited_page: Option<ObjectId>,
+    depth: usize,
+    out: &mut Vec<StructElement>,
+    walk: &mut StructWalk,
+) {
+    if walk.exhausted() {
+        return;
+    }
+    if let Some(id) = ref_id {
+        if !walk.active.insert(id) {
+            walk.note_skipped();
+            return; // cycle: this object is already on the current path
+        }
+    }
+    parse_struct_element_dict(doc, dict, role_map, inherited_page, depth + 1, out, walk);
+    if let Some(id) = ref_id {
+        walk.active.remove(&id);
+    }
+}
+
 /// Check if dict has `/Type /MCR`.
 fn is_mcr_dict(dict: &lopdf::Dictionary) -> bool {
     dict.get(b"Type")
@@ -901,6 +1171,7 @@ fn contains_bytes(haystack: &[u8], needle: &[u8]) -> bool {
 #[cfg(test)]
 mod tests {
     use super::*;
+    use lopdf::dictionary;
 
     #[test]
     fn non_heading_content_roles() {
@@ -1231,4 +1502,395 @@ mod tests {
         let role_map = tree.mcid_to_roles(&page_ids);
         assert!(!role_map.is_empty(), "Should have MCID→role mappings");
     }
+
+    fn count_nodes(elems: &[StructElement]) -> usize {
+        elems.iter().map(|e| 1 + count_nodes(&e.children)).sum()
+    }
+
+    /// Wrap already-created struct elements under a `/StructTreeRoot` and
+    /// `/Catalog`, returning a document ready for [`StructTree::from_doc`].
+    /// `root_kid` is the top-level element the root's `/K` points at.
+    fn finalize_tagged_doc(mut doc: Document, root_kid: ObjectId) -> Document {
+        let root_id = doc.add_object(dictionary! {
+            "Type" => "StructTreeRoot",
+            "K" => vec![Object::Reference(root_kid)],
+        });
+        let catalog_id = doc.add_object(dictionary! {
+            "Type" => "Catalog",
+            "StructTreeRoot" => Object::Reference(root_id),
+        });
+        doc.trailer.set("Root", Object::Reference(catalog_id));
+        doc
+    }
+
+    #[test]
+    fn struct_tree_self_alias_kids_terminates() {
+        // A struct element that lists itself twice in `/K` (`/K [n 0 R n 0 R]`)
+        // must not expand into an exponential tree.
+        let mut doc = Document::new();
+        let elem = doc.new_object_id();
+        doc.set_object(
+            elem,
+            dictionary! {
+                "Type" => "StructElem",
+                "S" => "Div",
+                "K" => vec![Object::Reference(elem), Object::Reference(elem)],
+            },
+        );
+        let doc = finalize_tagged_doc(doc, elem);
+
+        let tree = StructTree::from_doc(&doc).expect("tree should parse");
+        let n = count_nodes(&tree.children);
+        assert!(
+            n < 10,
+            "self-alias must not explode; materialized {n} nodes"
+        );
+    }
+
+    #[test]
+    fn struct_tree_mutual_alias_kids_terminates() {
+        // A → B → A cycle via `/K` must terminate.
+        let mut doc = Document::new();
+        let a = doc.new_object_id();
+        let b = doc.new_object_id();
+        doc.set_object(
+            a,
+            dictionary! {
+                "Type" => "StructElem",
+                "S" => "Div",
+                "K" => vec![Object::Reference(b), Object::Reference(b)],
+            },
+        );
+        doc.set_object(
+            b,
+            dictionary! {
+                "Type" => "StructElem",
+                "S" => "Div",
+                "K" => vec![Object::Reference(a), Object::Reference(a)],
+            },
+        );
+        let doc = finalize_tagged_doc(doc, a);
+
+        let tree = StructTree::from_doc(&doc).expect("tree should parse");
+        let n = count_nodes(&tree.children);
+        assert!(
+            n < 100,
+            "mutual alias must terminate small; materialized {n} nodes"
+        );
+    }
+
+    #[test]
+    fn struct_tree_aliased_dag_respects_node_budget() {
+        // Distinct elements, each aliased twice in the next level's `/K`, form a
+        // DAG that would expand to 2^depth nodes (the per-path cycle guard does
+        // not catch this since every id is on the path only once). The global
+        // node budget must cap total materialization.
+        let mut doc = Document::new();
+        let levels = 22; // 2^22 ≈ 4.2M unbounded, well past the budget
+        let ids: Vec<ObjectId> = (0..=levels).map(|_| doc.new_object_id()).collect();
+        for i in 0..levels {
+            doc.set_object(
+                ids[i],
+                dictionary! {
+                    "Type" => "StructElem",
+                    "S" => "Div",
+                    "K" => vec![Object::Reference(ids[i + 1]), Object::Reference(ids[i + 1])],
+                },
+            );
+        }
+        doc.set_object(
+            ids[levels],
+            dictionary! { "Type" => "StructElem", "S" => "P" },
+        );
+        let root_id = doc.add_object(dictionary! {
+            "Type" => "StructTreeRoot",
+            "K" => vec![Object::Reference(ids[0])],
+        });
+        let catalog_id = doc.add_object(dictionary! {
+            "Type" => "Catalog",
+            "StructTreeRoot" => Object::Reference(root_id),
+        });
+        doc.trailer.set("Root", Object::Reference(catalog_id));
+
+        let tree = StructTree::from_doc(&doc).expect("tree should parse");
+        let n = count_nodes(&tree.children);
+        assert!(
+            n <= MAX_STRUCT_NODES,
+            "node count {n} exceeded budget {MAX_STRUCT_NODES}"
+        );
+    }
+
+    #[test]
+    fn struct_tree_wide_mcid_array_respects_budget() {
+        // A single struct element with a `/K` array of bare MCIDs wider than the
+        // budget must not allocate `content_refs` without bound — each array item
+        // is charged, so materialized marked-content refs stay within the budget.
+        let mut doc = Document::new();
+        let elem = doc.new_object_id();
+        let kids: Vec<Object> = (0..(MAX_STRUCT_NODES as i64 + 100))
+            .map(Object::Integer)
+            .collect();
+        doc.set_object(
+            elem,
+            dictionary! {
+                "Type" => "StructElem",
+                "S" => "P",
+                "K" => kids,
+            },
+        );
+        let doc = finalize_tagged_doc(doc, elem);
+
+        let tree = StructTree::from_doc(&doc).expect("tree should parse");
+        assert!(
+            tree.mcid_count() <= MAX_STRUCT_NODES,
+            "content_refs unbounded: {} > {MAX_STRUCT_NODES}",
+            tree.mcid_count()
+        );
+    }
+
+    #[test]
+    fn budget_charge_flags_truncation_once_exhausted() {
+        let mut walk = StructWalk::new();
+        walk.budget = 1;
+        assert!(walk.charge(), "should spend the last unit");
+        assert!(!walk.truncated, "not truncated while budget remained");
+        assert!(!walk.charge(), "budget exhausted");
+        assert!(walk.truncated, "exhaustion must set the truncation flag");
+        // Stays exhausted/flagged on subsequent calls.
+        assert!(!walk.charge());
+        assert!(walk.truncated);
+    }
+
+    #[test]
+    fn exhausted_flags_truncation_after_budget_spent_by_charge() {
+        // The dominant truncation path: the budget is driven to 0 by a
+        // successful `charge()` (which does not set the flag), and remaining
+        // items are then dropped by an `exhausted()` guard — which must flag it.
+        let mut walk = StructWalk::new();
+        walk.budget = 1;
+        assert!(walk.charge());
+        assert!(
+            !walk.truncated,
+            "spending the last unit is not truncation yet"
+        );
+        assert!(walk.exhausted(), "budget is now spent");
+        assert!(
+            walk.truncated,
+            "the guard that skips work must flag truncation"
+        );
+    }
+
+    #[test]
+    fn wide_kids_array_flags_truncation_via_parser() {
+        // Reproduce the reviewer's scenario through the real parser: a `/K`
+        // array wider than the budget drives the budget to 0 via `charge()`,
+        // then the loop guard drops the rest — the truncation flag must be set
+        // (so `from_doc` logs it) rather than staying silently false.
+        let mut doc = Document::new();
+        let elem = doc.new_object_id();
+        let kids: Vec<Object> = (0..20i64).map(Object::Integer).collect();
+        doc.set_object(
+            elem,
+            dictionary! { "Type" => "StructElem", "S" => "P", "K" => kids },
+        );
+        let dict = doc.get_dictionary(elem).unwrap().clone();
+
+        let mut walk = StructWalk::new();
+        walk.budget = 5; // smaller than the 20-item `/K` array
+        let role_map = HashMap::new();
+        let mut out = Vec::new();
+        parse_struct_element_dict(&doc, &dict, &role_map, None, 0, &mut out, &mut walk);
+        assert!(
+            walk.truncated,
+            "a `/K` array wider than the budget must flag truncation"
+        );
+    }
+
+    #[test]
+    fn cycle_skip_flags_truncation() {
+        // A `/K` reference cycle is dropped rather than expanded; that skip must
+        // still flag truncation so the one-shot warning fires for malformed
+        // trees, not only for budget exhaustion.
+        let mut doc = Document::new();
+        let elem = doc.new_object_id();
+        doc.set_object(
+            elem,
+            dictionary! {
+                "Type" => "StructElem",
+                "S" => "Div",
+                "K" => vec![Object::Reference(elem), Object::Reference(elem)],
+            },
+        );
+        let dict = doc.get_dictionary(elem).unwrap().clone();
+
+        let mut walk = StructWalk::new();
+        walk.active.insert(elem); // simulate `elem` already on the DFS path
+        let role_map = HashMap::new();
+        let mut out = Vec::new();
+        parse_struct_element_dict(&doc, &dict, &role_map, None, 0, &mut out, &mut walk);
+        assert!(
+            walk.truncated,
+            "a cycle-skipped `/K` child must flag truncation"
+        );
+    }
+
+    #[test]
+    fn bare_mcid_charges_node_and_reference() {
+        // A bare MCID `/K` child becomes a wrapper node carrying one content
+        // reference — two materialized items — so it must charge two budget
+        // units, not one.
+        let doc = Document::new();
+        let obj = Object::Integer(7);
+        let role_map = HashMap::new();
+        let mut out = Vec::new();
+        let mut walk = StructWalk::new();
+        let before = walk.budget;
+        parse_kid(&doc, &obj, &role_map, None, 0, &mut out, &mut walk);
+        assert_eq!(
+            out.len(),
+            1,
+            "bare MCID should materialize one wrapper node"
+        );
+        assert_eq!(
+            before - walk.budget,
+            2,
+            "bare MCID must charge for both the node and its content reference"
+        );
+    }
+
+    #[test]
+    fn mcr_dict_charges_node_and_reference() {
+        // A top-level MCR `/K` dict materializes the same wrapper node + content
+        // reference as a bare MCID, so it must charge the same two budget units
+        // (not one), keeping the per-item budgeting uniform.
+        let doc = Document::new();
+        let obj = Object::Dictionary(dictionary! { "Type" => "MCR", "MCID" => 3 });
+        let role_map = HashMap::new();
+        let mut out = Vec::new();
+        let mut walk = StructWalk::new();
+        let before = walk.budget;
+        parse_kid(&doc, &obj, &role_map, None, 0, &mut out, &mut walk);
+        assert_eq!(out.len(), 1, "MCR dict should materialize one wrapper node");
+        assert_eq!(
+            before - walk.budget,
+            2,
+            "MCR dict must charge for both the node and its content reference"
+        );
+    }
+
+    #[test]
+    fn leaf_wrappers_reserve_both_units_atomically() {
+        // With only one unit left, a two-item leaf wrapper (bare MCID or MCR
+        // dict) must consume nothing and flag truncation, leaving the unit for a
+        // later single-item element instead of half-charging.
+        let doc = Document::new();
+        let role_map = HashMap::new();
+
+        // Bare MCID via parse_kid.
+        let mut walk = StructWalk::new();
+        walk.budget = 1;
+        let mut out = Vec::new();
+        parse_kid(
+            &doc,
+            &Object::Integer(5),
+            &role_map,
+            None,
+            0,
+            &mut out,
+            &mut walk,
+        );
+        assert!(out.is_empty(), "bare MCID must not partially materialize");
+        assert_eq!(walk.budget, 1, "the leftover unit must be preserved");
+        assert!(walk.truncated);
+
+        // MCR dict via parse_struct_element_dict.
+        let mcr = dictionary! { "Type" => "MCR", "MCID" => 1 };
+        let mut walk = StructWalk::new();
+        walk.budget = 1;
+        let mut out = Vec::new();
+        parse_struct_element_dict(&doc, &mcr, &role_map, None, 0, &mut out, &mut walk);
+        assert!(out.is_empty(), "MCR dict must not partially materialize");
+        assert_eq!(walk.budget, 1, "the leftover unit must be preserved");
+        assert!(walk.truncated);
+    }
+
+    #[test]
+    fn insufficient_reservation_stops_the_scan() {
+        // A one-unit budget is not "exhausted" for a one-unit item, but once a
+        // two-unit leaf reservation fails, the walk is stalled so enclosing `/K`
+        // loops stop instead of scanning the rest of a wide array.
+        let mut walk = StructWalk::new();
+        walk.budget = 1;
+        assert!(
+            !walk.exhausted(),
+            "one unit left must still allow a one-unit item"
+        );
+        assert!(!walk.charge_n(2), "cannot reserve two units from one");
+        assert!(
+            walk.exhausted(),
+            "an insufficient reservation must stop the loop"
+        );
+        assert!(walk.truncated);
+    }
+
+    #[test]
+    fn work_budget_bounds_examined_items() {
+        let mut walk = StructWalk::new();
+        walk.work = 2;
+        assert!(walk.spend_work());
+        assert!(walk.spend_work());
+        assert!(!walk.spend_work(), "traversal budget exhausted");
+        assert!(walk.truncated);
+    }
+
+    #[test]
+    fn wide_unsupported_kids_stop_at_work_budget() {
+        // A wide `/K` array of unsupported values (nulls) materializes nothing;
+        // it must stop at the traversal budget instead of scanning every entry.
+        let mut doc = Document::new();
+        let elem = doc.new_object_id();
+        let kids: Vec<Object> = (0..1000).map(|_| Object::Null).collect();
+        doc.set_object(
+            elem,
+            dictionary! { "Type" => "StructElem", "S" => "P", "K" => kids },
+        );
+        let dict = doc.get_dictionary(elem).unwrap().clone();
+
+        let mut walk = StructWalk::new();
+        walk.work = 10; // far smaller than the 1000-entry array
+        let role_map = HashMap::new();
+        let mut out = Vec::new();
+        parse_struct_element_dict(&doc, &dict, &role_map, None, 0, &mut out, &mut walk);
+        assert!(
+            walk.truncated,
+            "a wide unsupported `/K` array must hit the work budget"
+        );
+    }
+
+    #[test]
+    fn non_materializing_dicts_do_not_charge_node_budget() {
+        let doc = Document::new();
+        let role_map = HashMap::new();
+
+        // OBJR dict: materializes no node, so it must not spend the node budget.
+        let objr = dictionary! { "Type" => "OBJR" };
+        let mut walk = StructWalk::new();
+        let before = walk.budget;
+        let mut out = Vec::new();
+        parse_struct_element_dict(&doc, &objr, &role_map, None, 0, &mut out, &mut walk);
+        assert!(out.is_empty());
+        assert_eq!(walk.budget, before, "OBJR must not spend the node budget");
+
+        // A struct dict without a valid /S also materializes nothing.
+        let no_s = dictionary! { "Type" => "StructElem" };
+        let mut walk = StructWalk::new();
+        let before = walk.budget;
+        let mut out = Vec::new();
+        parse_struct_element_dict(&doc, &no_s, &role_map, None, 0, &mut out, &mut walk);
+        assert!(out.is_empty());
+        assert_eq!(
+            walk.budget, before,
+            "a dict without /S must not spend the node budget"
+        );
+    }
 }

src/tables/format.rs

@@ -123,6 +123,7 @@ fn format_toc_as_list(cells: &[Vec<String>], footnotes: &[String]) -> String {
 
 /// True when the cell looks like a page number.  Accepts:
 ///   - plain digit tokens: "42", "86 86"
+///   - canonical roman numerals (front-matter pages): "vii", "ix", "xii"
 ///   - dashed section-page IDs: "5-21", "A-1", "B--3", "TC-2" (common in
 ///     technical manuals)
 fn is_page_number_cell(cell: &str) -> bool {
@@ -138,6 +139,9 @@ fn is_page_number_cell(cell: &str) -> bool {
         if all_digits {
             return t.len() <= 4;
         }
+        if super::canonical_roman_value(t).is_some() {
+            return true;
+        }
         // Section-page form: uppercase letters, digits, dashes; at least
         // one digit present.
         t.chars()
@@ -184,6 +188,21 @@ fn starts_with_numbered_label(cell: &str) -> bool {
             .is_some_and(|c| matches!(c, '.' | ')' | '-' | ':'))
 }
 
+fn starts_with_hierarchical_numbered_label(cell: &str) -> bool {
+    let token = cell
+        .split_whitespace()
+        .next()
+        .unwrap_or("")
+        .trim_end_matches(['.', ')', ':', '-']);
+    let levels: Vec<&str> = token.split('.').collect();
+    (2..=4).contains(&levels.len())
+        && levels.iter().all(|level| {
+            !level.is_empty()
+                && level.len() <= 3
+                && level.chars().all(|character| character.is_ascii_digit())
+        })
+}
+
 fn alpha_word_count(cell: &str) -> usize {
     cell.split_whitespace()
         .filter(|word| word.chars().any(|c| c.is_alphabetic()))
@@ -337,11 +356,12 @@ fn clean_table_cells(cells: &[Vec<String>]) -> (Vec<Vec<String>>, Vec<String>) {
         // mid-sentence/lowercase ("continued text here", "with 3.5%...") or
         // carry lowercase fragments in the later cells, so keep those mergeable.
         let looks_like_hierarchical_subrow = first_cell.is_empty()
-            && row.len() >= 3
             && first_non_empty_col == Some(1)
             && looks_like_compact_entry_label(first_non_empty_cell)
-            && ((non_first_cells.len() >= 2 && title_like_later_cells > 0)
+            && ((row.len() == 2 && starts_with_hierarchical_numbered_label(first_non_empty_cell))
+                || (row.len() >= 3 && non_first_cells.len() >= 2 && title_like_later_cells > 0)
                 || (non_first_cells.len() == 1
+                    && row.len() >= 3
                     && prev_first_cell_empty
                     && alpha_word_count(first_non_empty_cell) >= 2));
         let looks_like_new_first_column_entry = !first_cell.is_empty()
@@ -684,6 +704,46 @@ mod tests {
         assert_eq!(cleaned[4][1], "Model training");
     }
 
+    #[test]
+    fn test_clean_table_cells_two_column_numbered_subrows_not_merged() {
+        let cells = vec![
+            vec!["Area".into(), "Competence".into()],
+            vec![
+                "1. Embodying sustainability values".into(),
+                "1.1 Valuing sustainability".into(),
+            ],
+            vec!["".into(), "1.2 Supporting fairness".into()],
+            vec!["".into(), "1.3 Promoting nature".into()],
+            vec![
+                "2. Embracing complexity".into(),
+                "2.1 Systems thinking".into(),
+            ],
+            vec!["".into(), "2.2 Critical thinking".into()],
+        ];
+        let (cleaned, _) = clean_table_cells(&cells);
+
+        assert_eq!(cleaned.len(), 6);
+        assert_eq!(cleaned[2], vec!["", "1.2 Supporting fairness"]);
+        assert_eq!(cleaned[3], vec!["", "1.3 Promoting nature"]);
+        assert_eq!(cleaned[5], vec!["", "2.2 Critical thinking"]);
+    }
+
+    #[test]
+    fn test_clean_table_cells_two_column_numbered_continuation_merges() {
+        let cells = vec![
+            vec!["Area".into(), "Requirement".into()],
+            vec!["Safety".into(), "The program includes".into()],
+            vec!["".into(), "1. First requirement for every operator".into()],
+        ];
+        let (cleaned, _) = clean_table_cells(&cells);
+
+        assert_eq!(cleaned.len(), 2);
+        assert_eq!(
+            cleaned[1][1],
+            "The program includes 1. First requirement for every operator"
+        );
+    }
+
     #[test]
     fn test_clean_table_cells_partial_hierarchical_subrow_not_merged() {
         let cells = vec![

src/tables/grid.rs

@@ -436,7 +436,8 @@ pub(crate) fn recover_header_row(
         .iter()
         .enumerate()
         .filter(|(_, item)| {
-            item.font_size > small_font_threshold
+            !item.is_strikeout
+                && item.font_size > small_font_threshold
                 && item.y > first_row_y
                 && item.y <= first_row_y + row_gap_limit
         })
@@ -804,6 +805,31 @@ mod tests {
         assert_eq!(table.rows.len(), rows_before);
     }
 
+    #[test]
+    fn test_recover_header_row_skips_strikeout_candidates() {
+        let mut old_col1 = make_item("Old Col1", 100.0, 520.0, 12.0);
+        old_col1.is_strikeout = true;
+        let mut old_col2 = make_item("Old Col2", 200.0, 520.0, 12.0);
+        old_col2.is_strikeout = true;
+        let all_items = vec![
+            old_col1,
+            old_col2,
+            make_item("A", 100.0, 500.0, 8.0),
+            make_item("B", 200.0, 500.0, 8.0),
+        ];
+        let mut table = Table {
+            columns: vec![100.0, 200.0],
+            rows: vec![500.0, 480.0],
+            cells: vec![vec!["A".into(), "B".into()], vec!["C".into(), "D".into()]],
+            item_indices: vec![2, 3],
+            kind: TableKind::Data,
+        };
+
+        recover_header_row(&mut table, &all_items, 9.0);
+        assert_eq!(table.rows.len(), 2);
+        assert_eq!(table.cells[0], vec!["A", "B"]);
+    }
+
     #[test]
     fn test_recover_header_row_too_far_above() {
         let all_items = vec![

src/tables/mod.rs

@@ -4,7 +4,7 @@
 
 mod detect_heuristic;
 mod detect_lines;
-mod detect_rects;
+pub(crate) mod detect_rects;
 mod detect_struct;
 mod financial;
 mod format;
@@ -12,10 +12,15 @@ mod grid;
 pub mod structured;
 
 pub use detect_heuristic::detect_tables;
-pub(crate) use detect_heuristic::is_table_of_contents;
+pub(crate) use detect_heuristic::{
+    content_width, detect_tables_with_page_width, is_table_of_contents,
+};
 pub use detect_lines::detect_tables_from_lines;
+pub(crate) use detect_lines::{
+    detect_dense_line_chart_regions, detect_vector_grid_tables_from_lines,
+};
 pub(crate) use detect_rects::cluster_rects;
-pub use detect_rects::{detect_tables_from_rects, RectHintRegion};
+pub use detect_rects::{detect_chart_regions, detect_tables_from_rects, RectHintRegion};
 pub use detect_struct::detect_tables_from_struct_tree;
 pub use format::table_to_markdown;
 pub use structured::{cells_to_markdown, StructuredCell};
@@ -177,6 +182,60 @@ pub(crate) fn try_build_rect_guided_table(
     ))
 }
 
+/// Canonical lowercase roman numeral for `n` (the i/v/x/l/c range).
+pub(super) fn to_roman_lower(mut n: u32) -> String {
+    const TABLE: [(u32, &str); 9] = [
+        (100, "c"),
+        (90, "xc"),
+        (50, "l"),
+        (40, "xl"),
+        (10, "x"),
+        (9, "ix"),
+        (5, "v"),
+        (4, "iv"),
+        (1, "i"),
+    ];
+    let mut out = String::new();
+    for (val, sym) in TABLE {
+        while n >= val {
+            out.push_str(sym);
+            n -= val;
+        }
+    }
+    out
+}
+
+/// Parse a *canonical* roman numeral (i/v/x/l/c range, ≤8 chars) to its value.
+/// Returns `None` for non-canonical strings, so ordinary words made of those
+/// letters — "civil", "mix", "ill" — are not mistaken for numbers. Shared by
+/// the TOC detector and the TOC formatter so the two stay in sync.
+pub(super) fn canonical_roman_value(token: &str) -> Option<u32> {
+    let lower = token.trim().to_ascii_lowercase();
+    if lower.is_empty() || lower.len() > 8 || !lower.chars().all(|c| "ivxlc".contains(c)) {
+        return None;
+    }
+    let mut total = 0i32;
+    let mut prev = 0i32;
+    for c in lower.chars().rev() {
+        let v = match c {
+            'i' => 1,
+            'v' => 5,
+            'x' => 10,
+            'l' => 50,
+            'c' => 100,
+            _ => return None,
+        };
+        if v < prev {
+            total -= v;
+        } else {
+            total += v;
+            prev = v;
+        }
+    }
+    let value = u32::try_from(total).ok().filter(|&n| n > 0)?;
+    (to_roman_lower(value) == lower).then_some(value)
+}
+
 /// Split a TextItem whose text contains multiple whitespace-separated tokens
 /// (like "10 11 12 ... 31") into individual TextItems, each assigned to the
 /// nearest column boundary.

src/text_utils.rs

@@ -7,6 +7,83 @@
 use crate::types::TextItem;
 use unicode_normalization::UnicodeNormalization;
 
+/// Return whether text is an explicit page-number expression.
+///
+/// This strict form is suitable before layout, where removing one numeric item
+/// from substantive text such as `Page 42 explains the result` would lose data.
+pub(crate) fn is_explicit_page_number_expression(text: &str) -> bool {
+    let trimmed = text.trim();
+    if trimmed.is_empty() {
+        return false;
+    }
+
+    let is_number = |value: &str| {
+        !value.is_empty() && value.chars().all(|character| character.is_ascii_digit())
+    };
+
+    if trimmed.len() <= 4 && is_number(trimmed) {
+        return true;
+    }
+
+    if trimmed.len() >= 3 && trimmed.starts_with('-') && trimmed.ends_with('-') {
+        let inner = trimmed[1..trimmed.len() - 1].trim();
+        if is_number(inner) {
+            return true;
+        }
+    }
+
+    let lowercase = trimmed.to_ascii_lowercase();
+    if let Some(rest) = lowercase.strip_prefix("page") {
+        let words: Vec<&str> = rest.split_whitespace().collect();
+        if words.len() >= 3 && is_number(words[0]) && words[1] == "of" && is_number(words[2]) {
+            return true;
+        }
+        if words.len() >= 2 && words[0] == "of" && is_number(words[1]) {
+            return true;
+        }
+        return match words.as_slice() {
+            [] | ["of"] => true,
+            [number] => is_number(number),
+            ["of", total] => is_number(total),
+            [number, "of", total] => is_number(number) && is_number(total),
+            _ => false,
+        };
+    }
+
+    let words: Vec<&str> = lowercase.split_whitespace().collect();
+    match words.as_slice() {
+        [number, "of", total] => is_number(number) && is_number(total),
+        _ => false,
+    }
+}
+
+/// Return whether a completed Markdown line looks like a page number or a
+/// labeled running header.
+///
+/// At this stage the complete line and surrounding breaks are available, so a
+/// leading `Page N` remains compatible with the existing header cleanup even
+/// when the PDF appends a chapter or document title.
+pub(crate) fn is_page_number_line(text: &str) -> bool {
+    if is_explicit_page_number_expression(text) {
+        return true;
+    }
+
+    let lowercase = text.trim().to_ascii_lowercase();
+    lowercase.strip_prefix("page").is_some_and(|rest| {
+        let mut characters = rest.trim_start().chars().peekable();
+        let mut has_page_number = false;
+        while characters
+            .peek()
+            .is_some_and(|character| character.is_ascii_digit())
+        {
+            has_page_number = true;
+            characters.next();
+        }
+
+        has_page_number && characters.next().is_none_or(char::is_whitespace)
+    })
+}
+
 /// Check if a character is CJK (Chinese, Japanese, Korean).
 /// CJK languages don't use spaces between words, so word-boundary
 /// heuristics should not apply when CJK characters are involved.

src/tounicode.rs

@@ -4,11 +4,17 @@
 
 use log::{debug, warn};
 use lopdf::{Document, Object, ObjectId};
+use std::borrow::Cow;
 use std::collections::{HashMap, HashSet};
+#[cfg(not(target_arch = "wasm32"))]
 use std::path::{Path, PathBuf};
 
 use crate::glyph_names::glyph_to_char;
 
+#[cfg(target_arch = "wasm32")]
+static BUILTIN_CMAPS: include_dir::Dir<'_> =
+    include_dir::include_dir!("$CARGO_MANIFEST_DIR/external/bcmaps");
+
 /// A parsed ToUnicode CMap mapping CIDs to Unicode strings
 #[derive(Debug, Default, Clone)]
 pub struct ToUnicodeCMap {
@@ -874,6 +880,25 @@ fn try_remap_subset_cmap(
         None => return (cmap, None),
     };
 
+    // Both repair paths below assume CIDs are glyph indices that a subsetter can
+    // renumber, which is only true for CIDFontType2 (TrueType). For CIDFontType0
+    // (CFF), CIDs are resolved through the CFF charset, so a valid CMap stays valid
+    // after subsetting and renumbering it corrupts otherwise-correct text.
+    // CIDToGIDMap is likewise CIDFontType2-only (PDF 32000-1:2008, 9.7.4.2), so this
+    // also ignores a CIDToGIDMap that a malformed producer attached to a CFF font.
+    // /Subtype may be an indirect reference, so resolve it through the document.
+    // Only bail out when the descendant is *explicitly* something other than
+    // CIDFontType2: a missing or unresolvable /Subtype keeps the previous
+    // behaviour rather than silently disabling the repair.
+    let subtype = cid_font_dict.get(b"Subtype").ok().and_then(|o| match o {
+        Object::Reference(r) => doc.get_object(*r).ok().and_then(|o| o.as_name().ok()),
+        other => other.as_name().ok(),
+    });
+    if subtype.is_some_and(|name| name != b"CIDFontType2") {
+        debug!("Subset remap skipped for obj={obj_num}: descendant is not CIDFontType2");
+        return (cmap, None);
+    }
+
     // If there's an explicit CIDToGIDMap, build a repaired CMap using it.
     if let Some(cid_to_gid) = get_cid_to_gid_map(cid_font_dict, doc) {
         if let Some(repaired) = build_cmap_with_cid_to_gid_map(&cmap, &cid_to_gid) {
@@ -1149,9 +1174,7 @@ fn build_gid_to_unicode(face: &ttf_parser::Face<'_>) -> Option<HashMap<u16, char
 /// Build a ToUnicodeCMap from pdf.js built-in binary CMaps (bcmaps).
 fn build_cmap_from_builtin_cmap(ordering: &str) -> Option<ToUnicodeCMap> {
     let name = format!("Adobe-{}-UCS2.bcmap", ordering);
-    let dir = find_bcmaps_dir()?;
-    let path = dir.join(name);
-    let data = std::fs::read(&path).ok()?;
+    let data = read_builtin_cmap_file(&name)?;
     let mut cmap = parse_binary_cmap(&data).ok()?;
     if cmap.char_map.is_empty() && cmap.ranges.is_empty() {
         return None;
@@ -1159,13 +1182,14 @@ fn build_cmap_from_builtin_cmap(ordering: &str) -> Option<ToUnicodeCMap> {
     cmap.code_byte_length = 2;
     debug!(
         "Built-in CMap {}: char_map={} ranges={}",
-        path.display(),
+        name,
         cmap.char_map.len(),
         cmap.ranges.len()
     );
     Some(cmap)
 }
 
+#[cfg(not(target_arch = "wasm32"))]
 fn find_bcmaps_dir() -> Option<PathBuf> {
     if let Ok(dir) = std::env::var("PDF_INSPECTOR_BCMAPS_DIR") {
         let p = PathBuf::from(dir);
@@ -1182,6 +1206,18 @@ fn find_bcmaps_dir() -> Option<PathBuf> {
     None
 }
 
+#[cfg(not(target_arch = "wasm32"))]
+fn read_builtin_cmap_file(name: &str) -> Option<Cow<'static, [u8]>> {
+    let path = find_bcmaps_dir()?.join(name);
+    std::fs::read(path).ok().map(Cow::Owned)
+}
+
+#[cfg(target_arch = "wasm32")]
+fn read_builtin_cmap_file(name: &str) -> Option<Cow<'static, [u8]>> {
+    let file = BUILTIN_CMAPS.get_file(name)?;
+    Some(Cow::Borrowed(file.contents()))
+}
+
 fn parse_binary_cmap(data: &[u8]) -> Result<ToUnicodeCMap, String> {
     let mut stream = BinaryCMapStream::new(data);
     let _header = stream.read_byte().ok_or("unexpected EOF in bcmap header")?;
@@ -1492,9 +1528,7 @@ fn parse_encoding_cmap_object(obj: &Object, doc: &Document) -> Option<EncodingCM
 }
 
 fn load_builtin_encoding_cmap(name: &str) -> Option<EncodingCMap> {
-    let dir = find_bcmaps_dir()?;
-    let path = dir.join(format!("{}.bcmap", name));
-    let data = std::fs::read(&path).ok()?;
+    let data = read_builtin_cmap_file(&format!("{}.bcmap", name))?;
     parse_binary_cmap_encoding(&data).ok()
 }
 
@@ -1779,9 +1813,7 @@ fn load_builtin_cmap_by_name(name: &str) -> Option<ToUnicodeCMap> {
     if !name.ends_with("UCS2") {
         return None;
     }
-    let dir = find_bcmaps_dir()?;
-    let path = dir.join(format!("{}.bcmap", name));
-    let data = std::fs::read(&path).ok()?;
+    let data = read_builtin_cmap_file(&format!("{}.bcmap", name))?;
     let mut cmap = parse_binary_cmap(&data).ok()?;
     if cmap.char_map.is_empty() && cmap.ranges.is_empty() {
         return None;
@@ -3146,4 +3178,106 @@ endbfrange
             "Remap must fire when CMap's CIDs are outside W array coverage"
         );
     }
+
+    #[test]
+    fn test_try_remap_skipped_for_cid_font_type0() {
+        // Same W/CMap mismatch as the CIDFontType2 case above, but the descendant is
+        // CIDFontType0 (CFF). There CIDs are resolved through the CFF charset, so the
+        // ToUnicode CIDs stay valid after subsetting and must not be renumbered.
+        // Real-world case: Japanese Adobe-Japan1 PDFs (e.g. National Diet Library
+        // minutes) where remapping turned correct text into unrelated glyphs.
+        let cmap_content = r#"
+1 begincodespacerange
+<0000><FFFF>
+endcodespacerange
+1 beginbfrange
+<0200> <0220> <0410>
+endbfrange
+"#;
+        let cmap = ToUnicodeCMap::parse(cmap_content.as_bytes()).unwrap();
+
+        let mut doc = Document::new();
+
+        // CIDToGIDMap is CIDFontType2-only, but a malformed producer can still emit
+        // one on a CFF font. Use a real stream (not /Identity, which is treated as
+        // "no map") so this also fails if the guard is moved back below the
+        // CIDToGIDMap branch: cid 1 -> gid 0x0200, which the CMap resolves.
+        let mut cid_to_gid = vec![0u8; 68];
+        cid_to_gid[2] = 0x02;
+        cid_to_gid[3] = 0x00;
+        let cid_to_gid_id =
+            doc.add_object(lopdf::Stream::new(lopdf::Dictionary::new(), cid_to_gid));
+
+        let mut cid_font = lopdf::Dictionary::new();
+        cid_font.set("Subtype", lopdf::Object::Name(b"CIDFontType0".to_vec()));
+        cid_font.set("CIDToGIDMap", lopdf::Object::Reference(cid_to_gid_id));
+        cid_font.set(
+            "W",
+            lopdf::Object::Array(vec![
+                lopdf::Object::Integer(1),
+                lopdf::Object::Array(vec![lopdf::Object::Integer(500); 34]),
+            ]),
+        );
+        let cid_font_id = doc.add_object(cid_font);
+
+        let mut font_dict = lopdf::Dictionary::new();
+        font_dict.set("Encoding", lopdf::Object::Name(b"Identity-H".to_vec()));
+        font_dict.set(
+            "DescendantFonts",
+            lopdf::Object::Array(vec![lopdf::Object::Reference(cid_font_id)]),
+        );
+
+        let (primary, remapped) = try_remap_subset_cmap(cmap, &font_dict, &doc, 789);
+        assert!(
+            remapped.is_none(),
+            "Remap must be skipped for CIDFontType0 (CFF) descendants, including a \
+             CIDToGIDMap a malformed producer attached to one"
+        );
+        // The original CMap must still resolve its own CIDs.
+        assert_eq!(primary.lookup(0x0200), Some("\u{0410}".to_string()));
+    }
+
+    #[test]
+    fn test_try_remap_resolves_indirect_subtype() {
+        // /Subtype may be stored as an indirect reference. A genuine CIDFontType2
+        // font must still get the repair, so the guard has to dereference it rather
+        // than treat the unresolved value as "not CIDFontType2".
+        let cmap_content = r#"
+1 begincodespacerange
+<0000><FFFF>
+endcodespacerange
+1 beginbfrange
+<0200> <0220> <0410>
+endbfrange
+"#;
+        let cmap = ToUnicodeCMap::parse(cmap_content.as_bytes()).unwrap();
+
+        let mut doc = Document::new();
+        let subtype_id = doc.add_object(lopdf::Object::Name(b"CIDFontType2".to_vec()));
+
+        let mut cid_font = lopdf::Dictionary::new();
+        cid_font.set("Subtype", lopdf::Object::Reference(subtype_id));
+        cid_font.set("CIDToGIDMap", lopdf::Object::Name(b"Identity".to_vec()));
+        cid_font.set(
+            "W",
+            lopdf::Object::Array(vec![
+                lopdf::Object::Integer(0),
+                lopdf::Object::Array(vec![lopdf::Object::Integer(500); 34]),
+            ]),
+        );
+        let cid_font_id = doc.add_object(cid_font);
+
+        let mut font_dict = lopdf::Dictionary::new();
+        font_dict.set("Encoding", lopdf::Object::Name(b"Identity-H".to_vec()));
+        font_dict.set(
+            "DescendantFonts",
+            lopdf::Object::Array(vec![lopdf::Object::Reference(cid_font_id)]),
+        );
+
+        let (_primary, remapped) = try_remap_subset_cmap(cmap, &font_dict, &doc, 790);
+        assert!(
+            remapped.is_some(),
+            "An indirect /Subtype naming CIDFontType2 must still reach the remap"
+        );
+    }
 }

src/types.rs

@@ -148,14 +148,17 @@ impl TextLine {
         self.text_with_formatting(false, false, false)
     }
 
-    /// Get text with optional bold/italic/underline markdown formatting
+    /// Get text with optional bold/italic/decorative markdown formatting.
+    ///
+    /// `format_decorations` enables both geometrically detected source
+    /// decorations: underline (`<u>`) and strikeout (`<s>`).
     pub fn text_with_formatting(
         &self,
         format_bold: bool,
         format_italic: bool,
-        format_underline: bool,
+        format_decorations: bool,
     ) -> String {
-        if !format_bold && !format_italic && !format_underline {
+        if !format_bold && !format_italic && !format_decorations {
             return self.text_plain();
         }
 
@@ -165,6 +168,7 @@ impl TextLine {
         let mut current_bold = false;
         let mut current_italic = false;
         let mut current_underline = false;
+        let mut current_strikeout = false;
 
         for (i, item) in self.items.iter().enumerate() {
             let text = item.text.as_str();
@@ -190,13 +194,16 @@ impl TextLine {
             // we push text_trimmed below (which strips it).
             let has_leading_space = text.starts_with(' ');
 
-            // Check for style changes. Underline is exclusive: `<u>` content
-            // stays free of `**`/`*` markers — consumers (and the eval
-            // harnesses this feeds) match the tag content literally, and
-            // mixed `<u>**x**</u>` nesting breaks that.
-            let item_underline = format_underline && item.is_underline;
-            let item_bold = format_bold && item.is_bold && !item_underline;
-            let item_italic = format_italic && item.is_italic && !item_underline;
+            // Check for style changes. Source decorations are exclusive:
+            // `<u>`/`<s>` content stays free of `**`/`*` markers — consumers
+            // (and the eval harnesses this feeds) match tag content literally,
+            // and mixed nesting breaks that. A struck-and-underlined item is
+            // emitted as struck text because deletion is the stronger semantic
+            // distinction in redline documents.
+            let item_strikeout = format_decorations && item.is_strikeout;
+            let item_underline = format_decorations && item.is_underline && !item_strikeout;
+            let item_bold = format_bold && item.is_bold && !item_underline && !item_strikeout;
+            let item_italic = format_italic && item.is_italic && !item_underline && !item_strikeout;
 
             // Close previous styles if they change
             if current_italic && !item_italic {
@@ -211,6 +218,10 @@ impl TextLine {
                 result.push_str("</u>");
                 current_underline = false;
             }
+            if current_strikeout && !item_strikeout {
+                result.push_str("</s>");
+                current_strikeout = false;
+            }
 
             // Add space: either from spacing logic or preserved from item text
             if needs_space || (has_leading_space && !result.is_empty() && !result.ends_with(' ')) {
@@ -222,6 +233,10 @@ impl TextLine {
                 result.push_str("<u>");
                 current_underline = true;
             }
+            if item_strikeout && !current_strikeout {
+                result.push_str("<s>");
+                current_strikeout = true;
+            }
             if item_bold && !current_bold {
                 result.push_str("**");
                 current_bold = true;
@@ -244,6 +259,9 @@ impl TextLine {
         if current_underline {
             result.push_str("</u>");
         }
+        if current_strikeout {
+            result.push_str("</s>");
+        }
 
         result
     }
@@ -309,3 +327,88 @@ impl TextLine {
             || space_already_exists)
     }
 }
+
+#[cfg(test)]
+mod formatting_tests {
+    use super::{ItemType, TextItem, TextLine};
+
+    fn item(text: &str, x: f32, width: f32, strikeout: bool) -> TextItem {
+        TextItem {
+            text: text.to_string(),
+            x,
+            y: 100.0,
+            width,
+            height: 12.0,
+            font: "F1".to_string(),
+            font_size: 12.0,
+            page: 1,
+            is_bold: false,
+            is_italic: false,
+            is_underline: false,
+            is_strikeout: strikeout,
+            item_type: ItemType::Text,
+            mcid: None,
+        }
+    }
+
+    fn line(items: Vec<TextItem>) -> TextLine {
+        TextLine {
+            items,
+            y: 100.0,
+            page: 1,
+            adaptive_threshold: 0.1,
+        }
+    }
+
+    #[test]
+    fn formatting_emits_semantic_strikeout() {
+        let line = line(vec![item("deleted", 10.0, 42.0, true)]);
+
+        assert_eq!(
+            line.text_with_formatting(true, true, true),
+            "<s>deleted</s>"
+        );
+    }
+
+    #[test]
+    fn formatting_closes_strikeout_before_live_text() {
+        let line = line(vec![
+            item("keep", 10.0, 24.0, false),
+            item("remove", 40.0, 42.0, true),
+            item("keep", 88.0, 24.0, false),
+        ]);
+
+        assert_eq!(
+            line.text_with_formatting(true, true, true),
+            "keep <s>remove</s> keep"
+        );
+    }
+
+    #[test]
+    fn formatting_coalesces_adjacent_struck_items() {
+        let line = line(vec![
+            item("deleted", 10.0, 42.0, true),
+            item("words", 58.0, 30.0, true),
+        ]);
+
+        assert_eq!(
+            line.text_with_formatting(true, true, true),
+            "<s>deleted words</s>"
+        );
+    }
+
+    #[test]
+    fn strikeout_takes_precedence_over_other_styles() {
+        let mut decorated = item("deleted", 10.0, 42.0, true);
+        decorated.is_bold = true;
+        decorated.is_italic = true;
+        decorated.is_underline = true;
+        let line = line(vec![decorated]);
+
+        assert_eq!(
+            line.text_with_formatting(true, true, true),
+            "<s>deleted</s>"
+        );
+        assert_eq!(line.text(), "deleted");
+    }
+}

tests/fixtures/broken_startxref_pointer.pdf

@@ -0,0 +1,68 @@
+%PDF-1.3
+%“Œ‹ž ReportLab Generated PDF document (opensource)
+1 0 obj
+<<
+/F1 2 0 R
+>>
+endobj
+2 0 obj
+<<
+/BaseFont /Helvetica /Encoding /WinAnsiEncoding /Name /F1 /Subtype /Type1 /Type /Font
+>>
+endobj
+3 0 obj
+<<
+/Contents 7 0 R /MediaBox [ 0 0 612 792 ] /Parent 6 0 R /Resources <<
+/Font 1 0 R /ProcSet [ /PDF /Text /ImageB /ImageC /ImageI ]
+>> /Rotate 0 /Trans <<
+
+>> 
+  /Type /Page
+>>
+endobj
+4 0 obj
+<<
+/PageMode /UseNone /Pages 6 0 R /Type /Catalog
+>>
+endobj
+5 0 obj
+<<
+/Author (anonymous) /CreationDate (D:20260803112923+00'00') /Creator (anonymous) /Keywords () /ModDate (D:20260803112923+00'00') /Producer (ReportLab PDF Library - \(opensource\)) 
+  /Subject (unspecified) /Title (untitled) /Trapped /False
+>>
+endobj
+6 0 obj
+<<
+/Count 1 /Kids [ 3 0 R ] /Type /Pages
+>>
+endobj
+7 0 obj
+<<
+/Filter [ /ASCII85Decode /FlateDecode ] /Length 202
+>>
+stream
+GarW05mr9@&;9NOME,dW.,B;'jAjYq0S4Z`*D9aMA;]$5J)A/$3lESen1?F)ZJsa4$4&N%-%cs)#qW5EVhhbPiRDrAV>MC%.spto@CU"ZdipR'TtFiMR_%m*Hm$N%qL7a"ckkp9T/s[N2"Og377mP*M^akb2XQZ@'l*qT(9bVtDb5+)S&Q.#%E)<]Ao`TSk2AE'/E\fn~>endstream
+endobj
+xref
+0 8
+0000000000 65535 f 
+0000000061 00000 n 
+0000000092 00000 n 
+0000000199 00000 n 
+0000000392 00000 n 
+0000000460 00000 n 
+0000000721 00000 n 
+0000000780 00000 n 
+trailer
+<<
+/ID 
+[<6d7ea1213c5974c78613d5d2a08423b5><6d7ea1213c5974c78613d5d2a08423b5>]
+% ReportLab generated PDF document -- digest (opensource)
+
+/Info 5 0 R
+/Root 4 0 R
+/Size 8
+>>
+startxref
+9072
+%%EOF

tests/integration_tests.rs

@@ -8,12 +8,13 @@ use pdf_inspector::{
     detect_pdf_type, detect_vector_grid_in_region_mem, extract_pages_markdown,
     extract_pages_markdown_mem, extract_tables_in_regions_mem, extract_text,
     extract_text_in_regions_mem, extract_text_with_positions, extract_text_with_positions_mem,
-    process_pdf_mem, process_pdf_with_options, to_markdown, MarkdownOptions, PdfError, PdfOptions,
+    process_pdf_mem, process_pdf_mem_with_options, process_pdf_with_options, to_markdown,
+    to_markdown_from_items_with_rects_and_page_count, MarkdownOptions, PdfError, PdfOptions,
     PdfType, TextItem,
 };
 use std::collections::HashSet;
 
-fn make_minimal_text_pdf() -> Vec<u8> {
+fn make_text_pdf(content: &str, media_box: &str) -> Vec<u8> {
     let mut pdf = b"%PDF-1.4\n".to_vec();
     let mut offsets = vec![0usize];
 
@@ -40,10 +41,11 @@ fn make_minimal_text_pdf() -> Vec<u8> {
         &mut pdf,
         &mut offsets,
         3,
-        "<< /Type /Page /Parent 2 0 R /MediaBox [0 0 612 792] /Resources << /Font << /F1 5 0 R >> >> /Contents 4 0 R >>",
+        &format!(
+            "<< /Type /Page /Parent 2 0 R /MediaBox [{media_box}] /Resources << /Font << /F1 5 0 R >> >> /Contents 4 0 R >>"
+        ),
     );
 
-    let content = "BT /F1 12 Tf 100 700 Td (Hello World) Tj 0 -14 Td (Second Line) Tj 0 -14 Td (Third Line) Tj ET";
     add_object(
         &mut pdf,
         &mut offsets,
@@ -79,6 +81,186 @@ fn make_minimal_text_pdf() -> Vec<u8> {
     pdf
 }
 
+fn make_recurring_contextual_folio_pdf() -> 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");
+    }
+
+    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 5 0 R 7 0 R 9 0 R] /Count 4 >>",
+    );
+    for page_index in 0..4 {
+        let page_id = 3 + page_index * 2;
+        let content_id = page_id + 1;
+        add_object(
+            &mut pdf,
+            &mut offsets,
+            page_id,
+            &format!(
+                "<< /Type /Page /Parent 2 0 R /MediaBox [0 0 612 792] /Resources << /Font << /F1 11 0 R >> >> /Contents {content_id} 0 R >>"
+            ),
+        );
+        let page_number = page_index + 1;
+        let content = format!(
+            "BT /F1 12 Tf 1 0 0 1 25 30 Tm ({page_number}) Tj 1 0 0 1 41 30 Tm (Company report footer) Tj 1 0 0 1 72 700 Tm (Body page {page_number}) Tj ET"
+        );
+        add_object(
+            &mut pdf,
+            &mut offsets,
+            content_id,
+            &format!(
+                "<< /Length {} >>\nstream\n{}\nendstream",
+                content.len(),
+                content
+            ),
+        );
+    }
+    add_object(
+        &mut pdf,
+        &mut offsets,
+        11,
+        "<< /Type /Font /Subtype /Type1 /BaseFont /Helvetica >>",
+    );
+
+    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_malformed_unselected_page() -> 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");
+    }
+
+    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 5 0 R] /Count 2 >>",
+    );
+    add_object(
+        &mut pdf,
+        &mut offsets,
+        3,
+        "<< /Type /Page /Parent 2 0 R /MediaBox [0 0 612 792] /Resources << /Font << /F1 7 0 R >> >> /Contents 4 0 R >>",
+    );
+    let content = "BT /F1 12 Tf 1 0 0 1 25 30 Tm (1) Tj 1 0 0 1 41 30 Tm (Company report footer) Tj 1 0 0 1 72 700 Tm (Selected page text) Tj 0 -16 Td (More selected text) Tj 0 -16 Td (Still selected text) Tj ET";
+    add_object(
+        &mut pdf,
+        &mut offsets,
+        4,
+        &format!(
+            "<< /Length {} >>\nstream\n{}\nendstream",
+            content.len(),
+            content
+        ),
+    );
+    add_object(
+        &mut pdf,
+        &mut offsets,
+        5,
+        "<< /Type /Page /Parent 2 0 R /MediaBox [0 0 612 792] /Resources << /Font << /F1 7 0 R >> >> /Contents 6 0 R >>",
+    );
+    add_object(
+        &mut pdf,
+        &mut offsets,
+        6,
+        "<< /Length 3 >>\nstream\nBI \nendstream",
+    );
+    add_object(
+        &mut pdf,
+        &mut offsets,
+        7,
+        "<< /Type /Font /Subtype /Type1 /BaseFont /Helvetica >>",
+    );
+
+    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_minimal_text_pdf() -> Vec<u8> {
+    make_text_pdf(
+        "BT /F1 12 Tf 100 700 Td (Hello World) Tj 0 -14 Td (Second Line) Tj 0 -14 Td (Third Line) Tj ET",
+        "0 0 612 792",
+    )
+}
+
+fn make_digit_run_repro_pdf() -> Vec<u8> {
+    let content = r#"BT
+/F1 12 Tf
+1 0 0 1 72 780 Tm (A\)) Tj
+1 0 0 1 96 780 Tm (The) Tj
+1 0 0 1 126 780 Tm (total) Tj
+1 0 0 1 166 780 Tm (of) Tj
+1 0 0 1 186 780 Tm (730) Tj
+1 0 0 1 220 780 Tm (seats) Tj
+1 0 0 1 262 780 Tm (was) Tj
+1 0 0 1 296 780 Tm (approved.) Tj
+1 0 0 1 72 755 Tm (B\)) Tj
+1 0 0 1 96 755 Tm (let) Tj
+1 0 0 1 120 755 Tm (log) Tj
+1 0 0 1 150 755 Tm (2) Tj
+1 0 0 1 164 755 Tm (=) Tj
+1 0 0 1 180 755 Tm (a) Tj
+1 0 0 1 72 720 Tm (C\) Control: The total of 730 seats was approved. let log 2 = a) Tj
+ET"#;
+    make_text_pdf(content, "0 0 595 842")
+}
+
 fn truncate_eof_marker(mut pdf: Vec<u8>) -> Vec<u8> {
     assert!(pdf.ends_with(b"%%EOF"));
     pdf.pop();
@@ -330,6 +512,21 @@ fn test_group_into_lines_sorting_by_x() {
     assert_eq!(lines[0].text(), "First Second Third");
 }
 
+#[test]
+fn test_digit_only_text_runs_are_preserved_in_markdown() {
+    let pdf = make_digit_run_repro_pdf();
+
+    let items = extract_text_with_positions_mem(&pdf).expect("extract positioned text");
+    assert!(items.iter().any(|item| item.text == "730"));
+    assert!(items.iter().any(|item| item.text == "2"));
+
+    let result = process_pdf_mem(&pdf).expect("convert PDF to markdown");
+    assert_eq!(
+        result.markdown.expect("markdown output").trim(),
+        "A) The total of 730 seats was approved.\nB) let log 2 = a\nC) Control: The total of 730 seats was approved. let log 2 = a"
+    );
+}
+
 // ============================================================================
 // MarkdownOptions Tests
 // ============================================================================
@@ -337,6 +534,7 @@ fn test_group_into_lines_sorting_by_x() {
 #[test]
 fn test_markdown_options_default() {
     let opts = MarkdownOptions::default();
+    assert_eq!(opts.profile, pdf_inspector::MarkdownProfile::Fidelity);
     assert!(opts.detect_headers);
     assert!(opts.detect_lists);
     assert!(opts.detect_code);
@@ -346,6 +544,7 @@ fn test_markdown_options_default() {
 #[test]
 fn test_markdown_options_custom() {
     let opts = MarkdownOptions {
+        profile: pdf_inspector::MarkdownProfile::Compact,
         detect_headers: false,
         detect_lists: true,
         detect_code: false,
@@ -361,6 +560,7 @@ fn test_markdown_options_custom() {
         ..Default::default()
     };
     assert!(!opts.detect_headers);
+    assert_eq!(opts.profile, pdf_inspector::MarkdownProfile::Compact);
     assert!(opts.detect_lists);
     assert!(!opts.detect_code);
     assert_eq!(opts.base_font_size, Some(14.0));
@@ -544,6 +744,30 @@ fn test_markdown_from_items_page_breaks() {
     assert!(md.contains("Content on second page"));
 }
 
+#[test]
+fn test_markdown_page_count_overload_includes_trailing_blank_pages_in_folio_coverage() {
+    let mut items = Vec::new();
+    for (page, value) in [(1, "1"), (2, "2"), (3, "3"), (4, "4")] {
+        items.push(make_text_item(value, 25.0, 30.0, 12.0, page));
+        items.push(make_text_item(
+            "Company report footer",
+            41.0,
+            30.0,
+            12.0,
+            page,
+        ));
+    }
+    let options = MarkdownOptions {
+        strip_headers_footers: false,
+        ..MarkdownOptions::default()
+    };
+
+    let md = to_markdown_from_items_with_rects_and_page_count(items, options, &[], 20);
+
+    assert!(md.contains("1 Company report footer"));
+    assert!(md.contains("4 Company report footer"));
+}
+
 // ============================================================================
 // Markdown From Lines Tests
 // ============================================================================
@@ -1086,6 +1310,18 @@ fn test_snapshot_2013_app2() {
     assert_snapshot("2013-app2");
 }
 
+/// First two pages of Shannon's "A Mathematical Theory of Communication"
+/// (1998 dvips 5.58 → Distiller 3 retypesetting). Canonical legacy-TeX PDF:
+/// non-embedded base-14 fonts with no /Widths (exercises the built-in AFM
+/// metrics fallback), Type3 PK bitmap math fonts with FontMatrix
+/// [1 0 0 -1 0 0] (exercises visual-size scaling), a two-line embedded drop
+/// cap, indent-only paragraph breaks, and display math that must not be
+/// detected as tables or headings.
+#[test]
+fn test_snapshot_shannon_entropy() {
+    assert_snapshot("shannon-entropy-p1-2");
+}
+
 // ============================================================================
 // Pages Needing OCR Tests
 // ============================================================================
@@ -2912,6 +3148,57 @@ fn test_extract_pages_markdown_basic() {
     assert!(!result.pages[0].needs_ocr);
 }
 
+#[test]
+fn test_extract_pages_markdown_uses_document_wide_folio_context() {
+    let pdf = make_recurring_contextual_folio_pdf();
+    let result = extract_pages_markdown_mem(&pdf, None).unwrap();
+
+    assert_eq!(result.pages.len(), 4);
+    for (index, page) in result.pages.iter().enumerate() {
+        assert!(page.markdown.contains("Company report footer"));
+        assert!(
+            !page
+                .markdown
+                .contains(&format!("{} Company report footer", index + 1)),
+            "recurring contextual folio survived on page {}: {}",
+            index + 1,
+            page.markdown
+        );
+    }
+}
+
+#[test]
+fn test_process_pdf_page_filter_uses_document_wide_folio_context() {
+    let pdf = make_recurring_contextual_folio_pdf();
+    let result = process_pdf_mem_with_options(&pdf, PdfOptions::new().pages([1])).unwrap();
+    let markdown = result.markdown.unwrap();
+
+    assert!(markdown.contains("Company report footer"));
+    assert!(!markdown.contains("1 Company report footer"), "{markdown}");
+    assert!(markdown.contains("Body page 1"));
+    assert!(!markdown.contains("Body page 2"));
+}
+
+#[test]
+fn test_selected_page_ignores_context_only_extraction_failure() {
+    let pdf = make_pdf_with_malformed_unselected_page();
+
+    let pages = extract_pages_markdown_mem(&pdf, Some(&[0])).unwrap();
+    assert_eq!(pages.pages.len(), 1);
+    assert!(pages.pages[0].markdown.contains("Selected page text"));
+
+    let result = process_pdf_mem_with_options(&pdf, PdfOptions::new().pages([1])).unwrap();
+    let markdown = result.markdown.unwrap();
+    assert!(markdown.contains("Selected page text"));
+}
+
+#[test]
+fn test_requested_page_extraction_failure_remains_fatal() {
+    let pdf = make_pdf_with_malformed_unselected_page();
+
+    assert!(extract_pages_markdown_mem(&pdf, Some(&[1])).is_err());
+}
+
 #[test]
 fn test_extract_pages_markdown_page_ordering() {
     let buf = std::fs::read("tests/fixtures/nexo-price-en.pdf").unwrap();
@@ -3639,6 +3926,65 @@ fn encrypted_pdf_decrypts_with_correct_password() {
     );
 }
 
+/// Regression for the #231 review finding: `extract_pages_markdown`'s
+/// `has_template_image` check must be gated the same way
+/// `classify_pdf`/`detect_pdf_type` gates it (image_count <= 1, few text
+/// ops, low alphanumeric diversity) — not treated as sufficient on its
+/// own. The fixture is a real text page with substantial, richly varied
+/// body text (>=50 Tj ops) drawn over a full-bleed background image
+/// (e.g. letterhead/watermark). Before the fix, has_template_image alone
+/// forced needs_ocr=true and discarded the page's clean markdown; now the
+/// page must extract normally.
+#[test]
+fn test_extract_pages_markdown_does_not_ocr_text_page_with_watermark_image() {
+    let buf = std::fs::read("tests/fixtures/text_page_with_watermark_image.pdf").unwrap();
+
+    let ext = extract_pages_markdown_mem(&buf, None).expect("fixture should extract");
+    let page = &ext.pages[0];
+    assert!(
+        !page.needs_ocr,
+        "a text page with substantial real text should not be routed to OCR \
+         just because it has a background image"
+    );
+    assert!(
+        page.markdown.contains("watermark"),
+        "expected the page's real body text to be preserved, got: {:?}",
+        page.markdown
+    );
+}
+
+/// Regression for the #231 review finding: `extract_pages_markdown` never
+/// checked `has_vector_text` at all, even though `detect_from_document`'s
+/// Mixed-type per-page routing always sends vector-outlined-text pages to
+/// OCR (outlined glyphs can't be extracted as text). A page with massive
+/// path ops (outlined decorative text) plus a short genuine caption would
+/// extract that caption cleanly — non-empty, non-garbled — so the
+/// existing empty/garbage-text checks alone couldn't catch it.
+#[test]
+fn test_extract_pages_markdown_ocrs_page_with_vector_outlined_text() {
+    let buf = std::fs::read("tests/fixtures/vector_outlined_text_with_caption.pdf").unwrap();
+
+    let cls = pdf_inspector::detector::detect_pdf_type_mem(&buf).expect("fixture should classify");
+    assert!(
+        cls.pages_needing_ocr.contains(&1),
+        "classify_pdf should flag page 1 as needing OCR (vector-outlined text), got: {:?}",
+        cls.pages_needing_ocr
+    );
+
+    let ext = extract_pages_markdown_mem(&buf, None).expect("fixture should extract");
+    let page = &ext.pages[0];
+    assert!(
+        page.needs_ocr,
+        "extract_pages_markdown must agree with classify_pdf that this page needs OCR"
+    );
+    assert!(
+        page.markdown.is_empty(),
+        "a page flagged needs_ocr must not return markdown as if extraction were \
+         trustworthy, got: {:?}",
+        page.markdown
+    );
+}
+
 #[test]
 fn pdf_options_debug_redacts_password() {
     let opts = PdfOptions::new().password("secret123");
@@ -3649,3 +3995,60 @@ fn pdf_options_debug_redacts_password() {
     );
     assert!(dbg.contains("REDACTED"), "expected redaction marker: {dbg}");
 }
+
+/// Regression for #228: a `startxref` pointer corrupted to point at the
+/// wrong byte offset (a single flipped digit — a real, common writer bug)
+/// must not make the whole file unprocessable. The real classic xref table
+/// is still present and findable by scanning for the `xref` keyword; both
+/// pypdf and pdfium recover the same way. Before this fix, every entry
+/// point raised "Invalid PDF structure" on a file whose object data was
+/// otherwise completely intact.
+#[test]
+fn test_process_pdf_recovers_corrupted_startxref_pointer() {
+    let result = process_pdf_with_options(
+        "tests/fixtures/broken_startxref_pointer.pdf",
+        PdfOptions::new(),
+    )
+    .expect("a corrupted startxref pointer should be recoverable, like pypdf/pdfium");
+
+    assert_eq!(result.page_count, 1);
+    let md = result.markdown.unwrap_or_default();
+    assert!(
+        md.contains("Order Detail Report by Account") && md.contains("WIDGET ASSEMBLY"),
+        "recovered document should extract its real text, got: {md:?}"
+    );
+}
+
+/// Regression for #227: `extract_pages_markdown`'s per-page `needs_ocr`
+/// must agree with `classify_pdf`/`detect_pdf_type` on the same page. The
+/// fixture is a full-page raster "scan" with a single line of genuine
+/// native text drawn over it (a header) — the native text extracts
+/// perfectly cleanly (no decoding issues, non-empty), so a needs_ocr
+/// computation based on text-quality signals alone says `false`, while
+/// detection correctly sees a dominant background image and says the page
+/// needs OCR. Both must now agree it needs OCR, and the markdown must not
+/// be returned as if the extraction were trustworthy.
+#[test]
+fn test_extract_pages_markdown_agrees_with_classify_on_scan_with_native_header() {
+    let buf = std::fs::read("tests/fixtures/scan_with_native_header_text.pdf").unwrap();
+
+    let cls = pdf_inspector::detector::detect_pdf_type_mem(&buf).expect("fixture should classify");
+    assert!(
+        cls.pages_needing_ocr.contains(&1),
+        "classify_pdf should flag page 1 as needing OCR (image-dominated), got: {:?}",
+        cls.pages_needing_ocr
+    );
+
+    let ext = extract_pages_markdown_mem(&buf, None).expect("fixture should extract");
+    let page = &ext.pages[0];
+    assert!(
+        page.needs_ocr,
+        "extract_pages_markdown must agree with classify_pdf that this page needs OCR"
+    );
+    assert!(
+        page.markdown.is_empty(),
+        "a page flagged needs_ocr must not return markdown as if extraction were \
+         trustworthy, got: {:?}",
+        page.markdown
+    );
+}

tests/snapshots/p1244-1996.md

@@ -22,7 +22,9 @@ Name and address of employee
 
 **Publication 1244 (Rev. 7-96)** Cat. No. 44472W
 
-**Instructions** You must keep sufficient proof to show the amount of your tip income for the year. A daily record of your tip income is considered sufficient proof. Keep a daily record for each workday showing the amount of cash and credit card tips received directly from customers or other employees. Also keep a record of the amount of tips, if any, you paid to other employees through tip sharing, tip pooling or other arrangements, and the names of employees to whom you paid tips. Show the date that each entry is made. This date should be on or near the date you received the tip income. You may use **Form 4070A**, Employee’s Daily Record of Tips, or any other daily record to record your tips. **Reporting Tips to Your Employer.—**If you receive tips that total $20 or more for any month while working for one employer, you must report the tips to your employer. Tips include cash left by customers, tips customers add to credit card charges, and tips you receive from other employees. You must report your tips for any one month by the 10th day of the next month. If the 10th day falls on a Saturday, Sunday, or legal holiday, you may give the report to your employer on the next business day that is not a Saturday, Sunday, or legal holiday. You must report tips that total $20 or more every month regardless of your total wages and tips for the year. You may use **Form 4070**, Employee’s Report of Tips to Employer, to report your tips to your employer. See the instructions on the back of Form 4070. You must include all tips, including tips not reported to your employer, as wages on your income tax return. You may use the last page of this publication to total your tips for the year. Your employer must withhold income, social security, and Medicare (or railroad retirement) taxes on tips you report. Your employer usually deducts the withholding due on tips from your regular wages.
+### Instructions
+
+You must keep sufficient proof to show the amount of your tip income for the year. A daily record of your tip income is considered sufficient proof. Keep a daily record for each workday showing the amount of cash and credit card tips received directly from customers or other employees. Also keep a record of the amount of tips, if any, you paid to other employees through tip sharing, tip pooling or other arrangements, and the names of employees to whom you paid tips. Show the date that each entry is made. This date should be on or near the date you received the tip income. You may use **Form 4070A**, Employee’s Daily Record of Tips, or any other daily record to record your tips. **Reporting Tips to Your Employer.—**If you receive tips that total $20 or more for any month while working for one employer, you must report the tips to your employer. Tips include cash left by customers, tips customers add to credit card charges, and tips you receive from other employees. You must report your tips for any one month by the 10th day of the next month. If the 10th day falls on a Saturday, Sunday, or legal holiday, you may give the report to your employer on the next business day that is not a Saturday, Sunday, or legal holiday. You must report tips that total $20 or more every month regardless of your total wages and tips for the year. You may use **Form 4070**, Employee’s Report of Tips to Employer, to report your tips to your employer. See the instructions on the back of Form 4070. You must include all tips, including tips not reported to your employer, as wages on your income tax return. You may use the last page of this publication to total your tips for the year. Your employer must withhold income, social security, and Medicare (or railroad retirement) taxes on tips you report. Your employer usually deducts the withholding due on tips from your regular wages.
 
 *(continued on inside of back cover)*
 
@@ -54,7 +56,7 @@ tips of directly from customers received other employees paid tips rec’d. entr
 **2.** Report total credit card tips (col. **b**) on Form 4070, line **2.**
 **3.** Report total tips paid out (col. **c**) on Form 4070, line **3.** **Page 4**
 
-Form Employee’s Report (Rev. July 1996)
+Form **4070** Employee’s Report (Rev. July 1996)
 
 ## of Tips to EmployerOMB No. 1545-0065
 
@@ -72,7 +74,7 @@ Month or shorter period in which tips were received **4** Net tips (lines **1 +
 
 forms simpler, we would be happy to hear from you. You can write to the Tax Forms Committee, Western Area Distribution Center, Rancho Cordova, CA 95743-0001. **Purpose.—**Use this form to report tips you receive to your employer. This includes cash tips, tips you receive from other employees, and credit card tips. You must report tips every month regardless of your total wages and tips for the year. However, you do not have to report tips to your employer for any month you received less than $20 in tips while working for that employer. Report tips by the 10th day of the month following the month that you receive them. If the 10th day is a Saturday, Sunday, or legal holiday, report tips by the next day that is not a Saturday, Sunday, or legal holiday. See **Pub. 531**, Reporting Tip Income, for more information. You can get additional copies of **Pub. 1244**, Employee’s Daily Record of Tips and Report to Employer, which contains both Forms 4070A and 4070, by calling 1-800-TAX-FORM (1-800-829-3676).
 
-**Instructions** *(continued)*
+<u>Instructions (continued)</u>
 
 **Unreported Tips.—**If you received tips of $20 or more for any month while working for one employer but did not report them to your employer, you must figure and pay social security and Medicare taxes on the unreported tips when you file your tax return. If you have unreported tips, you **must** use Form 1040 and **Form 4137,** Social Security and Medicare Tax on Unreported Tip Income, to report them. You may **not** use Form 1040A or 1040EZ. Employees subject to the Railroad Retirement Tax Act **cannot** use Form 4137 to pay railroad retirement tax on unreported tips. To get railroad retirement credit, you must report tips to your employer. If you do not report tips to your employer as required, you may be charged a penalty of 50% of the social security and Medicare taxes (or railroad retirement tax) due on the unreported tips unless there was reasonable cause for not reporting them. **Additional Information.—**Get **Pub. 531,** Reporting Tip Income, and Form 4137 for more information on tips. If you are an employee of certain large food or beverage establishments, see Pub. 531 for tip allocation rules. **Recordkeeping.—**If you do not keep a daily record of tips, you must keep other reliable proof of the tip income you received. This proof includes copies of restaurant bills and credit card charges that show amounts customers added as tips. Keep your tip income records for as long as the information on them may be needed in the administration of any Internal Revenue law.
 

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

@@ -0,0 +1,43 @@
+Reprinted with corrections from *The Bell System Technical Journal,* Vol. 27, pp. 379–423, 623–656, July, October, 1948.
+
+## A Mathematical Theory of Communication
+
+### By C. E. SHANNON
+
+INTRODUCTION
+
+HE recent development of various methods of modulation such as PCM and PPM which exchange
+
+# Tbandwidth for signal-to-noise ratio has intensified the interest in a general theory of communication. A
+
+basis for such a theory is contained in the important papers of Nyquist¹ and Hartley² on this subject. In the present paper we will extend the theory to include a number of new factors, in particular the effect of noise in the channel, and the savings possible due to the statistical structure of the original message and due to the nature of the final destination of the information. The fundamental problem of communication is that of reproducing at one point either exactly or ap- proximately a message selected at another point. Frequently the messages have *meaning*; that is they refer to or are correlated according to some system with certain physical or conceptual entities. These semantic aspects of communication are irrelevant to the engineering problem. The significant aspect is that the actual message is one *selected from a set* of possible messages. The system must be designed to operate for each possible selection, not just the one which will actually be chosen since this is unknown at the time of design. If the number of messages in the set is finite then this number or any monotonic function of this number can be regarded as a measure of the information produced when one message is chosen from the set, all choices being equally likely. As was pointed out by Hartley the most natural choice is the logarithmic function. Although this definition must be generalized considerably when we consider the influence of the statistics of the message and when we have a continuous range of messages, we will in all cases use an essentially logarithmic measure. The logarithmic measure is more convenient for various reasons:
+
+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
+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
+*N* and log₂2 *N* = *N*. If the base 10 is used the units may be called decimal digits. Since
+
+log₂*M* = log₁₀*M*= log₁₀2 = 3:32 log₁₀*M*;
+
+1 Nyquist, H., “Certain Factors Affecting Telegraph Speed,” *Bell System Technical Journal,* April 1924, p. 324; “Certain Topics in Telegraph Transmission Theory,” *A.I.E.E. Trans.,* v. 47, April 1928, p. 617. 2 Hartley, R. V. L., “Transmission of Information,” *Bell System Technical Journal,* July 1928, p. 535.
+
+INFORMATION SOURCE TRANSMITTER RECEIVER DESTINATION
+
+SIGNAL RECEIVED SIGNAL MESSAGE MESSAGE
+
+NOISE SOURCE
+
+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.
+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
+necessary to represent the various elements involved as mathematical entities, suitably idealized from their
+

tests/snapshots/td9264.md

@@ -1,17 +1,8 @@
-(e) [Reserved]. For further guidance, see §1.1563-3T(e)(1). Par. 50. Section 1.1563-3T is added to read as follows:
-<u>§1.1563-3T Rules for determining stock ownership (temporary)</u>.
-
-(a) through (d)(2)(iii) [Reserved]. For further guidance, see §1.1563-3(a)
-through (d)(2)(iii). (iv) <u>Statement</u>. If the application of paragraph (d)(2)(ii) or (iii) of §1.1563-3 does not result in a corporation being treated as a component member of only one controlled group of corporations on a December 31, then such corporation will be treated as a component member of only one such group on such date. Such corporation may elect the group in which it is to be included by including on or with its income tax return a statement entitled, “STATEMENT TO ELECT CONTROLLED GROUP PURSUANT TO §1.1563-3T(d)(2)(iv).” The statement must include--
-
-(A) A description of each of the controlled groups in which the corporation
-could be included. The description must include the name and employer identification number of each component member of each such group and the stock ownership of the component members of each such group; and
-
-(B) The following representation: [INSERT NAME AND EMPLOYER
-IDENTIFICATION NUMBER OF CORPORATION] ELECTS TO BE TREATED AS A COMPONENT MEMBER OF THE [INSERT DESIGNATION OF GROUP].
-
-(v) <u>Election</u>-- (A) <u>Election filed</u>. An election filed under paragraph (d)(2)(iv) of
-this section is irrevocable and effective until paragraph (d)(2)(ii) or (iii) of §1.1563-3 applies or until a change in the stock ownership of the corporation results in
+||||(e) [Reserved]. For further guidance, see §1.1563-3T(e)(1). Par. 50. Section 1.1563-3T is added to read as follows: §1.1563-3T Rules for determining stock ownership (temporary). (a) through (d)(2)(iii) [Reserved]. For further guidance, see §1.1563-3(a)|
+|---|---|---|---|
+||through (d)(2)(iii).|||
+||(iv)|Statement|. If the application of paragraph (d)(2)(ii) or (iii) of §1.1563-3 does not result in a corporation being treated as a component member of only one controlled group of corporations on a December 31, then such corporation will be treated as a component member of only one such group on such date. Such corporation may elect the group in which it is to be included by including on or with its income tax return a statement entitled, “STATEMENT TO ELECT CONTROLLED GROUP PURSUANT TO §1.1563-3T(d)(2)(iv).” The statement must include-- (A) A description of each of the controlled groups in which the corporation could be included. The description must include the name and employer identification number of each component member of each such group and the stock ownership of the component members of each such group; and (B) The following representation: [INSERT NAME AND EMPLOYER IDENTIFICATION NUMBER OF CORPORATION] ELECTS TO BE TREATED AS A COMPONENT MEMBER OF THE [INSERT DESIGNATION OF GROUP].|
+||(v)|Election|-- (A) Election filed. An election filed under paragraph (d)(2)(iv) of this section is irrevocable and effective until paragraph (d)(2)(ii) or (iii) of §1.1563-3 applies or until a change in the stock ownership of the corporation results in|
 
 |termination of membership in the controlled group in which such corporation has||
 |---|---|
@@ -29,7 +20,7 @@ this section is irrevocable and effective until paragraph (d)(2)(ii) or (iii) of
 Federal income tax return (including any amended return filed on or before the due date (including extensions) of such original return) timely filed on or after May 30,
 
 2006.
-(2) Expiration date. The applicability of this section will expire on May 26,
+(2) <u>Expiration date</u>. The applicability of this section will expire on May 26,
 2009. Par. 51. Section 1.6012-2 is amended by revising paragraph (c) and adding paragraph (k) to read as follows: <u>§1.6012-2 Corporations required to make returns of income</u>.
 * * * * *
 (c) [Reserved]. For further guidance, see §1.6012-2T(c).

tests/snapshots/thermo-freon12.md

@@ -6,20 +6,22 @@
 
 #### Thermodynamic Properties
 
-**of**
-
-®
+**of** ®
 
 # Freon 12
 
-**(R-12)** **Technical Information** **Technical Information**
+##### (R-12)
+
+##### Technical Information Technical Information
 
 **®** **Thermodynamic Properties of Freon 12 Refrigerant** **(R-12)** **SI Units**
 
 Tables of the thermodynamic **Units** properties of R-12 have been developed and are presented here. P = Pressure in kPa. Absolute This information is based on values calculated using the NIST REFPROP T = Temperature in Celcius Database (McLinden, M.O., Klein,
 
 S.A., Lemmon, E.W., and Peskin, Vf = Fluid (liquid) specific volume
-A.P., NIST Standard Reference in cubic meters per kilogram Database 23, NIST thermodynamic and transport properties of Vg = Vapour (gas) specific volume refrigerants and refrigerant in cubic meters per kilogram mixtures – REFPROP version 6.01, Standard Reference Data Program, df and dg = Fluid and Vapour National Institute of Standards and (respectively) densities in Technology, 1998). kilograms per cubic meter
+A.P., NIST Standard Reference in cubic meters per kilogram Database 23, NIST thermodynamic and transport properties of Vg = Vapour (gas) specific volume refrigerants and refrigerant in cubic meters per kilogram mixtures – REFPROP version 6.01, Standard Reference Data Program, df and dg = Fluid and Vapour National Institute of Standards and (respectively) densities in Technology, 1998).
+kilograms per cubic meter
+
 ##### H = Enthalpy (kJ/kg)
 
 ##### S = Entropy (kJ/kg.K)
@@ -43,9 +45,9 @@ l
 
 **Freon** **®** **12 Saturation Properties-Temperature Table**
 
-|Temp|Pressure||Volume|||Density||Enthalpy|||Entropy|Temp|
-|---|---|---|---|---|---|---|---|---|---|---|---|---|
-|°C|[kPa]|[m³ Liquid v f|/kg]|Vapour v g|Liquid d f|[kg/m³] Vapour d g|Liquid H f|[kJ/kg] Latent H fg|Vapour H g|Liquid S f|[kJ/K-kg] Vapour S g|°C|
+|Temp|Pressure||Volume||Density||Enthalpy|||Entropy|Temp|
+|---|---|---|---|---|---|---|---|---|---|---|---|
+|°C|[kPa]|[m³ Liquid v f|/kg] Vapour v g|[kg/m³ Liquid d f|] Vapour d g|Liquid H f|[kJ/kg] Latent H fg|Vapour H g|Liquid S f|[kJ/K-kg] Vapour S g|°C|
 
 |-100|1.2|0.0006|10.0000|1679.0|0.100|113.3|192.8|306.1|0.6077|1.7210|-100|
 |---|---|---|---|---|---|---|---|---|---|---|---|

wasm/Cargo.toml

@@ -0,0 +1,37 @@
+[package]
+name = "pdf-inspector-wasm"
+version = "0.1.3"
+edition = "2021"
+authors = ["Firecrawl Team"]
+description = "Browser WebAssembly bindings for pdf-inspector"
+license = "MIT"
+repository = "https://github.com/firecrawl/pdf-inspector"
+homepage = "https://github.com/firecrawl/pdf-inspector"
+readme = "README.md"
+publish = false
+
+[lib]
+crate-type = ["cdylib", "rlib"]
+
+[dependencies]
+console_error_panic_hook = "0.1"
+js-sys = "0.3"
+pdf-inspector = { path = ".." }
+serde = { version = "1", features = ["derive"] }
+serde-wasm-bindgen = "0.6"
+wasm-bindgen = "0.2"
+
+[dev-dependencies]
+wasm-bindgen-test = "0.3"
+
+[profile.release]
+codegen-units = 1
+lto = true
+opt-level = "s"
+strip = true
+
+[package.metadata.wasm-pack.profile.release]
+# Rust 1.95 emits bulk-memory instructions that the binaryen bundled with
+# wasm-pack 0.15.0 does not yet validate. rustc still performs the release,
+# size, and LTO optimizations above.
+wasm-opt = false

wasm/LICENSE

@@ -0,0 +1,58 @@
+MIT License
+
+Copyright (c) 2026 Firecrawl
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+
+Third-party notices
+===================
+
+Adobe CMaps
+-----------
+
+The WebAssembly binary embeds binary CMaps derived from Adobe CMap resources.
+
+Copyright 1990-2009 Adobe Systems Incorporated.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+Redistributions of source code must retain the above copyright notice, this
+list of conditions and the following disclaimer.
+
+Redistributions in binary form must reproduce the above copyright notice,
+this list of conditions and the following disclaimer in the documentation
+and/or other materials provided with the distribution.
+
+Neither the name of Adobe Systems Incorporated nor the names of its
+contributors may be used to endorse or promote products derived from this
+software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.

wasm/README.md

@@ -0,0 +1,60 @@
+# @firecrawl/pdf-inspector-wasm
+
+Browser WebAssembly bindings for [pdf-inspector](https://github.com/firecrawl/pdf-inspector). Classify PDFs and extract structured Markdown locally from a `Uint8Array`, using the same Rust core as the native Node.js, Python, and Rust packages.
+
+## Install
+
+```bash
+npm install @firecrawl/pdf-inspector-wasm
+```
+
+## Usage
+
+```ts
+import init, { processPdf } from "@firecrawl/pdf-inspector-wasm";
+
+await init();
+
+const response = await fetch("/annual-report.pdf");
+const pdf = new Uint8Array(await response.arrayBuffer());
+const result = processPdf(pdf);
+
+console.log(result.pdfType);
+console.log(result.markdown);
+```
+
+Pass options when you need selected pages or compact Markdown:
+
+```ts
+const result = processPdf(pdf, {
+  pages: [1, 3, 5],
+  profile: "compact",
+  includePageMarkers: true,
+});
+```
+
+The package also exports:
+
+- `detectPdf(pdf, options?)` for detection without extraction.
+- `classifyPdf(pdf)` for the lightweight result shape shared with the native Node.js API.
+- `extractText(pdf)` for plain text.
+- `version()` for the WASM package version.
+
+## Browser behavior
+
+- Parsing runs locally. PDF bytes are not uploaded anywhere.
+- The build is single-threaded and does not require cross-origin isolation.
+- CMaps are embedded so CJK font decoding does not depend on a filesystem.
+- Extraction is synchronous after `init()`. For large documents, call it from a Web Worker to keep the UI responsive.
+- Image-only documents still require a separate OCR step.
+
+## Build from source
+
+```bash
+cargo install wasm-pack --version 0.15.0 --locked
+wasm-pack build wasm --target web --scope firecrawl --release
+```
+
+## License
+
+MIT

wasm/src/lib.rs

@@ -0,0 +1,440 @@
+use pdf_inspector::{
+    LayoutComplexity, MarkdownProfile, PageOcrReasons, PdfOptions, PdfProcessResult, PdfType,
+    ProcessMode,
+};
+use serde::{Deserialize, Serialize};
+use wasm_bindgen::prelude::*;
+
+#[wasm_bindgen(typescript_custom_section)]
+const TYPESCRIPT_TYPES: &str = r#"
+export type PdfType = "TextBased" | "Scanned" | "ImageBased" | "Mixed";
+export type MarkdownProfile = "fidelity" | "compact";
+
+export interface ProcessOptions {
+  /** Restrict extraction to these 1-indexed page numbers. */
+  pages?: number[];
+  /** Password for an encrypted PDF. */
+  password?: string;
+  /** Source-faithful output by default, or compact output for fewer tokens. */
+  profile?: MarkdownProfile;
+  /** Insert `<!-- Page N -->` markers between pages. */
+  includePageMarkers?: boolean;
+  /** Include image placeholders in Markdown output. */
+  includeImages?: boolean;
+}
+
+export interface PageOcrReasons {
+  /** 1-indexed page number. */
+  page: number;
+  reasons: string[];
+}
+
+export interface LayoutComplexity {
+  isComplex: boolean;
+  /** 1-indexed page numbers. */
+  pagesWithTables: number[];
+  /** 1-indexed page numbers. */
+  pagesWithColumns: number[];
+}
+
+export interface PdfProcessResult {
+  pdfType: PdfType;
+  markdown?: string;
+  pageCount: number;
+  processingTimeMs: number;
+  /** 1-indexed page numbers. */
+  pagesNeedingOcr: number[];
+  ocrReasonsByPage: PageOcrReasons[];
+  title?: string;
+  confidence: number;
+  layout: LayoutComplexity;
+  hasEncodingIssues: boolean;
+}
+
+export interface PdfClassification {
+  pdfType: PdfType;
+  pageCount: number;
+  /** 0-indexed page numbers, matching the native Node.js API. */
+  pagesNeedingOcr: number[];
+  confidence: number;
+}
+
+export function processPdf(data: Uint8Array, options?: ProcessOptions): PdfProcessResult;
+export function detectPdf(data: Uint8Array, options?: Pick<ProcessOptions, "password">): PdfProcessResult;
+export function classifyPdf(data: Uint8Array): PdfClassification;
+export function extractText(data: Uint8Array): string;
+export function version(): string;
+"#;
+
+#[derive(Debug, Default, Deserialize)]
+#[serde(default, rename_all = "camelCase", deny_unknown_fields)]
+struct WasmProcessOptions {
+    pages: Option<Vec<u32>>,
+    password: Option<String>,
+    profile: Option<WasmMarkdownProfile>,
+    include_page_markers: Option<bool>,
+    include_images: Option<bool>,
+}
+
+#[derive(Debug, Deserialize)]
+#[serde(rename_all = "lowercase")]
+enum WasmMarkdownProfile {
+    Fidelity,
+    Compact,
+}
+
+#[derive(Serialize)]
+#[serde(rename_all = "camelCase")]
+struct WasmPageOcrReasons {
+    page: u32,
+    reasons: Vec<String>,
+}
+
+impl From<PageOcrReasons> for WasmPageOcrReasons {
+    fn from(value: PageOcrReasons) -> Self {
+        Self {
+            page: value.page,
+            reasons: value.reasons,
+        }
+    }
+}
+
+#[derive(Serialize)]
+#[serde(rename_all = "camelCase")]
+struct WasmLayoutComplexity {
+    is_complex: bool,
+    pages_with_tables: Vec<u32>,
+    pages_with_columns: Vec<u32>,
+}
+
+impl From<LayoutComplexity> for WasmLayoutComplexity {
+    fn from(value: LayoutComplexity) -> Self {
+        Self {
+            is_complex: value.is_complex,
+            pages_with_tables: value.pages_with_tables,
+            pages_with_columns: value.pages_with_columns,
+        }
+    }
+}
+
+#[derive(Serialize)]
+#[serde(rename_all = "camelCase")]
+struct WasmPdfProcessResult {
+    pdf_type: &'static str,
+    markdown: Option<String>,
+    page_count: u32,
+    processing_time_ms: f64,
+    pages_needing_ocr: Vec<u32>,
+    ocr_reasons_by_page: Vec<WasmPageOcrReasons>,
+    title: Option<String>,
+    confidence: f64,
+    layout: WasmLayoutComplexity,
+    has_encoding_issues: bool,
+}
+
+impl From<PdfProcessResult> for WasmPdfProcessResult {
+    fn from(value: PdfProcessResult) -> Self {
+        Self {
+            pdf_type: pdf_type_name(value.pdf_type),
+            markdown: value.markdown,
+            page_count: value.page_count,
+            processing_time_ms: value.processing_time_ms as f64,
+            pages_needing_ocr: value.pages_needing_ocr,
+            ocr_reasons_by_page: value
+                .ocr_reasons_by_page
+                .into_iter()
+                .map(Into::into)
+                .collect(),
+            title: value.title,
+            confidence: value.confidence as f64,
+            layout: value.layout.into(),
+            has_encoding_issues: value.has_encoding_issues,
+        }
+    }
+}
+
+#[derive(Serialize)]
+#[serde(rename_all = "camelCase")]
+struct WasmPdfClassification {
+    pdf_type: &'static str,
+    page_count: u32,
+    pages_needing_ocr: Vec<u32>,
+    confidence: f64,
+}
+
+fn pdf_type_name(pdf_type: PdfType) -> &'static str {
+    match pdf_type {
+        PdfType::TextBased => "TextBased",
+        PdfType::Scanned => "Scanned",
+        PdfType::ImageBased => "ImageBased",
+        PdfType::Mixed => "Mixed",
+    }
+}
+
+fn js_error(context: &str, error: impl std::fmt::Display) -> JsValue {
+    js_sys::Error::new(&format!("{context}: {error}")).into()
+}
+
+fn deserialize_options(value: JsValue) -> Result<WasmProcessOptions, JsValue> {
+    if value.is_undefined() || value.is_null() {
+        return Ok(WasmProcessOptions::default());
+    }
+
+    serde_wasm_bindgen::from_value(value).map_err(|error| js_error("invalid options", error))
+}
+
+fn build_options(value: JsValue, mode: ProcessMode) -> Result<PdfOptions, JsValue> {
+    let options = deserialize_options(value)?;
+    if options
+        .pages
+        .as_ref()
+        .is_some_and(|pages| pages.contains(&0))
+    {
+        return Err(js_error(
+            "invalid options",
+            "pages are 1-indexed; page 0 is invalid",
+        ));
+    }
+
+    let mut result = PdfOptions::new().mode(mode);
+    if let Some(pages) = options.pages {
+        result = result.pages(pages);
+    }
+    if let Some(password) = options.password {
+        result = result.password(password);
+    }
+    if let Some(profile) = options.profile {
+        result.markdown.profile = match profile {
+            WasmMarkdownProfile::Fidelity => MarkdownProfile::Fidelity,
+            WasmMarkdownProfile::Compact => MarkdownProfile::Compact,
+        };
+    }
+    if let Some(include_page_markers) = options.include_page_markers {
+        result.markdown.include_page_numbers = include_page_markers;
+    }
+    if let Some(include_images) = options.include_images {
+        result.markdown.include_images = include_images;
+    }
+    Ok(result)
+}
+
+fn serialize<T: Serialize>(value: &T) -> Result<JsValue, JsValue> {
+    serde_wasm_bindgen::to_value(value).map_err(|error| js_error("serialize result", error))
+}
+
+fn initialize() {
+    console_error_panic_hook::set_once();
+}
+
+/// Process PDF bytes entirely inside WebAssembly.
+#[wasm_bindgen(js_name = processPdf, skip_typescript)]
+pub fn process_pdf(data: &[u8], options: JsValue) -> Result<JsValue, JsValue> {
+    initialize();
+    let options = build_options(options, ProcessMode::Full)?;
+    let started = js_sys::Date::now();
+    let mut result = pdf_inspector::process_pdf_mem_with_options(data, options)
+        .map_err(|error| js_error("process PDF", error))?;
+    result.processing_time_ms = (js_sys::Date::now() - started).max(0.0) as u64;
+    serialize(&WasmPdfProcessResult::from(result))
+}
+
+/// Classify PDF bytes without extracting text or producing Markdown.
+#[wasm_bindgen(js_name = detectPdf, skip_typescript)]
+pub fn detect_pdf(data: &[u8], options: JsValue) -> Result<JsValue, JsValue> {
+    initialize();
+    let options = build_options(options, ProcessMode::DetectOnly)?;
+    let started = js_sys::Date::now();
+    let mut result = pdf_inspector::process_pdf_mem_with_options(data, options)
+        .map_err(|error| js_error("detect PDF", error))?;
+    result.processing_time_ms = (js_sys::Date::now() - started).max(0.0) as u64;
+    serialize(&WasmPdfProcessResult::from(result))
+}
+
+/// Return the lightweight classification shape used by the native Node API.
+#[wasm_bindgen(js_name = classifyPdf, skip_typescript)]
+pub fn classify_pdf(data: &[u8]) -> Result<JsValue, JsValue> {
+    initialize();
+    let result =
+        pdf_inspector::classify_pdf_mem(data).map_err(|error| js_error("classify PDF", error))?;
+    serialize(&WasmPdfClassification {
+        pdf_type: pdf_type_name(result.pdf_type),
+        page_count: result.page_count,
+        pages_needing_ocr: result.pages_needing_ocr,
+        confidence: result.confidence as f64,
+    })
+}
+
+/// Extract plain text from PDF bytes without Markdown conversion.
+#[wasm_bindgen(js_name = extractText, skip_typescript)]
+pub fn extract_text(data: &[u8]) -> Result<String, JsValue> {
+    initialize();
+    let items = pdf_inspector::extractor::extract_text_with_positions_mem(data)
+        .map_err(|error| js_error("extract text", error))?;
+    Ok(
+        pdf_inspector::extractor::group_into_lines_preserving_all_text(items)
+            .into_iter()
+            .map(|line| line.text())
+            .filter(|line| !line.trim().is_empty())
+            .collect::<Vec<_>>()
+            .join("\n"),
+    )
+}
+
+/// Return the WebAssembly package version.
+#[wasm_bindgen(skip_typescript)]
+pub fn version() -> String {
+    env!("CARGO_PKG_VERSION").to_string()
+}
+
+#[cfg(all(test, target_arch = "wasm32"))]
+mod tests {
+    use super::*;
+    use js_sys::Reflect;
+    use wasm_bindgen_test::*;
+
+    const TEXT_PDF: &[u8] = include_bytes!("../../tests/fixtures/thermo-freon12.pdf");
+    const ENCRYPTED_PDF: &[u8] = include_bytes!("../../tests/fixtures/encrypted-secret123.pdf");
+
+    fn synthetic_korea1_pdf() -> 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");
+        }
+
+        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 0 200 200] /Resources << /Font << /F0 5 0 R >> >> /Contents 4 0 R >>",
+        );
+
+        // Adobe-Korea1 CID 1086 (0x043E) maps to U+AC00 (Korean syllable GA).
+        // There is deliberately no ToUnicode stream: decoding must use the
+        // embedded predefined CMap rather than lopdf's plain-text fallback.
+        // Korea1 CIDs 21 and 19 map to ASCII "4" and "2". Place them near
+        // the bottom edge so they look exactly like a numeric page footer.
+        let content = "BT /F0 12 Tf 50 100 Td <043E> Tj 0 -60 Td <00150013> Tj ET";
+        add_object(
+            &mut pdf,
+            &mut offsets,
+            4,
+            &format!(
+                "<< /Length {} >>\nstream\n{}\nendstream",
+                content.len(),
+                content
+            ),
+        );
+        add_object(
+            &mut pdf,
+            &mut offsets,
+            5,
+            "<< /Type /Font /Subtype /Type0 /BaseFont /SyntheticKorea1 /Encoding /Identity-H /DescendantFonts [6 0 R] >>",
+        );
+        add_object(
+            &mut pdf,
+            &mut offsets,
+            6,
+            "<< /Type /Font /Subtype /CIDFontType2 /BaseFont /SyntheticKorea1 /CIDSystemInfo << /Registry (Adobe) /Ordering (Korea1) /Supplement 2 >> /FontDescriptor 7 0 R /DW 1000 >>",
+        );
+        add_object(
+            &mut pdf,
+            &mut offsets,
+            7,
+            "<< /Type /FontDescriptor /FontName /SyntheticKorea1 /Flags 4 /FontBBox [-100 -200 1000 900] /ItalicAngle 0 /Ascent 800 /Descent -200 /CapHeight 700 /StemV 80 >>",
+        );
+
+        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
+    }
+
+    #[wasm_bindgen_test]
+    fn processes_pdf_to_markdown() {
+        let result = process_pdf(TEXT_PDF, JsValue::UNDEFINED).expect("process PDF");
+        let pdf_type = Reflect::get(&result, &JsValue::from_str("pdfType"))
+            .expect("pdfType")
+            .as_string()
+            .expect("pdfType string");
+        let markdown = Reflect::get(&result, &JsValue::from_str("markdown"))
+            .expect("markdown")
+            .as_string()
+            .expect("markdown string");
+
+        assert_eq!(pdf_type, "TextBased");
+        assert!(!markdown.is_empty());
+    }
+
+    #[wasm_bindgen_test]
+    fn rejects_non_pdf_bytes() {
+        assert!(process_pdf(b"not a PDF", JsValue::UNDEFINED).is_err());
+    }
+
+    #[wasm_bindgen_test]
+    fn classifies_and_extracts_plain_text() {
+        let classification = classify_pdf(TEXT_PDF).expect("classify PDF");
+        let pdf_type = Reflect::get(&classification, &JsValue::from_str("pdfType"))
+            .expect("pdfType")
+            .as_string()
+            .expect("pdfType string");
+        let text = extract_text(TEXT_PDF).expect("extract text");
+
+        assert_eq!(pdf_type, "TextBased");
+        assert!(!text.is_empty());
+    }
+
+    #[wasm_bindgen_test]
+    fn extracts_cjk_and_preserves_numeric_page_footer() {
+        let text = extract_text(&synthetic_korea1_pdf()).expect("extract predefined CMap text");
+
+        assert_eq!(text, "가\n42");
+    }
+
+    #[wasm_bindgen_test]
+    fn opens_encrypted_pdf_with_password() {
+        assert!(process_pdf(ENCRYPTED_PDF, JsValue::UNDEFINED).is_err());
+
+        let options = js_sys::Object::new();
+        Reflect::set(
+            &options,
+            &JsValue::from_str("password"),
+            &JsValue::from_str("secret123"),
+        )
+        .expect("set password");
+        let result = process_pdf(ENCRYPTED_PDF, options.into()).expect("process encrypted PDF");
+        let markdown = Reflect::get(&result, &JsValue::from_str("markdown"))
+            .expect("markdown")
+            .as_string()
+            .expect("markdown string");
+
+        assert!(!markdown.is_empty());
+    }
+}