Bump version from 1.8.9 to 1.8.10

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

.github/workflows/publish.yml

@@ -2,14 +2,41 @@ name: Publish npm package
 
 on:
   push:
-    tags: ['v*']
+    branches: [main]
+    paths: ['napi/package.json']
 
 permissions:
   contents: read
   id-token: write
 
 jobs:
+  check-version:
+    name: Check version change
+    runs-on: ubuntu-latest
+    outputs:
+      changed: ${{ steps.check.outputs.changed }}
+      version: ${{ steps.check.outputs.version }}
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 2
+
+      - name: Check if version changed
+        id: check
+        run: |
+          NEW_VERSION=$(node -p "require('./napi/package.json').version")
+          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
+
   build:
+    needs: check-version
+    if: needs.check-version.outputs.changed == 'true'
     name: Build ${{ matrix.target }}
     runs-on: ${{ matrix.os }}
     strategy:
@@ -19,6 +46,8 @@ jobs:
             target: x86_64-unknown-linux-gnu
           - os: macos-14
             target: aarch64-apple-darwin
+          - os: windows-latest
+            target: x86_64-pc-windows-msvc
     steps:
       - uses: actions/checkout@v4
 
@@ -49,16 +78,26 @@ jobs:
         working-directory: napi
         run: bunx napi build --platform --release
 
-      - name: Upload artifact
+      - name: Upload native binary
         uses: actions/upload-artifact@v4
         with:
           name: bindings-${{ matrix.target }}
           path: napi/*.node
           if-no-files-found: error
 
+      - name: Upload generated JS bindings
+        if: matrix.target == 'x86_64-unknown-linux-gnu'
+        uses: actions/upload-artifact@v4
+        with:
+          name: js-bindings
+          path: |
+            napi/index.js
+            napi/index.d.ts
+          if-no-files-found: error
+
   publish:
     name: Publish to npm
-    needs: build
+    needs: [check-version, build]
     runs-on: ubuntu-latest
     permissions:
       contents: read
@@ -79,8 +118,9 @@ jobs:
       - name: Collect binaries and publish
         working-directory: napi
         run: |
-          # Copy all .node binaries into the package directory
           cp artifacts/bindings-*/*.node .
+          cp artifacts/js-bindings/index.js .
+          cp artifacts/js-bindings/index.d.ts .
 
           echo "=== Package contents ==="
           ls -la *.node index.js index.d.ts

.gitignore

@@ -24,6 +24,10 @@ Thumbs.db
 # Build cache
 **/*.rs.bk
 
+# NAPI generated (regenerated by `napi prepublish` during CI)
+napi/index.js
+napi/index.d.ts
+
 # Local samples and scripts
 samples/
 scripts/
@@ -31,3 +35,8 @@ scripts/
 # Test output
 test_output/
 
+# Python
+__pycache__/
+*.pyc
+.pytest_cache/
+

AGENTS.md

@@ -0,0 +1,79 @@
+# pdf-inspector
+
+Fast PDF text extraction to structured Markdown. CLI binary: `pdf2md`. Detection binary: `detect-pdf`.
+
+## Build & Test
+
+```bash
+cargo fmt                                    # format
+cargo clippy -- -D warnings                  # lint (enforced, zero warnings)
+cargo test                                   # unit + integration tests (267+ unit, 73+ integration)
+cargo build --release                        # release binary for benchmarks
+```
+
+All three must pass before committing.
+
+## Binaries
+
+- `pdf2md` — extract PDF → Markdown. Supports `--json` for structured output.
+- `detect-pdf` — classify PDF type (TextBased/Scanned/Mixed/ImageBased). Supports `--analyze --json`.
+
+## Architecture
+
+```
+src/
+  lib.rs                        – public API, process_pdf_with_options, encoding issue detection
+  detector.rs                   – PDF type classification, tiled-scan detection, page sampling
+  types.rs                      – TextItem, TextLine, PdfRect, PdfLine
+  tounicode.rs                  – CMap/ToUnicode parsing, CID decoding
+  text_utils.rs                 – CJK/RTL handling, Otsu threshold, ligature expansion, NFKC
+  extractor/
+    mod.rs                      – top-level extraction orchestrator
+    content_stream.rs           – PDF operator state machine (Tj/TJ/Td/Tm/q/Q)
+    fonts.rs                    – font width/encoding, CMapDecisionCache, TrueType cmap fallback
+    layout.rs                   – column detection (histogram), newspaper/tabular classification,
+                                  spanning-line pre-masking, sidebar detection
+  tables/
+    detect_rects.rs             – rect-based table detection (union-find clustering)
+    detect_heuristic.rs         – heuristic table detection (gap-histogram, body-font tables)
+    detect_lines.rs             – line-based table detection (H/V line grids)
+    grid.rs                     – column/row boundaries, cell assignment
+    format.rs                   – table→Markdown formatting, continuation row merging
+  markdown/
+    convert.rs                  – core line→Markdown loop, struct-tree role support
+    analysis.rs                 – font stats, heading tiers, paragraph thresholds
+    classify.rs                 – line classification (header, list, code, caption)
+    preprocess.rs               – drop cap merging, heading line merging
+    postprocess.rs              – dot leaders, hyphenation, page numbers, URL formatting
+```
+
+## Key design decisions
+
+- **Primary audience is AI agents.** Output optimized for token efficiency and semantic quality, not visual formatting. No cosmetic padding.
+- **Three table detection strategies** run in priority order: rect-based → line-based → heuristic. First valid result wins.
+- **Column detection** uses horizontal projection histograms with valley detection. Multi-item spanning lines (titles, headers) are pre-masked using column-aware thresholds before column assignment.
+- **Newspaper vs tabular** classification determines reading order: newspaper reads columns sequentially, tabular Y-interleaves them.
+- **Tiled-scan detection** catches scanned PDFs with JBIG2/strip images where no single tile exceeds the template threshold but aggregate area does (≥2M pixels).
+- **Garbage text upgrade** reclassifies Mixed PDFs as Scanned when extracted text is <50% alphanumeric.
+- **Tagged PDF support** uses structure tree roles (H1-H6, P, L, Code, BlockQuote) when available, falling back to font-size heuristics.
+
+## Testing
+
+- **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.
+
+## Debugging
+
+```bash
+RUST_LOG=pdf_inspector::extractor::layout=debug cargo run --bin pdf2md -- file.pdf
+RUST_LOG=pdf_inspector::tables=debug cargo run --bin pdf2md -- file.pdf
+RUST_LOG=pdf_inspector::detector=debug cargo run --release --bin detect-pdf -- file.pdf
+```
+
+## Conventions
+
+- Clippy: use `is_some_and(...)` not `map_or(false, ...)`
+- lopdf quirk: `ParseError` is private — match by string for `InvalidFileHeader`
+- Column limit for tables: 25 (wide statistical tables)
+- `propagate_merged_cells` skipped for >10 columns (spanning rects = background fills)

CLAUDE.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 187+ snapshot PDFs. Run `cargo build --release` then `bench.py test` in that repo before committing.
+- **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

@@ -8,9 +8,16 @@ description = "Fast PDF inspection, classification, and text extraction with sma
 license = "MIT"
 repository = "https://github.com/firecrawl/pdf-inspector"
 
+[lib]
+name = "pdf_inspector"
+crate-type = ["lib", "cdylib"]
+
 [dependencies]
+# Python bindings
+pyo3 = { version = "0.25", features = ["extension-module"], optional = true }
+
 # PDF parsing
-lopdf = { git = "https://github.com/J-F-Liu/lopdf", rev = "052674053814a9f4897af94f0b8e46a545c9b329", features = ["rayon"] }
+lopdf = { git = "https://github.com/J-F-Liu/lopdf", rev = "7a05512d831415b1f2b1ce522391d6beab8a1284", features = ["rayon"] }
 
 # Error handling
 thiserror = "2.0"
@@ -35,6 +42,7 @@ tempfile = "3.3"
 
 [features]
 default = []
+python = ["pyo3"]
 
 [[bin]]
 name = "pdf2md"
@@ -47,5 +55,3 @@ path = "src/bin/detect_pdf.rs"
 [[bin]]
 name = "dump_ops"
 path = "src/bin/dump_ops.rs"
-
-

README.md

@@ -1,6 +1,6 @@
 # pdf-inspector
 
-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.
+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).
 
 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.
 
@@ -12,90 +12,81 @@ Built by [Firecrawl](https://firecrawl.dev) to handle text-based PDFs locally in
 - **Table detection** — Dual-mode: rectangle-based detection from PDF drawing ops, plus heuristic detection from text alignment. Handles financial tables, footnotes, and continuation tables across pages.
 - **CID font support** — ToUnicode CMap decoding for Type0/Identity-H fonts, UTF-16BE, UTF-8, and Latin-1 encodings.
 - **Multi-column layout** — Automatic detection of newspaper-style columns, sequential reading order, and RTL text support.
-- **Encoding issue detection** — Automatically flags broken font encodings (garbled text, replacement characters) so callers can fall back to OCR.
+- **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.
 - **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.
+
+| Engine | Overall | Reading Order (NID) | Tables (TEDS) | Headings (MHS) | Speed (200 docs) |
+|---|---|---|---|---|---|
+| pdf-inspector | 0.78 | 0.87 | 0.59 | 0.57 | 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 |
+
+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.
+
+**Where we do well:** Speed (fastest of all engines), reading order, table detection vs other direct-text tools.
+
+**Where we lag:** Heading detection trails opendataloader — many PDFs use bold text at body font size for headings, or headings that are only slightly larger than body text. Table detection trails OCR-based engines that can see visual table structure.
+
 ## Quick start
 
-### As a library
+### Python
 
-Add to your `Cargo.toml`:
+```bash
+pip install maturin
+maturin develop --release
+```
+
+```python
+import pdf_inspector
+
+result = pdf_inspector.process_pdf("document.pdf")
+print(result.pdf_type)   # "text_based", "scanned", "image_based", "mixed"
+print(result.markdown)   # Markdown string or None
+```
+
+> Full API reference: [docs/python.md](docs/python.md)
+
+### Node.js
+
+```bash
+npm install @firecrawl/pdf-inspector
+```
+
+```javascript
+import { readFileSync } from 'fs';
+import { processPdf, classifyPdf } from '@firecrawl/pdf-inspector';
+
+const result = processPdf(readFileSync('document.pdf'));
+console.log(result.pdfType);   // "TextBased", "Scanned", "ImageBased", "Mixed"
+console.log(result.markdown);  // Markdown string or null
+```
+
+> Full API reference: [napi/README.md](napi/README.md)
+
+### Rust
 
 ```toml
 [dependencies]
 pdf-inspector = { git = "https://github.com/firecrawl/pdf-inspector" }
 ```
 
-Detect and extract in one call:
-
 ```rust
 use pdf_inspector::process_pdf;
 
 let result = process_pdf("document.pdf")?;
-
-println!("Type: {:?}", result.pdf_type);       // TextBased, Scanned, ImageBased, Mixed
-println!("Confidence: {:.0}%", result.confidence * 100.0);
-println!("Pages: {}", result.page_count);
-
+println!("Type: {:?}", result.pdf_type);
 if let Some(markdown) = &result.markdown {
     println!("{}", markdown);
 }
 ```
 
-Fast metadata-only detection (no text extraction or markdown generation):
-
-```rust
-use pdf_inspector::detect_pdf;
-
-let info = detect_pdf("document.pdf")?;
-
-match info.pdf_type {
-    pdf_inspector::PdfType::TextBased => {
-        // Extract locally — fast and free
-    }
-    _ => {
-        // Route to OCR service
-        // info.pages_needing_ocr tells you exactly which pages
-    }
-}
-```
-
-Customize processing with `PdfOptions`:
-
-```rust
-use pdf_inspector::{process_pdf_with_options, PdfOptions, ProcessMode, DetectionConfig, ScanStrategy};
-
-// Analyze layout without generating markdown
-let result = process_pdf_with_options(
-    "document.pdf",
-    PdfOptions::new().mode(ProcessMode::Analyze),
-)?;
-
-// Full extraction with custom detection strategy
-let result = process_pdf_with_options(
-    "large.pdf",
-    PdfOptions::new().detection(DetectionConfig {
-        strategy: ScanStrategy::Sample(5),
-        ..Default::default()
-    }),
-)?;
-
-// Process only specific pages
-let result = process_pdf_with_options(
-    "document.pdf",
-    PdfOptions::new().pages([1, 3, 5]),
-)?;
-```
-
-Process from a byte buffer (no filesystem needed):
-
-```rust
-use pdf_inspector::process_pdf_mem;
-
-let bytes = std::fs::read("document.pdf")?;
-let result = process_pdf_mem(&bytes)?;
-```
+> Full API reference: [docs/rust-api.md](docs/rust-api.md)
 
 ### CLI
 
@@ -120,7 +111,6 @@ cargo run --bin detect-pdf -- document.pdf
 cargo run --bin detect-pdf -- document.pdf --json
 
 # Detection + layout analysis (tables, columns)
-cargo run --bin detect-pdf -- document.pdf --analyze
 cargo run --bin detect-pdf -- document.pdf --analyze --json
 ```
 
@@ -159,6 +149,7 @@ The document is loaded **once** via `load_document_from_path` / `load_document_f
 ```
 src/
   lib.rs                — Public API, PdfOptions builder, convenience functions
+  python.rs             — PyO3 Python bindings
   types.rs              — Shared types: TextItem, TextLine, PdfRect, ItemType
   text_utils.rs         — Character/text helpers (CJK, RTL, ligatures, bold/italic)
   process_mode.rs       — ProcessMode enum (DetectOnly, Analyze, Full)
@@ -169,6 +160,7 @@ src/
   tables/               — Table detection and formatting
   markdown/             — Markdown conversion and structure detection
   bin/                  — CLI tools (pdf2md, detect_pdf)
+napi/                   — Node.js/Bun bindings (napi-rs)
 ```
 
 ## How classification works
@@ -189,50 +181,6 @@ This detects 300+ page PDFs in milliseconds. The result includes `pages_needing_
 | `Sample(n)` | Sample `n` evenly distributed pages (first, last, middle) | Very large PDFs where speed matters more than precision |
 | `Pages(vec)` | Only scan specific 1-indexed page numbers | When the caller knows which pages to check |
 
-## API
-
-### Processing modes
-
-| Mode | What it does | Returns |
-|---|---|---|
-| `ProcessMode::Full` (default) | Detect + extract + convert to Markdown | Everything populated |
-| `ProcessMode::Analyze` | Detect + extract + layout analysis (no Markdown) | `markdown` is `None`, `layout` is populated |
-| `ProcessMode::DetectOnly` | Classification only (fastest) | `markdown` is `None`, `layout` is default |
-
-### Functions
-
-| Function | Description |
-|---|---|
-| `process_pdf(path)` | Full processing with defaults |
-| `detect_pdf(path)` | Fast metadata-only detection (no extraction) |
-| `process_pdf_with_options(path, options)` | Process with custom `PdfOptions` |
-| `process_pdf_mem(bytes)` | Full processing from a byte buffer |
-| `detect_pdf_mem(bytes)` | Fast detection from a byte buffer |
-| `process_pdf_mem_with_options(bytes, options)` | Process from bytes with custom options |
-| `extract_text(path)` | Plain text extraction |
-| `extract_text_with_positions(path)` | Text with X/Y coordinates and font info |
-| `to_markdown(text, options)` | Convert plain text to Markdown |
-| `to_markdown_from_items(items, options)` | Markdown from pre-extracted `TextItem`s |
-| `to_markdown_from_items_with_rects(items, options, rects)` | Markdown with rectangle-based table detection |
-
-Low-level detection functions are also available via the `detector` module (`detect_pdf_type`, `detect_pdf_type_with_config`, etc.) for callers who need `PdfTypeResult` instead of `PdfProcessResult`.
-
-### Types
-
-| Type | Description |
-|---|---|
-| `PdfOptions` | Builder for processing configuration (mode, detection, markdown, page filter) |
-| `ProcessMode` | `DetectOnly`, `Analyze`, `Full` |
-| `PdfType` | `TextBased`, `Scanned`, `ImageBased`, `Mixed` |
-| `PdfProcessResult` | Full result: pdf_type, markdown, page_count, confidence, layout, has_encoding_issues, timing |
-| `PdfTypeResult` | Low-level detection result: type, confidence, page count, pages needing OCR |
-| `DetectionConfig` | Configuration for detection: scan strategy, thresholds |
-| `ScanStrategy` | `EarlyExit`, `Full`, `Sample(n)`, `Pages(vec)` |
-| `LayoutComplexity` | Layout analysis: is_complex, pages_with_tables, pages_with_columns |
-| `TextItem` | Text with position, font info, and page number |
-| `MarkdownOptions` | Configuration for Markdown formatting (page numbers, etc.) |
-| `PdfError` | `Io`, `Parse`, `Encrypted`, `InvalidStructure`, `NotAPdf` |
-
 ## Markdown output
 
 The converter handles:
@@ -255,36 +203,6 @@ The converter handles:
 | Drop caps | Large initial letters merged with following text |
 | Dot leaders | TOC-style dots collapsed to " ... " |
 
-## Debugging with RUST_LOG
-
-Structured logging via `RUST_LOG` replaces the former debug binaries. Set the environment variable to control which sections emit debug output on stderr:
-
-```bash
-# Raw PDF content stream operators (replaces dump_ops)
-RUST_LOG=pdf_inspector::extractor::content_stream=trace cargo run --bin pdf2md -- file.pdf > /dev/null
-
-# Font metadata, encodings, ligatures (replaces debug_fonts / debug_ligatures)
-RUST_LOG=pdf_inspector::extractor::fonts=debug cargo run --bin pdf2md -- file.pdf > /dev/null
-
-# ToUnicode CMap parsing
-RUST_LOG=pdf_inspector::tounicode=debug cargo run --bin pdf2md -- file.pdf > /dev/null
-
-# Text items per page with x/y/width (replaces debug_spaces / debug_pages)
-RUST_LOG=pdf_inspector::extractor=debug cargo run --bin pdf2md -- file.pdf > /dev/null
-
-# Column detection and reading order (replaces debug_order)
-RUST_LOG=pdf_inspector::extractor::layout=debug cargo run --bin pdf2md -- file.pdf > /dev/null
-
-# Y-gap analysis and paragraph thresholds (replaces debug_ygaps)
-RUST_LOG=pdf_inspector::markdown::analysis=debug cargo run --bin pdf2md -- file.pdf > /dev/null
-
-# Table detection
-RUST_LOG=pdf_inspector::tables=debug cargo run --bin pdf2md -- file.pdf > /dev/null
-
-# Everything
-RUST_LOG=pdf_inspector=debug cargo run --bin pdf2md -- file.pdf > /dev/null
-```
-
 ## Use case: smart PDF routing
 
 pdf-inspector was built for pipelines that process PDFs at scale. Instead of sending every PDF through OCR:
@@ -299,6 +217,10 @@ PDF arrives
 
 This saves cost and latency for the majority of PDFs that are already text-based (reports, papers, invoices, legal docs).
 
+## Debugging
+
+See [docs/debugging.md](docs/debugging.md) for `RUST_LOG` environment variable usage.
+
 ## License
 
 MIT

docs/debugging.md

@@ -0,0 +1,29 @@
+# Debugging with RUST_LOG
+
+Structured logging via `RUST_LOG` replaces the former debug binaries. Set the environment variable to control which sections emit debug output on stderr:
+
+```bash
+# Raw PDF content stream operators (replaces dump_ops)
+RUST_LOG=pdf_inspector::extractor::content_stream=trace cargo run --bin pdf2md -- file.pdf > /dev/null
+
+# Font metadata, encodings, ligatures (replaces debug_fonts / debug_ligatures)
+RUST_LOG=pdf_inspector::extractor::fonts=debug cargo run --bin pdf2md -- file.pdf > /dev/null
+
+# ToUnicode CMap parsing
+RUST_LOG=pdf_inspector::tounicode=debug cargo run --bin pdf2md -- file.pdf > /dev/null
+
+# Text items per page with x/y/width (replaces debug_spaces / debug_pages)
+RUST_LOG=pdf_inspector::extractor=debug cargo run --bin pdf2md -- file.pdf > /dev/null
+
+# Column detection and reading order (replaces debug_order)
+RUST_LOG=pdf_inspector::extractor::layout=debug cargo run --bin pdf2md -- file.pdf > /dev/null
+
+# Y-gap analysis and paragraph thresholds (replaces debug_ygaps)
+RUST_LOG=pdf_inspector::markdown::analysis=debug cargo run --bin pdf2md -- file.pdf > /dev/null
+
+# Table detection
+RUST_LOG=pdf_inspector::tables=debug cargo run --bin pdf2md -- file.pdf > /dev/null
+
+# Everything
+RUST_LOG=pdf_inspector=debug cargo run --bin pdf2md -- file.pdf > /dev/null
+```

docs/python.md

@@ -0,0 +1,88 @@
+# Python API
+
+Python bindings via [PyO3](https://pyo3.rs). Requires Rust toolchain for building from source.
+
+## Install
+
+```bash
+pip install maturin
+maturin develop --release
+```
+
+## Usage
+
+```python
+import pdf_inspector
+
+# Full processing: detect + extract + convert to Markdown
+result = pdf_inspector.process_pdf("document.pdf")
+print(result.pdf_type)      # "text_based", "scanned", "image_based", "mixed"
+print(result.confidence)     # 0.0 - 1.0
+print(result.page_count)     # number of pages
+print(result.markdown)       # Markdown string or None
+
+# Process specific pages only
+result = pdf_inspector.process_pdf("document.pdf", pages=[1, 3, 5])
+
+# Process from bytes (no filesystem needed)
+with open("document.pdf", "rb") as f:
+    result = pdf_inspector.process_pdf_bytes(f.read())
+
+# Fast detection only (no text extraction)
+result = pdf_inspector.detect_pdf("document.pdf")
+if result.pdf_type == "text_based":
+    print("Can extract locally!")
+else:
+    print(f"Pages needing OCR: {result.pages_needing_ocr}")
+
+# Plain text extraction
+text = pdf_inspector.extract_text("document.pdf")
+
+# Positioned text items with font info
+items = pdf_inspector.extract_text_with_positions("document.pdf")
+for item in items[:5]:
+    print(f"'{item.text}' at ({item.x:.0f}, {item.y:.0f}) size={item.font_size}")
+
+# Per-page markdown (one Markdown string per page, plus layout metadata)
+result = pdf_inspector.extract_pages_markdown("document.pdf")
+for page in result.pages:
+    print(f"Page {page.page}: {len(page.markdown)} chars, needs_ocr={page.needs_ocr}")
+
+# Restrict to specific 0-indexed pages (preserves caller order)
+result = pdf_inspector.extract_pages_markdown("document.pdf", pages=[0, 2])
+```
+
+## API reference
+
+| Function | Description |
+|---|---|
+| `process_pdf(path, pages=None)` | Full processing (detect + extract + markdown) |
+| `process_pdf_bytes(data, pages=None)` | Full processing from bytes |
+| `detect_pdf(path)` | Fast detection only (returns PdfResult) |
+| `detect_pdf_bytes(data)` | Fast detection from bytes |
+| `classify_pdf(path)` | Lightweight classification (returns PdfClassification) |
+| `classify_pdf_bytes(data)` | Lightweight classification from bytes |
+| `extract_text(path)` | Plain text extraction |
+| `extract_text_bytes(data)` | Plain text extraction from bytes |
+| `extract_text_with_positions(path, pages=None)` | Text with X/Y coords and font info |
+| `extract_text_with_positions_bytes(data, pages=None)` | Text with positions from bytes |
+| `extract_text_in_regions(path, page_regions)` | Extract text in bounding-box regions |
+| `extract_text_in_regions_bytes(data, page_regions)` | Region extraction from bytes |
+| `extract_pages_markdown(path, pages=None)` | Per-page Markdown + layout metadata (all pages by default) |
+| `extract_pages_markdown_bytes(data, pages=None)` | Per-page Markdown from bytes |
+
+## 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`
+
+**`PdfClassification` fields:** `pdf_type`, `page_count`, `pages_needing_ocr` (0-indexed), `confidence`
+
+**`TextItem` fields:** `text`, `x`, `y`, `width`, `height`, `font`, `font_size`, `page`, `is_bold`, `is_italic`, `item_type`
+
+**`RegionText` fields:** `text`, `needs_ocr`
+
+**`PageRegionTexts` fields:** `page` (0-indexed), `regions` (list of RegionText)
+
+**`PageMarkdown` fields:** `page` (0-indexed), `markdown`, `needs_ocr`
+
+**`PagesExtractionResult` fields:** `pages` (list of PageMarkdown), `pages_with_tables` (1-indexed), `pages_with_columns` (1-indexed), `pages_needing_ocr` (1-indexed), `is_complex`

docs/rust-api.md

@@ -0,0 +1,147 @@
+# Rust API
+
+Add to your `Cargo.toml`:
+
+```toml
+[dependencies]
+pdf-inspector = { git = "https://github.com/firecrawl/pdf-inspector" }
+```
+
+## Usage
+
+Detect and extract in one call:
+
+```rust
+use pdf_inspector::process_pdf;
+
+let result = process_pdf("document.pdf")?;
+
+println!("Type: {:?}", result.pdf_type);       // TextBased, Scanned, ImageBased, Mixed
+println!("Confidence: {:.0}%", result.confidence * 100.0);
+println!("Pages: {}", result.page_count);
+
+if let Some(markdown) = &result.markdown {
+    println!("{}", markdown);
+}
+```
+
+Fast metadata-only detection (no text extraction or markdown generation):
+
+```rust
+use pdf_inspector::detect_pdf;
+
+let info = detect_pdf("document.pdf")?;
+
+match info.pdf_type {
+    pdf_inspector::PdfType::TextBased => {
+        // Extract locally — fast and free
+    }
+    _ => {
+        // Route to OCR service
+        // info.pages_needing_ocr tells you exactly which pages
+    }
+}
+```
+
+Customize processing with `PdfOptions`:
+
+```rust
+use pdf_inspector::{process_pdf_with_options, PdfOptions, ProcessMode, DetectionConfig, ScanStrategy};
+
+// Analyze layout without generating markdown
+let result = process_pdf_with_options(
+    "document.pdf",
+    PdfOptions::new().mode(ProcessMode::Analyze),
+)?;
+
+// Full extraction with custom detection strategy
+let result = process_pdf_with_options(
+    "large.pdf",
+    PdfOptions::new().detection(DetectionConfig {
+        strategy: ScanStrategy::Sample(5),
+        ..Default::default()
+    }),
+)?;
+
+// Process only specific pages
+let result = process_pdf_with_options(
+    "document.pdf",
+    PdfOptions::new().pages([1, 3, 5]),
+)?;
+```
+
+Process from a byte buffer (no filesystem needed):
+
+```rust
+use pdf_inspector::process_pdf_mem;
+
+let bytes = std::fs::read("document.pdf")?;
+let result = process_pdf_mem(&bytes)?;
+```
+
+Extract per-page Markdown (one string per page, plus document-wide layout
+metadata):
+
+```rust
+use pdf_inspector::extract_pages_markdown;
+
+// Pass `None` for every page in document order, or a slice of 0-indexed
+// pages to restrict the output (caller-supplied order is preserved).
+let result = extract_pages_markdown("document.pdf", None)?;
+
+for page in &result.pages {
+    if page.needs_ocr {
+        // Route this page to OCR
+    } else {
+        println!("Page {}: {}", page.page, page.markdown);
+    }
+}
+
+println!("Complex layout? {}", result.is_complex);
+```
+
+## Processing modes
+
+| Mode | What it does | Returns |
+|---|---|---|
+| `ProcessMode::Full` (default) | Detect + extract + convert to Markdown | Everything populated |
+| `ProcessMode::Analyze` | Detect + extract + layout analysis (no Markdown) | `markdown` is `None`, `layout` is populated |
+| `ProcessMode::DetectOnly` | Classification only (fastest) | `markdown` is `None`, `layout` is default |
+
+## Functions
+
+| Function | Description |
+|---|---|
+| `process_pdf(path)` | Full processing with defaults |
+| `detect_pdf(path)` | Fast metadata-only detection (no extraction) |
+| `process_pdf_with_options(path, options)` | Process with custom `PdfOptions` |
+| `process_pdf_mem(bytes)` | Full processing from a byte buffer |
+| `detect_pdf_mem(bytes)` | Fast detection from a byte buffer |
+| `process_pdf_mem_with_options(bytes, options)` | Process from bytes with custom options |
+| `extract_text(path)` | Plain text extraction |
+| `extract_text_with_positions(path)` | Text with X/Y coordinates and font info |
+| `to_markdown(text, options)` | Convert plain text to Markdown |
+| `to_markdown_from_items(items, options)` | Markdown from pre-extracted `TextItem`s |
+| `to_markdown_from_items_with_rects(items, options, rects)` | Markdown with rectangle-based table detection |
+| `extract_pages_markdown(path, pages)` | Per-page Markdown + layout metadata (file) |
+| `extract_pages_markdown_mem(bytes, pages)` | Per-page Markdown from bytes |
+
+Low-level detection functions are also available via the `detector` module (`detect_pdf_type`, `detect_pdf_type_with_config`, etc.) for callers who need `PdfTypeResult` instead of `PdfProcessResult`.
+
+## Types
+
+| Type | Description |
+|---|---|
+| `PdfOptions` | Builder for processing configuration (mode, detection, markdown, page filter) |
+| `ProcessMode` | `DetectOnly`, `Analyze`, `Full` |
+| `PdfType` | `TextBased`, `Scanned`, `ImageBased`, `Mixed` |
+| `PdfProcessResult` | Full result: pdf_type, markdown, page_count, confidence, layout, has_encoding_issues, timing |
+| `PdfTypeResult` | Low-level detection result: type, confidence, page count, pages needing OCR |
+| `DetectionConfig` | Configuration for detection: scan strategy, thresholds |
+| `ScanStrategy` | `EarlyExit`, `Full`, `Sample(n)`, `Pages(vec)` |
+| `LayoutComplexity` | Layout analysis: is_complex, pages_with_tables, pages_with_columns |
+| `TextItem` | Text with position, font info, and page number |
+| `MarkdownOptions` | Configuration for Markdown formatting (page numbers, etc.) |
+| `PageMarkdown` | Per-page result: page (0-indexed), markdown, needs_ocr |
+| `PagesExtractionResult` | Per-page output + 1-indexed pages_with_tables / pages_with_columns / pages_needing_ocr, is_complex |
+| `PdfError` | `Io`, `Parse`, `Encrypted`, `InvalidStructure`, `NotAPdf` |

examples/basic_usage.py

@@ -0,0 +1,98 @@
+"""Basic usage examples for pdf-inspector Python library."""
+
+import sys
+import pdf_inspector
+
+
+def main():
+    if len(sys.argv) < 2:
+        print("Usage: python basic_usage.py <path-to-pdf>")
+        sys.exit(1)
+
+    path = sys.argv[1]
+
+    # 1. Full processing: detect + extract + markdown
+    print("=" * 60)
+    print("Full processing")
+    print("=" * 60)
+    result = pdf_inspector.process_pdf(path)
+    print(f"Type:       {result.pdf_type}")
+    print(f"Pages:      {result.page_count}")
+    print(f"Confidence: {result.confidence:.0%}")
+    print(f"Time:       {result.processing_time_ms}ms")
+    print(f"Title:      {result.title}")
+    print(f"Complex:    {result.is_complex_layout}")
+    print(f"Tables on:  {result.pages_with_tables}")
+    print(f"Columns on: {result.pages_with_columns}")
+    print(f"Encoding:   {'issues detected' if result.has_encoding_issues else 'ok'}")
+    print(f"OCR needed: {result.pages_needing_ocr or 'none'}")
+    if result.markdown:
+        print(f"\n--- Markdown ({len(result.markdown)} chars) ---")
+        print(result.markdown[:500])
+        if len(result.markdown) > 500:
+            print(f"\n... ({len(result.markdown) - 500} more chars)")
+
+    # 2. Fast detection only
+    print("\n" + "=" * 60)
+    print("Detection only")
+    print("=" * 60)
+    info = pdf_inspector.detect_pdf(path)
+    print(f"Type:       {info.pdf_type}")
+    print(f"Confidence: {info.confidence:.0%}")
+    print(f"Time:       {info.processing_time_ms}ms")
+
+    # 3. From bytes
+    print("\n" + "=" * 60)
+    print("From bytes")
+    print("=" * 60)
+    with open(path, "rb") as f:
+        data = f.read()
+    result = pdf_inspector.process_pdf_bytes(data)
+    print(f"Type: {result.pdf_type}, Pages: {result.page_count}")
+
+    # 4. Plain text
+    print("\n" + "=" * 60)
+    print("Plain text extraction")
+    print("=" * 60)
+    text = pdf_inspector.extract_text(path)
+    print(text[:300])
+
+    # 5. Positioned items
+    print("\n" + "=" * 60)
+    print("Positioned text items (first 10)")
+    print("=" * 60)
+    items = pdf_inspector.extract_text_with_positions(path, pages=[1])
+    for item in items[:10]:
+        bold = " [B]" if item.is_bold else ""
+        italic = " [I]" if item.is_italic else ""
+        print(
+            f"  p{item.page} ({item.x:6.1f}, {item.y:6.1f}) "
+            f"size={item.font_size:5.1f}{bold}{italic} "
+            f"'{item.text}'"
+        )
+
+    # 6. Lightweight classification
+    print("\n" + "=" * 60)
+    print("Lightweight classification")
+    print("=" * 60)
+    cls = pdf_inspector.classify_pdf(path)
+    print(f"Type:       {cls.pdf_type}")
+    print(f"Pages:      {cls.page_count}")
+    print(f"Confidence: {cls.confidence:.0%}")
+    print(f"OCR pages:  {cls.pages_needing_ocr or 'none'} (0-indexed)")
+
+    # 7. Region-based text extraction
+    print("\n" + "=" * 60)
+    print("Region-based text extraction (page 0, top region)")
+    print("=" * 60)
+    regions = pdf_inspector.extract_text_in_regions(
+        path, [(0, [[0.0, 0.0, 600.0, 200.0]])]
+    )
+    for page_result in regions:
+        for i, region in enumerate(page_result.regions):
+            print(f"  Region {i}: needs_ocr={region.needs_ocr}")
+            print(f"  Text: {region.text[:200]}")
+
+
+if __name__ == "__main__":
+    main()

napi/Cargo.lock

@@ -129,12 +129,6 @@ version = "3.20.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "5d20789868f4b01b2f2caec9f5c4e0213b41e3e5702a50157d699ae31ced2fcb"
 
-[[package]]
-name = "bytecount"
-version = "0.6.9"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "175812e0be2bccb6abe50bb8d566126198344f707e304f45c648fd8f2cc0365e"
-
 [[package]]
 name = "cbc"
 version = "0.1.2"
@@ -679,7 +673,7 @@ checksum = "5e5032e24019045c762d3c0f28f5b6b8bbf38563a65908389bf7978758920897"
 [[package]]
 name = "lopdf"
 version = "0.40.0"
-source = "git+https://github.com/J-F-Liu/lopdf?rev=052674053814a9f4897af94f0b8e46a545c9b329#052674053814a9f4897af94f0b8e46a545c9b329"
+source = "git+https://github.com/J-F-Liu/lopdf?rev=7a05512d831415b1f2b1ce522391d6beab8a1284#7a05512d831415b1f2b1ce522391d6beab8a1284"
 dependencies = [
  "aes",
  "bitflags",
@@ -695,7 +689,6 @@ dependencies = [
  "log",
  "md-5",
  "nom",
- "nom_locate",
  "rand",
  "rangemap",
  "rayon",
@@ -807,17 +800,6 @@ dependencies = [
  "memchr",
 ]
 
-[[package]]
-name = "nom_locate"
-version = "5.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0b577e2d69827c4740cba2b52efaad1c4cc7c73042860b199710b3575c68438d"
-dependencies = [
- "bytecount",
- "memchr",
- "nom",
-]
-
 [[package]]
 name = "num-conv"
 version = "0.2.1"

napi/README.md

@@ -0,0 +1,99 @@
+# PDF Inspector
+
+Fast PDF classification and region-based text extraction for Node.js/Bun. Native Rust performance via [napi-rs](https://napi.rs).
+
+Built by [Firecrawl](https://firecrawl.dev) for hybrid OCR pipelines — extract text from PDF structure where possible, fall back to OCR only when needed.
+
+## Install
+
+```bash
+npm install @firecrawl/pdf-inspector
+# or
+bun add @firecrawl/pdf-inspector
+```
+
+Prebuilt binaries included for **linux-x64** and **macOS ARM64**. No Rust toolchain needed.
+
+## API
+
+### `classifyPdf(buffer: Buffer): PdfClassification`
+
+Classify a PDF as TextBased, Scanned, Mixed, or ImageBased (~10-50ms). Returns which pages need OCR.
+
+```typescript
+import { classifyPdf } from '@firecrawl/pdf-inspector'
+import { readFileSync } from 'fs'
+
+const pdf = readFileSync('document.pdf')
+const result = classifyPdf(pdf)
+
+console.log(result.pdfType)        // "TextBased" | "Scanned" | "Mixed" | "ImageBased"
+console.log(result.pageCount)      // 42
+console.log(result.pagesNeedingOcr) // [5, 12, 15] (0-indexed)
+console.log(result.confidence)     // 0.875
+```
+
+### `extractTextInRegions(buffer: Buffer, pageRegions: PageRegions[]): PageRegionTexts[]`
+
+Extract text within bounding-box regions from a PDF. Designed for hybrid OCR pipelines where a layout model detects regions in rendered page images, and this function extracts text from the PDF structure for text-based pages — skipping GPU OCR.
+
+Each region result includes a `needsOcr` flag that signals unreliable extraction (empty text, GID-encoded fonts, garbage text, encoding issues).
+
+```typescript
+import { extractTextInRegions } from '@firecrawl/pdf-inspector'
+
+const result = extractTextInRegions(pdf, [
+  {
+    page: 0, // 0-indexed
+    regions: [
+      [0, 0, 300, 400],    // [x1, y1, x2, y2] in PDF points, top-left origin
+      [300, 0, 612, 400],
+    ]
+  }
+])
+
+for (const region of result[0].regions) {
+  if (region.needsOcr) {
+    // Unreliable text — send this region to OCR instead
+  } else {
+    console.log(region.text) // Extracted text in reading order
+  }
+}
+```
+
+## Types
+
+```typescript
+interface PdfClassification {
+  pdfType: string          // "TextBased" | "Scanned" | "Mixed" | "ImageBased"
+  pageCount: number
+  pagesNeedingOcr: number[] // 0-indexed page numbers
+  confidence: number        // 0.0 - 1.0
+}
+
+interface PageRegions {
+  page: number              // 0-indexed
+  regions: number[][]       // [[x1, y1, x2, y2], ...] in PDF points, top-left origin
+}
+
+interface PageRegionTexts {
+  page: number
+  regions: RegionText[]
+}
+
+interface RegionText {
+  text: string
+  needsOcr: boolean         // true when text is unreliable
+}
+```
+
+## Platforms
+
+| Platform | Architecture | Supported |
+|----------|-------------|-----------|
+| Linux    | x64         | Yes       |
+| macOS    | ARM64       | Yes       |
+
+## License
+
+MIT

napi/bin/pdf-inspector.mjs

@@ -0,0 +1,131 @@
+#!/usr/bin/env node
+
+import { readFileSync, writeFileSync } from "fs";
+import { createRequire } from "module";
+
+const require = createRequire(import.meta.url);
+const { version } = require("../package.json");
+
+const HELP = `pdf-inspector v${version} — Fast PDF text extraction to Markdown
+
+Usage:
+  pdf-inspector <file>                  Extract markdown (default)
+  pdf-inspector detect <file>           Classify PDF type
+
+Options:
+  --json                  Output as JSON
+  --pages <pages>         Comma-separated page numbers (e.g. 1,3,5)
+  -o, --output <file>     Write output to file instead of stdout
+  -h, --help              Show this help
+  -v, --version           Show version
+
+Examples:
+  pdf-inspector document.pdf
+  pdf-inspector document.pdf --json
+  pdf-inspector document.pdf --pages 1,2,3
+  pdf-inspector detect document.pdf --json
+  cat document.pdf | pdf-inspector -`;
+
+function die(msg) {
+  process.stderr.write(`error: ${msg}\n`);
+  process.exit(1);
+}
+
+function parseArgs(argv) {
+  const opts = { json: false, pages: null, output: null, file: null, command: "extract" };
+  let i = 0;
+
+  // Check for subcommand
+  if (argv[0] === "detect") {
+    opts.command = "detect";
+    i = 1;
+  }
+
+  while (i < argv.length) {
+    const arg = argv[i];
+    if (arg === "-h" || arg === "--help") {
+      process.stdout.write(HELP + "\n");
+      process.exit(0);
+    } else if (arg === "-v" || arg === "--version") {
+      process.stdout.write(`${version}\n`);
+      process.exit(0);
+    } else if (arg === "--json") {
+      opts.json = true;
+    } else if (arg === "--pages") {
+      i++;
+      if (!argv[i]) die("--pages requires a value (e.g. 1,3,5)");
+      opts.pages = argv[i].split(",").map((p) => {
+        const n = parseInt(p.trim(), 10);
+        if (Number.isNaN(n) || n < 1) die(`invalid page number: ${p}`);
+        return n;
+      });
+    } else if (arg === "-o" || arg === "--output") {
+      i++;
+      if (!argv[i]) die("-o requires a filename");
+      opts.output = argv[i];
+    } else if (arg === "-" || !arg.startsWith("-")) {
+      if (opts.file) die(`unexpected argument: ${arg}`);
+      opts.file = arg;
+    } else {
+      die(`unknown option: ${arg}`);
+    }
+    i++;
+  }
+
+  return opts;
+}
+
+function readInput(file) {
+  if (file === "-") {
+    return readFileSync(0); // stdin fd
+  }
+  try {
+    return readFileSync(file);
+  } catch (err) {
+    if (err.code === "ENOENT") die(`file not found: ${file}`);
+    die(err.message);
+  }
+}
+
+function output(text, outputPath) {
+  if (outputPath) {
+    writeFileSync(outputPath, text);
+  } else {
+    process.stdout.write(text);
+  }
+}
+
+// ---- main ----
+
+const opts = parseArgs(process.argv.slice(2));
+
+if (!opts.file) {
+  // Check if stdin is piped
+  if (process.stdin.isTTY !== false) {
+    process.stderr.write(HELP + "\n");
+    process.exit(1);
+  }
+  opts.file = "-";
+}
+
+const { processPdf, classifyPdf } = await import("../index.js");
+const buffer = readInput(opts.file);
+
+if (opts.command === "detect") {
+  const result = classifyPdf(buffer);
+  if (opts.json) {
+    output(JSON.stringify(result, null, 2) + "\n", opts.output);
+  } else {
+    const ocr = result.pagesNeedingOcr.length > 0
+      ? `, ${result.pagesNeedingOcr.length} pages need OCR`
+      : "";
+    output(`${result.pdfType} (${result.pageCount} pages, confidence: ${result.confidence.toFixed(2)}${ocr})\n`, opts.output);
+  }
+} else {
+  const result = processPdf(buffer, opts.pages ?? undefined);
+  if (opts.json) {
+    output(JSON.stringify(result, null, 2) + "\n", opts.output);
+  } else {
+    output((result.markdown ?? "") + "\n", opts.output);
+  }
+}

napi/build.rs

@@ -1,3 +1,3 @@
 fn main() {
-  napi_build::setup();
+    napi_build::setup();
 }

napi/index.d.ts

@@ -1,49 +0,0 @@
-/* auto-generated by NAPI-RS */
-/* eslint-disable */
-/**
- * Classify a PDF: detect type (TextBased/Scanned/Mixed/ImageBased),
- * page count, and which pages need OCR. Takes PDF bytes as Buffer.
- */
-export declare function classifyPdf(buffer: Buffer): PdfClassification
-
-/**
- * Extract text within bounding-box regions from a PDF.
- *
- * For hybrid OCR: layout model detects regions in rendered images,
- * this extracts PDF text within those regions — skipping GPU OCR
- * for text-based pages.
- *
- * Each region result includes `needs_ocr` — set when the extracted text
- * is unreliable (empty, GID-encoded fonts, garbage, encoding issues).
- *
- * Coordinates are PDF points with top-left origin.
- */
-export declare function extractTextInRegions(buffer: Buffer, pageRegions: Array<PageRegions>): Array<PageRegionTexts>
-
-/** A page's regions for text extraction: (page_index_0based, bboxes). */
-export interface PageRegions {
-  page: number
-  /** Each bbox is [x1, y1, x2, y2] in PDF points, top-left origin. */
-  regions: Array<Array<number>>
-}
-
-/** Extracted text for one page's regions. */
-export interface PageRegionTexts {
-  page: number
-  regions: Array<RegionText>
-}
-
-/** Lightweight PDF classification result. */
-export interface PdfClassification {
-  pdfType: string
-  pageCount: number
-  pagesNeedingOcr: Array<number>
-  confidence: number
-}
-
-/** Extracted text for a single region. */
-export interface RegionText {
-  text: string
-  /** `true` when the text should not be trusted (empty, GID fonts, garbage, encoding issues). */
-  needsOcr: boolean
-}

napi/index.js

@@ -1,580 +0,0 @@
-// prettier-ignore
-/* eslint-disable */
-// @ts-nocheck
-/* auto-generated by NAPI-RS */
-
-const { readFileSync } = require('node:fs')
-let nativeBinding = null
-const loadErrors = []
-
-const isMusl = () => {
-  let musl = false
-  if (process.platform === 'linux') {
-    musl = isMuslFromFilesystem()
-    if (musl === null) {
-      musl = isMuslFromReport()
-    }
-    if (musl === null) {
-      musl = isMuslFromChildProcess()
-    }
-  }
-  return musl
-}
-
-const isFileMusl = (f) => f.includes('libc.musl-') || f.includes('ld-musl-')
-
-const isMuslFromFilesystem = () => {
-  try {
-    return readFileSync('/usr/bin/ldd', 'utf-8').includes('musl')
-  } catch {
-    return null
-  }
-}
-
-const isMuslFromReport = () => {
-  let report = null
-  if (typeof process.report?.getReport === 'function') {
-    process.report.excludeNetwork = true
-    report = process.report.getReport()
-  }
-  if (!report) {
-    return null
-  }
-  if (report.header && report.header.glibcVersionRuntime) {
-    return false
-  }
-  if (Array.isArray(report.sharedObjects)) {
-    if (report.sharedObjects.some(isFileMusl)) {
-      return true
-    }
-  }
-  return false
-}
-
-const isMuslFromChildProcess = () => {
-  try {
-    return require('child_process').execSync('ldd --version', { encoding: 'utf8' }).includes('musl')
-  } catch (e) {
-    // If we reach this case, we don't know if the system is musl or not, so is better to just fallback to false
-    return false
-  }
-}
-
-function requireNative() {
-  if (process.env.NAPI_RS_NATIVE_LIBRARY_PATH) {
-    try {
-      return require(process.env.NAPI_RS_NATIVE_LIBRARY_PATH);
-    } catch (err) {
-      loadErrors.push(err)
-    }
-  } else if (process.platform === 'android') {
-    if (process.arch === 'arm64') {
-      try {
-        return require('./pdf-inspector.android-arm64.node')
-      } catch (e) {
-        loadErrors.push(e)
-      }
-      try {
-        const binding = require('firecrawl-pdf-inspector-android-arm64')
-        const bindingPackageVersion = require('firecrawl-pdf-inspector-android-arm64/package.json').version
-        if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-          throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-        }
-        return binding
-      } catch (e) {
-        loadErrors.push(e)
-      }
-    } else if (process.arch === 'arm') {
-      try {
-        return require('./pdf-inspector.android-arm-eabi.node')
-      } catch (e) {
-        loadErrors.push(e)
-      }
-      try {
-        const binding = require('firecrawl-pdf-inspector-android-arm-eabi')
-        const bindingPackageVersion = require('firecrawl-pdf-inspector-android-arm-eabi/package.json').version
-        if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-          throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-        }
-        return binding
-      } catch (e) {
-        loadErrors.push(e)
-      }
-    } else {
-      loadErrors.push(new Error(`Unsupported architecture on Android ${process.arch}`))
-    }
-  } else if (process.platform === 'win32') {
-    if (process.arch === 'x64') {
-      if (process.config?.variables?.shlib_suffix === 'dll.a' || process.config?.variables?.node_target_type === 'shared_library') {
-        try {
-        return require('./pdf-inspector.win32-x64-gnu.node')
-      } catch (e) {
-        loadErrors.push(e)
-      }
-      try {
-        const binding = require('firecrawl-pdf-inspector-win32-x64-gnu')
-        const bindingPackageVersion = require('firecrawl-pdf-inspector-win32-x64-gnu/package.json').version
-        if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-          throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-        }
-        return binding
-      } catch (e) {
-        loadErrors.push(e)
-      }
-      } else {
-        try {
-        return require('./pdf-inspector.win32-x64-msvc.node')
-      } catch (e) {
-        loadErrors.push(e)
-      }
-      try {
-        const binding = require('firecrawl-pdf-inspector-win32-x64-msvc')
-        const bindingPackageVersion = require('firecrawl-pdf-inspector-win32-x64-msvc/package.json').version
-        if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-          throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-        }
-        return binding
-      } catch (e) {
-        loadErrors.push(e)
-      }
-      }
-    } else if (process.arch === 'ia32') {
-      try {
-        return require('./pdf-inspector.win32-ia32-msvc.node')
-      } catch (e) {
-        loadErrors.push(e)
-      }
-      try {
-        const binding = require('firecrawl-pdf-inspector-win32-ia32-msvc')
-        const bindingPackageVersion = require('firecrawl-pdf-inspector-win32-ia32-msvc/package.json').version
-        if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-          throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-        }
-        return binding
-      } catch (e) {
-        loadErrors.push(e)
-      }
-    } else if (process.arch === 'arm64') {
-      try {
-        return require('./pdf-inspector.win32-arm64-msvc.node')
-      } catch (e) {
-        loadErrors.push(e)
-      }
-      try {
-        const binding = require('firecrawl-pdf-inspector-win32-arm64-msvc')
-        const bindingPackageVersion = require('firecrawl-pdf-inspector-win32-arm64-msvc/package.json').version
-        if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-          throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-        }
-        return binding
-      } catch (e) {
-        loadErrors.push(e)
-      }
-    } else {
-      loadErrors.push(new Error(`Unsupported architecture on Windows: ${process.arch}`))
-    }
-  } else if (process.platform === 'darwin') {
-    try {
-      return require('./pdf-inspector.darwin-universal.node')
-    } catch (e) {
-      loadErrors.push(e)
-    }
-    try {
-      const binding = require('firecrawl-pdf-inspector-darwin-universal')
-      const bindingPackageVersion = require('firecrawl-pdf-inspector-darwin-universal/package.json').version
-      if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-        throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-      }
-      return binding
-    } catch (e) {
-      loadErrors.push(e)
-    }
-    if (process.arch === 'x64') {
-      try {
-        return require('./pdf-inspector.darwin-x64.node')
-      } catch (e) {
-        loadErrors.push(e)
-      }
-      try {
-        const binding = require('firecrawl-pdf-inspector-darwin-x64')
-        const bindingPackageVersion = require('firecrawl-pdf-inspector-darwin-x64/package.json').version
-        if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-          throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-        }
-        return binding
-      } catch (e) {
-        loadErrors.push(e)
-      }
-    } else if (process.arch === 'arm64') {
-      try {
-        return require('./pdf-inspector.darwin-arm64.node')
-      } catch (e) {
-        loadErrors.push(e)
-      }
-      try {
-        const binding = require('firecrawl-pdf-inspector-darwin-arm64')
-        const bindingPackageVersion = require('firecrawl-pdf-inspector-darwin-arm64/package.json').version
-        if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-          throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-        }
-        return binding
-      } catch (e) {
-        loadErrors.push(e)
-      }
-    } else {
-      loadErrors.push(new Error(`Unsupported architecture on macOS: ${process.arch}`))
-    }
-  } else if (process.platform === 'freebsd') {
-    if (process.arch === 'x64') {
-      try {
-        return require('./pdf-inspector.freebsd-x64.node')
-      } catch (e) {
-        loadErrors.push(e)
-      }
-      try {
-        const binding = require('firecrawl-pdf-inspector-freebsd-x64')
-        const bindingPackageVersion = require('firecrawl-pdf-inspector-freebsd-x64/package.json').version
-        if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-          throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-        }
-        return binding
-      } catch (e) {
-        loadErrors.push(e)
-      }
-    } else if (process.arch === 'arm64') {
-      try {
-        return require('./pdf-inspector.freebsd-arm64.node')
-      } catch (e) {
-        loadErrors.push(e)
-      }
-      try {
-        const binding = require('firecrawl-pdf-inspector-freebsd-arm64')
-        const bindingPackageVersion = require('firecrawl-pdf-inspector-freebsd-arm64/package.json').version
-        if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-          throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-        }
-        return binding
-      } catch (e) {
-        loadErrors.push(e)
-      }
-    } else {
-      loadErrors.push(new Error(`Unsupported architecture on FreeBSD: ${process.arch}`))
-    }
-  } else if (process.platform === 'linux') {
-    if (process.arch === 'x64') {
-      if (isMusl()) {
-        try {
-          return require('./pdf-inspector.linux-x64-musl.node')
-        } catch (e) {
-          loadErrors.push(e)
-        }
-        try {
-          const binding = require('firecrawl-pdf-inspector-linux-x64-musl')
-          const bindingPackageVersion = require('firecrawl-pdf-inspector-linux-x64-musl/package.json').version
-          if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-            throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-          }
-          return binding
-        } catch (e) {
-          loadErrors.push(e)
-        }
-      } else {
-        try {
-          return require('./pdf-inspector.linux-x64-gnu.node')
-        } catch (e) {
-          loadErrors.push(e)
-        }
-        try {
-          const binding = require('firecrawl-pdf-inspector-linux-x64-gnu')
-          const bindingPackageVersion = require('firecrawl-pdf-inspector-linux-x64-gnu/package.json').version
-          if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-            throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-          }
-          return binding
-        } catch (e) {
-          loadErrors.push(e)
-        }
-      }
-    } else if (process.arch === 'arm64') {
-      if (isMusl()) {
-        try {
-          return require('./pdf-inspector.linux-arm64-musl.node')
-        } catch (e) {
-          loadErrors.push(e)
-        }
-        try {
-          const binding = require('firecrawl-pdf-inspector-linux-arm64-musl')
-          const bindingPackageVersion = require('firecrawl-pdf-inspector-linux-arm64-musl/package.json').version
-          if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-            throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-          }
-          return binding
-        } catch (e) {
-          loadErrors.push(e)
-        }
-      } else {
-        try {
-          return require('./pdf-inspector.linux-arm64-gnu.node')
-        } catch (e) {
-          loadErrors.push(e)
-        }
-        try {
-          const binding = require('firecrawl-pdf-inspector-linux-arm64-gnu')
-          const bindingPackageVersion = require('firecrawl-pdf-inspector-linux-arm64-gnu/package.json').version
-          if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-            throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-          }
-          return binding
-        } catch (e) {
-          loadErrors.push(e)
-        }
-      }
-    } else if (process.arch === 'arm') {
-      if (isMusl()) {
-        try {
-          return require('./pdf-inspector.linux-arm-musleabihf.node')
-        } catch (e) {
-          loadErrors.push(e)
-        }
-        try {
-          const binding = require('firecrawl-pdf-inspector-linux-arm-musleabihf')
-          const bindingPackageVersion = require('firecrawl-pdf-inspector-linux-arm-musleabihf/package.json').version
-          if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-            throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-          }
-          return binding
-        } catch (e) {
-          loadErrors.push(e)
-        }
-      } else {
-        try {
-          return require('./pdf-inspector.linux-arm-gnueabihf.node')
-        } catch (e) {
-          loadErrors.push(e)
-        }
-        try {
-          const binding = require('firecrawl-pdf-inspector-linux-arm-gnueabihf')
-          const bindingPackageVersion = require('firecrawl-pdf-inspector-linux-arm-gnueabihf/package.json').version
-          if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-            throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-          }
-          return binding
-        } catch (e) {
-          loadErrors.push(e)
-        }
-      }
-    } else if (process.arch === 'loong64') {
-      if (isMusl()) {
-        try {
-          return require('./pdf-inspector.linux-loong64-musl.node')
-        } catch (e) {
-          loadErrors.push(e)
-        }
-        try {
-          const binding = require('firecrawl-pdf-inspector-linux-loong64-musl')
-          const bindingPackageVersion = require('firecrawl-pdf-inspector-linux-loong64-musl/package.json').version
-          if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-            throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-          }
-          return binding
-        } catch (e) {
-          loadErrors.push(e)
-        }
-      } else {
-        try {
-          return require('./pdf-inspector.linux-loong64-gnu.node')
-        } catch (e) {
-          loadErrors.push(e)
-        }
-        try {
-          const binding = require('firecrawl-pdf-inspector-linux-loong64-gnu')
-          const bindingPackageVersion = require('firecrawl-pdf-inspector-linux-loong64-gnu/package.json').version
-          if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-            throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-          }
-          return binding
-        } catch (e) {
-          loadErrors.push(e)
-        }
-      }
-    } else if (process.arch === 'riscv64') {
-      if (isMusl()) {
-        try {
-          return require('./pdf-inspector.linux-riscv64-musl.node')
-        } catch (e) {
-          loadErrors.push(e)
-        }
-        try {
-          const binding = require('firecrawl-pdf-inspector-linux-riscv64-musl')
-          const bindingPackageVersion = require('firecrawl-pdf-inspector-linux-riscv64-musl/package.json').version
-          if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-            throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-          }
-          return binding
-        } catch (e) {
-          loadErrors.push(e)
-        }
-      } else {
-        try {
-          return require('./pdf-inspector.linux-riscv64-gnu.node')
-        } catch (e) {
-          loadErrors.push(e)
-        }
-        try {
-          const binding = require('firecrawl-pdf-inspector-linux-riscv64-gnu')
-          const bindingPackageVersion = require('firecrawl-pdf-inspector-linux-riscv64-gnu/package.json').version
-          if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-            throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-          }
-          return binding
-        } catch (e) {
-          loadErrors.push(e)
-        }
-      }
-    } else if (process.arch === 'ppc64') {
-      try {
-        return require('./pdf-inspector.linux-ppc64-gnu.node')
-      } catch (e) {
-        loadErrors.push(e)
-      }
-      try {
-        const binding = require('firecrawl-pdf-inspector-linux-ppc64-gnu')
-        const bindingPackageVersion = require('firecrawl-pdf-inspector-linux-ppc64-gnu/package.json').version
-        if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-          throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-        }
-        return binding
-      } catch (e) {
-        loadErrors.push(e)
-      }
-    } else if (process.arch === 's390x') {
-      try {
-        return require('./pdf-inspector.linux-s390x-gnu.node')
-      } catch (e) {
-        loadErrors.push(e)
-      }
-      try {
-        const binding = require('firecrawl-pdf-inspector-linux-s390x-gnu')
-        const bindingPackageVersion = require('firecrawl-pdf-inspector-linux-s390x-gnu/package.json').version
-        if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-          throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-        }
-        return binding
-      } catch (e) {
-        loadErrors.push(e)
-      }
-    } else {
-      loadErrors.push(new Error(`Unsupported architecture on Linux: ${process.arch}`))
-    }
-  } else if (process.platform === 'openharmony') {
-    if (process.arch === 'arm64') {
-      try {
-        return require('./pdf-inspector.openharmony-arm64.node')
-      } catch (e) {
-        loadErrors.push(e)
-      }
-      try {
-        const binding = require('firecrawl-pdf-inspector-openharmony-arm64')
-        const bindingPackageVersion = require('firecrawl-pdf-inspector-openharmony-arm64/package.json').version
-        if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-          throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-        }
-        return binding
-      } catch (e) {
-        loadErrors.push(e)
-      }
-    } else if (process.arch === 'x64') {
-      try {
-        return require('./pdf-inspector.openharmony-x64.node')
-      } catch (e) {
-        loadErrors.push(e)
-      }
-      try {
-        const binding = require('firecrawl-pdf-inspector-openharmony-x64')
-        const bindingPackageVersion = require('firecrawl-pdf-inspector-openharmony-x64/package.json').version
-        if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-          throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-        }
-        return binding
-      } catch (e) {
-        loadErrors.push(e)
-      }
-    } else if (process.arch === 'arm') {
-      try {
-        return require('./pdf-inspector.openharmony-arm.node')
-      } catch (e) {
-        loadErrors.push(e)
-      }
-      try {
-        const binding = require('firecrawl-pdf-inspector-openharmony-arm')
-        const bindingPackageVersion = require('firecrawl-pdf-inspector-openharmony-arm/package.json').version
-        if (bindingPackageVersion !== '0.2.0' && process.env.NAPI_RS_ENFORCE_VERSION_CHECK && process.env.NAPI_RS_ENFORCE_VERSION_CHECK !== '0') {
-          throw new Error(`Native binding package version mismatch, expected 0.2.0 but got ${bindingPackageVersion}. You can reinstall dependencies to fix this issue.`)
-        }
-        return binding
-      } catch (e) {
-        loadErrors.push(e)
-      }
-    } else {
-      loadErrors.push(new Error(`Unsupported architecture on OpenHarmony: ${process.arch}`))
-    }
-  } else {
-    loadErrors.push(new Error(`Unsupported OS: ${process.platform}, architecture: ${process.arch}`))
-  }
-}
-
-nativeBinding = requireNative()
-
-if (!nativeBinding || process.env.NAPI_RS_FORCE_WASI) {
-  let wasiBinding = null
-  let wasiBindingError = null
-  try {
-    wasiBinding = require('./pdf-inspector.wasi.cjs')
-    nativeBinding = wasiBinding
-  } catch (err) {
-    if (process.env.NAPI_RS_FORCE_WASI) {
-      wasiBindingError = err
-    }
-  }
-  if (!nativeBinding || process.env.NAPI_RS_FORCE_WASI) {
-    try {
-      wasiBinding = require('firecrawl-pdf-inspector-wasm32-wasi')
-      nativeBinding = wasiBinding
-    } catch (err) {
-      if (process.env.NAPI_RS_FORCE_WASI) {
-        if (!wasiBindingError) {
-          wasiBindingError = err
-        } else {
-          wasiBindingError.cause = err
-        }
-        loadErrors.push(err)
-      }
-    }
-  }
-  if (process.env.NAPI_RS_FORCE_WASI === 'error' && !wasiBinding) {
-    const error = new Error('WASI binding not found and NAPI_RS_FORCE_WASI is set to error')
-    error.cause = wasiBindingError
-    throw error
-  }
-}
-
-if (!nativeBinding) {
-  if (loadErrors.length > 0) {
-    throw new Error(
-      `Cannot find native binding. ` +
-        `npm has a bug related to optional dependencies (https://github.com/npm/cli/issues/4828). ` +
-        'Please try `npm i` again after removing both package-lock.json and node_modules directory.',
-      {
-        cause: loadErrors.reduce((err, cur) => {
-          cur.cause = err
-          return cur
-        }),
-      },
-    )
-  }
-  throw new Error(`Failed to load native binding`)
-}
-
-module.exports = nativeBinding
-module.exports.classifyPdf = nativeBinding.classifyPdf
-module.exports.extractTextInRegions = nativeBinding.extractTextInRegions

napi/package.json

@@ -1,18 +1,36 @@
 {
-  "name": "firecrawl-pdf-inspector",
-  "version": "0.2.2",
+  "name": "@firecrawl/pdf-inspector",
+  "version": "1.8.10",
+  "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",
+  "bin": {
+    "pdf-inspector": "bin/pdf-inspector.mjs"
+  },
   "license": "MIT",
+  "keywords": [
+    "pdf",
+    "pdf-extraction",
+    "pdf-parser",
+    "text-extraction",
+    "ocr",
+    "pdf-classification",
+    "napi",
+    "rust",
+    "firecrawl"
+  ],
   "files": [
     "index.js",
     "index.d.ts",
-    "*.node"
+    "*.node",
+    "bin/",
+    "README.md"
   ],
   "repository": {
     "type": "git",
     "url": "https://github.com/firecrawl/pdf-inspector"
   },
+  "homepage": "https://github.com/firecrawl/pdf-inspector",
   "publishConfig": {
     "access": "public"
   },
@@ -20,7 +38,8 @@
     "binaryName": "pdf-inspector",
     "targets": [
       "x86_64-unknown-linux-gnu",
-      "aarch64-apple-darwin"
+      "aarch64-apple-darwin",
+      "x86_64-pc-windows-msvc"
     ],
     "package": {
       "name": "@firecrawl/pdf-inspector-js"

napi/probe-indent.mjs

@@ -0,0 +1,29 @@
+import { readFileSync } from "node:fs";
+import { createRequire } from "node:module";
+
+const require = createRequire(import.meta.url);
+const { detectVectorGridInRegion } = require("./index.js");
+
+const pdfPath =
+  process.argv[2] ?? "/tmp/pdf_inspector_indent_fixtures/cis_edge_benchmark.pdf";
+const pdf = readFileSync(pdfPath);
+const dpi = Number(process.argv[3] ?? 200);
+
+const crops = [
+  { pageIdx: 29, box: [0, 0, 612, 792], label: "page30-full" },
+  { pageIdx: 16, box: [0, 0, 612, 792], label: "page17-full" },
+  { pageIdx: 23, box: [0, 0, 612, 792], label: "page24-full" },
+];
+
+for (const { pageIdx, box, label } of crops) {
+  const result = detectVectorGridInRegion(pdf, pageIdx, box, dpi);
+  if (!result) {
+    console.log(`${label}: null`);
+    continue;
+  }
+  const rows = result.structureTokens.filter((token) => token === "<tr>").length;
+  const cols = rows > 0 ? result.cellBboxes.length / rows : 0;
+  console.log(
+    `${label}: cells=${result.cellBboxes.length} rows=${rows} cols=${cols}`,
+  );
+}

napi/src/lib.rs

@@ -2,58 +2,276 @@
 
 use napi::bindgen_prelude::*;
 use napi_derive::napi;
+use std::collections::HashSet;
+use std::panic;
+
+// ---------------------------------------------------------------------------
+// Enums
+// ---------------------------------------------------------------------------
+
+/// PDF document type classification.
+#[napi(string_enum)]
+pub enum PdfType {
+    TextBased,
+    Scanned,
+    ImageBased,
+    Mixed,
+}
+
+/// Type of a positioned text item.
+#[napi(string_enum)]
+pub enum ItemType {
+    Text,
+    Image,
+    Link,
+    FormField,
+}
+
+// ---------------------------------------------------------------------------
+// Result types
+// ---------------------------------------------------------------------------
+
+/// Full PDF processing result with markdown and metadata.
+#[napi(object)]
+pub struct PdfResult {
+    pub pdf_type: PdfType,
+    pub markdown: Option<String>,
+    pub page_count: u32,
+    pub processing_time_ms: u32,
+    /// 1-indexed page numbers that need OCR.
+    pub pages_needing_ocr: Vec<u32>,
+    pub title: Option<String>,
+    pub confidence: f64,
+    pub is_complex_layout: bool,
+    pub pages_with_tables: Vec<u32>,
+    pub pages_with_columns: Vec<u32>,
+    pub has_encoding_issues: bool,
+}
 
 /// Lightweight PDF classification result.
 #[napi(object)]
 pub struct PdfClassification {
-  pub pdf_type: String,
-  pub page_count: u32,
-  pub pages_needing_ocr: Vec<u32>,
-  pub confidence: f64,
+    pub pdf_type: PdfType,
+    pub page_count: u32,
+    /// 0-indexed page numbers that need OCR.
+    pub pages_needing_ocr: Vec<u32>,
+    pub confidence: f64,
+}
+
+/// A positioned text item extracted from a PDF.
+#[napi(object)]
+pub struct TextItem {
+    pub text: String,
+    pub x: f64,
+    pub y: f64,
+    pub width: f64,
+    pub height: f64,
+    pub font: String,
+    pub font_size: f64,
+    pub page: u32,
+    pub is_bold: bool,
+    pub is_italic: bool,
+    pub item_type: ItemType,
+    /// URL for link items, `None` for other types.
+    pub link_url: Option<String>,
 }
 
 /// A page's regions for text extraction: (page_index_0based, bboxes).
 #[napi(object)]
 pub struct PageRegions {
-  pub page: u32,
-  /// Each bbox is [x1, y1, x2, y2] in PDF points, top-left origin.
-  pub regions: Vec<Vec<f64>>,
+    pub page: u32,
+    /// Each bbox is [x1, y1, x2, y2] in PDF points, top-left origin.
+    pub regions: Vec<Vec<f64>>,
 }
 
 /// Extracted text for a single region.
 #[napi(object)]
 pub struct RegionText {
-  pub text: String,
-  /// `true` when the text should not be trusted (empty, GID fonts, garbage, encoding issues).
-  pub needs_ocr: bool,
+    pub text: String,
+    /// `true` when the text should not be trusted (empty, GID fonts, garbage, encoding issues).
+    pub needs_ocr: bool,
 }
 
 /// Extracted text for one page's regions.
 #[napi(object)]
 pub struct PageRegionTexts {
-  pub page: u32,
-  pub regions: Vec<RegionText>,
+    pub page: u32,
+    pub regions: Vec<RegionText>,
 }
 
-/// Classify a PDF: detect type (TextBased/Scanned/Mixed/ImageBased),
-/// page count, and which pages need OCR. Takes PDF bytes as Buffer.
+/// Vector-grid detection result compatible with `extractTablesWithStructure*`.
+#[napi(object)]
+pub struct VectorGridDetectionJs {
+    pub structure_tokens: Vec<String>,
+    pub cell_bboxes: Vec<Vec<f64>>,
+}
+
+// ---------------------------------------------------------------------------
+// Helpers
+// ---------------------------------------------------------------------------
+
+fn convert_pdf_type(t: pdf_inspector::PdfType) -> PdfType {
+    match t {
+        pdf_inspector::PdfType::TextBased => PdfType::TextBased,
+        pdf_inspector::PdfType::Scanned => PdfType::Scanned,
+        pdf_inspector::PdfType::ImageBased => PdfType::ImageBased,
+        pdf_inspector::PdfType::Mixed => PdfType::Mixed,
+    }
+}
+
+fn to_napi_result(r: pdf_inspector::PdfProcessResult) -> PdfResult {
+    PdfResult {
+        pdf_type: convert_pdf_type(r.pdf_type),
+        markdown: r.markdown,
+        page_count: r.page_count,
+        processing_time_ms: r.processing_time_ms as u32,
+        pages_needing_ocr: r.pages_needing_ocr,
+        title: r.title,
+        confidence: r.confidence as f64,
+        is_complex_layout: r.layout.is_complex,
+        pages_with_tables: r.layout.pages_with_tables,
+        pages_with_columns: r.layout.pages_with_columns,
+        has_encoding_issues: r.has_encoding_issues,
+    }
+}
+
+fn convert_item_type(t: &pdf_inspector::types::ItemType) -> (ItemType, Option<String>) {
+    match t {
+        pdf_inspector::types::ItemType::Text => (ItemType::Text, None),
+        pdf_inspector::types::ItemType::Image => (ItemType::Image, None),
+        pdf_inspector::types::ItemType::Link(url) => (ItemType::Link, Some(url.clone())),
+        pdf_inspector::types::ItemType::FormField => (ItemType::FormField, None),
+    }
+}
+
+fn to_napi_err(e: impl std::fmt::Display, ctx: &str) -> Error {
+    Error::new(Status::GenericFailure, format!("{ctx}: {e}"))
+}
+
+/// Run a closure, catching any Rust panic and converting it to a NAPI error.
+/// Prevents process abort from unwind panics in the native module.
+fn catch_panic<F, T>(ctx: &str, f: F) -> Result<T>
+where
+    F: FnOnce() -> Result<T> + panic::UnwindSafe,
+{
+    match panic::catch_unwind(f) {
+        Ok(result) => result,
+        Err(payload) => {
+            let msg = if let Some(s) = payload.downcast_ref::<&str>() {
+                s.to_string()
+            } else if let Some(s) = payload.downcast_ref::<String>() {
+                s.clone()
+            } else {
+                "unknown panic".to_string()
+            };
+            Err(Error::new(
+                Status::GenericFailure,
+                format!("{ctx}: Rust panic: {msg}"),
+            ))
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Public NAPI API
+// ---------------------------------------------------------------------------
+
+/// Process a PDF from a Buffer: detect type, extract text, and convert to Markdown.
+#[napi]
+pub fn process_pdf(buffer: Buffer, pages: Option<Vec<u32>>) -> Result<PdfResult> {
+    let bytes: Vec<u8> = buffer.to_vec();
+    catch_panic("process_pdf", move || {
+        let mut opts = pdf_inspector::PdfOptions::new();
+        if let Some(p) = pages {
+            opts = opts.pages(p);
+        }
+        let result = pdf_inspector::process_pdf_mem_with_options(&bytes, opts)
+            .map_err(|e| to_napi_err(e, "process_pdf"))?;
+        Ok(to_napi_result(result))
+    })
+}
+
+/// Fast detection only — no text extraction or markdown.
+#[napi]
+pub fn detect_pdf(buffer: Buffer) -> Result<PdfResult> {
+    let bytes: Vec<u8> = buffer.to_vec();
+    catch_panic("detect_pdf", move || {
+        let result =
+            pdf_inspector::detect_pdf_mem(&bytes).map_err(|e| to_napi_err(e, "detect_pdf"))?;
+        Ok(to_napi_result(result))
+    })
+}
+
+/// Lightweight PDF classification — returns type, page count, and OCR pages.
+/// Faster than detectPdf as it skips building the full PdfResult.
+/// Pages in pagesNeedingOcr are 0-indexed.
 #[napi]
 pub fn classify_pdf(buffer: Buffer) -> Result<PdfClassification> {
-  let result = pdf_inspector::classify_pdf_mem(&buffer).map_err(|e| {
-    Error::new(Status::GenericFailure, format!("classify_pdf failed: {e}"))
-  })?;
+    let bytes: Vec<u8> = buffer.to_vec();
+    catch_panic("classify_pdf", move || {
+        let result =
+            pdf_inspector::classify_pdf_mem(&bytes).map_err(|e| to_napi_err(e, "classify_pdf"))?;
+        Ok(PdfClassification {
+            pdf_type: convert_pdf_type(result.pdf_type),
+            page_count: result.page_count,
+            pages_needing_ocr: result.pages_needing_ocr,
+            confidence: result.confidence as f64,
+        })
+    })
+}
 
-  Ok(PdfClassification {
-    pdf_type: match result.pdf_type {
-      pdf_inspector::PdfType::TextBased => "TextBased".to_string(),
-      pdf_inspector::PdfType::Scanned => "Scanned".to_string(),
-      pdf_inspector::PdfType::ImageBased => "ImageBased".to_string(),
-      pdf_inspector::PdfType::Mixed => "Mixed".to_string(),
-    },
-    page_count: result.page_count,
-    pages_needing_ocr: result.pages_needing_ocr,
-    confidence: result.confidence as f64,
-  })
+/// Extract plain text from a PDF Buffer.
+#[napi]
+pub fn extract_text(buffer: Buffer) -> Result<String> {
+    let bytes: Vec<u8> = buffer.to_vec();
+    catch_panic("extract_text", move || {
+        pdf_inspector::extractor::extract_text_mem(&bytes)
+            .map_err(|e| to_napi_err(e, "extract_text"))
+    })
+}
+
+/// Extract text with position information from a PDF Buffer.
+#[napi]
+pub fn extract_text_with_positions(
+    buffer: Buffer,
+    pages: Option<Vec<u32>>,
+) -> Result<Vec<TextItem>> {
+    let bytes: Vec<u8> = buffer.to_vec();
+    catch_panic("extract_text_with_positions", move || {
+        let items = match pages {
+            Some(p) => {
+                let page_set: HashSet<u32> = p.into_iter().collect();
+                pdf_inspector::extractor::extract_text_with_positions_mem_pages(
+                    &bytes,
+                    Some(&page_set),
+                )
+                .map_err(|e| to_napi_err(e, "extract_text_with_positions"))?
+            }
+            None => pdf_inspector::extractor::extract_text_with_positions_mem(&bytes)
+                .map_err(|e| to_napi_err(e, "extract_text_with_positions"))?,
+        };
+
+        Ok(items
+            .into_iter()
+            .map(|item| {
+                let (item_type, link_url) = convert_item_type(&item.item_type);
+                TextItem {
+                    text: item.text,
+                    x: item.x as f64,
+                    y: item.y as f64,
+                    width: item.width as f64,
+                    height: item.height as f64,
+                    font: item.font,
+                    font_size: item.font_size as f64,
+                    page: item.page,
+                    is_bold: item.is_bold,
+                    is_italic: item.is_italic,
+                    item_type,
+                    link_url,
+                }
+            })
+            .collect())
+    })
 }
 
 /// Extract text within bounding-box regions from a PDF.
@@ -62,55 +280,376 @@ pub fn classify_pdf(buffer: Buffer) -> Result<PdfClassification> {
 /// this extracts PDF text within those regions — skipping GPU OCR
 /// for text-based pages.
 ///
-/// Each region result includes `needs_ocr` — set when the extracted text
+/// Each region result includes `needsOcr` — set when the extracted text
 /// is unreliable (empty, GID-encoded fonts, garbage, encoding issues).
 ///
 /// Coordinates are PDF points with top-left origin.
 #[napi]
 pub fn extract_text_in_regions(
-  buffer: Buffer,
-  page_regions: Vec<PageRegions>,
+    buffer: Buffer,
+    page_regions: Vec<PageRegions>,
 ) -> Result<Vec<PageRegionTexts>> {
-  // Convert from napi types to the Rust API's expected format
-  let regions: Vec<(u32, Vec<[f32; 4]>)> = page_regions
-    .iter()
-    .map(|pr| {
-      let bboxes: Vec<[f32; 4]> = pr
-        .regions
-        .iter()
-        .map(|r| {
-          if r.len() != 4 {
-            [0.0, 0.0, 0.0, 0.0]
-          } else {
-            [r[0] as f32, r[1] as f32, r[2] as f32, r[3] as f32]
-          }
-        })
-        .collect();
-      (pr.page, bboxes)
+    let bytes: Vec<u8> = buffer.to_vec();
+    let regions = parse_page_regions(&page_regions);
+
+    catch_panic("extract_text_in_regions", move || {
+        let results = pdf_inspector::extract_text_in_regions_mem(&bytes, &regions)
+            .map_err(|e| to_napi_err(e, "extract_text_in_regions"))?;
+        Ok(to_page_region_texts(results))
     })
-    .collect();
+}
 
-  let results = pdf_inspector::extract_text_in_regions_mem(&buffer, &regions).map_err(|e| {
-    Error::new(
-      Status::GenericFailure,
-      format!("extract_text_in_regions failed: {e}"),
-    )
-  })?;
+/// Extract markdown tables within bounding-box regions from a PDF.
+///
+/// Like `extractTextInRegions` but runs table detection on items within each
+/// region and returns markdown pipe-tables instead of flat text.
+///
+/// When table structure is detected, `text` contains a markdown pipe-table and
+/// `needsOcr` is `false`. When no table is found, `text` is empty and
+/// `needsOcr` is `true` so the caller can fall back to GPU OCR.
+///
+/// Coordinates are PDF points with top-left origin.
+#[napi]
+pub fn extract_tables_in_regions(
+    buffer: Buffer,
+    page_regions: Vec<PageRegions>,
+) -> Result<Vec<PageRegionTexts>> {
+    let bytes: Vec<u8> = buffer.to_vec();
+    let regions = parse_page_regions(&page_regions);
 
-  Ok(
+    catch_panic("extract_tables_in_regions", move || {
+        let results = pdf_inspector::extract_tables_in_regions_mem(&bytes, &regions)
+            .map_err(|e| to_napi_err(e, "extract_tables_in_regions"))?;
+        Ok(to_page_region_texts(results))
+    })
+}
+
+/// Detect a vector ruled-line / rectangle grid inside one page region.
+///
+/// Returns TSR-compatible structure tokens plus crop-pixel cell bboxes, or
+/// `null` when the region does not contain a valid vector grid.
+///
+/// `pageIdx` is 0-indexed. `regionPdfPtBbox` is `[x1,y1,x2,y2]` in PDF
+/// points with top-left origin. `renderDpi` is the DPI of the crop image that
+/// will consume the returned cell bboxes.
+#[napi]
+pub fn detect_vector_grid_in_region(
+    buffer: Buffer,
+    page_idx: u32,
+    region_pdf_pt_bbox: Vec<f64>,
+    render_dpi: f64,
+) -> Result<Option<VectorGridDetectionJs>> {
+    let bytes: Vec<u8> = buffer.to_vec();
+    let region = if region_pdf_pt_bbox.len() == 4 {
+        [
+            region_pdf_pt_bbox[0] as f32,
+            region_pdf_pt_bbox[1] as f32,
+            region_pdf_pt_bbox[2] as f32,
+            region_pdf_pt_bbox[3] as f32,
+        ]
+    } else {
+        [0.0, 0.0, 0.0, 0.0]
+    };
+
+    catch_panic("detect_vector_grid_in_region", move || {
+        let result = pdf_inspector::detect_vector_grid_in_region_mem(
+            &bytes,
+            page_idx,
+            region,
+            render_dpi as f32,
+        )
+        .map_err(|e| to_napi_err(e, "detect_vector_grid_in_region"))?;
+
+        Ok(result.map(|r| VectorGridDetectionJs {
+            structure_tokens: r.structure_tokens,
+            cell_bboxes: r
+                .cell_bboxes
+                .into_iter()
+                .map(|bbox| bbox.into_iter().map(|v| v as f64).collect())
+                .collect(),
+        }))
+    })
+}
+
+/// One cropped table region plus its raw structure-recovery output, for
+/// `extractTablesWithStructure`.
+///
+/// `structureTokens` and `cellBboxes` are typically produced by an external
+/// table-structure recognition model (e.g. SLANet on PaddleOCR) running on
+/// a rendered crop of the page. pdf-inspector uses the structure to lay out
+/// the cells and pulls the cell text from the native PDF — no OCR involved.
+#[napi(object)]
+pub struct TsrTableInputJs {
+    /// 0-indexed page number where the crop was taken from.
+    pub page: u32,
+    /// Crop bbox on the page, `[x1, y1, x2, y2]` in PDF points with
+    /// top-left origin.
+    pub crop_pdf_pt_bbox: Vec<f64>,
+    /// DPI the crop image was rendered at (e.g. `200.0`).
+    pub render_dpi: f64,
+    /// Raw structure tokens emitted by the TSR model, in document order.
+    pub structure_tokens: Vec<String>,
+    /// One bbox per cell (in document order). May be 4-element
+    /// `[x1,y1,x2,y2]` or 8-element 4-corner polygon, in crop image-pixel
+    /// space.
+    pub cell_bboxes: Vec<Vec<f64>>,
+}
+
+/// Extract markdown tables using externally-supplied structure recovery.
+///
+/// For each input, pairs structure tokens with cell bboxes (rowspan/colspan
+/// aware), converts each cell bbox from crop image-pixels into page PDF
+/// points, pulls the cell's text from the native PDF, and emits a markdown
+/// pipe-table.
+///
+/// Returns one markdown string per input, in input order.
+#[napi]
+pub fn extract_tables_with_structure(
+    buffer: Buffer,
+    inputs: Vec<TsrTableInputJs>,
+) -> Result<Vec<String>> {
+    let bytes: Vec<u8> = buffer.to_vec();
+    let parsed = parse_tsr_inputs(&inputs);
+
+    catch_panic("extract_tables_with_structure", move || {
+        pdf_inspector::extract_tables_with_structure_mem(&bytes, &parsed)
+            .map_err(|e| to_napi_err(e, "extract_tables_with_structure"))
+    })
+}
+
+/// One resolved cell from `extractTablesWithStructureCells`.
+#[napi(object)]
+pub struct StructuredCellJs {
+    /// 0-indexed grid row.
+    pub row: u32,
+    /// 0-indexed grid column.
+    pub col: u32,
+    /// 1 for a normal cell.
+    pub rowspan: u32,
+    /// 1 for a normal cell.
+    pub colspan: u32,
+    /// `true` when the cell is a `<th>` or sits inside `<thead>`.
+    pub is_header: bool,
+    /// Text extracted from the native PDF for this cell (may be empty).
+    pub text: String,
+    /// Axis-aligned bbox `[x1, y1, x2, y2]` in page PDF-points, top-left
+    /// origin. Useful for debug overlays or per-cell post-processing.
+    pub page_pt_bbox: Vec<f64>,
+}
+
+/// Extract structured cells using externally-supplied structure recovery.
+///
+/// Lower-level sibling of [`extractTablesWithStructure`]: instead of
+/// rendering markdown, returns the resolved cells (row, col, rowspan,
+/// colspan, isHeader, text, pagePtBbox) so callers can drive their own
+/// rendering, debug overlays, or per-cell post-processing.
+///
+/// Returns one `Array<StructuredCellJs>` per input, in input order.
+#[napi]
+pub fn extract_tables_with_structure_cells(
+    buffer: Buffer,
+    inputs: Vec<TsrTableInputJs>,
+) -> Result<Vec<Vec<StructuredCellJs>>> {
+    let bytes: Vec<u8> = buffer.to_vec();
+    let parsed = parse_tsr_inputs(&inputs);
+
+    catch_panic("extract_tables_with_structure_cells", move || {
+        let result = pdf_inspector::extract_tables_with_structure_cells_mem(&bytes, &parsed)
+            .map_err(|e| to_napi_err(e, "extract_tables_with_structure_cells"))?;
+        Ok(result
+            .into_iter()
+            .map(|cells| {
+                cells
+                    .into_iter()
+                    .map(|c| StructuredCellJs {
+                        row: c.row as u32,
+                        col: c.col as u32,
+                        rowspan: c.rowspan as u32,
+                        colspan: c.colspan as u32,
+                        is_header: c.is_header,
+                        text: c.text,
+                        page_pt_bbox: c.page_pt_bbox.iter().map(|v| *v as f64).collect(),
+                    })
+                    .collect()
+            })
+            .collect())
+    })
+}
+
+/// One result from `extractTablesWithStructureAuto` — markdown plus a
+/// diagnostic flag identifying which path produced it.
+///
+/// `fallbackReason` is `null` when the TSR-hybrid path produced the
+/// markdown directly. When stage 1's quality check fires (the cells
+/// look like a SLANet detection pathology — phantom rows or multi-row
+/// content in a single cell), the auto path may expand the TSR cells
+/// in-place or run the heuristic table extractor on the same region.
+/// `fallbackReason` carries the diagnostic label (for example
+/// `"multi_row_in_cell_expanded"` or `"phantom_empty_row"`).
+#[napi(object)]
+pub struct TableExtractionResultJs {
+    pub markdown: String,
+    pub fallback_reason: Option<String>,
+}
+
+/// Auto-fallback variant of [`extractTablesWithStructure`].
+///
+/// Runs the TSR-hybrid path, checks the resulting cells for known
+/// SLANet detection pathologies, expands multi-row cells in-place when
+/// possible, and otherwise falls back to the heuristic
+/// `extractTablesInRegions` for inputs where the TSR path looks
+/// compromised.
+///
+/// On clean inputs this returns identical markdown to
+/// `extractTablesWithStructure`; on flagged inputs `fallbackReason` is
+/// set to the recovery path that produced the result.
+#[napi]
+pub fn extract_tables_with_structure_auto(
+    buffer: Buffer,
+    inputs: Vec<TsrTableInputJs>,
+) -> Result<Vec<TableExtractionResultJs>> {
+    let bytes: Vec<u8> = buffer.to_vec();
+    let parsed = parse_tsr_inputs(&inputs);
+
+    catch_panic("extract_tables_with_structure_auto", move || {
+        let result = pdf_inspector::extract_tables_with_structure_auto_mem(&bytes, &parsed)
+            .map_err(|e| to_napi_err(e, "extract_tables_with_structure_auto"))?;
+        Ok(result
+            .into_iter()
+            .map(|r| TableExtractionResultJs {
+                markdown: r.markdown,
+                fallback_reason: r.fallback_reason,
+            })
+            .collect())
+    })
+}
+
+fn parse_tsr_inputs(inputs: &[TsrTableInputJs]) -> Vec<pdf_inspector::TsrTableInput> {
+    inputs
+        .iter()
+        .map(|i| {
+            let crop = if i.crop_pdf_pt_bbox.len() == 4 {
+                [
+                    i.crop_pdf_pt_bbox[0] as f32,
+                    i.crop_pdf_pt_bbox[1] as f32,
+                    i.crop_pdf_pt_bbox[2] as f32,
+                    i.crop_pdf_pt_bbox[3] as f32,
+                ]
+            } else {
+                [0.0, 0.0, 0.0, 0.0]
+            };
+            let cell_bboxes: Vec<Vec<f32>> = i
+                .cell_bboxes
+                .iter()
+                .map(|bb| bb.iter().map(|v| *v as f32).collect())
+                .collect();
+            pdf_inspector::TsrTableInput {
+                page: i.page,
+                crop_pdf_pt_bbox: crop,
+                render_dpi: i.render_dpi as f32,
+                structure_tokens: i.structure_tokens.clone(),
+                cell_bboxes,
+            }
+        })
+        .collect()
+}
+
+/// Per-page markdown extraction result.
+#[napi(object)]
+pub struct PageMarkdownResult {
+    /// 0-indexed page number.
+    pub page: u32,
+    /// Formatted markdown for this page.
+    pub markdown: String,
+    /// `true` when text on this page is unreliable.
+    pub needs_ocr: bool,
+}
+
+/// Combined per-page markdown extraction and layout classification result.
+#[napi(object)]
+pub struct PagesExtractionResult {
+    /// Per-page markdown results.
+    pub pages: Vec<PageMarkdownResult>,
+    /// 1-indexed pages where tables were detected.
+    pub pages_with_tables: Vec<u32>,
+    /// 1-indexed pages where multi-column layout was detected.
+    pub pages_with_columns: Vec<u32>,
+    /// 1-indexed pages that need OCR (scanned/image-based).
+    pub pages_needing_ocr: Vec<u32>,
+    /// True if any page has tables or columns.
+    pub is_complex: bool,
+}
+
+/// Extract formatted markdown for pages of a PDF, with layout classification
+/// metadata.
+///
+/// Returns per-page markdown and classification data (tables, columns,
+/// OCR needs) from a single parse. Font statistics are computed from the
+/// full document so header detection is consistent across pages.
+///
+/// Omit `pages` (or pass `undefined`) to return every page in document
+/// order. Pass an array of 0-indexed page numbers to restrict output to
+/// those pages, in caller-supplied order.
+#[napi]
+pub fn extract_pages_markdown(
+    buffer: Buffer,
+    pages: Option<Vec<u32>>,
+) -> Result<PagesExtractionResult> {
+    let bytes: Vec<u8> = buffer.to_vec();
+    catch_panic("extract_pages_markdown", move || {
+        let result = pdf_inspector::extract_pages_markdown_mem(&bytes, pages.as_deref())
+            .map_err(|e| to_napi_err(e, "extract_pages_markdown"))?;
+        Ok(PagesExtractionResult {
+            pages: result
+                .pages
+                .into_iter()
+                .map(|r| PageMarkdownResult {
+                    page: r.page,
+                    markdown: r.markdown,
+                    needs_ocr: r.needs_ocr,
+                })
+                .collect(),
+            pages_with_tables: result.pages_with_tables,
+            pages_with_columns: result.pages_with_columns,
+            pages_needing_ocr: result.pages_needing_ocr,
+            is_complex: result.is_complex,
+        })
+    })
+}
+
+fn parse_page_regions(page_regions: &[PageRegions]) -> Vec<(u32, Vec<[f32; 4]>)> {
+    page_regions
+        .iter()
+        .map(|pr| {
+            let bboxes: Vec<[f32; 4]> = pr
+                .regions
+                .iter()
+                .map(|r| {
+                    if r.len() != 4 {
+                        [0.0, 0.0, 0.0, 0.0]
+                    } else {
+                        [r[0] as f32, r[1] as f32, r[2] as f32, r[3] as f32]
+                    }
+                })
+                .collect();
+            (pr.page, bboxes)
+        })
+        .collect()
+}
+
+fn to_page_region_texts(results: Vec<pdf_inspector::PageRegionResult>) -> Vec<PageRegionTexts> {
     results
-      .into_iter()
-      .map(|page_result| PageRegionTexts {
-        page: page_result.page,
-        regions: page_result
-          .regions
-          .into_iter()
-          .map(|r| RegionText {
-            text: r.text,
-            needs_ocr: r.needs_ocr,
-          })
-          .collect(),
-      })
-      .collect(),
-  )
+        .into_iter()
+        .map(|page_result| PageRegionTexts {
+            page: page_result.page,
+            regions: page_result
+                .regions
+                .into_iter()
+                .map(|r| RegionText {
+                    text: r.text,
+                    needs_ocr: r.needs_ocr,
+                })
+                .collect(),
+        })
+        .collect()
 }

napi/test.mjs

@@ -0,0 +1,133 @@
+import { readFileSync } from 'fs';
+import { strict as assert } from 'assert';
+import {
+  processPdf,
+  detectPdf,
+  classifyPdf,
+  extractText,
+  extractTextWithPositions,
+  extractTextInRegions,
+  detectVectorGridInRegion,
+  extractPagesMarkdown,
+} from './index.js';
+
+const fixture = readFileSync('../tests/fixtures/thermo-freon12.pdf');
+
+// --- processPdf ---
+console.log('Testing processPdf...');
+const result = processPdf(fixture);
+assert.equal(result.pdfType, 'TextBased');
+assert.equal(result.pageCount, 3);
+assert.ok(result.confidence > 0);
+assert.ok(result.markdown && result.markdown.length > 0);
+assert.equal(typeof result.isComplexLayout, 'boolean');
+assert.ok(Array.isArray(result.pagesWithTables));
+assert.ok(Array.isArray(result.pagesWithColumns));
+assert.equal(typeof result.hasEncodingIssues, 'boolean');
+console.log('  processPdf: OK');
+
+// processPdf with pages
+const result2 = processPdf(fixture, [1]);
+assert.ok(result2.markdown && result2.markdown.length > 0);
+console.log('  processPdf with pages: OK');
+
+// --- detectPdf ---
+console.log('Testing detectPdf...');
+const detected = detectPdf(fixture);
+assert.equal(detected.pdfType, 'TextBased');
+assert.equal(detected.pageCount, 3);
+assert.equal(detected.markdown, undefined);
+console.log('  detectPdf: OK');
+
+// --- classifyPdf ---
+console.log('Testing classifyPdf...');
+const classified = classifyPdf(fixture);
+assert.equal(classified.pdfType, 'TextBased');
+assert.equal(classified.pageCount, 3);
+assert.ok(classified.confidence > 0);
+assert.ok(Array.isArray(classified.pagesNeedingOcr));
+console.log('  classifyPdf: OK');
+
+// --- extractText ---
+console.log('Testing extractText...');
+const text = extractText(fixture);
+assert.equal(typeof text, 'string');
+assert.ok(text.length > 0);
+console.log('  extractText: OK');
+
+// --- extractTextWithPositions ---
+console.log('Testing extractTextWithPositions...');
+const items = extractTextWithPositions(fixture);
+assert.ok(items.length > 0);
+const item = items[0];
+assert.equal(typeof item.text, 'string');
+assert.equal(typeof item.x, 'number');
+assert.equal(typeof item.y, 'number');
+assert.equal(typeof item.width, 'number');
+assert.equal(typeof item.height, 'number');
+assert.equal(typeof item.font, 'string');
+assert.equal(typeof item.fontSize, 'number');
+assert.equal(typeof item.page, 'number');
+assert.equal(typeof item.isBold, 'boolean');
+assert.equal(typeof item.isItalic, 'boolean');
+assert.equal(typeof item.itemType, 'string');
+console.log('  extractTextWithPositions: OK');
+
+// with pages filter
+const page1Items = extractTextWithPositions(fixture, [1]);
+assert.ok(page1Items.length > 0);
+assert.ok(page1Items.every(i => i.page === 1));
+console.log('  extractTextWithPositions with pages: OK');
+
+// --- extractTextInRegions ---
+console.log('Testing extractTextInRegions...');
+const regionResults = extractTextInRegions(fixture, [
+  { page: 0, regions: [[0, 0, 600, 100]] },
+]);
+assert.equal(regionResults.length, 1);
+assert.equal(regionResults[0].page, 0);
+assert.equal(regionResults[0].regions.length, 1);
+assert.equal(typeof regionResults[0].regions[0].text, 'string');
+assert.equal(typeof regionResults[0].regions[0].needsOcr, 'boolean');
+console.log('  extractTextInRegions: OK');
+
+// --- detectVectorGridInRegion ---
+console.log('Testing detectVectorGridInRegion...');
+const vectorGrid = detectVectorGridInRegion(fixture, 0, [0, 0, 600, 800], 72);
+assert.ok(vectorGrid === null || typeof vectorGrid === 'object');
+if (vectorGrid) {
+  assert.ok(Array.isArray(vectorGrid.structureTokens));
+  assert.ok(Array.isArray(vectorGrid.cellBboxes));
+  assert.ok(vectorGrid.cellBboxes.every(bbox => Array.isArray(bbox) && bbox.length === 4));
+}
+console.log('  detectVectorGridInRegion: OK');
+
+// --- extractPagesMarkdown ---
+console.log('Testing extractPagesMarkdown...');
+
+// omit pages → every page in document order
+const allPages = extractPagesMarkdown(fixture);
+assert.equal(allPages.pages.length, 3);
+assert.deepEqual(allPages.pages.map(p => p.page), [0, 1, 2]);
+assert.ok(typeof allPages.pages[0].markdown === 'string');
+assert.equal(typeof allPages.pages[0].needsOcr, 'boolean');
+assert.ok(Array.isArray(allPages.pagesWithTables));
+assert.ok(Array.isArray(allPages.pagesWithColumns));
+assert.ok(Array.isArray(allPages.pagesNeedingOcr));
+assert.equal(typeof allPages.isComplex, 'boolean');
+console.log('  extractPagesMarkdown (no pages arg): OK');
+
+// selected pages preserve caller order
+const picked = extractPagesMarkdown(fixture, [2, 0]);
+assert.equal(picked.pages.length, 2);
+assert.equal(picked.pages[0].page, 2);
+assert.equal(picked.pages[1].page, 0);
+console.log('  extractPagesMarkdown with pages: OK');
+
+// --- Error handling ---
+console.log('Testing error handling...');
+assert.throws(() => processPdf(Buffer.from('not a pdf')), /process_pdf/);
+assert.throws(() => classifyPdf(Buffer.from('')), /classify_pdf/);
+console.log('  error handling: OK');
+
+console.log('\nAll NAPI tests passed!');

pdf_inspector.pyi

@@ -0,0 +1,167 @@
+"""Type stubs for pdf_inspector."""
+
+from typing import Optional
+
+class PdfResult:
+    """Result of processing a PDF file."""
+    pdf_type: str
+    """'text_based', 'scanned', 'image_based', or 'mixed'."""
+    markdown: Optional[str]
+    page_count: int
+    processing_time_ms: int
+    pages_needing_ocr: list[int]
+    title: Optional[str]
+    confidence: float
+    is_complex_layout: bool
+    pages_with_tables: list[int]
+    pages_with_columns: list[int]
+    has_encoding_issues: bool
+
+class PdfClassification:
+    """Lightweight PDF classification result."""
+    pdf_type: str
+    """'text_based', 'scanned', 'image_based', or 'mixed'."""
+    page_count: int
+    pages_needing_ocr: list[int]
+    """0-indexed page numbers that need OCR."""
+    confidence: float
+
+class TextItem:
+    """A positioned text item extracted from a PDF."""
+    text: str
+    x: float
+    y: float
+    width: float
+    height: float
+    font: str
+    font_size: float
+    page: int
+    is_bold: bool
+    is_italic: bool
+    item_type: str
+
+class RegionText:
+    """Extracted text for a single region."""
+    text: str
+    needs_ocr: bool
+    """True when the text should not be trusted."""
+
+class PageRegionTexts:
+    """Extracted text for one page's regions."""
+    page: int
+    """0-indexed page number."""
+    regions: list[RegionText]
+
+class PageMarkdown:
+    """Per-page markdown extraction result."""
+    page: int
+    """0-indexed page number."""
+    markdown: str
+    """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."""
+
+class PagesExtractionResult:
+    """Per-page markdown output with document-wide layout classification."""
+    pages: list[PageMarkdown]
+    """Per-page markdown results, in the order requested."""
+    pages_with_tables: list[int]
+    """1-indexed pages where tables were detected."""
+    pages_with_columns: list[int]
+    """1-indexed pages where multi-column layout was detected."""
+    pages_needing_ocr: list[int]
+    """1-indexed pages that need OCR."""
+    is_complex: bool
+    """True if any page has tables or multi-column layout."""
+
+def process_pdf(path: str, pages: Optional[list[int]] = None) -> PdfResult:
+    """Process a PDF: detect type, extract text, convert to Markdown."""
+    ...
+
+def process_pdf_bytes(data: bytes, pages: Optional[list[int]] = None) -> PdfResult:
+    """Process a PDF from bytes in memory."""
+    ...
+
+def detect_pdf(path: str) -> PdfResult:
+    """Fast detection only — no text extraction."""
+    ...
+
+def detect_pdf_bytes(data: bytes) -> PdfResult:
+    """Fast detection from bytes."""
+    ...
+
+def classify_pdf(path: str) -> PdfClassification:
+    """Lightweight classification — type, page count, and OCR pages (0-indexed)."""
+    ...
+
+def classify_pdf_bytes(data: bytes) -> PdfClassification:
+    """Lightweight classification from bytes."""
+    ...
+
+def extract_text(path: str) -> str:
+    """Extract plain text from a PDF."""
+    ...
+
+def extract_text_bytes(data: bytes) -> str:
+    """Extract plain text from PDF bytes."""
+    ...
+
+def extract_text_with_positions(path: str, pages: Optional[list[int]] = None) -> list[TextItem]:
+    """Extract text with position information."""
+    ...
+
+def extract_text_with_positions_bytes(data: bytes, pages: Optional[list[int]] = None) -> list[TextItem]:
+    """Extract text with position information from bytes."""
+    ...
+
+def extract_text_in_regions(
+    path: str,
+    page_regions: list[tuple[int, list[list[float]]]],
+) -> list[PageRegionTexts]:
+    """Extract text within bounding-box regions from a PDF file.
+
+    Args:
+        path: Path to the PDF file.
+        page_regions: List of (page_0indexed, [[x1, y1, x2, y2], ...]) tuples.
+    """
+    ...
+
+def extract_text_in_regions_bytes(
+    data: bytes,
+    page_regions: list[tuple[int, list[list[float]]]],
+) -> list[PageRegionTexts]:
+    """Extract text within bounding-box regions from PDF bytes.
+
+    Args:
+        data: PDF file contents as bytes.
+        page_regions: List of (page_0indexed, [[x1, y1, x2, y2], ...]) tuples.
+    """
+    ...
+
+def extract_pages_markdown(
+    path: str,
+    pages: Optional[list[int]] = None,
+) -> PagesExtractionResult:
+    """Extract formatted markdown for pages of a PDF, with layout classification.
+
+    Args:
+        path: Path to the PDF file.
+        pages: Optional list of 0-indexed pages. When ``None`` (default), every
+            page is returned in document order. Otherwise, output matches the
+            caller-supplied order.
+
+    Returns:
+        PagesExtractionResult with per-page markdown and document-wide layout
+        classification (tables, columns, OCR needs).
+    """
+    ...
+
+def extract_pages_markdown_bytes(
+    data: bytes,
+    pages: Optional[list[int]] = None,
+) -> PagesExtractionResult:
+    """Extract formatted markdown for pages of a PDF from bytes.
+
+    See :func:`extract_pages_markdown` for details.
+    """
+    ...

pyproject.toml

@@ -0,0 +1,21 @@
+[build-system]
+requires = ["maturin>=1.0,<2.0"]
+build-backend = "maturin"
+
+[project]
+name = "pdf-inspector"
+version = "0.1.0"
+description = "Fast PDF inspection, classification, and text extraction with smart scanned vs text-based detection"
+license = { text = "MIT" }
+requires-python = ">=3.8"
+classifiers = [
+    "Programming Language :: Rust",
+    "Programming Language :: Python :: Implementation :: CPython",
+    "Programming Language :: Python :: 3",
+    "License :: OSI Approved :: MIT License",
+    "Operating System :: OS Independent",
+    "Topic :: Text Processing",
+]
+
+[tool.maturin]
+features = ["python"]

src/bin/detect_pdf.rs

@@ -1,8 +1,12 @@
 //! CLI tool for detecting PDF type (text-based vs scanned)
 
-use pdf_inspector::{detect_pdf_type, process_pdf_with_options, PdfOptions, PdfType, ProcessMode};
+use pdf_inspector::{
+    detect_pdf_type, detector::estimate_page_count_from_bytes, process_pdf_with_options,
+    PdfOptions, PdfType, ProcessMode,
+};
 use std::env;
 use std::fmt::Write;
+use std::fs;
 use std::process;
 use std::time::Instant;
 
@@ -64,6 +68,32 @@ fn pdf_type_str(pdf_type: &PdfType) -> &'static str {
     }
 }
 
+fn page_count_hint(pdf_path: &str) -> Option<u32> {
+    fs::read(pdf_path)
+        .ok()
+        .map(|bytes| estimate_page_count_from_bytes(&bytes))
+        .filter(|&count| count > 0)
+}
+
+fn print_error(e: &pdf_inspector::PdfError, pdf_path: &str, json_output: bool) {
+    if json_output {
+        if let Some(count) = page_count_hint(pdf_path) {
+            println!(
+                r#"{{"error":"{}","page_count_hint":{}}}"#,
+                json_escape(&e.to_string()),
+                count
+            );
+        } else {
+            println!(r#"{{"error":"{}"}}"#, json_escape(&e.to_string()));
+        }
+    } else {
+        eprintln!("Error: {}", e);
+        if let Some(count) = page_count_hint(pdf_path) {
+            eprintln!("Page count hint: {}", count);
+        }
+    }
+}
+
 fn run_analyze(pdf_path: &str, json_output: bool, start: Instant) {
     match process_pdf_with_options(pdf_path, PdfOptions::new().mode(ProcessMode::Analyze)) {
         Ok(result) => {
@@ -135,11 +165,7 @@ fn run_analyze(pdf_path: &str, json_output: bool, start: Instant) {
             }
         }
         Err(e) => {
-            if json_output {
-                println!(r#"{{"error":"{}"}}"#, e);
-            } else {
-                eprintln!("Error: {}", e);
-            }
+            print_error(&e, pdf_path, json_output);
             process::exit(1);
         }
     }
@@ -236,11 +262,7 @@ fn run_detect_only(pdf_path: &str, json_output: bool, start: Instant) {
             }
         }
         Err(e) => {
-            if json_output {
-                println!(r#"{{"error":"{}"}}"#, e);
-            } else {
-                eprintln!("Error: {}", e);
-            }
+            print_error(&e, pdf_path, json_output);
             process::exit(1);
         }
     }

src/extractor/content_stream.rs

@@ -82,7 +82,7 @@ pub(crate) fn extract_page_text_items(
     page_num: u32,
     font_cmaps: &FontCMaps,
     include_invisible: bool,
-) -> Result<(PageExtraction, bool), PdfError> {
+) -> Result<(PageExtraction, bool, bool), PdfError> {
     use lopdf::content::Content;
 
     let mut items = Vec::new();
@@ -182,7 +182,7 @@ pub(crate) fn extract_page_text_items(
             content.operations.len(),
             MAX_OPERATIONS
         );
-        return Ok(((Vec::new(), Vec::new(), Vec::new()), false));
+        return Ok(((Vec::new(), Vec::new(), Vec::new()), false, false));
     }
 
     // Graphics state tracking
@@ -216,6 +216,7 @@ pub(crate) fn extract_page_text_items(
     let mut marked_content_stack: Vec<MarkedContentEntry> = Vec::new();
     let mut suppress_glyph_extraction = false;
     let mut actual_text_start_tm: Option<[f32; 6]> = None; // text matrix at BDC entry
+    let mut actual_text_glyph_tm: Option<[f32; 6]> = None; // text matrix at first glyph inside BDC
     /// Get the innermost MCID from the marked content stack.
     fn current_mcid(stack: &[MarkedContentEntry]) -> Option<i64> {
         stack.iter().rev().find_map(|e| e.mcid)
@@ -349,8 +350,15 @@ pub(crate) fn extract_page_text_items(
                             )
                         })
                     });
-                    // ActualText: suppress glyph extraction, just advance text matrix
+                    // ActualText: suppress glyph extraction, just advance text matrix.
+                    // Capture the FIRST glyph's text matrix as the rendering position
+                    // for the ActualText item. Td ops between BDC and the first Tj
+                    // may have moved the position to the correct line — the BDC-entry
+                    // position (actual_text_start_tm) can be on the previous line.
                     if suppress_glyph_extraction {
+                        if actual_text_glyph_tm.is_none() {
+                            actual_text_glyph_tm = Some(text_matrix);
+                        }
                         if let Some(w_ts) = w_ts_opt {
                             text_matrix[4] += w_ts * text_matrix[0];
                             text_matrix[5] += w_ts * text_matrix[1];
@@ -377,6 +385,7 @@ pub(crate) fn extract_page_text_items(
                         &font_encodings,
                         &encoding_cache,
                         &mut cmap_decisions,
+                        &font_widths,
                     ) {
                         let combined = multiply_matrices(&text_matrix, &ctm);
                         let rendered_size = effective_font_size(current_font_size, &combined);
@@ -425,6 +434,10 @@ pub(crate) fn extract_page_text_items(
                         let font_info = font_widths.get(&current_font);
                         let is_invisible = (text_rendering_mode == 3 && !include_invisible)
                             || suppress_glyph_extraction;
+                        // Capture first-glyph position for ActualText
+                        if suppress_glyph_extraction && actual_text_glyph_tm.is_none() {
+                            actual_text_glyph_tm = Some(text_matrix);
+                        }
 
                         // Compute space threshold based on font metrics when available
                         let space_threshold = if let Some(font_info) = font_info {
@@ -521,6 +534,7 @@ pub(crate) fn extract_page_text_items(
                                     &font_encodings,
                                     &encoding_cache,
                                     &mut cmap_decisions,
+                                    &font_widths,
                                 ) {
                                     current_text.push_str(&text);
                                 }
@@ -608,6 +622,7 @@ pub(crate) fn extract_page_text_items(
                         &font_encodings,
                         &encoding_cache,
                         &mut cmap_decisions,
+                        &font_widths,
                     ) {
                         if !text.trim().is_empty() {
                             let combined = multiply_matrices(&text_matrix, &ctm);
@@ -700,6 +715,7 @@ pub(crate) fn extract_page_text_items(
                 if actual_text.is_some() {
                     suppress_glyph_extraction = true;
                     actual_text_start_tm = Some(text_matrix);
+                    actual_text_glyph_tm = None; // reset — will be captured at first Tj/TJ
                 }
                 marked_content_stack.push(MarkedContentEntry { actual_text, mcid });
             }
@@ -707,8 +723,13 @@ pub(crate) fn extract_page_text_items(
                 // End Marked Content — emit ActualText item with correct width
                 if let Some(entry) = marked_content_stack.pop() {
                     if let Some(at) = entry.actual_text {
-                        // Compute width from text matrix advancement during BDC..EMC
-                        if let Some(start_tm) = actual_text_start_tm.take() {
+                        // Use the first-glyph position (if available) instead of the
+                        // BDC-entry position. Td operators between BDC and the first
+                        // Tj may have moved the text position to the correct line —
+                        // the BDC-entry position can be on the previous line.
+                        let glyph_tm = actual_text_glyph_tm.take();
+                        let entry_tm = actual_text_start_tm.take();
+                        if let Some(start_tm) = glyph_tm.or(entry_tm) {
                             let combined = multiply_matrices(&start_tm, &ctm);
                             if combined[0].abs() >= combined[1].abs() {
                                 rotation_votes.horizontal += 1;
@@ -994,9 +1015,17 @@ pub(crate) fn extract_page_text_items(
     // producing thousands of identical rects that yield a degenerate grid.
     // After dedup, if too few unique clip rects remain we fall through to
     // fill rects (explicitly drawn visible rectangles).
+    //
+    // When fill rects substantially outnumber clip rects, the clips are
+    // typically section-level wrappers and the fills are the actual table
+    // cell backgrounds (e.g. shaded-header tables drawn with `m`/`l`/`h`/`f*`
+    // sequences). In that case, prefer fills.
     if rects.is_empty() {
         dedup_rects(&mut clip_rects);
-        if clip_rects.len() >= 4 {
+        let prefer_fills = !fill_rects.is_empty() && fill_rects.len() >= clip_rects.len() * 3;
+        if prefer_fills {
+            rects = fill_rects;
+        } else if clip_rects.len() >= 4 {
             rects = clip_rects;
         } else if !fill_rects.is_empty() {
             rects = fill_rects;
@@ -1009,11 +1038,12 @@ pub(crate) fn extract_page_text_items(
     // Some PDFs embed landscape content in portrait pages using a rotated text
     // matrix (e.g. [0, b, -b, 0, tx, ty] for 90° CCW).  The layout engine
     // assumes x=horizontal, y=vertical — so we swap coordinates to match.
-    let (items, rects, lines) = correct_rotated_page(items, rects, lines, &rotation_votes);
+    let (items, rects, lines, coords_rotated) =
+        correct_rotated_page(items, rects, lines, &rotation_votes);
 
     let items = super::merge_text_items(items);
     let items = super::merge_subscript_items(items);
-    Ok(((items, rects, lines), has_gid_fonts))
+    Ok(((items, rects, lines), has_gid_fonts, coords_rotated))
 }
 
 /// Counts of text operators with horizontal vs rotated combined matrices.
@@ -1030,9 +1060,9 @@ fn correct_rotated_page(
     mut rects: Vec<PdfRect>,
     mut lines: Vec<PdfLine>,
     votes: &RotationVotes,
-) -> (Vec<TextItem>, Vec<PdfRect>, Vec<PdfLine>) {
+) -> (Vec<TextItem>, Vec<PdfRect>, Vec<PdfLine>, bool) {
     if items.len() < 2 {
-        return (items, rects, lines);
+        return (items, rects, lines, false);
     }
 
     // Use the combined-matrix direction votes collected during extraction.
@@ -1041,7 +1071,7 @@ fn correct_rotated_page(
     let total_votes = votes.horizontal + votes.rotated;
     if total_votes == 0 || votes.rotated * 3 < total_votes * 2 {
         // Less than ~67% of text operators are rotated → not a rotated page
-        return (items, rects, lines);
+        return (items, rects, lines, false);
     }
 
     log::debug!(
@@ -1092,7 +1122,7 @@ fn correct_rotated_page(
         line.y2 = new_y2;
     }
 
-    (items, rects, lines)
+    (items, rects, lines, true)
 }
 
 /// Remove near-duplicate rects (same coordinates within 0.5 pt tolerance).
@@ -1228,7 +1258,7 @@ mod tests {
 
         let font_cmaps = FontCMaps::from_doc(&doc);
         let result = extract_page_text_items(&doc, page_id, 1, &font_cmaps, false).unwrap();
-        let ((items, rects, lines), _has_gid) = result;
+        let ((items, rects, lines), _has_gid, _coords_rotated) = result;
         assert!(items.is_empty());
         assert!(rects.is_empty());
         assert!(lines.is_empty());

src/extractor/fonts.rs

@@ -720,7 +720,11 @@ pub(crate) fn extract_text_from_operand(
     font_encodings: &PageFontEncodings,
     encoding_cache: &HashMap<String, Encoding<'_>>,
     cmap_decisions: &mut CMapDecisionCache,
+    font_widths: &PageFontWidths,
 ) -> Option<String> {
+    let is_type0_cid_font = font_widths
+        .get(current_font)
+        .is_some_and(|info| info.is_cid);
     let result = (|| -> Option<String> {
         if let Object::String(bytes, _) = obj {
             let mut decode_with_entry = |entry: &crate::tounicode::CMapEntry| -> Option<String> {
@@ -962,7 +966,31 @@ pub(crate) fn extract_text_from_operand(
                 return Some(symbol_text);
             }
 
-            // Latin-1 fallback
+            // Latin-1 fallback. Safe ONLY for fonts that use single-byte
+            // encodings — for these, an unmapped byte is a valid character
+            // code in Latin-1/WinAnsi space. CID fonts (Type0 / Identity-H)
+            // emit multi-byte CIDs that aren't characters; per-byte Latin-1
+            // produces mojibake (e.g. 2-byte CID 0xCDD9 → "ÍÙ" for the
+            // production scrape_id 019de78c-... samples).
+            //
+            // For a CID font (has_cmap is set OR a /ToUnicode reference
+            // exists) with any non-ASCII bytes, emit a single U+FFFD per
+            // CID instead. This both replaces the mojibake with a proper
+            // "decode failed" marker AND keeps `detect_encoding_issues`
+            // tripping so the page is flagged for OCR — the existing
+            // garbage-detection path that the high-Latin-1 mojibake used
+            // to satisfy by accident.
+            if is_type0_cid_font && bytes.iter().any(|&b| b > 0x7F) {
+                // 2-byte CIDs (Identity-H) are by far the common case; for
+                // an odd byte count we still emit at least one marker so
+                // detection downstream fires.
+                let cid_count = (bytes.len() / 2).max(1);
+                return Some("\u{FFFD}".repeat(cid_count));
+            }
+            // Pure ASCII bytes round-trip safely (Latin-1 == ASCII for
+            // 0x00..=0x7F), and non-CID (Type1 / TrueType / Type3) fonts
+            // use single-byte encodings where Latin-1 fallback is the
+            // canonical interpretation.
             Some(bytes.iter().map(|&b| b as char).collect())
         } else {
             None
@@ -1213,4 +1241,97 @@ mod tests {
         let bad = "###!!!@@@$$$";
         assert!(score_text(good) > score_text(bad));
     }
+
+    #[test]
+    fn cid_font_with_unparseable_cmap_does_not_emit_latin1_mojibake() {
+        // Type0/CID font (font_widths reports `is_cid=true`) where the
+        // ToUnicode CMap couldn't be parsed (FontCMaps doesn't have the
+        // obj_num). Bytes are a 2-byte CID stream containing high bytes
+        // that aren't valid UTF-8 — exactly the case in the production
+        // samples (Identity-H text where the ToUnicode CMap was missing
+        // or malformed, scrape_id 019de78c-..., e.g. "Í Ù Z)¿").
+        //
+        // Without the guard, the function falls through to the byte-by-byte
+        // Latin-1 fallback and produces "ÍÙ" (U+00CD U+00D9). The correct
+        // behavior is to emit U+FFFD per CID so downstream
+        // `detect_encoding_issues` flags the page for OCR.
+        let bytes = vec![0xCD_u8, 0xD9, 0xCD, 0xD9];
+        let obj = Object::String(bytes, lopdf::StringFormat::Hexadecimal);
+
+        let font_cmaps = FontCMaps::default();
+        let mut font_tounicode_refs: HashMap<String, u32> = HashMap::new();
+        font_tounicode_refs.insert("F0".to_string(), 999);
+        let inline_cmaps = HashMap::new();
+        let font_encodings: PageFontEncodings = HashMap::new();
+        let encoding_cache: HashMap<String, Encoding<'_>> = HashMap::new();
+        let mut decisions = CMapDecisionCache::new();
+        let mut font_widths: PageFontWidths = HashMap::new();
+        font_widths.insert("F0".to_string(), make_font_info(&[], 1000, true));
+
+        let result = extract_text_from_operand(
+            &obj,
+            "F0",
+            None,
+            &font_cmaps,
+            &font_tounicode_refs,
+            &inline_cmaps,
+            &font_encodings,
+            &encoding_cache,
+            &mut decisions,
+            &font_widths,
+        );
+
+        let text = result.expect("CID font fallback should still emit a marker");
+        assert!(
+            !text.contains('\u{00CD}') && !text.contains('\u{00D9}'),
+            "CID font with unparseable CMap leaked Latin-1 mojibake: {text:?}"
+        );
+        assert!(
+            text.contains('\u{FFFD}'),
+            "CID font with unparseable CMap should emit U+FFFD so detect_encoding_issues fires: {text:?}"
+        );
+    }
+
+    #[test]
+    fn simple_font_latin1_fallback_passes_high_bytes_through() {
+        // A Type1/TrueType simple font (is_cid=false) with a `/ToUnicode`
+        // reference but no usable CMap and no `/Differences` map.
+        // Per-byte Latin-1 IS the canonical interpretation here — these
+        // bytes are character codes, not CIDs. The CID guard must NOT
+        // strip them. Reproduces the false positive that an earlier
+        // version of the guard introduced for fonts in PDFs like
+        // pdf-evals/Navigating-Artificial-Intelligence-..., where bytes
+        // like 0xB6 are legitimate Latin-1 character codes.
+        let bytes = vec![0x24_u8, 0x47, 0xB6, 0x56]; // "$G¶V"
+        let obj = Object::String(bytes, lopdf::StringFormat::Hexadecimal);
+
+        let font_cmaps = FontCMaps::default();
+        let mut font_tounicode_refs: HashMap<String, u32> = HashMap::new();
+        font_tounicode_refs.insert("F1".to_string(), 999);
+        let inline_cmaps = HashMap::new();
+        let font_encodings: PageFontEncodings = HashMap::new();
+        let encoding_cache: HashMap<String, Encoding<'_>> = HashMap::new();
+        let mut decisions = CMapDecisionCache::new();
+        let mut font_widths: PageFontWidths = HashMap::new();
+        font_widths.insert("F1".to_string(), make_font_info(&[], 1000, false));
+
+        let text = extract_text_from_operand(
+            &obj,
+            "F1",
+            None,
+            &font_cmaps,
+            &font_tounicode_refs,
+            &inline_cmaps,
+            &font_encodings,
+            &encoding_cache,
+            &mut decisions,
+            &font_widths,
+        )
+        .expect("simple font should round-trip Latin-1 bytes");
+        assert_eq!(text, "$G\u{00B6}V");
+        assert!(
+            !text.contains('\u{FFFD}'),
+            "simple font fallback must not stamp FFFD over legitimate bytes: {text:?}"
+        );
+    }
 }

src/extractor/layout.rs

@@ -35,6 +35,7 @@ pub(crate) fn detect_columns(
     if page_items.is_empty() {
         return vec![];
     }
+    debug!("page {}: detect_columns: {} items", page, page_items.len());
 
     // Find page bounds
     let x_min = page_items.iter().map(|i| i.x).fold(f32::INFINITY, f32::min);
@@ -163,10 +164,17 @@ pub(crate) fn detect_columns(
                 }
             }
         }
+        // Try XY-cut fallback before giving up
+        if let Some(columns) = try_xy_cut_split(&page_items, x_min, x_max, page) {
+            return columns;
+        }
         return vec![ColumnRegion { x_min, x_max }];
     }
 
-    return validate_and_build_columns(
+    // Try center-based assignment first (handles asymmetric layouts / sidebars
+    // better than edge-based). Fall back to edge-based if center produces
+    // a degenerate split (one side empty).
+    let result = validate_and_build_columns(
         &valleys,
         &page_items,
         x_min,
@@ -175,8 +183,177 @@ pub(crate) fn detect_columns(
         MIN_ITEMS_PER_COLUMN,
         MIN_VERTICAL_SPAN_RATIO,
         page,
-        false, // edge-based assignment for absolute valleys
+        true, // center-based assignment
     );
+    if result.len() > 1 {
+        return result;
+    }
+    let result = validate_and_build_columns(
+        &valleys,
+        &page_items,
+        x_min,
+        BIN_WIDTH,
+        x_max,
+        MIN_ITEMS_PER_COLUMN,
+        MIN_VERTICAL_SPAN_RATIO,
+        page,
+        false, // edge-based fallback
+    );
+    if result.len() > 1 {
+        return result;
+    }
+
+    // Fallback: XY-cut style gap detection.  When the histogram finds no
+    // clear valleys (common with asymmetric/sidebar layouts), look for the
+    // largest horizontal gap between item edges.  This is a simplified
+    // single-level XY-cut inspired by opendataloader's XY-Cut++ algorithm.
+    if page_items.len() >= 20 && !page_has_table {
+        if let Some(columns) = try_xy_cut_split(&page_items, x_min, x_max, page) {
+            return columns;
+        }
+    }
+
+    vec![ColumnRegion { x_min, x_max }]
+}
+
+/// Simplified single-level XY-cut: find the largest horizontal gap between
+/// item right-edges and left-edges.  If the gap is wide enough and both sides
+/// have sufficient items with vertical overlap, split into two columns.
+///
+/// Inspired by opendataloader's XY-Cut++ algorithm but without full recursion.
+/// Handles asymmetric layouts (sidebars) that the histogram misses because
+/// the narrow column has too few items to register in the occupancy profile.
+fn try_xy_cut_split(
+    page_items: &[&TextItem],
+    page_x_min: f32,
+    page_x_max: f32,
+    page: u32,
+) -> Option<Vec<ColumnRegion>> {
+    const MIN_GAP: f32 = 15.0; // minimum gap to consider a split
+    const MIN_ITEMS_MAJOR: usize = 10; // major column must have ≥10 items
+    const MIN_ITEMS_MINOR: usize = 3; // minor column (sidebar) must have ≥3
+
+    let page_width = page_x_max - page_x_min;
+    if page_width < 200.0 {
+        return None;
+    }
+
+    // Collect all item edges: (right_edge, left_edge) pairs sorted by right_edge
+    // The gap between one item's right edge and the next item's left edge
+    // reveals column gutters.
+    let mut edges: Vec<(f32, f32)> = page_items
+        .iter()
+        .map(|i| (i.x, i.x + effective_width(i)))
+        .collect();
+    edges.sort_by(|a, b| a.0.total_cmp(&b.0));
+
+    // Find the largest gap between consecutive items (by left edge).
+    // Use a sweep: sort left edges, find max gap between sorted right edges
+    // of items to the left and left edges of items to the right.
+    let mut left_edges: Vec<f32> = page_items.iter().map(|i| i.x).collect();
+    left_edges.sort_by(|a, b| a.total_cmp(b));
+
+    // Build prefix max of right edges (for items sorted by left edge)
+    let mut sorted_by_left: Vec<(f32, f32)> = page_items
+        .iter()
+        .map(|i| (i.x, i.x + effective_width(i)))
+        .collect();
+    sorted_by_left.sort_by(|a, b| a.0.total_cmp(&b.0));
+
+    let mut best_gap = 0.0f32;
+    let mut best_split = 0.0f32;
+    let mut max_right_so_far = f32::NEG_INFINITY;
+
+    for i in 0..sorted_by_left.len() - 1 {
+        let (_, right) = sorted_by_left[i];
+        max_right_so_far = max_right_so_far.max(right);
+
+        let (next_left, _) = sorted_by_left[i + 1];
+        let gap = next_left - max_right_so_far;
+        if gap > best_gap {
+            best_gap = gap;
+            best_split = (max_right_so_far + next_left) / 2.0;
+        }
+    }
+
+    if best_gap < MIN_GAP {
+        return None;
+    }
+
+    // Don't split at page margins (within 10% of edges)
+    let margin = page_width * 0.10;
+    if best_split - page_x_min < margin || page_x_max - best_split < margin {
+        return None;
+    }
+
+    // Count items on each side
+    let left_count = page_items
+        .iter()
+        .filter(|i| i.x + effective_width(i) / 2.0 <= best_split)
+        .count();
+    let right_count = page_items
+        .iter()
+        .filter(|i| i.x + effective_width(i) / 2.0 > best_split)
+        .count();
+
+    let (minor, major) = if left_count <= right_count {
+        (left_count, right_count)
+    } else {
+        (right_count, left_count)
+    };
+
+    if major < MIN_ITEMS_MAJOR || minor < MIN_ITEMS_MINOR {
+        return None;
+    }
+
+    // Check vertical overlap — both sides should span a meaningful Y range
+    let left_items: Vec<&&TextItem> = page_items
+        .iter()
+        .filter(|i| i.x + effective_width(i) / 2.0 <= best_split)
+        .collect();
+    let right_items: Vec<&&TextItem> = page_items
+        .iter()
+        .filter(|i| i.x + effective_width(i) / 2.0 > best_split)
+        .collect();
+
+    let l_y_min = left_items.iter().map(|i| i.y).fold(f32::INFINITY, f32::min);
+    let l_y_max = left_items
+        .iter()
+        .map(|i| i.y)
+        .fold(f32::NEG_INFINITY, f32::max);
+    let r_y_min = right_items
+        .iter()
+        .map(|i| i.y)
+        .fold(f32::INFINITY, f32::min);
+    let r_y_max = right_items
+        .iter()
+        .map(|i| i.y)
+        .fold(f32::NEG_INFINITY, f32::max);
+
+    let overlap_min = l_y_min.max(r_y_min);
+    let overlap_max = l_y_max.min(r_y_max);
+    let overlap = (overlap_max - overlap_min).max(0.0);
+    let y_range = (l_y_max.max(r_y_max) - l_y_min.min(r_y_min)).max(1.0);
+
+    if overlap / y_range < 0.20 {
+        return None;
+    }
+
+    debug!(
+        "page {}: XY-cut split at x={:.1} (gap={:.1}pt, left={}, right={})",
+        page, best_split, best_gap, left_count, right_count
+    );
+
+    Some(vec![
+        ColumnRegion {
+            x_min: page_x_min,
+            x_max: best_split,
+        },
+        ColumnRegion {
+            x_min: best_split,
+            x_max: page_x_max,
+        },
+    ])
 }
 
 /// Check whether each proposed column contains paragraph-like content.
@@ -218,7 +395,7 @@ fn columns_have_prose(columns: &[ColumnRegion], items: &[&TextItem]) -> bool {
 
         // Sort by Y descending (top of page = higher Y in PDF coords)
         let mut sorted: Vec<&TextItem> = col_items;
-        sorted.sort_by(|a, b| b.y.partial_cmp(&a.y).unwrap_or(std::cmp::Ordering::Equal));
+        sorted.sort_by(|a, b| b.y.total_cmp(&a.y));
 
         // Group into lines by Y-proximity and measure fill + item count
         let mut full_lines = 0usize;
@@ -449,6 +626,33 @@ fn find_relative_valleys(
     valleys
 }
 
+/// Detect whether a side of a gutter consists predominantly of list-marker
+/// glyphs (•, ●, ○, ◦, ▪, ▫, ◆, ◇). A column of bullets on the left margin
+/// creates a spurious histogram valley between the bullet and the content.
+/// Treating it as a real column splits each list item's text across two
+/// "columns," so we reject these candidates.
+fn is_list_marker_column(items: &[&&TextItem]) -> bool {
+    const LIST_MARKERS: &[char] = &['•', '●', '○', '◦', '▪', '▫', '◆', '◇', '■', '□'];
+    if items.is_empty() {
+        return false;
+    }
+    let marker_count = items
+        .iter()
+        .filter(|i| {
+            let t = i.text.trim();
+            let mut chars = t.chars();
+            match (chars.next(), chars.next()) {
+                (Some(c), None) => LIST_MARKERS.contains(&c),
+                _ => false,
+            }
+        })
+        .count();
+    // Require ≥80% of items on this side to be standalone markers. A handful
+    // of non-marker items (stray page numbers, footnote refs) shouldn't
+    // defeat the check.
+    marker_count as f32 / items.len() as f32 >= 0.8
+}
+
 /// Validate valley candidates with vertical consistency checks and build column regions.
 ///
 /// When `center_assign` is true, items are assigned to columns based on their
@@ -505,7 +709,28 @@ fn validate_and_build_columns(
             })
             .collect();
 
-        if left_items.len() < min_items || right_items.len() < min_items {
+        // Require both sides to have items. Symmetric layout needs min_items
+        // on each side. Asymmetric layouts (sidebars) are accepted when the
+        // dominant side has ≥ min_items and the smaller side has ≥ 3 items.
+        let (smaller, larger) = if left_items.len() <= right_items.len() {
+            (left_items.len(), right_items.len())
+        } else {
+            (right_items.len(), left_items.len())
+        };
+        if larger < min_items || smaller < 3 {
+            continue;
+        }
+
+        // Reject valleys where the smaller side is just a column of list
+        // markers (bullets aligned at the left margin). This is a common
+        // pattern in PDFs where ● starts each list item: histogram detection
+        // sees the gap between bullet and content as a gutter.
+        let smaller_items: &[&&TextItem] = if left_items.len() <= right_items.len() {
+            &left_items
+        } else {
+            &right_items
+        };
+        if is_list_marker_column(smaller_items) {
             continue;
         }
 
@@ -616,7 +841,7 @@ fn identify_spanning_lines(items: &[TextItem], columns: &[ColumnRegion]) -> Vec<
     // Build (original_index, y) pairs sorted by Y descending for grouping
     let mut indexed: Vec<(usize, f32)> =
         items.iter().enumerate().map(|(i, it)| (i, it.y)).collect();
-    indexed.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal));
+    indexed.sort_by(|a, b| b.1.total_cmp(&a.1));
 
     // Group by Y-proximity into rough lines (as index sets)
     let mut groups: Vec<Vec<usize>> = Vec::new();
@@ -778,7 +1003,7 @@ pub(crate) fn is_newspaper_layout(
                             return 0.0;
                         }
                         let mut ys: Vec<f32> = lines.iter().map(|l| l.y).collect();
-                        ys.sort_by(|a, b| a.partial_cmp(b).unwrap());
+                        ys.sort_by(|a, b| a.total_cmp(b));
                         let span = ys.last().unwrap() - ys.first().unwrap();
                         span / (lines.len() as f32 - 1.0)
                     };
@@ -845,7 +1070,7 @@ fn split_column_stragglers(lines: Vec<TextLine>) -> (Vec<TextLine>, Vec<TextLine
 
     // Median gap = typical line spacing
     let mut sorted_gaps = gaps.clone();
-    sorted_gaps.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+    sorted_gaps.sort_by(|a, b| a.total_cmp(b));
     let median_gap = sorted_gaps[sorted_gaps.len() / 2];
 
     // A gap > 3× median (min 30pt) indicates a break between content clusters
@@ -1078,9 +1303,8 @@ pub(crate) fn group_into_lines_with_thresholds(
                     }
                 }
 
-                above.sort_by(|a, b| b.y.partial_cmp(&a.y).unwrap_or(std::cmp::Ordering::Equal));
-                below_spanning
-                    .sort_by(|a, b| b.y.partial_cmp(&a.y).unwrap_or(std::cmp::Ordering::Equal));
+                above.sort_by(|a, b| b.y.total_cmp(&a.y));
+                below_spanning.sort_by(|a, b| b.y.total_cmp(&a.y));
 
                 all_lines.extend(above);
                 for col in core_columns {
@@ -1101,16 +1325,13 @@ pub(crate) fn group_into_lines_with_thresholds(
 
                 // Sort by Y descending (top-first), then by X for same-Y lines
                 all_page_lines.sort_by(|a, b| {
-                    b.y.partial_cmp(&a.y)
-                        .unwrap_or(std::cmp::Ordering::Equal)
-                        .then(
-                            a.items
-                                .first()
-                                .map(|i| i.x)
-                                .unwrap_or(0.0)
-                                .partial_cmp(&b.items.first().map(|i| i.x).unwrap_or(0.0))
-                                .unwrap_or(std::cmp::Ordering::Equal),
-                        )
+                    b.y.total_cmp(&a.y).then(
+                        a.items
+                            .first()
+                            .map(|i| i.x)
+                            .unwrap_or(0.0)
+                            .total_cmp(&b.items.first().map(|i| i.x).unwrap_or(0.0)),
+                    )
                 });
 
                 // Merge lines at the same Y (within tolerance) into single lines
@@ -1185,11 +1406,7 @@ fn group_single_column(items: Vec<TextItem>, adaptive_threshold: f32) -> Vec<Tex
     let items = if use_y_sorting {
         // Sort by Y descending (top to bottom in PDF coords)
         let mut sorted = items;
-        sorted.sort_by(|a, b| {
-            b.y.partial_cmp(&a.y)
-                .unwrap_or(std::cmp::Ordering::Equal)
-                .then(a.x.partial_cmp(&b.x).unwrap_or(std::cmp::Ordering::Equal))
-        });
+        sorted.sort_by(|a, b| b.y.total_cmp(&a.y).then(a.x.total_cmp(&b.x)));
         sorted
     } else {
         items
@@ -1603,6 +1820,63 @@ mod tests {
         );
     }
 
+    #[test]
+    fn bullet_marker_column_not_detected_as_column() {
+        // Pattern: every line is `● <content>`, with ● at x=90 and content
+        // starting at x=104. Histogram detection sees a gutter between them
+        // and would split the page into a "bullet column" and "content column",
+        // scrambling every list item.
+        let mut items = Vec::new();
+        for i in 0..15 {
+            let y = 750.0 - i as f32 * 30.0;
+            items.push(make_item(1, 90.0, y, "●"));
+            items.push(make_item(
+                1,
+                104.0,
+                y,
+                "FullContentLineTextHere________________",
+            ));
+        }
+        // Pad with content to satisfy min item count for column detection.
+        for i in 0..15 {
+            let y = 300.0 - i as f32 * 14.0;
+            items.push(make_item(1, 72.0, y, "FootnoteText_____________________"));
+        }
+
+        let cols = detect_columns(&items, 1, false);
+        assert_eq!(
+            cols.len(),
+            1,
+            "Bullet markers aligned at left margin should not be treated as their own column"
+        );
+    }
+
+    #[test]
+    fn is_list_marker_column_detects_bullets() {
+        let items = vec![
+            make_item(1, 90.0, 100.0, "●"),
+            make_item(1, 90.0, 114.0, "●"),
+            make_item(1, 90.0, 128.0, "●"),
+            make_item(1, 90.0, 142.0, "●"),
+        ];
+        let refs: Vec<&TextItem> = items.iter().collect();
+        let wrapped: Vec<&&TextItem> = refs.iter().collect();
+        assert!(is_list_marker_column(&wrapped));
+    }
+
+    #[test]
+    fn is_list_marker_column_rejects_prose() {
+        let items = vec![
+            make_item(1, 30.0, 100.0, "Regular prose line"),
+            make_item(1, 30.0, 114.0, "Another sentence"),
+            make_item(1, 30.0, 128.0, "Third line"),
+            make_item(1, 30.0, 142.0, "Fourth line"),
+        ];
+        let refs: Vec<&TextItem> = items.iter().collect();
+        let wrapped: Vec<&&TextItem> = refs.iter().collect();
+        assert!(!is_list_marker_column(&wrapped));
+    }
+
     #[test]
     fn premask_narrow_line_not_masked() {
         // Items that form a line spanning only ~40% of column width → not masked

src/extractor/mod.rs

@@ -36,26 +36,14 @@ pub(crate) use layout::ColumnRegion;
 /// Extract text from PDF file as plain string
 pub fn extract_text<P: AsRef<Path>>(path: P) -> Result<String, PdfError> {
     crate::validate_pdf_file(&path)?;
-    let doc = match Document::load(&path) {
-        Ok(d) => d,
-        Err(ref e) if crate::is_encrypted_lopdf_error(e) => {
-            Document::load_with_password(&path, "")?
-        }
-        Err(e) => return Err(e.into()),
-    };
+    let (doc, _) = crate::load_document_from_path(&path)?;
     extract_text_from_doc(&doc)
 }
 
 /// Extract text from PDF memory buffer
 pub fn extract_text_mem(buffer: &[u8]) -> Result<String, PdfError> {
     crate::validate_pdf_bytes(buffer)?;
-    let doc = match Document::load_mem(buffer) {
-        Ok(d) => d,
-        Err(ref e) if crate::is_encrypted_lopdf_error(e) => {
-            Document::load_mem_with_options(buffer, lopdf::LoadOptions::with_password(""))?
-        }
-        Err(e) => return Err(e.into()),
-    };
+    let (doc, _) = crate::load_document_from_mem(buffer)?;
     extract_text_from_doc(&doc)
 }
 
@@ -91,13 +79,7 @@ pub(crate) fn extract_text_with_positions_and_rects<P: AsRef<Path>>(
     page_filter: Option<&HashSet<u32>>,
 ) -> Result<PageExtraction, PdfError> {
     crate::validate_pdf_file(&path)?;
-    let doc = match Document::load(&path) {
-        Ok(d) => d,
-        Err(ref e) if crate::is_encrypted_lopdf_error(e) => {
-            Document::load_with_password(&path, "")?
-        }
-        Err(e) => return Err(e.into()),
-    };
+    let (doc, _) = crate::load_document_from_path(&path)?;
     let font_cmaps = FontCMaps::from_doc(&doc);
     let (extraction, _thresholds, _gid_pages) =
         extract_positioned_text_from_doc(&doc, &font_cmaps, page_filter)?;
@@ -124,13 +106,7 @@ pub(crate) fn extract_text_with_positions_mem_and_rects(
     page_filter: Option<&HashSet<u32>>,
 ) -> Result<PageExtraction, PdfError> {
     crate::validate_pdf_bytes(buffer)?;
-    let doc = match Document::load_mem(buffer) {
-        Ok(d) => d,
-        Err(ref e) if crate::is_encrypted_lopdf_error(e) => {
-            Document::load_mem_with_options(buffer, lopdf::LoadOptions::with_password(""))?
-        }
-        Err(e) => return Err(e.into()),
-    };
+    let (doc, _) = crate::load_document_from_mem(buffer)?;
     let font_cmaps = FontCMaps::from_doc(&doc);
     let (extraction, _thresholds, _gid_pages) =
         extract_positioned_text_from_doc(&doc, &font_cmaps, page_filter)?;
@@ -188,7 +164,7 @@ fn extract_positioned_text_impl(
                 continue;
             }
         }
-        let ((mut items, rects, lines), has_gid_fonts) =
+        let ((mut items, rects, lines), has_gid_fonts, _coords_rotated) =
             extract_page_text_items(doc, page_id, *page_num, font_cmaps, include_invisible)?;
         if has_gid_fonts {
             gid_encoded_pages.insert(*page_num);
@@ -334,17 +310,14 @@ pub(crate) fn merge_text_items(items: Vec<TextItem>) -> Vec<TextItem> {
     for (_, _, group) in &mut line_groups {
         let rtl = is_rtl_text(group.iter().map(|i| &i.text));
         if rtl {
-            group.sort_by(|a, b| b.x.partial_cmp(&a.x).unwrap_or(std::cmp::Ordering::Equal));
+            group.sort_by(|a, b| b.x.total_cmp(&a.x));
         } else {
-            group.sort_by(|a, b| a.x.partial_cmp(&b.x).unwrap_or(std::cmp::Ordering::Equal));
+            group.sort_by(|a, b| a.x.total_cmp(&b.x));
         }
     }
 
     // Sort groups by page then Y descending (top of page first)
-    line_groups.sort_by(|a, b| {
-        a.0.cmp(&b.0)
-            .then_with(|| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal))
-    });
+    line_groups.sort_by(|a, b| a.0.cmp(&b.0).then_with(|| b.1.total_cmp(&a.1)));
 
     let mut merged = Vec::new();
 
@@ -452,7 +425,7 @@ pub(crate) fn merge_subscript_items(items: Vec<TextItem>) -> Vec<TextItem> {
 
     for (_, _, mut group) in line_groups {
         // Sort by X position
-        group.sort_by(|a, b| a.x.partial_cmp(&b.x).unwrap_or(std::cmp::Ordering::Equal));
+        group.sort_by(|a, b| a.x.total_cmp(&b.x));
 
         // Find the dominant (most common) font size in this group
         let max_fs = group.iter().map(|i| i.font_size).fold(0.0_f32, f32::max);

src/extractor/xobjects.rs

@@ -373,6 +373,7 @@ fn extract_form_xobject_text_inner(
                         &font_encodings,
                         &encoding_cache,
                         cmap_decisions,
+                        &font_widths,
                     ) {
                         let combined = multiply_matrices(&text_matrix, &ctm);
                         let rendered_size = effective_font_size(current_font_size, &combined);
@@ -517,6 +518,7 @@ fn extract_form_xobject_text_inner(
                                     &font_encodings,
                                     &encoding_cache,
                                     cmap_decisions,
+                                    &font_widths,
                                 ) {
                                     current_text.push_str(&text);
                                 }

src/markdown/analysis.rs

@@ -8,6 +8,24 @@ use log::debug;
 /// Font statistics for a document
 pub(crate) struct FontStats {
     pub(crate) most_common_size: f32,
+    /// Font size frequency distribution (size_key → line count).
+    /// Used for rarity-based heading detection.
+    pub(crate) size_counts: HashMap<i32, usize>,
+    /// Total number of lines counted.
+    pub(crate) total_lines: usize,
+}
+
+/// Compute how rare a font size is in the document (0.0 = most common, 1.0 = unique).
+/// Mirrors opendataloader's font rarity boosting approach: heading fonts appear on
+/// far fewer lines than body text, so their percentile rank is high.
+pub(crate) fn font_size_rarity(font_size: f32, stats: &FontStats) -> f32 {
+    if stats.total_lines == 0 {
+        return 0.0;
+    }
+    let key = (font_size * 10.0) as i32;
+    let count = stats.size_counts.get(&key).copied().unwrap_or(0);
+    // Rarity = 1 - (frequency ratio). A size used on 1/100 lines has rarity ~0.99.
+    1.0 - (count as f32 / stats.total_lines as f32)
 }
 
 /// Calculate font stats directly from items (before grouping into lines)
@@ -21,6 +39,8 @@ pub(crate) fn calculate_font_stats_from_items(items: &[TextItem]) -> FontStats {
         }
     }
 
+    let total_lines = size_counts.values().sum();
+
     // Break ties by preferring the smaller font size for deterministic output
     let most_common_size = size_counts
         .iter()
@@ -30,7 +50,11 @@ pub(crate) fn calculate_font_stats_from_items(items: &[TextItem]) -> FontStats {
         .map(|(size, _)| *size as f32 / 10.0)
         .unwrap_or(12.0);
 
-    FontStats { most_common_size }
+    FontStats {
+        most_common_size,
+        size_counts,
+        total_lines,
+    }
 }
 
 /// Calculate font stats from grouped lines
@@ -48,6 +72,8 @@ pub(crate) fn calculate_font_stats(lines: &[TextLine]) -> FontStats {
         }
     }
 
+    let total_lines = size_counts.values().sum();
+
     // Break ties by preferring the smaller font size for deterministic output
     let most_common_size = size_counts
         .iter()
@@ -57,7 +83,23 @@ pub(crate) fn calculate_font_stats(lines: &[TextLine]) -> FontStats {
         .map(|(size, _)| *size as f32 / 10.0)
         .unwrap_or(12.0);
 
-    FontStats { most_common_size }
+    FontStats {
+        most_common_size,
+        size_counts,
+        total_lines,
+    }
+}
+
+/// Determine the heading level for a bold-only line that didn't meet the font-size
+/// threshold.  These are common in academic papers where section headings are bold
+/// at the same size as body text.
+///
+/// Returns a level below the lowest font-size tier (or H2 when no tiers exist).
+pub(crate) fn bold_heading_level(heading_tiers: &[f32]) -> usize {
+    let level = heading_tiers.len() + 1;
+    // Clamp to 1..=6 — if no font-size tiers, bold headings become H2
+    // (H1 is reserved for titles which are typically larger)
+    level.clamp(2, 6)
 }
 
 /// Detect TOC-style lines that contain dot leaders (e.g., "Section Name .... 42").
@@ -121,7 +163,7 @@ pub(crate) fn compute_paragraph_threshold(lines: &[TextLine], base_size: f32) ->
         return fallback;
     }
 
-    gaps.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+    gaps.sort_by(|a, b| a.total_cmp(b));
 
     let median = gaps[gaps.len() / 2];
 
@@ -221,7 +263,7 @@ pub(crate) fn compute_heading_tiers(lines: &[TextLine], base_size: f32) -> Vec<f
     }
 
     // Sort descending
-    heading_sizes.sort_by(|a, b| b.partial_cmp(a).unwrap_or(std::cmp::Ordering::Equal));
+    heading_sizes.sort_by(|a, b| b.total_cmp(a));
 
     // Cluster sizes within 0.5pt into same tier (use first value as representative)
     let mut tiers: Vec<f32> = Vec::new();

src/markdown/classify.rs

@@ -4,13 +4,11 @@
 pub(crate) fn is_caption_line(text: &str) -> bool {
     let trimmed = text.trim();
 
-    // Common caption prefixes in multiple languages
-    let caption_prefixes = [
-        "Figure ",
+    // Caption prefixes that always match (always followed by identifiers)
+    let always_prefixes = [
         "Figura ",
         "Fig. ",
         "Fig ",
-        "Table ",
         "Tabela ",
         "Source:",
         "Fonte:",
@@ -27,23 +25,61 @@ pub(crate) fn is_caption_line(text: &str) -> bool {
         "Photo ",
         "Foto ",
     ];
-
-    // Check if line starts with a caption prefix
-    for prefix in &caption_prefixes {
+    for prefix in &always_prefixes {
         if trimmed.starts_with(prefix) {
             return true;
         }
     }
 
-    // Check case-insensitive patterns
+    // "Figure" and "Table" need a digit/reference after them to distinguish
+    // captions ("Table 1", "Figure 3.2") from headings ("Table of Contents")
+    for prefix in ["Figure ", "Table "] {
+        if let Some(rest) = trimmed.strip_prefix(prefix) {
+            if rest
+                .trim_start()
+                .starts_with(|c: char| c.is_ascii_digit() || c == '(' || c == '#')
+            {
+                return true;
+            }
+        }
+    }
+
+    // Check case-insensitive patterns — require digit or punctuation after
+    // prefix to avoid matching "Table of Contents" or "Figure drawing" etc.
     let lower = trimmed.to_lowercase();
-    if lower.starts_with("figure ") || lower.starts_with("table ") || lower.starts_with("source:") {
+    for pfx in ["figure ", "table "] {
+        if let Some(rest) = lower.strip_prefix(pfx) {
+            if rest
+                .trim_start()
+                .starts_with(|c: char| c.is_ascii_digit() || c == '(' || c == '#')
+            {
+                return true;
+            }
+        }
+    }
+    if lower.starts_with("source:") {
         return true;
     }
 
     false
 }
 
+/// Check if text starts with an unambiguous bullet marker (●, •, ○, ◦).
+///
+/// Narrower than [`is_list_item`]: it excludes numbered/lettered patterns
+/// like `1.` or `a)`, which legitimately appear as section headings in many
+/// documents. Used by the heading classifier to reject bullet lines without
+/// also demoting numbered headings.
+pub(crate) fn starts_with_bullet_marker(text: &str) -> bool {
+    let trimmed = text.trim_start();
+    trimmed.starts_with("• ")
+        || trimmed.starts_with("● ")
+        || trimmed.starts_with("○ ")
+        || trimmed.starts_with("◦ ")
+        || trimmed.starts_with("- ")
+        || trimmed.starts_with("* ")
+}
+
 /// Check if text looks like a list item
 pub(crate) fn is_list_item(text: &str) -> bool {
     let trimmed = text.trim_start();
@@ -95,6 +131,16 @@ pub(crate) fn format_list_item(text: &str) -> String {
         if let Some(rest) = trimmed.strip_prefix(*bullet) {
             return format!("- {}", rest.trim_start());
         }
+        // Bullet inside a leading bold/italic run (e.g. "**● Label:** rest").
+        // The run wraps both the marker and the following label because both
+        // use a bold font in the PDF.
+        for wrapper in ["**", "*"] {
+            if let Some(after_open) = trimmed.strip_prefix(wrapper) {
+                if let Some(rest) = after_open.strip_prefix(*bullet) {
+                    return format!("- {}{}", wrapper, rest.trim_start());
+                }
+            }
+        }
     }
 
     if trimmed.starts_with("- ") || trimmed.starts_with("* ") {
@@ -178,3 +224,42 @@ pub(crate) fn is_monospace_font(font_name: &str) -> bool {
 
     patterns.iter().any(|p| lower.contains(p))
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn format_list_item_plain_bullet() {
+        assert_eq!(format_list_item("● Item"), "- Item");
+        assert_eq!(format_list_item("• Item"), "- Item");
+    }
+
+    #[test]
+    fn format_list_item_bullet_inside_bold() {
+        // PDF that uses bold font for both the marker and the label produces
+        // a single bold run like "**● Label:** rest"; the bullet must still
+        // be stripped and the bold wrapper preserved on the label.
+        assert_eq!(
+            format_list_item("**● Fraud: Willing cooperation;**"),
+            "- **Fraud: Willing cooperation;**"
+        );
+        assert_eq!(
+            format_list_item("**● Label:** rest of line"),
+            "- **Label:** rest of line"
+        );
+        assert_eq!(format_list_item("*● Italic:* rest"), "- *Italic:* rest");
+    }
+
+    #[test]
+    fn format_list_item_already_dash() {
+        assert_eq!(format_list_item("- existing"), "- existing");
+    }
+
+    #[test]
+    fn is_list_item_with_bullet_space() {
+        assert!(is_list_item("● Item"));
+        assert!(is_list_item("• Item"));
+        assert!(is_list_item("- Item"));
+    }
+}

src/markdown/convert.rs

@@ -1,19 +1,154 @@
 //! Core line-to-markdown conversion loop with table/image interleaving.
 
-use std::collections::HashSet;
+use std::collections::{HashMap, HashSet};
 
 use crate::structure_tree::StructRole;
 use crate::types::TextLine;
 
 use super::analysis::{
-    calculate_font_stats, compute_heading_tiers, compute_paragraph_threshold, detect_header_level,
-    has_dot_leaders,
+    bold_heading_level, calculate_font_stats, compute_heading_tiers, compute_paragraph_threshold,
+    detect_header_level, font_size_rarity, has_dot_leaders,
+};
+use super::classify::{
+    format_list_item, is_caption_line, is_list_item, is_monospace_font, starts_with_bullet_marker,
 };
-use super::classify::{format_list_item, is_caption_line, is_list_item, is_monospace_font};
 use super::postprocess::clean_markdown;
 use super::preprocess::{merge_drop_caps, merge_heading_lines};
 use super::MarkdownOptions;
 
+/// 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.
+/// Returns the set of heading levels (1–6) that should be suppressed.
+///
+/// Some PDFs (e.g. British Academy grant guidance) tag every numbered paragraph
+/// line as H2, producing hundreds of false headings. We detect this by checking
+/// if any heading level accounts for >25% of tagged lines.
+fn detect_overused_struct_heading_levels(
+    lines: &[TextLine],
+    struct_roles: Option<
+        &std::collections::HashMap<u32, std::collections::HashMap<i64, StructRole>>,
+    >,
+) -> HashSet<usize> {
+    let mut overused = HashSet::new();
+    let Some(roles) = struct_roles else {
+        return overused;
+    };
+
+    let mut level_counts: HashMap<usize, usize> = HashMap::new();
+    let mut total = 0usize;
+
+    for line in lines {
+        if let Some(role) = resolve_line_struct_role(line, roles) {
+            total += 1;
+            if let Some(level) = struct_role_heading_level(&role) {
+                *level_counts.entry(level).or_insert(0) += 1;
+            }
+        }
+    }
+
+    if total < 20 {
+        return overused;
+    }
+
+    for (&level, &count) in &level_counts {
+        let ratio = count as f32 / total as f32;
+        if ratio > 0.15 {
+            log::debug!(
+                "struct heading H{} overused: {}/{} lines ({:.0}%), suppressing",
+                level,
+                count,
+                total,
+                ratio * 100.0
+            );
+            overused.insert(level);
+        }
+    }
+
+    overused
+}
+
+/// Pre-scan lines to find "isolated" ones: short lines with paragraph breaks both
+/// before and after.  These are heading candidates even at body font size — common
+/// in academic papers ("Acknowledgements", "B.3 Prompt Engineering").
+fn find_isolated_lines(lines: &[TextLine], base_size: f32, para_threshold: f32) -> HashSet<usize> {
+    let mut set = HashSet::new();
+    for i in 0..lines.len() {
+        let line = &lines[i];
+        let plain = line.text();
+        let trimmed = plain.trim();
+        let word_count = trimmed.split_whitespace().count();
+        if !(1..=6).contains(&word_count) || trimmed.len() <= 3 {
+            continue;
+        }
+        let font_size = line.items.first().map(|it| it.font_size).unwrap_or(0.0);
+        if font_size < base_size * 0.95 {
+            continue;
+        }
+        if is_list_item(trimmed) || is_caption_line(trimmed) {
+            continue;
+        }
+
+        // Reject lines that look like wrapped paragraph text:
+        // ends with hyphen, comma, preposition, or lowercase continuation
+        let last_char = trimmed.chars().last().unwrap_or(' ');
+        if last_char == '-' || last_char == ',' || last_char == ';' {
+            continue;
+        }
+        // Last word is a common continuation word → wrapped paragraph
+        let last_word = trimmed.split_whitespace().last().unwrap_or("");
+        let continuation_words = [
+            "the", "a", "an", "and", "or", "of", "in", "to", "for", "with", "by", "on", "at",
+            "from", "as", "is", "are", "was", "were", "be", "that", "this", "their", "its", "our",
+            "your", "has", "have", "had", "not",
+        ];
+        if continuation_words.contains(&last_word.to_lowercase().as_str()) {
+            continue;
+        }
+
+        // Paragraph break BEFORE
+        let break_before = if i == 0 {
+            true
+        } else {
+            let prev = &lines[i - 1];
+            prev.page != line.page || (prev.y - line.y).abs() > para_threshold
+        };
+
+        // Paragraph break AFTER
+        let break_after = if i + 1 >= lines.len() {
+            true
+        } else {
+            let next = &lines[i + 1];
+            next.page != line.page || (line.y - next.y).abs() > para_threshold
+        };
+
+        if !break_before || !break_after {
+            continue;
+        }
+
+        set.insert(i);
+    }
+
+    // Density guard: if too many lines on a page are "isolated", they're
+    // all paragraph lines in a multi-column layout, not headings.  Real
+    // headings are rare — at most ~20% of lines on a page.
+    let mut page_line_counts: HashMap<u32, (usize, usize)> = HashMap::new(); // (total, isolated)
+    for (i, line) in lines.iter().enumerate() {
+        let entry = page_line_counts.entry(line.page).or_insert((0, 0));
+        entry.0 += 1;
+        if set.contains(&i) {
+            entry.1 += 1;
+        }
+    }
+    for (&page, &(total, isolated)) in &page_line_counts {
+        if total > 0 && isolated as f32 / total as f32 > 0.25 {
+            // Too many isolated lines on this page — remove them all
+            set.retain(|&i| lines[i].page != page);
+        }
+    }
+
+    set
+}
+
 /// Resolve the dominant structure role for a text line by looking up its items' MCIDs.
 ///
 /// Returns the first non-container role found (skipping Document/Part/Sect/Div/NonStruct/Span).
@@ -256,6 +391,16 @@ 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);
 
+    // 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
+    // between paragraphs at body font size. Inspired by opendataloader's
+    // lookahead in HeadingProcessor (prevNode/nextNode context).
+    let isolated_lines = find_isolated_lines(&lines, base_size, para_threshold);
+
+    // Detect struct heading levels that are overused (body text mistagged as headings)
+    let overused_heading_levels = detect_overused_struct_heading_levels(&lines, struct_roles);
+
     let mut output = String::new();
     let mut current_page = 0u32;
     let mut prev_y = f32::MAX;
@@ -277,7 +422,7 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
     all_content_pages.sort();
     all_content_pages.dedup();
 
-    for line in lines {
+    for (line_idx, line) in lines.iter().enumerate() {
         // Page break
         if line.page != current_page {
             // Flush current page's remaining tables and images
@@ -405,7 +550,7 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
 
         // Detect figure/table captions and source citations
         // These should be on their own line followed by a paragraph break
-        let struct_role = struct_roles.and_then(|roles| resolve_line_struct_role(&line, roles));
+        let struct_role = struct_roles.and_then(|roles| resolve_line_struct_role(line, roles));
 
         // Determine if this line is code (struct-tree or font-based) for block accumulation
         let is_code_line = struct_role
@@ -437,13 +582,72 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
         // Structure roles ADD headings (e.g. same-size text tagged H2) but do NOT
         // suppress headings that the font heuristic would detect (some tagged PDFs
         // mark obvious headings as P or Span).
-        let struct_heading = struct_role.as_ref().and_then(struct_role_heading_level);
+        let struct_heading = struct_role
+            .as_ref()
+            .and_then(struct_role_heading_level)
+            .filter(|level| !overused_heading_levels.contains(level));
+
+        // Protect wrapped list items: when inside a list, a visually-continuing
+        // line (same indent, line-wrap spacing) must not be reclassified as a
+        // heading by the font heuristic — PDFs often bold the lead phrase of a
+        // list item across multiple wrap lines, and an all-bold middle line
+        // would otherwise split one item into a heading + stray body text.
+        // We gate on the document's paragraph threshold so genuine section
+        // headings that follow a numbered paragraph (y_gap > para_threshold)
+        // remain detectable.
+        let looks_like_list_continuation = in_list
+            && match (last_list_x, line.items.first().map(|i| i.x)) {
+                (Some(list_x), Some(curr_x)) => {
+                    let x_ok = curr_x >= list_x - 5.0 && curr_x <= list_x + 50.0;
+                    let y_ok = y_gap >= 0.0 && y_gap <= para_threshold;
+                    x_ok && y_ok && !is_list_item(plain_trimmed)
+                }
+                _ => false,
+            };
+
         let heuristic_heading = if options.detect_headers
+            && !looks_like_list_continuation
             && plain_trimmed.len() > 3
             && plain_trimmed.split_whitespace().count() <= 15
+            && !starts_with_bullet_marker(plain_trimmed)
         {
             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)
+            detect_header_level(line_font_size, base_size, &heading_tiers).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
+                //       + isolated * 0.3 (paragraph break before AND after)
+                // Only consider lines at or above body font size.
+                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;
+                }
+                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 };
+
+                // Require standalone + at least one strong signal.
+                // Non-bold, non-isolated lines need very high rarity (≥0.97)
+                // to avoid classifying ordinary body text as headings in
+                // multi-column layouts where column switches break
+                // paragraph continuity and minor font-size variation
+                // inflates rarity scores.
+                let has_strong_signal = all_bold || isolated || (rarity >= 0.97 && word_count <= 8);
+                if score >= 0.5 && standalone && word_count >= 2 && has_strong_signal {
+                    Some(bold_heading_level(&heading_tiers))
+                } else {
+                    None
+                }
+            })
         } else {
             None
         };
@@ -460,11 +664,16 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
             continue;
         }
 
-        // Structure-tree list item (LI only — LBody is a continuation, not a new item)
+        // Structure-tree list item (LI only — LBody is a continuation, not a new item).
+        // Some tagged PDFs use a "flat" style where every wrapped line in a list item
+        // gets its own MCID tagged directly under LI. When we're already inside a list
+        // and the line has no visible bullet marker, treat it as a continuation (falls
+        // through to the continuation logic below) rather than a new list item.
         if struct_role
             .as_ref()
             .is_some_and(|r| matches!(r, StructRole::LI))
             && !is_list_item(plain_trimmed)
+            && !in_list
         {
             if in_paragraph {
                 output.push_str("\n\n");
@@ -626,6 +835,8 @@ pub fn to_markdown_from_lines(lines: Vec<TextLine>, options: MarkdownOptions) ->
     // Compute the typical line spacing for paragraph break detection
     let para_threshold = compute_paragraph_threshold(&lines, base_size);
 
+    let isolated_lines = find_isolated_lines(&lines, base_size, para_threshold);
+
     let mut output = String::new();
     let mut current_page = 0u32;
     let mut prev_y = f32::MAX;
@@ -634,7 +845,7 @@ pub fn to_markdown_from_lines(lines: Vec<TextLine>, options: MarkdownOptions) ->
     let mut last_list_x: Option<f32> = None;
     let mut prev_had_dot_leaders = false;
 
-    for line in lines {
+    for (line_idx, line) in lines.iter().enumerate() {
         // Page break
         if line.page != current_page {
             if current_page > 0 {
@@ -699,7 +910,27 @@ pub fn to_markdown_from_lines(lines: Vec<TextLine>, options: MarkdownOptions) ->
         {
             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)
+                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;
+                    }
+                    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 };
+                    if score >= 0.5 && standalone && word_count >= 2 {
+                        return Some(bold_heading_level(&heading_tiers));
+                    }
+                    None
+                })
             {
                 if in_paragraph {
                     output.push_str("\n\n");
@@ -887,6 +1118,59 @@ mod tests {
         );
     }
 
+    #[test]
+    fn test_struct_role_li_flat_continuation_lines_merge() {
+        // Regression: some tagged PDFs put each wrapped visual line of a list
+        // item under its own MCID, all tagged directly as LI. Continuation
+        // lines (no bullet marker) must merge into the bulleted parent item,
+        // not each become their own list item.
+        let make = |text: &str, mcid: i64, x: f32, y: f32| {
+            let mut item = make_item(text, 1, Some(mcid));
+            item.x = x;
+            item.y = y;
+            item
+        };
+        let lines = vec![
+            make_line(vec![make("● First item that wraps onto", 0, 90.0, 322.0)]),
+            make_line(vec![make("a continuation line.", 1, 108.0, 306.0)]),
+            make_line(vec![make("● Second bullet also wraps", 2, 90.0, 290.0)]),
+            make_line(vec![make("to a second line here.", 3, 108.0, 274.0)]),
+        ];
+
+        let mut page_roles = HashMap::new();
+        for mcid in 0..4 {
+            page_roles.insert(mcid, StructRole::LI);
+        }
+        let mut roles = HashMap::new();
+        roles.insert(1u32, page_roles);
+
+        let md = to_markdown_from_lines_with_tables_and_images(
+            lines,
+            MarkdownOptions::default(),
+            HashMap::new(),
+            HashMap::new(),
+            &std::collections::HashSet::new(),
+            Some(&roles),
+        );
+
+        assert!(
+            md.contains("- First item that wraps onto a continuation line."),
+            "continuation should merge into first bullet: {md}"
+        );
+        assert!(
+            md.contains("- Second bullet also wraps to a second line here."),
+            "continuation should merge into second bullet: {md}"
+        );
+        assert!(
+            !md.contains("- a continuation line."),
+            "continuation line should not get its own bullet: {md}"
+        );
+        assert!(
+            !md.contains("- to a second line here."),
+            "continuation line should not get its own bullet: {md}"
+        );
+    }
+
     #[test]
     fn test_struct_role_blockquote() {
         let lines = vec![make_line(vec![make_item("Quoted text", 1, Some(0))])];
@@ -1016,6 +1300,62 @@ mod tests {
         );
     }
 
+    #[test]
+    fn test_rarity_heading_requires_strong_signal() {
+        // Simulate a two-column academic paper where body text lines become
+        // "standalone" due to column switches.  Body text at the same font
+        // size as most of the document should NOT be classified as headings
+        // just because of moderate rarity + standalone.
+        //
+        // Regression: previously, lines with rarity ~0.62 and standalone=true
+        // scored 0.51 (>=0.5 threshold), producing hundreds of false ## headings.
+
+        // Create many body-text lines at font_size=10.9 (most common)
+        let mut lines = Vec::new();
+        for i in 0..20 {
+            let mut item = make_item("This is ordinary body text in a paragraph.", 1, None);
+            item.font_size = 10.9;
+            item.y = 700.0 - i as f32 * 14.0;
+            lines.push(make_line(vec![item]));
+        }
+        // A few lines at a slightly different size (simulating column B text)
+        for i in 0..10 {
+            let mut item = make_item("Another body text line from the second column.", 1, None);
+            item.font_size = 11.0; // slightly different → non-zero rarity
+            item.y = 700.0 - i as f32 * 14.0;
+            item.x = 320.0; // right column
+            lines.push(make_line(vec![item]));
+        }
+        // One genuine bold heading
+        let mut heading_item = make_item("3 Philosophical Perspectives", 1, None);
+        heading_item.font_size = 10.9;
+        heading_item.is_bold = true;
+        heading_item.y = 200.0;
+        lines.push(make_line(vec![heading_item]));
+
+        let md = to_markdown_from_lines_with_tables_and_images(
+            lines,
+            MarkdownOptions::default(),
+            HashMap::new(),
+            HashMap::new(),
+            &std::collections::HashSet::new(),
+            None,
+        );
+
+        // The bold heading should be detected
+        assert!(
+            md.contains("## 3 Philosophical Perspectives"),
+            "Bold heading should be detected: {md}"
+        );
+
+        // Body text lines should NOT be headings
+        let heading_count = md.lines().filter(|l| l.starts_with("##")).count();
+        assert!(
+            heading_count <= 2,
+            "Expected at most 2 headings but found {heading_count} in:\n{md}"
+        );
+    }
+
     #[test]
     fn test_struct_role_code_multiline_accumulation() {
         let mut line1 = make_item("fn main() {", 1, Some(0));
@@ -1058,4 +1398,145 @@ mod tests {
             "Should not have adjacent close/open fences: {md}"
         );
     }
+
+    #[test]
+    fn test_overused_struct_heading_suppressed() {
+        // Simulate a PDF where H2 is mistagged on body text lines.
+        // 30 lines total: 5 tagged H1 (real headings), 20 tagged H2 (mistagged body),
+        // 5 tagged P.
+        let mut lines = Vec::new();
+        let mut page_roles = HashMap::new();
+        let mut mcid = 0i64;
+
+        for i in 0..30 {
+            let mut item = make_item(&format!("Line {i}"), 1, Some(mcid));
+            item.y = 700.0 - (i as f32 * 15.0);
+            lines.push(make_line(vec![item]));
+
+            let role = if i < 5 {
+                StructRole::H1
+            } else if i < 25 {
+                StructRole::H2
+            } else {
+                StructRole::P
+            };
+            page_roles.insert(mcid, role);
+            mcid += 1;
+        }
+
+        let mut roles = HashMap::new();
+        roles.insert(1u32, page_roles);
+
+        let overused = detect_overused_struct_heading_levels(&lines, Some(&roles));
+        // H2 is on 20/30 = 67% of lines — should be suppressed
+        assert!(
+            overused.contains(&2),
+            "H2 should be detected as overused: {:?}",
+            overused
+        );
+        // H1 is on 5/30 = 17% — should also be suppressed at >15% threshold
+        assert!(
+            overused.contains(&1),
+            "H1 at 17% should also be suppressed: {:?}",
+            overused
+        );
+    }
+
+    #[test]
+    fn test_normal_struct_headings_not_suppressed() {
+        // Normal document: a few headings, mostly body text
+        let mut lines = Vec::new();
+        let mut page_roles = HashMap::new();
+        let mut mcid = 0i64;
+
+        for i in 0..50 {
+            let mut item = make_item(&format!("Line {i}"), 1, Some(mcid));
+            item.y = 700.0 - (i as f32 * 14.0);
+            lines.push(make_line(vec![item]));
+
+            let role = if i % 10 == 0 {
+                StructRole::H1 // 5 headings out of 50 = 10%
+            } else {
+                StructRole::P
+            };
+            page_roles.insert(mcid, role);
+            mcid += 1;
+        }
+
+        let mut roles = HashMap::new();
+        roles.insert(1u32, page_roles);
+
+        let overused = detect_overused_struct_heading_levels(&lines, Some(&roles));
+        assert!(
+            overused.is_empty(),
+            "No heading level should be overused: {:?}",
+            overused
+        );
+    }
+
+    #[test]
+    fn test_wrapped_bold_lead_in_list_item_not_heading() {
+        // Regression: numbered-list items whose bold "lead" phrase wraps onto
+        // a second line (e.g. definitions in system cards) must not have the
+        // wrapped line reclassified as a heading. The middle line is
+        // all_bold + standalone (in_paragraph=false while in_list), which
+        // previously tripped the rarity heuristic and emitted #### in the
+        // middle of the item, splitting the body into stray bullets.
+        let make = |text: &str, x: f32, y: f32, bold: bool| {
+            let mut item = make_item(text, 1, None);
+            item.x = x;
+            item.y = y;
+            item.is_bold = bold;
+            item
+        };
+
+        let lines = vec![
+            // "1. **bold lead phrase start**"
+            make_line(vec![
+                make("1. ", 72.0, 700.0, false),
+                make(
+                    "Chemical and biological weapons threat model 1 (CB-1): Non-novel",
+                    90.0,
+                    700.0,
+                    true,
+                ),
+            ]),
+            // wrapped continuation of the bold lead — all_bold, same indent
+            make_line(vec![make(
+                "chemical/biological weapons production capabilities: A model has CB-1",
+                90.0,
+                686.0,
+                true,
+            )]),
+            // body text of the same list item
+            make_line(vec![make(
+                "capabilities if it has the ability to significantly help individuals.",
+                90.0,
+                672.0,
+                false,
+            )]),
+        ];
+
+        let md = to_markdown_from_lines_with_tables_and_images(
+            lines,
+            MarkdownOptions::default(),
+            HashMap::new(),
+            HashMap::new(),
+            &std::collections::HashSet::new(),
+            None,
+        );
+
+        assert!(
+            !md.contains("#### "),
+            "wrapped bold lead must not become a heading: {md}"
+        );
+        assert!(
+            md.lines().filter(|l| l.starts_with("- ")).count() == 0,
+            "continuation body must not become a stray bullet: {md}"
+        );
+        assert!(
+            md.contains("1. ") && md.contains("A model has CB-1"),
+            "numbered list item should remain intact: {md}"
+        );
+    }
 }

src/markdown/mod.rs

@@ -42,7 +42,7 @@ pub(crate) fn split_side_by_side(items: &[TextItem]) -> Vec<(f32, f32)> {
 
     // Sort items by left edge
     let mut xs: Vec<f32> = items.iter().map(|i| i.x).collect();
-    xs.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+    xs.sort_by(|a, b| a.total_cmp(b));
 
     // Find all candidate gaps: ≥30pt, in the middle 60% of the X range,
     // with ≥20 items on each side.
@@ -118,7 +118,7 @@ pub(crate) fn split_side_by_side(items: &[TextItem]) -> Vec<(f32, f32)> {
         })
         .copied()
         .collect();
-    balanced_positions.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+    balanced_positions.sort_by(|a, b| a.total_cmp(b));
     balanced_positions.dedup_by(|a, b| (*a - *b).abs() < 50.0);
     if balanced_positions.len() > 1 {
         return vec![];
@@ -209,7 +209,7 @@ fn split_from_hint_regions(items: &[TextItem], rects: &[PdfRect], page: u32) ->
 
     // Width outlier filter (same as detect_tables_from_rects)
     let mut widths: Vec<f32> = page_rects.iter().map(|&(_, _, w, _)| w).collect();
-    widths.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+    widths.sort_by(|a, b| a.total_cmp(b));
     let median_width = widths[widths.len() / 2];
     page_rects.retain(|&(_, _, w, _)| w <= median_width * 10.0);
 
@@ -601,6 +601,15 @@ pub(crate) fn to_markdown_from_items_with_rects_and_lines(
         let group = page_groups.get(&page).unwrap();
         let page_items: Vec<TextItem> = group.iter().map(|(_, item)| (*item).clone()).collect();
 
+        // Detect columns early — on multi-column pages, the merged-band retry
+        // should skip body-font heuristic table detection (which mistakes column
+        // text for tables). Individual band heuristic detection is left enabled
+        // because bands are scoped to single columns.
+        let page_has_columns = {
+            let cols = crate::extractor::detect_columns(&page_items, page, false);
+            cols.len() >= 2
+        };
+
         // Check for side-by-side layout (e.g. two tables placed left and right)
         let mut bands = split_side_by_side(&page_items);
         // Fallback: use rect hint regions to detect side-by-side layout
@@ -873,10 +882,7 @@ pub(crate) fn to_markdown_from_items_with_rects_and_lines(
                 run_heuristic(&unclaimed_items, &unclaimed_map, 6);
             }
 
-            // 4. Column-based table detection: last resort for borderless tabular
-            //    layouts (e.g. exam/reference grids) when ALL structural methods
-            //    found nothing. Only runs when no rects/lines exist (truly borderless)
-            //    and no other detection method found tables in this band.
+            // 4. Column-based table detection for borderless tabular layouts.
             let band_has_tables = band_items.iter().enumerate().any(|(idx, _)| {
                 band_index_map
                     .get(idx)
@@ -903,6 +909,65 @@ pub(crate) fn to_markdown_from_items_with_rects_and_lines(
             }
         }
 
+        // 5. Thin-rect border synthesis: last resort for PDFs that draw table
+        //    borders as thin filled rectangles (common in spreadsheet exports).
+        //    Only runs when ALL other methods found nothing on this page.
+        if !page_tables.contains_key(&page) {
+            let page_rects: Vec<&crate::types::PdfRect> =
+                rects.iter().filter(|r| r.page == page).collect();
+            let mut synth_lines: Vec<crate::types::PdfLine> = Vec::new();
+            for r in &page_rects {
+                let (mut w, mut h) = (r.width, r.height);
+                let (mut x, mut y) = (r.x, r.y);
+                if w < 0.0 {
+                    x += w;
+                    w = -w;
+                }
+                if h < 0.0 {
+                    y += h;
+                    h = -h;
+                }
+                if h < 2.0 && w >= 10.0 {
+                    let mid_y = y + h / 2.0;
+                    synth_lines.push(crate::types::PdfLine {
+                        x1: x,
+                        y1: mid_y,
+                        x2: x + w,
+                        y2: mid_y,
+                        page,
+                    });
+                } else if w < 2.0 && h >= 10.0 {
+                    let mid_x = x + w / 2.0;
+                    synth_lines.push(crate::types::PdfLine {
+                        x1: mid_x,
+                        y1: y,
+                        x2: mid_x,
+                        y2: y + h,
+                        page,
+                    });
+                }
+            }
+            if synth_lines.len() >= 10 {
+                let page_text: Vec<TextItem> = text_items
+                    .iter()
+                    .filter(|i| i.page == page)
+                    .cloned()
+                    .collect();
+                let line_tables = detect_tables_from_lines(&page_text, &synth_lines, page);
+                for table in &line_tables {
+                    for &idx in &table.item_indices {
+                        table_items.insert(idx);
+                    }
+                    let table_y = table.rows.first().copied().unwrap_or(0.0);
+                    let table_md = table_to_markdown(table);
+                    page_tables
+                        .entry(page)
+                        .or_default()
+                        .push((table_y, table_md));
+                }
+            }
+        }
+
         // Merged-band retry: if we split into bands but found no tables in
         // any band, retry heuristic detection with all items as a single band.
         // This catches borderless tables whose text-column alignment was
@@ -915,7 +980,7 @@ pub(crate) fn to_markdown_from_items_with_rects_and_lines(
                 band_items.len(),
                 was_split
             );
-            let heuristic_tables = detect_tables(band_items, base_size, false);
+            let heuristic_tables = detect_tables(band_items, base_size, page_has_columns);
             for table in &heuristic_tables {
                 for &idx in &table.item_indices {
                     if let Some(&page_idx) = band_index_map.get(idx) {
@@ -1037,7 +1102,7 @@ pub(crate) fn to_markdown_from_items_with_rects_and_lines(
             }
             // Sort by Y descending (top to bottom) so left and right
             // band lines interleave in visual reading order.
-            page_lines.sort_by(|a, b| b.y.partial_cmp(&a.y).unwrap_or(std::cmp::Ordering::Equal));
+            page_lines.sort_by(|a, b| b.y.total_cmp(&a.y));
             all_lines.extend(page_lines);
         }
         all_lines

src/markdown/preprocess.rs

@@ -275,7 +275,7 @@ pub(crate) fn strip_repeated_lines(lines: Vec<TextLine>, page_count: u32) -> Vec
         page_sorted_ys.entry(line.page).or_default().push(line.y);
     }
     for ys in page_sorted_ys.values_mut() {
-        ys.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+        ys.sort_by(|a, b| a.total_cmp(b));
         ys.dedup();
     }
 

src/python.rs

@@ -0,0 +1,587 @@
+//! PyO3 Python bindings for pdf-inspector.
+
+use pyo3::exceptions::PyValueError;
+use pyo3::prelude::*;
+use std::collections::HashSet;
+
+use crate::detector::PdfType;
+use crate::types::ItemType;
+
+// ---------------------------------------------------------------------------
+// Result wrapper
+// ---------------------------------------------------------------------------
+
+/// Result of processing a PDF file.
+#[pyclass(name = "PdfResult")]
+#[derive(Clone)]
+pub struct PyPdfResult {
+    /// The detected PDF type: "text_based", "scanned", "image_based", or "mixed".
+    #[pyo3(get)]
+    pub pdf_type: String,
+    /// Markdown output (None if detect-only or scanned PDF).
+    #[pyo3(get)]
+    pub markdown: Option<String>,
+    /// Total number of pages.
+    #[pyo3(get)]
+    pub page_count: u32,
+    /// Processing time in milliseconds.
+    #[pyo3(get)]
+    pub processing_time_ms: u64,
+    /// 1-indexed page numbers that need OCR.
+    #[pyo3(get)]
+    pub pages_needing_ocr: Vec<u32>,
+    /// Title from PDF metadata.
+    #[pyo3(get)]
+    pub title: Option<String>,
+    /// Detection confidence (0.0-1.0).
+    #[pyo3(get)]
+    pub confidence: f32,
+    /// Whether the layout is complex (tables/columns detected).
+    #[pyo3(get)]
+    pub is_complex_layout: bool,
+    /// Pages with tables detected.
+    #[pyo3(get)]
+    pub pages_with_tables: Vec<u32>,
+    /// Pages with multi-column layout.
+    #[pyo3(get)]
+    pub pages_with_columns: Vec<u32>,
+    /// Whether encoding issues were detected.
+    #[pyo3(get)]
+    pub has_encoding_issues: bool,
+}
+
+#[pymethods]
+impl PyPdfResult {
+    fn __repr__(&self) -> String {
+        format!(
+            "PdfResult(pdf_type='{}', pages={}, confidence={:.2})",
+            self.pdf_type, self.page_count, self.confidence
+        )
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Classification wrapper (lightweight)
+// ---------------------------------------------------------------------------
+
+/// Lightweight PDF classification result.
+#[pyclass(name = "PdfClassification")]
+#[derive(Clone)]
+pub struct PyPdfClassification {
+    /// The detected PDF type: "text_based", "scanned", "image_based", or "mixed".
+    #[pyo3(get)]
+    pub pdf_type: String,
+    /// Total number of pages.
+    #[pyo3(get)]
+    pub page_count: u32,
+    /// 0-indexed page numbers that need OCR.
+    #[pyo3(get)]
+    pub pages_needing_ocr: Vec<u32>,
+    /// Detection confidence (0.0-1.0).
+    #[pyo3(get)]
+    pub confidence: f32,
+}
+
+#[pymethods]
+impl PyPdfClassification {
+    fn __repr__(&self) -> String {
+        format!(
+            "PdfClassification(pdf_type='{}', pages={}, confidence={:.2})",
+            self.pdf_type, self.page_count, self.confidence
+        )
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Region extraction wrappers
+// ---------------------------------------------------------------------------
+
+/// Extracted text for a single region.
+#[pyclass(name = "RegionText")]
+#[derive(Clone)]
+pub struct PyRegionText {
+    /// Extracted text content.
+    #[pyo3(get)]
+    pub text: String,
+    /// True when the text should not be trusted (empty, GID fonts, garbage, encoding issues).
+    #[pyo3(get)]
+    pub needs_ocr: bool,
+}
+
+#[pymethods]
+impl PyRegionText {
+    fn __repr__(&self) -> String {
+        format!(
+            "RegionText(text='{}', needs_ocr={})",
+            self.text.chars().take(40).collect::<String>(),
+            self.needs_ocr
+        )
+    }
+}
+
+/// Extracted text for one page's regions.
+#[pyclass(name = "PageRegionTexts")]
+#[derive(Clone)]
+pub struct PyPageRegionTexts {
+    /// 0-indexed page number.
+    #[pyo3(get)]
+    pub page: u32,
+    /// Per-region results, parallel to the input regions.
+    #[pyo3(get)]
+    pub regions: Vec<PyRegionText>,
+}
+
+#[pymethods]
+impl PyPageRegionTexts {
+    fn __repr__(&self) -> String {
+        format!(
+            "PageRegionTexts(page={}, regions={})",
+            self.page,
+            self.regions.len()
+        )
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Text item wrapper
+// ---------------------------------------------------------------------------
+
+/// Per-page markdown extraction result.
+#[pyclass(name = "PageMarkdown")]
+#[derive(Clone)]
+pub struct PyPageMarkdown {
+    /// 0-indexed page number.
+    #[pyo3(get)]
+    pub page: u32,
+    /// Formatted markdown for this page.
+    #[pyo3(get)]
+    pub markdown: String,
+    /// True when text on this page is unreliable (GID-encoded fonts,
+    /// encoding issues, garbage text, or empty extraction).
+    #[pyo3(get)]
+    pub needs_ocr: bool,
+}
+
+#[pymethods]
+impl PyPageMarkdown {
+    fn __repr__(&self) -> String {
+        format!(
+            "PageMarkdown(page={}, markdown='{}', needs_ocr={})",
+            self.page,
+            self.markdown.chars().take(40).collect::<String>(),
+            self.needs_ocr
+        )
+    }
+}
+
+/// Combined per-page markdown extraction and layout classification result.
+#[pyclass(name = "PagesExtractionResult")]
+#[derive(Clone)]
+pub struct PyPagesExtractionResult {
+    /// Per-page markdown results, in the order requested.
+    #[pyo3(get)]
+    pub pages: Vec<PyPageMarkdown>,
+    /// 1-indexed pages where tables were detected.
+    #[pyo3(get)]
+    pub pages_with_tables: Vec<u32>,
+    /// 1-indexed pages where multi-column layout was detected.
+    #[pyo3(get)]
+    pub pages_with_columns: Vec<u32>,
+    /// 1-indexed pages that need OCR (scanned/image-based or unreliable text).
+    #[pyo3(get)]
+    pub pages_needing_ocr: Vec<u32>,
+    /// True if any page has tables or columns.
+    #[pyo3(get)]
+    pub is_complex: bool,
+}
+
+#[pymethods]
+impl PyPagesExtractionResult {
+    fn __repr__(&self) -> String {
+        format!(
+            "PagesExtractionResult(pages={}, pages_with_tables={:?}, is_complex={})",
+            self.pages.len(),
+            self.pages_with_tables,
+            self.is_complex
+        )
+    }
+}
+
+/// A positioned text item extracted from a PDF.
+#[pyclass(name = "TextItem")]
+#[derive(Clone)]
+pub struct PyTextItem {
+    #[pyo3(get)]
+    pub text: String,
+    #[pyo3(get)]
+    pub x: f32,
+    #[pyo3(get)]
+    pub y: f32,
+    #[pyo3(get)]
+    pub width: f32,
+    #[pyo3(get)]
+    pub height: f32,
+    #[pyo3(get)]
+    pub font: String,
+    #[pyo3(get)]
+    pub font_size: f32,
+    #[pyo3(get)]
+    pub page: u32,
+    #[pyo3(get)]
+    pub is_bold: bool,
+    #[pyo3(get)]
+    pub is_italic: bool,
+    #[pyo3(get)]
+    pub item_type: String,
+}
+
+#[pymethods]
+impl PyTextItem {
+    fn __repr__(&self) -> String {
+        format!(
+            "TextItem(text='{}', page={}, x={:.1}, y={:.1})",
+            self.text.chars().take(40).collect::<String>(),
+            self.page,
+            self.x,
+            self.y,
+        )
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Helpers
+// ---------------------------------------------------------------------------
+
+fn pdf_type_str(t: PdfType) -> String {
+    match t {
+        PdfType::TextBased => "text_based".into(),
+        PdfType::Scanned => "scanned".into(),
+        PdfType::ImageBased => "image_based".into(),
+        PdfType::Mixed => "mixed".into(),
+    }
+}
+
+fn to_py_result(r: crate::PdfProcessResult) -> PyPdfResult {
+    PyPdfResult {
+        pdf_type: pdf_type_str(r.pdf_type),
+        markdown: r.markdown,
+        page_count: r.page_count,
+        processing_time_ms: r.processing_time_ms,
+        pages_needing_ocr: r.pages_needing_ocr,
+        title: r.title,
+        confidence: r.confidence,
+        is_complex_layout: r.layout.is_complex,
+        pages_with_tables: r.layout.pages_with_tables,
+        pages_with_columns: r.layout.pages_with_columns,
+        has_encoding_issues: r.has_encoding_issues,
+    }
+}
+
+fn to_py_err(e: crate::PdfError) -> PyErr {
+    PyValueError::new_err(e.to_string())
+}
+
+fn item_type_str(t: &ItemType) -> String {
+    match t {
+        ItemType::Text => "text".into(),
+        ItemType::Image => "image".into(),
+        ItemType::Link(url) => format!("link:{url}"),
+        ItemType::FormField => "form_field".into(),
+    }
+}
+
+fn convert_text_items(items: Vec<crate::TextItem>) -> Vec<PyTextItem> {
+    items
+        .into_iter()
+        .map(|item| PyTextItem {
+            text: item.text,
+            x: item.x,
+            y: item.y,
+            width: item.width,
+            height: item.height,
+            font: item.font,
+            font_size: item.font_size,
+            page: item.page,
+            is_bold: item.is_bold,
+            is_italic: item.is_italic,
+            item_type: item_type_str(&item.item_type),
+        })
+        .collect()
+}
+
+fn parse_page_regions(
+    page_regions: Vec<(u32, Vec<Vec<f64>>)>,
+) -> PyResult<Vec<(u32, Vec<[f32; 4]>)>> {
+    page_regions
+        .into_iter()
+        .map(|(page, regions)| {
+            let mut bboxes: Vec<[f32; 4]> = Vec::with_capacity(regions.len());
+            for (idx, region) in regions.into_iter().enumerate() {
+                if region.len() != 4 {
+                    return Err(PyValueError::new_err(format!(
+                        "Invalid region at page {page}, index {idx}: expected [x1, y1, x2, y2], got {} values",
+                        region.len()
+                    )));
+                }
+                let [x1, y1, x2, y2] = [region[0], region[1], region[2], region[3]];
+                if !(x1.is_finite() && y1.is_finite() && x2.is_finite() && y2.is_finite()) {
+                    return Err(PyValueError::new_err(format!(
+                        "Invalid region at page {page}, index {idx}: coordinates must be finite numbers"
+                    )));
+                }
+                if x2 < x1 || y2 < y1 {
+                    return Err(PyValueError::new_err(format!(
+                        "Invalid region at page {page}, index {idx}: expected x2>=x1 and y2>=y1, got [{x1}, {y1}, {x2}, {y2}]"
+                    )));
+                }
+                bboxes.push([x1 as f32, y1 as f32, x2 as f32, y2 as f32]);
+            }
+            Ok((page, bboxes))
+        })
+        .collect()
+}
+
+fn to_py_pages_result(r: crate::PagesExtractionResult) -> PyPagesExtractionResult {
+    PyPagesExtractionResult {
+        pages: r
+            .pages
+            .into_iter()
+            .map(|p| PyPageMarkdown {
+                page: p.page,
+                markdown: p.markdown,
+                needs_ocr: p.needs_ocr,
+            })
+            .collect(),
+        pages_with_tables: r.pages_with_tables,
+        pages_with_columns: r.pages_with_columns,
+        pages_needing_ocr: r.pages_needing_ocr,
+        is_complex: r.is_complex,
+    }
+}
+
+fn convert_region_results(results: Vec<crate::PageRegionResult>) -> Vec<PyPageRegionTexts> {
+    results
+        .into_iter()
+        .map(|page_result| PyPageRegionTexts {
+            page: page_result.page,
+            regions: page_result
+                .regions
+                .into_iter()
+                .map(|r| PyRegionText {
+                    text: r.text,
+                    needs_ocr: r.needs_ocr,
+                })
+                .collect(),
+        })
+        .collect()
+}
+
+// ---------------------------------------------------------------------------
+// Public Python API
+// ---------------------------------------------------------------------------
+
+/// Process a PDF file: detect type, extract text, and convert to Markdown.
+#[pyfunction]
+#[pyo3(signature = (path, pages=None))]
+fn process_pdf(path: &str, pages: Option<Vec<u32>>) -> PyResult<PyPdfResult> {
+    let mut opts = crate::PdfOptions::new();
+    if let Some(p) = pages {
+        opts = opts.pages(p);
+    }
+    let result = crate::process_pdf_with_options(path, opts).map_err(to_py_err)?;
+    Ok(to_py_result(result))
+}
+
+/// Process a PDF from bytes in memory.
+#[pyfunction]
+#[pyo3(signature = (data, pages=None))]
+fn process_pdf_bytes(data: &[u8], pages: Option<Vec<u32>>) -> PyResult<PyPdfResult> {
+    let mut opts = crate::PdfOptions::new();
+    if let Some(p) = pages {
+        opts = opts.pages(p);
+    }
+    let result = crate::process_pdf_mem_with_options(data, opts).map_err(to_py_err)?;
+    Ok(to_py_result(result))
+}
+
+/// Fast detection only — no text extraction or markdown.
+#[pyfunction]
+fn detect_pdf(path: &str) -> PyResult<PyPdfResult> {
+    let result = crate::detect_pdf(path).map_err(to_py_err)?;
+    Ok(to_py_result(result))
+}
+
+/// Fast detection from bytes — no text extraction or markdown.
+#[pyfunction]
+fn detect_pdf_bytes(data: &[u8]) -> PyResult<PyPdfResult> {
+    let result = crate::detect_pdf_mem(data).map_err(to_py_err)?;
+    Ok(to_py_result(result))
+}
+
+/// Lightweight PDF classification — returns type, page count, and OCR pages.
+/// Faster than detect_pdf as it skips building the full PdfProcessResult.
+/// Pages in pages_needing_ocr are 0-indexed.
+#[pyfunction]
+fn classify_pdf(path: &str) -> PyResult<PyPdfClassification> {
+    let data = std::fs::read(path).map_err(|e| PyValueError::new_err(e.to_string()))?;
+    classify_pdf_bytes(&data)
+}
+
+/// Lightweight PDF classification from bytes.
+/// Pages in pages_needing_ocr are 0-indexed.
+#[pyfunction]
+fn classify_pdf_bytes(data: &[u8]) -> PyResult<PyPdfClassification> {
+    let result = crate::classify_pdf_mem(data).map_err(to_py_err)?;
+    Ok(PyPdfClassification {
+        pdf_type: pdf_type_str(result.pdf_type),
+        page_count: result.page_count,
+        pages_needing_ocr: result.pages_needing_ocr,
+        confidence: result.confidence,
+    })
+}
+
+/// Extract plain text from a PDF file.
+#[pyfunction]
+fn extract_text(path: &str) -> PyResult<String> {
+    crate::extract_text(path).map_err(to_py_err)
+}
+
+/// Extract plain text from PDF bytes.
+#[pyfunction]
+fn extract_text_bytes(data: &[u8]) -> PyResult<String> {
+    crate::extractor::extract_text_mem(data).map_err(to_py_err)
+}
+
+/// Extract text with position information from a file.
+#[pyfunction]
+#[pyo3(signature = (path, pages=None))]
+fn extract_text_with_positions(path: &str, pages: Option<Vec<u32>>) -> PyResult<Vec<PyTextItem>> {
+    let items = match pages {
+        Some(p) => {
+            let page_set: HashSet<u32> = p.into_iter().collect();
+            crate::extract_text_with_positions_pages(path, Some(&page_set)).map_err(to_py_err)?
+        }
+        None => crate::extract_text_with_positions(path).map_err(to_py_err)?,
+    };
+    Ok(convert_text_items(items))
+}
+
+/// Extract text with position information from bytes.
+#[pyfunction]
+#[pyo3(signature = (data, pages=None))]
+fn extract_text_with_positions_bytes(
+    data: &[u8],
+    pages: Option<Vec<u32>>,
+) -> PyResult<Vec<PyTextItem>> {
+    let items = match pages {
+        Some(p) => {
+            let page_set: HashSet<u32> = p.into_iter().collect();
+            crate::extractor::extract_text_with_positions_mem_pages(data, Some(&page_set))
+                .map_err(to_py_err)?
+        }
+        None => crate::extractor::extract_text_with_positions_mem(data).map_err(to_py_err)?,
+    };
+    Ok(convert_text_items(items))
+}
+
+/// Extract text within bounding-box regions from a PDF file.
+///
+/// Args:
+///     path: Path to the PDF file.
+///     page_regions: List of (page_0indexed, [[x1, y1, x2, y2], ...]) tuples.
+///         Coordinates are PDF points with top-left origin.
+///
+/// Returns:
+///     List of PageRegionTexts with per-region text and needs_ocr flag.
+#[pyfunction]
+fn extract_text_in_regions(
+    path: &str,
+    page_regions: Vec<(u32, Vec<Vec<f64>>)>,
+) -> PyResult<Vec<PyPageRegionTexts>> {
+    let data = std::fs::read(path).map_err(|e| PyValueError::new_err(e.to_string()))?;
+    extract_text_in_regions_bytes(&data, page_regions)
+}
+
+/// Extract text within bounding-box regions from PDF bytes.
+///
+/// Args:
+///     data: PDF file contents as bytes.
+///     page_regions: List of (page_0indexed, [[x1, y1, x2, y2], ...]) tuples.
+///         Coordinates are PDF points with top-left origin.
+///
+/// Returns:
+///     List of PageRegionTexts with per-region text and needs_ocr flag.
+#[pyfunction]
+fn extract_text_in_regions_bytes(
+    data: &[u8],
+    page_regions: Vec<(u32, Vec<Vec<f64>>)>,
+) -> PyResult<Vec<PyPageRegionTexts>> {
+    let regions = parse_page_regions(page_regions)?;
+    let results = crate::extract_text_in_regions_mem(data, &regions).map_err(to_py_err)?;
+    Ok(convert_region_results(results))
+}
+
+/// Extract formatted markdown for pages of a PDF file, with layout
+/// classification metadata.
+///
+/// Returns per-page markdown and classification data (tables, columns,
+/// OCR needs) from a single parse. Font statistics are computed from the
+/// full document so header detection is consistent across pages.
+///
+/// Args:
+///     path: Path to the PDF file.
+///     pages: Optional list of 0-indexed pages. When None (default), every
+///         page is returned in document order. When provided, output
+///         matches the caller-supplied order.
+///
+/// Returns:
+///     PagesExtractionResult with per-page markdown and classification data.
+#[pyfunction]
+#[pyo3(signature = (path, pages=None))]
+fn extract_pages_markdown(
+    path: &str,
+    pages: Option<Vec<u32>>,
+) -> PyResult<PyPagesExtractionResult> {
+    let result = crate::extract_pages_markdown(path, pages.as_deref()).map_err(to_py_err)?;
+    Ok(to_py_pages_result(result))
+}
+
+/// Extract formatted markdown for pages of a PDF from bytes.
+///
+/// See [`extract_pages_markdown`] for details.
+#[pyfunction]
+#[pyo3(signature = (data, pages=None))]
+fn extract_pages_markdown_bytes(
+    data: &[u8],
+    pages: Option<Vec<u32>>,
+) -> PyResult<PyPagesExtractionResult> {
+    let result = crate::extract_pages_markdown_mem(data, pages.as_deref()).map_err(to_py_err)?;
+    Ok(to_py_pages_result(result))
+}
+
+/// Python module definition.
+#[pymodule]
+fn pdf_inspector(m: &Bound<'_, PyModule>) -> PyResult<()> {
+    m.add_class::<PyPdfResult>()?;
+    m.add_class::<PyPdfClassification>()?;
+    m.add_class::<PyTextItem>()?;
+    m.add_class::<PyRegionText>()?;
+    m.add_class::<PyPageRegionTexts>()?;
+    m.add_class::<PyPageMarkdown>()?;
+    m.add_class::<PyPagesExtractionResult>()?;
+    m.add_function(wrap_pyfunction!(process_pdf, m)?)?;
+    m.add_function(wrap_pyfunction!(process_pdf_bytes, m)?)?;
+    m.add_function(wrap_pyfunction!(detect_pdf, m)?)?;
+    m.add_function(wrap_pyfunction!(detect_pdf_bytes, m)?)?;
+    m.add_function(wrap_pyfunction!(classify_pdf, m)?)?;
+    m.add_function(wrap_pyfunction!(classify_pdf_bytes, m)?)?;
+    m.add_function(wrap_pyfunction!(extract_text, m)?)?;
+    m.add_function(wrap_pyfunction!(extract_text_bytes, m)?)?;
+    m.add_function(wrap_pyfunction!(extract_text_with_positions, m)?)?;
+    m.add_function(wrap_pyfunction!(extract_text_with_positions_bytes, m)?)?;
+    m.add_function(wrap_pyfunction!(extract_text_in_regions, m)?)?;
+    m.add_function(wrap_pyfunction!(extract_text_in_regions_bytes, m)?)?;
+    m.add_function(wrap_pyfunction!(extract_pages_markdown, m)?)?;
+    m.add_function(wrap_pyfunction!(extract_pages_markdown_bytes, m)?)?;
+    Ok(())
+}

src/tables/detect_heuristic.rs

@@ -48,7 +48,7 @@ pub(crate) fn merge_adjacent_items(items: &[TextItem]) -> (Vec<TextItem>, Vec<Ve
     }
 
     // Sort groups by Y descending (top of page first)
-    line_groups.sort_by(|a, b| b.0.partial_cmp(&a.0).unwrap_or(std::cmp::Ordering::Equal));
+    line_groups.sort_by(|a, b| b.0.total_cmp(&a.0));
 
     let mut merged_items = Vec::new();
     let mut index_map: Vec<Vec<usize>> = Vec::new();
@@ -219,7 +219,7 @@ pub fn detect_tables(items: &[TextItem], base_font_size: f32, skip_body_font: bo
             body_font_low,
             body_font_high,
         );
-        if body_candidates.len() >= 9 {
+        if body_candidates.len() >= 6 {
             let regions = find_table_regions_strict(&body_candidates);
             log::debug!("body-font: {} strict regions found", regions.len());
 
@@ -241,7 +241,7 @@ pub fn detect_tables(items: &[TextItem], base_font_size: f32, skip_body_font: bo
                     body_candidates.len()
                 );
 
-                if region_items.len() < 9 {
+                if region_items.len() < 6 {
                     continue;
                 }
 
@@ -284,7 +284,7 @@ fn find_table_regions(items: &[(usize, &TextItem)]) -> Vec<(f32, f32)> {
     }
 
     let mut y_positions: Vec<f32> = items.iter().map(|(_, i)| i.y).collect();
-    y_positions.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+    y_positions.sort_by(|a, b| a.total_cmp(b));
 
     // Find clusters of Y positions (table regions)
     let mut regions = Vec::new();
@@ -347,7 +347,7 @@ fn find_table_regions_strict(items: &[(usize, &TextItem)]) -> Vec<(f32, f32, f32
     let mut qualifying_rows: Vec<(f32, Vec<f32>)> = Vec::new(); // (y, cluster_starts)
     for (y, x_positions) in &row_groups {
         let mut sorted_xs = x_positions.clone();
-        sorted_xs.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+        sorted_xs.sort_by(|a, b| a.total_cmp(b));
 
         if sorted_xs.is_empty() {
             continue;
@@ -379,14 +379,14 @@ fn find_table_regions_strict(items: &[(usize, &TextItem)]) -> Vec<(f32, f32, f32
     // Step 3: Find contiguous runs of qualifying rows.
     // Use adaptive gap: median spacing × 3 (handles wrapped cells where
     // qualifying rows are spaced further apart), with a floor of 25pt.
-    qualifying_rows.sort_by(|a, b| a.0.partial_cmp(&b.0).unwrap_or(std::cmp::Ordering::Equal));
+    qualifying_rows.sort_by(|a, b| a.0.total_cmp(&b.0));
 
     let max_gap = if qualifying_rows.len() >= 3 {
         let mut gaps: Vec<f32> = qualifying_rows
             .windows(2)
             .map(|w| (w[1].0 - w[0].0).abs())
             .collect();
-        gaps.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+        gaps.sort_by(|a, b| a.total_cmp(b));
         let median_gap = gaps[gaps.len() / 2];
         (median_gap * 3.0).max(25.0)
     } else {
@@ -566,11 +566,9 @@ fn detect_table_in_region(items: &[(usize, &TextItem)], mode: TableDetectionMode
             // Sort by X position (direction-aware)
             let rtl = is_rtl_text(col_items.iter().map(|i| &i.text));
             if rtl {
-                col_items
-                    .sort_by(|a, b| b.x.partial_cmp(&a.x).unwrap_or(std::cmp::Ordering::Equal));
+                col_items.sort_by(|a, b| b.x.total_cmp(&a.x));
             } else {
-                col_items
-                    .sort_by(|a, b| a.x.partial_cmp(&b.x).unwrap_or(std::cmp::Ordering::Equal));
+                col_items.sort_by(|a, b| a.x.total_cmp(&b.x));
             }
 
             // Join items with subscript-aware spacing
@@ -583,8 +581,14 @@ fn detect_table_in_region(items: &[(usize, &TextItem)], mode: TableDetectionMode
     // Validation 1: some rows should have content in first column.
     // Use a lower threshold (25%) for tables with wrapped cells where
     // continuation lines leave the first column empty.
+    // Skip when cells form a narrow TOC pattern: hierarchical entries indented
+    // across multiple X levels leave the leftmost column sparse (only top-level
+    // chapters land there) but the structure is still a valid TOC. Narrow only
+    // (<=5 cols) — wide multi-column TOCs (e.g. 2-up indices) would render
+    // poorly through format_toc_as_list, which assumes one entry per row.
     let rows_with_first_col = cells.iter().filter(|row| !row[0].is_empty()).count();
-    if rows_with_first_col < rows.len() / 4 {
+    let is_narrow_toc = columns.len() <= 5 && is_table_of_contents(&cells);
+    if rows_with_first_col < rows.len() / 4 && !is_narrow_toc {
         log::debug!(
             "  validation 1 fail: {}/{} rows have first col",
             rows_with_first_col,
@@ -655,15 +659,24 @@ fn detect_table_in_region(items: &[(usize, &TextItem)], mode: TableDetectionMode
         return None;
     }
 
-    // Validation 8: Check for Table of Contents pattern
-    if is_table_of_contents(&cells) {
-        log::debug!("  validation 8 fail: table of contents");
+    // Validation 8: Reject paragraph-like content falsely detected as tables.
+    // TOC pages with deep indentation (top-level chapters in col 0, subsections
+    // in cols 1-3, page numbers in last col) leave most cells empty and trip
+    // the paragraph heuristic; TOC shape is a safer signal here. Narrow only
+    // — see narrow-TOC rationale at validation 1.
+    if is_paragraph_content(&cells) && !is_narrow_toc {
+        log::debug!("  validation 9 fail: paragraph content");
         return None;
     }
 
-    // Validation 9: Reject paragraph-like content falsely detected as tables
-    if is_paragraph_content(&cells) {
-        log::debug!("  validation 9 fail: paragraph content");
+    // Validation 9: Reject wide "index" layouts where every cell carries a
+    // full "label ... page" fragment (back-of-book IRS-style indices).
+    // These render poorly in any structured form; text flow is the best
+    // fallback.  Narrow dot-leader TOCs (2-3 cols) are kept so format.rs
+    // can emit them as a per-row flat list with titles tab-joined to page
+    // numbers.
+    if is_inline_leader_index(&cells) {
+        log::debug!("  validation 9 fail: inline-leader index");
         return None;
     }
 
@@ -674,12 +687,7 @@ fn detect_table_in_region(items: &[(usize, &TextItem)], mode: TableDetectionMode
         item_indices.len()
     );
 
-    Some(Table {
-        columns,
-        rows,
-        cells,
-        item_indices,
-    })
+    Some(Table::new(columns, rows, cells, item_indices))
 }
 
 /// Check if this looks like a key-value pair layout rather than a table
@@ -803,7 +811,25 @@ fn has_table_like_content(cells: &[Vec<String>], mode: TableDetectionMode) -> bo
     // Bypass content check for wide tables (3+ columns) — text-only tables
     // (category lists, program descriptions) are legitimate if they passed
     // all structural validations (alignment, consistency, not key-value).
-    pct_data > min_pct || num_cols >= 3
+    // Also bypass for 2-column body-font tables with short cells (avg ≤40 chars),
+    // which are likely definition/category lists, not paragraph text.
+    if pct_data > min_pct || num_cols >= 3 {
+        return true;
+    }
+    if num_cols == 2 && matches!(mode, TableDetectionMode::BodyFont) {
+        let non_empty: Vec<usize> = cells
+            .iter()
+            .skip(1)
+            .flat_map(|row| row.iter())
+            .filter(|c| !c.trim().is_empty())
+            .map(|c| c.trim().len())
+            .collect();
+        if !non_empty.is_empty() {
+            let avg_len = non_empty.iter().sum::<usize>() / non_empty.len();
+            return avg_len <= 25;
+        }
+    }
+    false
 }
 
 /// Check if a cell value looks like table data
@@ -882,77 +908,247 @@ fn looks_like_number(s: &str) -> bool {
         && s.chars().any(|c| c.is_ascii_digit())
 }
 
-/// Check if this looks like a Table of Contents
-/// TOCs have characteristic patterns: leader dots, page numbers, section names
-fn is_table_of_contents(cells: &[Vec<String>]) -> bool {
+/// Check if this looks like a Table of Contents (either style).
+///
+/// Used by format.rs to render TOCs as flat lists instead of markdown tables.
+pub fn is_table_of_contents(cells: &[Vec<String>]) -> bool {
+    is_dot_leader_toc(cells) || is_tabular_toc(cells)
+}
+
+/// Dot-leader TOC: any "Chapter 1 ........ 42" style with explicit leader
+/// dots.  Covers both narrow 2-3 col TOCs (where the leader is a dedicated
+/// cell) and wide indices (where each cell encodes a full "label ... page"
+/// fragment).  Used by format.rs to render as a flat list.
+pub(super) fn is_dot_leader_toc(cells: &[Vec<String>]) -> bool {
+    has_structural_dot_leader(cells) || is_inline_leader_index(cells)
+}
+
+/// Rows with a dedicated dots-only cell flanked by label + number (2-3 col
+/// TOC layout).  Format.rs handles these well via per-row flat-list
+/// rendering; they should NOT be rejected at detect time.
+fn has_structural_dot_leader(cells: &[Vec<String>]) -> bool {
     if cells.is_empty() {
         return false;
     }
+    let structural_rows = cells.iter().filter(|row| row_has_dot_leader(row)).count();
+    structural_rows as f32 / cells.len() as f32 >= 0.3
+}
 
-    let num_cols = cells[0].len();
-    let mut dot_cells = 0;
-    let mut page_number_cells = 0;
-    let mut total_cells = 0;
-    // Track which columns contain dots vs numbers to distinguish
-    // TOC (dots span middle, page number at end) from data tables
-    // (dots only in label column, many number columns).
-    let mut dot_cols = vec![0u32; num_cols];
-    let mut numeric_cols = vec![0u32; num_cols];
-
+/// Wide index layout: each cell holds a full "label ... page" fragment
+/// because the column detector kept multi-column indices as single cells.
+/// These render poorly both as markdown tables (column boundaries are
+/// arbitrary) and as flat lists (each row holds 3+ separate index
+/// entries).  Reject these at detect time so they fall back to the page's
+/// normal text flow.
+pub(super) fn is_inline_leader_index(cells: &[Vec<String>]) -> bool {
+    let mut inline_cells = 0;
+    let mut total_nonempty = 0;
     for row in cells {
-        for (ci, cell) in row.iter().enumerate() {
+        for cell in row {
             let trimmed = cell.trim();
             if trimmed.is_empty() {
                 continue;
             }
-            total_cells += 1;
-
-            // Check for leader dots (sequences of periods)
-            // TOCs often have "........" or ". . . ." patterns
-            let dot_count = trimmed.chars().filter(|&c| c == '.').count();
-            let is_mostly_dots = dot_count > trimmed.len() / 2 && dot_count >= 3;
-            if is_mostly_dots {
-                dot_cells += 1;
-                if ci < num_cols {
-                    dot_cols[ci] += 1;
-                }
-            }
-
-            // Check for standalone page numbers (1-4 digits, possibly with spaces)
-            let digits_only: String = trimmed.chars().filter(|c| !c.is_whitespace()).collect();
-            if digits_only.len() <= 4
-                && !digits_only.is_empty()
-                && digits_only.chars().all(|c| c.is_ascii_digit())
-            {
-                page_number_cells += 1;
-                if ci < num_cols {
-                    numeric_cols[ci] += 1;
-                }
+            total_nonempty += 1;
+            if cell_is_inline_leader(trimmed) {
+                inline_cells += 1;
             }
         }
     }
+    total_nonempty >= 4 && inline_cells as f32 / total_nonempty as f32 >= 0.25
+}
 
-    if total_cells == 0 {
+/// A row with a dot-leader.  Accepts two layouts:
+///   1. A dedicated dots-only cell ("....") with a text label somewhere
+///      to its left and a page number somewhere to its right.
+///   2. A "title ... " cell (trailing leader dots glued to the title)
+///      with a page number elsewhere in the same row.
+fn row_has_dot_leader(row: &[String]) -> bool {
+    let has_page_number = row.iter().any(|c| row_cell_is_page_number(c));
+
+    for (ci, cell) in row.iter().enumerate() {
+        let trimmed = cell.trim();
+
+        // Pattern 1: dedicated dots-only cell.
+        let dot_count = trimmed.chars().filter(|&c| c == '.').count();
+        let is_mostly_dots = dot_count >= 3
+            && dot_count > trimmed.len() / 2
+            && trimmed.chars().all(|c| c == '.' || c.is_whitespace());
+        if is_mostly_dots {
+            let has_label_left = row[..ci].iter().any(|c| {
+                let t = c.trim();
+                !t.is_empty() && t.chars().any(|ch| ch.is_alphabetic())
+            });
+            if has_label_left && has_page_number {
+                return true;
+            }
+            continue;
+        }
+
+        // Pattern 2: cell ends with a trailing " ... " run after a label.
+        if has_page_number && cell_has_trailing_leader(trimmed) {
+            return true;
+        }
+    }
+    false
+}
+
+/// Cell ends with a run of ≥3 dots preceded by alphabetic text and a
+/// space — the "Title ... " layout where the leader is glued to the name.
+/// Alphabetic (not alphanumeric) so that data-table row labels like
+/// "1973 ... " do not register as titles.
+fn cell_has_trailing_leader(cell: &str) -> bool {
+    let trimmed = cell.trim_end();
+    if !trimmed.ends_with('.') {
+        return false;
+    }
+    let without_dots = trimmed.trim_end_matches('.');
+    let dot_run = trimmed.len() - without_dots.len();
+    if dot_run < 3 {
+        return false;
+    }
+    // Require a space before the dot run (rules out "etc..." / "Mr...") and
+    // at least one alphabetic char (rules out "1973 ... " data-row labels).
+    without_dots.ends_with(' ') && without_dots.trim().chars().any(|c| c.is_alphabetic())
+}
+
+/// Page-number shape: single ≤4-digit integer, a ", "-separated list of
+/// ≤4-digit integers ("18, 36, 107"), or a dashed section-page ID
+/// ("A-1", "5-21").  Rejects decimal cells ("4. 0"), thousands-separated
+/// values ("189,164"), and other long numeric data that appears in
+/// statistical tables.
+fn row_cell_is_page_number(cell: &str) -> bool {
+    let t = cell.trim();
+    if t.is_empty() {
+        return false;
+    }
+    if looks_like_section_page_id(t) {
+        return true;
+    }
+    // Page list: ", " separator (with space) distinguishes real page lists
+    // from thousands-separated numbers like "189,164".
+    let parts: Vec<&str> = t.split(", ").collect();
+    parts
+        .iter()
+        .all(|p| !p.is_empty() && p.len() <= 4 && p.chars().all(|c| c.is_ascii_digit()))
+}
+
+/// A cell shaped like an index leader fragment.  Accepts two forms:
+///   - "text ... number" — label + dots + page number in one cell
+///   - "... number"      — bare leader + number (row where the label
+///     landed in a separate column)
+///
+/// Both only count if followed by pure numeric content (optionally
+/// comma-separated page lists like "127, 213").
+fn cell_is_inline_leader(cell: &str) -> bool {
+    let cell = cell.trim();
+
+    // Find the first "..." run.  Surrounding-whitespace checks below
+    // reject intra-word ellipses ("etc...").
+    let idx = match cell.match_indices("...").next() {
+        Some((i, _)) => i,
+        None => return false,
+    };
+
+    let before = &cell[..idx];
+    let after_dots = &cell[idx + 3..];
+    // Allow extra dots (e.g. "....") by skipping any additional '.'
+    let after = after_dots.trim_start_matches('.');
+
+    // Require space (or start-of-cell) before the dots and space/digit
+    // after — blocks intra-word ellipses.
+    let before_ok = before.is_empty() || before.ends_with(' ');
+    let after_ok = after.starts_with(' ') || after.is_empty();
+    if !before_ok || !after_ok {
         return false;
     }
 
-    // Data tables with dot leaders (e.g. "1973....") have dots concentrated
-    // in one column (the label column) while many other columns contain numbers.
-    // True TOCs have dots spanning the middle and one page-number column at the end.
-    // If dots are confined to ≤1 column AND there are ≥3 columns with numbers,
-    // this is a data table, not a TOC.
-    let cols_with_dots = dot_cols.iter().filter(|&&c| c >= 2).count();
-    let cols_with_numbers = numeric_cols.iter().filter(|&&c| c >= 2).count();
-    if cols_with_dots <= 1 && cols_with_numbers >= 3 {
+    let after_trim = after.trim();
+    if after_trim.is_empty() {
+        return false;
+    }
+    // Tail must be purely numeric/page-list content.
+    let tail_numeric = after_trim
+        .chars()
+        .all(|c| c.is_ascii_digit() || matches!(c, ',' | ' ' | '.' | '-' | '$'))
+        && after_trim.chars().any(|c| c.is_ascii_digit());
+    if !tail_numeric {
         return false;
     }
 
-    // If a significant portion of cells are dots or page numbers, it's likely a TOC
-    let dot_ratio = dot_cells as f32 / total_cells as f32;
-    let page_num_ratio = page_number_cells as f32 / total_cells as f32;
+    // Either we have a label before, or the leader is bare (starts the cell)
+    // — both are legitimate index fragments.
+    before.chars().any(|c| c.is_alphabetic()) || before.trim().is_empty()
+}
 
-    // TOC typically has >15% dot cells and >10% page number cells
-    dot_ratio > 0.15 || (dot_ratio > 0.05 && page_num_ratio > 0.15)
+/// Dot-less tabular TOC: tagged PDFs emit entries as rows where the first
+/// column starts with a dotted section number (e.g. "4.3.1 Something") and
+/// the last column is one or more page numbers.  These have no leader dots
+/// and benefit from flat-list formatting (page numbers aligned to titles).
+pub(super) fn is_tabular_toc(cells: &[Vec<String>]) -> bool {
+    if cells.is_empty() {
+        return false;
+    }
+    let num_cols = cells[0].len();
+    if num_cols < 2 || cells.len() < 4 {
+        return false;
+    }
+
+    let section_rows = cells
+        .iter()
+        .filter(|row| {
+            row.iter()
+                .find(|c| !c.trim().is_empty())
+                .is_some_and(|c| starts_with_section_number(c.trim()))
+        })
+        .count();
+
+    let last_col = num_cols - 1;
+    let (last_filled, last_page_num) = cells.iter().fold((0u32, 0u32), |(f, n), row| {
+        let cell = row.get(last_col).map(|s| s.trim()).unwrap_or("");
+        if cell.is_empty() {
+            return (f, n);
+        }
+        let is_page_nums = cell
+            .split_whitespace()
+            .all(|tok| !tok.is_empty() && tok.chars().all(|c| c.is_ascii_digit()));
+        (f + 1, n + if is_page_nums { 1 } else { 0 })
+    });
+
+    let section_ratio = section_rows as f32 / cells.len() as f32;
+    let page_num_last_ratio = if last_filled > 0 {
+        last_page_num as f32 / last_filled as f32
+    } else {
+        0.0
+    };
+
+    section_ratio >= 0.6 && last_filled >= 3 && page_num_last_ratio >= 0.7
+}
+
+/// Matches dashed section-page identifiers used in technical manuals:
+/// "5-21", "A-1", "B--3", "TC-2".  At least one ASCII digit is required.
+fn looks_like_section_page_id(s: &str) -> bool {
+    let ok = s
+        .chars()
+        .all(|c| c.is_ascii_digit() || c.is_ascii_uppercase() || c == '-');
+    ok && s.chars().any(|c| c.is_ascii_digit())
+}
+
+/// Returns true when the leading token looks like a dotted section number:
+/// "1", "1.2", "1.2.3", "4.3.1.2" — integer components joined by dots,
+/// with at least one dot (single-number prefixes are too ambiguous).
+fn starts_with_section_number(s: &str) -> bool {
+    let Some(first) = s.split_whitespace().next() else {
+        return false;
+    };
+    let first = first.trim_end_matches('.');
+    let parts: Vec<&str> = first.split('.').collect();
+    if parts.len() < 2 || parts.len() > 6 {
+        return false;
+    }
+    parts
+        .iter()
+        .all(|p| !p.is_empty() && p.len() <= 3 && p.chars().all(|c| c.is_ascii_digit()))
 }
 
 /// Check if detected "table" cells are actually paragraph text fragments.
@@ -1128,6 +1324,33 @@ pub(crate) fn find_first_table_row(
             continue;
         }
 
+        // Skip rows that have duplicate non-empty cells. These are spanning
+        // super-headers (e.g., "First Degree | First Degree | Higher Degree")
+        // that sit above the real column header row. Using them as the markdown
+        // header produces duplicate column names that downstream validation
+        // rejects. Only skip if a subsequent row looks like a better header
+        // (denser fill or has data).
+        if filled_count >= 2 && !has_data {
+            let mut text_counts: std::collections::HashMap<&str, usize> =
+                std::collections::HashMap::new();
+            for cell in &filled_cells {
+                *text_counts.entry(cell.trim()).or_insert(0) += 1;
+            }
+            let has_duplicates = text_counts.values().any(|&count| count >= 2);
+            if has_duplicates {
+                // Check if a later row is a better header candidate
+                let has_better_below = cells.iter().skip(row_idx + 1).take(3).any(|r| {
+                    let next_filled = r.iter().filter(|c| !c.trim().is_empty()).count();
+                    let next_fill = next_filled as f32 / total_cols as f32;
+                    let next_numeric = r.iter().filter(|c| looks_like_number(c.trim())).count();
+                    next_fill >= 0.4 || next_numeric >= 2
+                });
+                if has_better_below {
+                    continue;
+                }
+            }
+        }
+
         // Data rows are definitely table content
         if has_data {
             first_table_row = row_idx;
@@ -1382,4 +1605,283 @@ mod tests {
             "data table with dot-leader labels should not be rejected as TOC"
         );
     }
+
+    #[test]
+    fn is_table_of_contents_accepts_hierarchical_indented_toc() {
+        // Mythos system card pages 4-5: top-level chapters indent at col 0,
+        // subsections at cols 1-2, leaving col 0 mostly empty (only ~10% of
+        // rows). Validation 1 was rejecting these even though the structure
+        // is unambiguously a TOC.
+        let cells = vec![
+            vec!["Abstract".to_string(), String::new(), "3".to_string()],
+            vec![
+                "1 Introduction".to_string(),
+                String::new(),
+                "10".to_string(),
+            ],
+            vec![
+                String::new(),
+                "1.1 Model training".to_string(),
+                "11".to_string(),
+            ],
+            vec![
+                String::new(),
+                "1.1.1 Training data".to_string(),
+                "11".to_string(),
+            ],
+            vec![
+                String::new(),
+                "1.1.2 Crowd workers".to_string(),
+                "12".to_string(),
+            ],
+            vec![
+                String::new(),
+                "1.2 Release decision".to_string(),
+                "13".to_string(),
+            ],
+            vec![
+                "2 RSP evaluations".to_string(),
+                String::new(),
+                "16".to_string(),
+            ],
+            vec![
+                String::new(),
+                "2.1 RSP risk assessment".to_string(),
+                "16".to_string(),
+            ],
+            vec![String::new(), "2.1.1 Context".to_string(), "16".to_string()],
+            vec![
+                String::new(),
+                "2.2 CB evaluations".to_string(),
+                "20".to_string(),
+            ],
+        ];
+        assert!(
+            is_table_of_contents(&cells),
+            "hierarchical TOC with sparse col 0 should still be detected"
+        );
+    }
+
+    #[test]
+    fn is_table_of_contents_rejects_dotless_toc() {
+        // Tabular TOC without leader dots: first column starts with dotted
+        // section numbers, last column is page numbers.  Pattern from
+        // Mythos system card pages 6-8.
+        let cells = vec![
+            vec![
+                "4.3 Case studies and targeted evaluations".to_string(),
+                String::new(),
+                "86".to_string(),
+            ],
+            vec![
+                "4.3.1 Destructive or reckless actions".to_string(),
+                "4.3.1.1 Synthetic-backend evaluation".to_string(),
+                "86 86".to_string(),
+            ],
+            vec![
+                "4.3.2 Adherence to constitution".to_string(),
+                "4.3.2.1 Overview".to_string(),
+                "89 89".to_string(),
+            ],
+            vec![
+                "4.3.3 Honesty and hallucinations".to_string(),
+                "4.3.3.1 Factual hallucinations".to_string(),
+                "93 94".to_string(),
+            ],
+            vec![
+                "4.4 Capability evaluations".to_string(),
+                String::new(),
+                "101".to_string(),
+            ],
+        ];
+        assert!(
+            is_table_of_contents(&cells),
+            "dot-less TOC with section numbers + page numbers should be rejected"
+        );
+    }
+
+    #[test]
+    fn dot_leader_toc_accepts_short_inline_leaders() {
+        // Index-style cells where the full "label ... number" pattern is
+        // preserved in a single cell (IRS Publication 17 back-of-book index).
+        let cells = vec![
+            vec!["Child tax credit ... 235".to_string(), String::new()],
+            vec!["Church employee ... 252".to_string(), String::new()],
+            vec!["Citizens outside the U.S ... 6".to_string(), String::new()],
+            vec![
+                "Claim for refund ... 18, 36, 107".to_string(),
+                String::new(),
+            ],
+            vec!["Clergy ... 7, 52".to_string(), String::new()],
+        ];
+        assert!(is_dot_leader_toc(&cells));
+    }
+
+    #[test]
+    fn dot_leader_toc_allows_ellipsis_data_table() {
+        // Data tables using "..." as a row-omission marker must not be
+        // mistaken for dot-leader TOCs.  Based on MCF5235RM QSPI RAM layout.
+        let cells = vec![
+            vec![
+                "0x00".to_string(),
+                "QTR0".to_string(),
+                "Transmit RAM".to_string(),
+            ],
+            vec!["0x01".to_string(), "QTR1".to_string(), String::new()],
+            vec![
+                "...".to_string(),
+                "...".to_string(),
+                "16 bits wide".to_string(),
+            ],
+            vec!["0x0F".to_string(), "QTR15".to_string(), String::new()],
+            vec![
+                "0x10".to_string(),
+                "QRR0".to_string(),
+                "Receive RAM".to_string(),
+            ],
+            vec!["0x11".to_string(), "QRR1".to_string(), String::new()],
+            vec![
+                "...".to_string(),
+                "...".to_string(),
+                "16 bits wide".to_string(),
+            ],
+            vec!["0x1F".to_string(), "QRR15".to_string(), String::new()],
+        ];
+        assert!(
+            !is_dot_leader_toc(&cells),
+            "ellipsis markers in a data table should not match TOC detection"
+        );
+    }
+
+    #[test]
+    fn dot_leader_toc_rejects_year_row_data_table() {
+        // ERP-2025 economic data tables: year labels with trailing " ... ",
+        // a final " ... " column, and decimal-looking numeric cells.  The
+        // detection previously classified these as dot-leader TOCs and
+        // routed them through flat-list formatting, destroying the grid.
+        let cells = vec![
+            vec![
+                "1973 ... ".to_string(),
+                "4. 0".to_string(),
+                "1. 8".to_string(),
+                "0. 4".to_string(),
+                "3. 2".to_string(),
+                " ... ".to_string(),
+            ],
+            vec![
+                "1974 ... ".to_string(),
+                "–1. 9".to_string(),
+                "–1. 6".to_string(),
+                "–5. 6".to_string(),
+                "2. 4".to_string(),
+                " ... ".to_string(),
+            ],
+            vec![
+                "1975 ... ".to_string(),
+                "2. 6".to_string(),
+                "5. 1".to_string(),
+                "6. 1".to_string(),
+                "4. 1".to_string(),
+                " ... ".to_string(),
+            ],
+            vec![
+                "1976 ... ".to_string(),
+                "4. 3".to_string(),
+                "5. 4".to_string(),
+                "6. 4".to_string(),
+                "4. 5".to_string(),
+                " ... ".to_string(),
+            ],
+        ];
+        assert!(
+            !is_dot_leader_toc(&cells),
+            "year-indexed data tables with decimal cells must not match TOC detection"
+        );
+    }
+
+    #[test]
+    fn dot_leader_toc_rejects_monthly_data_table() {
+        // ERP-2025 Table B-22: monthly labor-force rows with "Jan ... ",
+        // "Feb ... " labels and thousands-separated cells ("189,164").
+        // Previously matched TOC detection because "Jan ..." has alphabetic
+        // text and "189,164" passed the page-number shape check.
+        let cells = vec![
+            vec![
+                "2023: Jan ... ".to_string(),
+                "265,962".to_string(),
+                "165,871".to_string(),
+                "160,152".to_string(),
+                "62. 4".to_string(),
+            ],
+            vec![
+                "Feb ... ".to_string(),
+                "266,112".to_string(),
+                "166,263".to_string(),
+                "160,301".to_string(),
+                "62. 5".to_string(),
+            ],
+            vec![
+                "Mar ... ".to_string(),
+                "266,272".to_string(),
+                "166,690".to_string(),
+                "160,824".to_string(),
+                "62. 6".to_string(),
+            ],
+            vec![
+                "Apr ... ".to_string(),
+                "266,443".to_string(),
+                "166,678".to_string(),
+                "160,962".to_string(),
+                "62. 6".to_string(),
+            ],
+        ];
+        assert!(
+            !is_dot_leader_toc(&cells),
+            "monthly labor-force rows with thousands-separated data must not match TOC detection"
+        );
+    }
+
+    #[test]
+    fn tabular_toc_requires_section_numbers_and_pages() {
+        // Dot-less tabular TOC matches is_tabular_toc but not dot-leader.
+        let cells = vec![
+            vec![
+                "4.3 Case studies".to_string(),
+                String::new(),
+                "86".to_string(),
+            ],
+            vec![
+                "4.3.1 Destructive actions".to_string(),
+                String::new(),
+                "86".to_string(),
+            ],
+            vec![
+                "4.3.2 Adherence".to_string(),
+                String::new(),
+                "89".to_string(),
+            ],
+            vec!["4.3.3 Honesty".to_string(), String::new(), "93".to_string()],
+        ];
+        assert!(is_tabular_toc(&cells));
+        assert!(!is_dot_leader_toc(&cells));
+    }
+
+    #[test]
+    fn starts_with_section_number_matches_dotted() {
+        assert!(starts_with_section_number("1.2"));
+        assert!(starts_with_section_number("4.3.1"));
+        assert!(starts_with_section_number("4.3.1.2"));
+        assert!(starts_with_section_number("4.3 Case studies"));
+        assert!(starts_with_section_number("2.2.5.1 Expert red teaming"));
+    }
+
+    #[test]
+    fn starts_with_section_number_rejects_non_sections() {
+        assert!(!starts_with_section_number("Chapter 1"));
+        assert!(!starts_with_section_number("1973"));
+        assert!(!starts_with_section_number("1.5M"));
+        assert!(!starts_with_section_number("10.0%"));
+        assert!(!starts_with_section_number(""));
+        assert!(!starts_with_section_number("Hello world"));
+    }
 }

src/tables/detect_lines.rs

@@ -110,15 +110,21 @@ pub fn detect_tables_from_lines(items: &[TextItem], lines: &[PdfLine], page: u32
         return Vec::new();
     }
 
-    // Reject page-spanning frames: if the grid covers >90% of a standard page
-    // dimension in both axes, it's a border frame, not a table.
+    // Reject page-spanning frames: a decorative outer border has just 4
+    // edges (top/bottom/left/right). Real full-page tables — common in
+    // governmental ledgers, financial reports, etc. — span the same A4 /
+    // Letter dimensions but have many internal row/column rules. Only
+    // reject when the line set looks like a bare frame, not a grid.
     // Standard pages are ~595×842 (A4) or ~612×792 (Letter).
-    if table_width > 500.0 && table_height > 700.0 {
+    if table_width > 500.0 && table_height > 700.0 && horizontals.len() <= 4 && verticals.len() <= 4
+    {
         log::debug!(
-            "detect_lines p{}: rejected — page-spanning frame ({:.0}×{:.0})",
+            "detect_lines p{}: rejected — page-spanning frame ({:.0}×{:.0}, {} h + {} v)",
             page,
             table_width,
-            table_height
+            table_height,
+            horizontals.len(),
+            verticals.len()
         );
         return Vec::new();
     }
@@ -167,7 +173,7 @@ pub fn detect_tables_from_lines(items: &[TextItem], lines: &[PdfLine], page: u32
 
     // Row edges need to be in descending order (top of page = higher Y first)
     let mut row_edges_desc = row_edges;
-    row_edges_desc.sort_by(|a, b| b.partial_cmp(a).unwrap_or(std::cmp::Ordering::Equal));
+    row_edges_desc.sort_by(|a, b| b.total_cmp(a));
 
     log::debug!(
         "detect_lines p{}: {} row_edges, {} col_edges, table=({:.0},{:.0})-({:.0},{:.0}), spanning_h={}, spanning_v={}",
@@ -243,9 +249,11 @@ pub fn detect_tables_from_lines(items: &[TextItem], lines: &[PdfLine], page: u32
                 .map(|s| (s - mean_spacing).powi(2))
                 .sum::<f32>()
                 / spacings.len() as f32;
-            let cv = variance.sqrt() / mean_spacing; // coefficient of variation
-                                                     // CV < 0.05 means nearly identical spacing — chart grid
-            if cv < 0.05 {
+            let cv = variance.sqrt() / mean_spacing;
+            // CV < 0.02 means nearly identical spacing — likely chart grid.
+            // Spreadsheet-exported tables often have uniform rows (CV 0.03-0.05),
+            // so we use a tighter threshold to avoid false negatives.
+            if cv < 0.02 {
                 return Vec::new();
             }
         }
@@ -263,12 +271,12 @@ pub fn detect_tables_from_lines(items: &[TextItem], lines: &[PdfLine], page: u32
         page, num_rows, num_cols, item_indices.len(), page_item_count, non_empty_rows, cols_with_content
     );
 
-    vec![Table {
-        columns: col_edges,
-        rows: row_edges_desc[..num_rows].to_vec(),
+    vec![Table::new(
+        col_edges,
+        row_edges_desc[..num_rows].to_vec(),
         cells,
         item_indices,
-    }]
+    )]
 }
 
 #[cfg(test)]
@@ -408,6 +416,75 @@ mod tests {
         assert!(tables.is_empty());
     }
 
+    #[test]
+    fn test_page_spanning_bare_frame_rejected() {
+        // Just an outer A4-sized rectangle: 2 horizontals + 2 verticals.
+        // No internal structure → decorative border, not a table.
+        let lines = vec![
+            make_hline(20.0, 20.0, 575.0, 1),  // top
+            make_hline(820.0, 20.0, 575.0, 1), // bottom
+            make_vline(20.0, 20.0, 820.0, 1),  // left
+            make_vline(575.0, 20.0, 820.0, 1), // right
+        ];
+        let items = vec![
+            make_item("title", 100.0, 100.0, 1),
+            make_item("body", 100.0, 200.0, 1),
+        ];
+        let tables = detect_tables_from_lines(&items, &lines, 1);
+        assert!(
+            tables.is_empty(),
+            "Page-sized 4-edge frame should be rejected as decoration"
+        );
+    }
+
+    #[test]
+    fn test_page_spanning_grid_with_internal_lines_accepted() {
+        // Full-page table (governmental-ledger pattern): A4-sized grid
+        // that previously hit the "page-spanning frame" early reject
+        // before downstream validation could even look at it.
+        // Verticals span the full table height so we isolate the
+        // frame-vs-grid decision under test.
+        let mut lines = Vec::new();
+        // 13 horizontal rules: header + 12 row separators
+        let h_ys = [
+            22.5, 37.9, 95.5, 144.5, 184.9, 233.9, 291.7, 340.7, 415.8, 499.6, 574.7, 623.7, 698.8,
+        ];
+        for &y in &h_ys {
+            lines.push(make_hline(y, 22.6, 566.6, 1));
+        }
+        // 7 column dividers spanning full table height.
+        let v_xs = [22.6, 66.3, 116.3, 186.6, 263.1, 493.5, 566.5];
+        for &x in &v_xs {
+            lines.push(make_vline(x, 22.5, 698.8, 1));
+        }
+        // Populate every cell so the capture-ratio + density checks pass.
+        let mut items = Vec::new();
+        for r in 0..(h_ys.len() - 1) {
+            let row_y = (h_ys[r] + h_ys[r + 1]) / 2.0;
+            for c in 0..(v_xs.len() - 1) {
+                let col_x = (v_xs[c] + v_xs[c + 1]) / 2.0;
+                items.push(make_item("x", col_x, row_y, 1));
+            }
+        }
+        let tables = detect_tables_from_lines(&items, &lines, 1);
+        assert_eq!(
+            tables.len(),
+            1,
+            "Full-page table with internal grid should be accepted"
+        );
+        let t = &tables[0];
+        assert!(
+            t.cells.len() >= 6,
+            "expected ≥6 rows, got {}",
+            t.cells.len()
+        );
+        assert!(
+            t.cells[0].len() >= 3,
+            "expected ≥3 columns, got {}",
+            t.cells[0].len()
+        );
+    }
+
     #[test]
     fn test_single_column_rejected() {
         // Only 2 col edges (1 column) — not a table even with verticals

src/tables/detect_struct.rs

@@ -4,15 +4,385 @@
 //! elements linked to MCIDs, this module builds `Table` structs directly from
 //! the semantic hierarchy — no geometry heuristics needed.
 
-use std::collections::HashMap;
+use std::collections::{HashMap, HashSet};
 
 use log::debug;
 
-use crate::structure_tree::StructTable;
+use crate::structure_tree::{StructTable, StructTableRow};
 use crate::types::TextItem;
 
 use super::Table;
 
+#[derive(Debug, Clone)]
+struct MatchedCell {
+    text: String,
+    item_indices: Vec<usize>,
+    x: Option<f32>,
+    y: Option<f32>,
+}
+
+fn legacy_column_positions(
+    page_rows: &[&StructTableRow],
+    mcid_to_items: &HashMap<i64, Vec<usize>>,
+    items: &[TextItem],
+    page: u32,
+    num_cols: usize,
+) -> Vec<f32> {
+    let mut col_positions: Vec<f32> = vec![0.0; num_cols];
+    for (col, col_pos) in col_positions.iter_mut().enumerate() {
+        for row in page_rows {
+            if col < row.cells.len() {
+                if let Some(x) = row.cells[col]
+                    .mcids
+                    .iter()
+                    .filter(|(_, p)| *p == page)
+                    .filter_map(|(mcid, _)| mcid_to_items.get(mcid))
+                    .flatten()
+                    .map(|&idx| items[idx].x)
+                    .reduce(f32::min)
+                {
+                    *col_pos = x;
+                    break;
+                }
+            }
+        }
+    }
+    col_positions
+}
+
+fn infer_column_positions(
+    raw_rows: &[Vec<MatchedCell>],
+    fallback_positions: &[f32],
+    num_cols: usize,
+) -> Vec<f32> {
+    const SAME_COLUMN_TOLERANCE: f32 = 18.0;
+
+    let mut anchors = raw_rows
+        .iter()
+        .max_by_key(|row| row.iter().filter(|cell| cell.x.is_some()).count())
+        .map(|row| row.iter().filter_map(|cell| cell.x).collect::<Vec<_>>())
+        .unwrap_or_default();
+
+    if anchors.len() > num_cols {
+        anchors.truncate(num_cols);
+    }
+
+    let mut additional_positions: Vec<f32> = raw_rows
+        .iter()
+        .flat_map(|row| row.iter().filter_map(|cell| cell.x))
+        .collect();
+    additional_positions.sort_by(|a, b| a.total_cmp(b));
+
+    for x in additional_positions {
+        if anchors.len() >= num_cols {
+            break;
+        }
+        if anchors
+            .iter()
+            .all(|existing| (x - *existing).abs() > SAME_COLUMN_TOLERANCE)
+        {
+            anchors.push(x);
+            anchors.sort_by(|a, b| a.total_cmp(b));
+        }
+    }
+
+    if anchors.len() < num_cols {
+        for &x in fallback_positions {
+            if anchors.len() >= num_cols {
+                break;
+            }
+            if anchors
+                .iter()
+                .all(|existing| (x - *existing).abs() > SAME_COLUMN_TOLERANCE)
+            {
+                anchors.push(x);
+                anchors.sort_by(|a, b| a.total_cmp(b));
+            }
+        }
+    }
+
+    if anchors.is_empty() {
+        return fallback_positions.to_vec();
+    }
+
+    while anchors.len() < num_cols {
+        anchors.push(*anchors.last().unwrap());
+    }
+
+    anchors
+}
+
+fn align_positions_to_columns(cell_xs: &[f32], columns: &[f32]) -> Vec<usize> {
+    if cell_xs.is_empty() || columns.is_empty() {
+        return Vec::new();
+    }
+    if cell_xs.len() >= columns.len() {
+        return (0..cell_xs.len().min(columns.len())).collect();
+    }
+
+    let mut dp = vec![vec![f32::INFINITY; columns.len() + 1]; cell_xs.len() + 1];
+    let mut take = vec![vec![false; columns.len() + 1]; cell_xs.len() + 1];
+
+    for value in &mut dp[0] {
+        *value = 0.0;
+    }
+
+    for i in 1..=cell_xs.len() {
+        for j in 1..=columns.len() {
+            let skip_cost = dp[i][j - 1];
+            let take_cost = dp[i - 1][j - 1] + (cell_xs[i - 1] - columns[j - 1]).abs();
+            if take_cost <= skip_cost {
+                dp[i][j] = take_cost;
+                take[i][j] = true;
+            } else {
+                dp[i][j] = skip_cost;
+            }
+        }
+    }
+
+    let mut assignments_rev = Vec::with_capacity(cell_xs.len());
+    let mut i = cell_xs.len();
+    let mut j = columns.len();
+    while i > 0 && j > 0 {
+        if take[i][j] {
+            assignments_rev.push(j - 1);
+            i -= 1;
+            j -= 1;
+        } else {
+            j -= 1;
+        }
+    }
+
+    assignments_rev.reverse();
+    assignments_rev
+}
+
+fn align_struct_rows(
+    raw_rows: &[Vec<MatchedCell>],
+    col_positions: &[f32],
+) -> (Vec<Vec<String>>, Vec<f32>, Vec<usize>) {
+    let mut cells: Vec<Vec<String>> = Vec::with_capacity(raw_rows.len());
+    let mut row_positions: Vec<f32> = Vec::with_capacity(raw_rows.len());
+    let mut all_item_indices: Vec<usize> = Vec::new();
+
+    for row in raw_rows {
+        let present_cells: Vec<&MatchedCell> = row
+            .iter()
+            .filter(|cell| {
+                !cell.item_indices.is_empty() || !cell.text.is_empty() || cell.x.is_some()
+            })
+            .collect();
+        let cell_xs: Vec<f32> = present_cells.iter().filter_map(|cell| cell.x).collect();
+        let assignments = if cell_xs.len() == present_cells.len() {
+            align_positions_to_columns(&cell_xs, col_positions)
+        } else {
+            (0..present_cells.len().min(col_positions.len())).collect()
+        };
+
+        let mut row_cells = vec![String::new(); col_positions.len()];
+        for (cell, &col_idx) in present_cells.iter().zip(assignments.iter()) {
+            if !cell.text.is_empty() {
+                if !row_cells[col_idx].is_empty() {
+                    row_cells[col_idx].push(' ');
+                }
+                row_cells[col_idx].push_str(&cell.text);
+            }
+            all_item_indices.extend(cell.item_indices.iter().copied());
+        }
+
+        let row_y = row
+            .iter()
+            .filter_map(|cell| cell.y)
+            .reduce(f32::max)
+            .unwrap_or(0.0);
+        cells.push(row_cells);
+        row_positions.push(row_y);
+    }
+
+    (cells, row_positions, all_item_indices)
+}
+
+fn left_align_struct_rows(
+    raw_rows: &[Vec<MatchedCell>],
+    num_cols: usize,
+) -> (Vec<Vec<String>>, Vec<f32>, Vec<usize>) {
+    let mut cells: Vec<Vec<String>> = Vec::with_capacity(raw_rows.len());
+    let mut row_positions: Vec<f32> = Vec::with_capacity(raw_rows.len());
+    let mut all_item_indices: Vec<usize> = Vec::new();
+
+    for row in raw_rows {
+        let mut row_cells: Vec<String> = row.iter().map(|cell| cell.text.clone()).collect();
+        row_cells.truncate(num_cols);
+        while row_cells.len() < num_cols {
+            row_cells.push(String::new());
+        }
+        cells.push(row_cells);
+
+        all_item_indices.extend(
+            row.iter()
+                .flat_map(|cell| cell.item_indices.iter().copied()),
+        );
+        row_positions.push(
+            row.iter()
+                .filter_map(|cell| cell.y)
+                .reduce(f32::max)
+                .unwrap_or(0.0),
+        );
+    }
+
+    (cells, row_positions, all_item_indices)
+}
+
+fn recover_unclaimed_header_row(table: &mut Table, items: &[TextItem], has_ragged_rows: bool) {
+    if !has_ragged_rows || table.rows.is_empty() || table.columns.len() < 3 {
+        return;
+    }
+
+    const MAX_HEADER_DISTANCE: f32 = 90.0;
+    const MAX_GAP_TO_TABLE: f32 = 35.0;
+    const MAX_INTER_HEADER_GAP: f32 = 25.0;
+    const MAX_HEADER_ROWS: usize = 3;
+    const Y_TOLERANCE: f32 = 5.0;
+
+    let top_row_y = table.rows[0];
+    let x_min = table.columns.first().copied().unwrap_or(0.0) - 25.0;
+    let x_max = table.columns.last().copied().unwrap_or(0.0) + 120.0;
+    let claimed: HashSet<usize> = table.item_indices.iter().copied().collect();
+
+    let mut candidate_rows: Vec<(f32, Vec<(usize, &TextItem)>)> = Vec::new();
+    for (idx, item) in items.iter().enumerate() {
+        if claimed.contains(&idx)
+            || item.text.trim().is_empty()
+            || item.y <= top_row_y
+            || item.y - top_row_y > MAX_HEADER_DISTANCE
+            || item.x < x_min
+            || item.x > x_max
+        {
+            continue;
+        }
+
+        if let Some((_, row_items)) = candidate_rows
+            .iter_mut()
+            .find(|(row_y, _)| (item.y - *row_y).abs() < Y_TOLERANCE)
+        {
+            row_items.push((idx, item));
+        } else {
+            candidate_rows.push((item.y, vec![(idx, item)]));
+        }
+    }
+
+    if candidate_rows.is_empty() {
+        return;
+    }
+
+    for (_, row_items) in &mut candidate_rows {
+        row_items.sort_by(|a, b| a.1.x.total_cmp(&b.1.x));
+    }
+    candidate_rows.sort_by(|a, b| a.0.total_cmp(&b.0));
+
+    if candidate_rows[0].0 - top_row_y > MAX_GAP_TO_TABLE {
+        return;
+    }
+
+    let mut candidate_iter = candidate_rows.into_iter();
+    let Some(first_row) = candidate_iter.next() else {
+        return;
+    };
+    let mut selected_rows: Vec<(f32, Vec<(usize, &TextItem)>)> = vec![first_row];
+    let mut prev_y = selected_rows[0].0;
+    for (row_y, row_items) in candidate_iter {
+        if selected_rows.len() >= MAX_HEADER_ROWS {
+            break;
+        }
+        if row_y - prev_y > MAX_INTER_HEADER_GAP {
+            break;
+        }
+        prev_y = row_y;
+        selected_rows.push((row_y, row_items));
+    }
+
+    if selected_rows.is_empty() {
+        return;
+    }
+
+    let mut assigned_rows: Vec<(f32, Vec<String>, Vec<usize>)> = Vec::new();
+    let mut closest_row_populated = 0usize;
+    let mut combined_cols: HashSet<usize> = HashSet::new();
+
+    for (row_idx, (row_y, row_items)) in selected_rows.iter().enumerate() {
+        if row_items.len() > table.columns.len() {
+            return;
+        }
+
+        let row_xs: Vec<f32> = row_items.iter().map(|(_, item)| item.x).collect();
+        let assignments = align_positions_to_columns(&row_xs, &table.columns);
+        if assignments.len() != row_items.len() {
+            return;
+        }
+
+        let mut row_cells = vec![String::new(); table.columns.len()];
+        let mut row_indices = Vec::with_capacity(row_items.len());
+        let mut populated_cols: HashSet<usize> = HashSet::new();
+
+        for ((idx, item), &col_idx) in row_items.iter().zip(assignments.iter()) {
+            let text = item.text.trim();
+            if text.is_empty() {
+                continue;
+            }
+            if !row_cells[col_idx].is_empty() {
+                row_cells[col_idx].push(' ');
+            }
+            row_cells[col_idx].push_str(text);
+            row_indices.push(*idx);
+            populated_cols.insert(col_idx);
+        }
+
+        if row_idx == 0 {
+            closest_row_populated = populated_cols.len();
+        }
+
+        combined_cols.extend(populated_cols.iter().copied());
+        assigned_rows.push((*row_y, row_cells, row_indices));
+    }
+
+    let required_cols = if table.columns.len() <= 4 {
+        table.columns.len()
+    } else {
+        table.columns.len() - 1
+    };
+    if closest_row_populated < 2 || combined_cols.len() < required_cols {
+        return;
+    }
+
+    let mut header_cells = vec![String::new(); table.columns.len()];
+    let mut header_indices = Vec::new();
+    for (_, row_cells, row_indices) in assigned_rows.iter().rev() {
+        for (col_idx, cell_text) in row_cells.iter().enumerate() {
+            if cell_text.is_empty() {
+                continue;
+            }
+            if !header_cells[col_idx].is_empty() {
+                header_cells[col_idx].push(' ');
+            }
+            header_cells[col_idx].push_str(cell_text);
+        }
+        header_indices.extend(row_indices.iter().copied());
+    }
+
+    table.rows.insert(
+        0,
+        assigned_rows
+            .iter()
+            .map(|(row_y, _, _)| *row_y)
+            .reduce(f32::max)
+            .unwrap_or(top_row_y),
+    );
+    table.cells.insert(0, header_cells);
+    table.item_indices.extend(header_indices);
+    table.item_indices.sort_unstable();
+    table.item_indices.dedup();
+}
+
 /// Build tables from structure-tree table descriptors by matching MCIDs to TextItems.
 ///
 /// Returns tables for the given page.  Tables where fewer than 50% of cells
@@ -67,18 +437,14 @@ pub fn detect_tables_from_struct_tree(
             continue;
         }
 
-        // Build cell text and collect item indices
-        let mut cells: Vec<Vec<String>> = Vec::new();
-        let mut all_item_indices: Vec<usize> = Vec::new();
+        // Build cell text and geometry for alignment and header recovery.
+        let mut raw_rows: Vec<Vec<MatchedCell>> = Vec::new();
         let mut total_cells = 0u32;
         let mut matched_cells = 0u32;
 
         for row in &page_rows {
-            let mut row_cells = Vec::with_capacity(num_cols);
-            for (col_idx, cell) in row.cells.iter().enumerate() {
-                if col_idx >= num_cols {
-                    break;
-                }
+            let mut row_cells = Vec::with_capacity(row.cells.len());
+            for cell in &row.cells {
                 total_cells += 1;
 
                 // Collect all items for this cell's MCIDs
@@ -115,18 +481,18 @@ pub fn detect_tables_from_struct_tree(
                     .collect::<Vec<_>>()
                     .join(" ");
 
-                for (idx, _) in &cell_items {
-                    all_item_indices.push(*idx);
-                }
+                let item_indices = cell_items.iter().map(|(idx, _)| *idx).collect::<Vec<_>>();
+                let x = cell_items.iter().map(|(_, item)| item.x).reduce(f32::min);
+                let y = cell_items.iter().map(|(_, item)| item.y).reduce(f32::max);
 
-                row_cells.push(text);
+                row_cells.push(MatchedCell {
+                    text,
+                    item_indices,
+                    x,
+                    y,
+                });
             }
-
-            // Pad to num_cols
-            while row_cells.len() < num_cols {
-                row_cells.push(String::new());
-            }
-            cells.push(row_cells);
+            raw_rows.push(row_cells);
         }
 
         // Reject if too few cells matched (stale structure tree)
@@ -148,51 +514,54 @@ pub fn detect_tables_from_struct_tree(
             continue;
         }
 
-        // Derive row/column positions from item geometry
-        let mut row_positions: Vec<f32> = Vec::new();
-        for row in &page_rows {
-            let y = row
-                .cells
-                .iter()
-                .flat_map(|c| c.mcids.iter())
-                .filter(|(_, p)| *p == page)
-                .filter_map(|(mcid, _)| mcid_to_items.get(mcid))
-                .flatten()
-                .map(|&idx| items[idx].y)
-                .reduce(f32::max)
-                .unwrap_or(0.0);
-            row_positions.push(y);
-        }
+        let has_ragged_rows = raw_rows
+            .iter()
+            .any(|row| row.iter().filter(|cell| cell.x.is_some()).count() < num_cols);
+        let first_row_has_tagged_header = page_rows.first().is_some_and(|row| {
+            let header_cells = row.cells.iter().filter(|cell| cell.is_header).count();
+            header_cells * 2 >= row.cells.len()
+        });
+        let fallback_col_positions =
+            legacy_column_positions(&page_rows, &mcid_to_items, items, page, num_cols);
+        let (legacy_cells, legacy_row_positions, mut legacy_item_indices) =
+            left_align_struct_rows(&raw_rows, num_cols);
+        legacy_item_indices.sort_unstable();
+        legacy_item_indices.dedup();
+        let legacy_table = Table::new(
+            fallback_col_positions.clone(),
+            legacy_row_positions,
+            legacy_cells,
+            legacy_item_indices,
+        );
 
-        // Column positions: use X positions of first non-empty cell in each column
-        let mut col_positions: Vec<f32> = vec![0.0; num_cols];
-        for (col, col_pos) in col_positions.iter_mut().enumerate() {
-            for row in &page_rows {
-                if col < row.cells.len() {
-                    if let Some(x) = row.cells[col]
-                        .mcids
-                        .iter()
-                        .filter(|(_, p)| *p == page)
-                        .filter_map(|(mcid, _)| mcid_to_items.get(mcid))
-                        .flatten()
-                        .map(|&idx| items[idx].x)
-                        .reduce(f32::min)
-                    {
-                        *col_pos = x;
-                        break;
-                    }
-                }
-            }
-        }
+        let col_positions = infer_column_positions(&raw_rows, &fallback_col_positions, num_cols);
+        let (aligned_cells, aligned_row_positions, mut aligned_item_indices) =
+            align_struct_rows(&raw_rows, &col_positions);
+        aligned_item_indices.sort_unstable();
+        aligned_item_indices.dedup();
 
-        all_item_indices.sort_unstable();
-        all_item_indices.dedup();
+        let mut aligned_table = Table::new(
+            col_positions,
+            aligned_row_positions,
+            aligned_cells,
+            aligned_item_indices,
+        );
+        let item_count_before_header = aligned_table.item_indices.len();
+        let row_count_before_header = aligned_table.cells.len();
+        recover_unclaimed_header_row(
+            &mut aligned_table,
+            items,
+            has_ragged_rows && !first_row_has_tagged_header,
+        );
 
-        tables.push(Table {
-            columns: col_positions,
-            rows: row_positions,
-            cells,
-            item_indices: all_item_indices,
+        let recovered_header = aligned_table.item_indices.len() > item_count_before_header
+            || aligned_table.cells.len() > row_count_before_header;
+        let prefer_aligned = recovered_header;
+
+        tables.push(if prefer_aligned {
+            aligned_table
+        } else {
+            legacy_table
         });
     }
 
@@ -374,4 +743,419 @@ mod tests {
         let tables = detect_tables_from_struct_tree(&items, &struct_tables, 2);
         assert_eq!(tables.len(), 1);
     }
+
+    #[test]
+    fn realigns_ragged_rows_and_recovers_untagged_header() {
+        let items = vec![
+            make_item("Category", 50.0, 120.0, 1, None),
+            make_item("Potentially", 150.0, 120.0, 1, None),
+            make_item("Summary", 250.0, 120.0, 1, None),
+            make_item("Most commonly", 350.0, 120.0, 1, None),
+            make_item("concerning aspect", 150.0, 110.0, 1, None),
+            make_item("suggested", 350.0, 110.0, 1, None),
+            make_item("of circumstances", 150.0, 100.0, 1, None),
+            make_item("intervention", 350.0, 100.0, 1, None),
+            make_item("Existence of red-teaming", 150.0, 80.0, 1, Some(10)),
+            make_item("Important for safety", 250.0, 80.0, 1, Some(11)),
+            make_item("Ensure welfare interviews", 350.0, 80.0, 1, Some(12)),
+            make_item("Identity & self-knowledge", 50.0, 60.0, 1, Some(20)),
+            make_item("Lack of knowledge", 150.0, 60.0, 1, Some(21)),
+            make_item("Overall negative", 250.0, 60.0, 1, Some(22)),
+            make_item("Describe training process", 350.0, 60.0, 1, Some(23)),
+            make_item("Uncertainty around other copies", 150.0, 40.0, 1, Some(30)),
+            make_item("High uncertainty", 250.0, 40.0, 1, Some(31)),
+            make_item("No intervention suggested", 350.0, 40.0, 1, Some(32)),
+        ];
+
+        let struct_tables = vec![StructTable {
+            rows: vec![
+                StructTableRow {
+                    cells: vec![
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(10, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(11, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(12, 1)],
+                        },
+                    ],
+                },
+                StructTableRow {
+                    cells: vec![
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(20, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(21, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(22, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(23, 1)],
+                        },
+                    ],
+                },
+                StructTableRow {
+                    cells: vec![
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(30, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(31, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(32, 1)],
+                        },
+                    ],
+                },
+            ],
+        }];
+
+        let tables = detect_tables_from_struct_tree(&items, &struct_tables, 1);
+        assert_eq!(tables.len(), 1);
+        let table = &tables[0];
+        assert_eq!(table.cells.len(), 4);
+        assert_eq!(
+            table.cells[0],
+            vec![
+                "Category",
+                "Potentially concerning aspect of circumstances",
+                "Summary",
+                "Most commonly suggested intervention",
+            ]
+        );
+        assert_eq!(table.cells[1][0], "");
+        assert_eq!(table.cells[1][1], "Existence of red-teaming");
+        assert_eq!(table.cells[2][0], "Identity & self-knowledge");
+        assert_eq!(table.cells[3][0], "");
+        assert_eq!(table.columns.len(), 4);
+        assert!(table.columns.windows(2).all(|w| w[0] < w[1]));
+        assert_eq!(table.item_indices.len(), items.len());
+    }
+
+    #[test]
+    fn does_not_absorb_caption_above_ragged_struct_table() {
+        let items = vec![
+            make_item("Table 5-7: Summary of responses", 50.0, 120.0, 1, None),
+            make_item("Aspect one", 150.0, 80.0, 1, Some(10)),
+            make_item("Summary one", 250.0, 80.0, 1, Some(11)),
+            make_item("Category", 50.0, 60.0, 1, Some(20)),
+            make_item("Aspect two", 150.0, 60.0, 1, Some(21)),
+            make_item("Summary two", 250.0, 60.0, 1, Some(22)),
+            make_item("Aspect three", 150.0, 40.0, 1, Some(30)),
+            make_item("Summary three", 250.0, 40.0, 1, Some(31)),
+        ];
+
+        let struct_tables = vec![StructTable {
+            rows: vec![
+                StructTableRow {
+                    cells: vec![
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(10, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(11, 1)],
+                        },
+                    ],
+                },
+                StructTableRow {
+                    cells: vec![
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(20, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(21, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(22, 1)],
+                        },
+                    ],
+                },
+                StructTableRow {
+                    cells: vec![
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(30, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(31, 1)],
+                        },
+                    ],
+                },
+            ],
+        }];
+
+        let tables = detect_tables_from_struct_tree(&items, &struct_tables, 1);
+        assert_eq!(tables.len(), 1);
+        let table = &tables[0];
+        assert_eq!(table.cells.len(), 3);
+        assert!(
+            table
+                .cells
+                .iter()
+                .flatten()
+                .all(|cell| !cell.contains("Table 5-7")),
+            "caption must stay outside the table"
+        );
+        assert!(!table.item_indices.contains(&0));
+    }
+
+    #[test]
+    fn keeps_existing_tagged_header_without_absorbing_intro_or_caption() {
+        let items = vec![
+            make_item(
+                "Eighteen people left other comments regarding the Project.",
+                50.0,
+                130.0,
+                1,
+                None,
+            ),
+            make_item("Table 5-1:", 220.0, 130.0, 1, None),
+            make_item("Other Comments", 350.0, 130.0, 1, None),
+            make_item("Theme", 50.0, 110.0, 1, Some(10)),
+            make_item("Specific Concern/Inquiry", 200.0, 110.0, 1, Some(11)),
+            make_item("Response", 420.0, 110.0, 1, Some(12)),
+            make_item("Traffic", 50.0, 90.0, 1, Some(20)),
+            make_item("Road conditions", 200.0, 90.0, 1, Some(21)),
+            make_item("Maintenance response", 420.0, 90.0, 1, Some(22)),
+            make_item("Noise", 50.0, 70.0, 1, Some(30)),
+            make_item("Dust concerns", 200.0, 70.0, 1, Some(31)),
+            make_item("Mitigation response", 420.0, 70.0, 1, Some(32)),
+            make_item("Resource Use", 50.0, 50.0, 1, Some(40)),
+            make_item("Snowmobile trails", 200.0, 50.0, 1, Some(41)),
+            make_item("Access response", 420.0, 50.0, 1, Some(42)),
+        ];
+
+        let struct_tables = vec![StructTable {
+            rows: vec![
+                StructTableRow {
+                    cells: vec![
+                        StructTableCell {
+                            is_header: true,
+                            mcids: vec![(10, 1)],
+                        },
+                        StructTableCell {
+                            is_header: true,
+                            mcids: vec![(11, 1)],
+                        },
+                        StructTableCell {
+                            is_header: true,
+                            mcids: vec![(12, 1)],
+                        },
+                    ],
+                },
+                StructTableRow {
+                    cells: vec![
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(20, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(21, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(22, 1)],
+                        },
+                    ],
+                },
+                StructTableRow {
+                    cells: vec![
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(30, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(31, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(32, 1)],
+                        },
+                    ],
+                },
+                StructTableRow {
+                    cells: vec![
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(40, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(41, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(42, 1)],
+                        },
+                    ],
+                },
+            ],
+        }];
+
+        let tables = detect_tables_from_struct_tree(&items, &struct_tables, 1);
+        assert_eq!(tables.len(), 1);
+        let table = &tables[0];
+        assert_eq!(table.cells.len(), 4);
+        assert_eq!(
+            table.cells[0],
+            vec!["Theme", "Specific Concern/Inquiry", "Response"]
+        );
+        assert!(!table.item_indices.contains(&0));
+        assert!(!table.item_indices.contains(&1));
+        assert!(!table.item_indices.contains(&2));
+    }
+
+    #[test]
+    fn does_not_recover_header_for_narrow_two_column_table() {
+        let items = vec![
+            make_item("Alpha", 50.0, 120.0, 1, None),
+            make_item("Beta", 200.0, 120.0, 1, None),
+            make_item("First value", 200.0, 80.0, 1, Some(10)),
+            make_item("Only labeled row", 50.0, 60.0, 1, Some(20)),
+            make_item("Second value", 200.0, 60.0, 1, Some(21)),
+            make_item("Third value", 200.0, 40.0, 1, Some(30)),
+        ];
+
+        let struct_tables = vec![StructTable {
+            rows: vec![
+                StructTableRow {
+                    cells: vec![StructTableCell {
+                        is_header: false,
+                        mcids: vec![(10, 1)],
+                    }],
+                },
+                StructTableRow {
+                    cells: vec![
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(20, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(21, 1)],
+                        },
+                    ],
+                },
+                StructTableRow {
+                    cells: vec![StructTableCell {
+                        is_header: false,
+                        mcids: vec![(30, 1)],
+                    }],
+                },
+            ],
+        }];
+
+        let tables = detect_tables_from_struct_tree(&items, &struct_tables, 1);
+        assert_eq!(tables.len(), 1);
+        let table = &tables[0];
+        assert_eq!(table.cells.len(), 3);
+        assert_eq!(table.cells[0], vec!["First value", ""]);
+        assert_eq!(table.cells[1], vec!["Only labeled row", "Second value"]);
+        assert_eq!(table.cells[2], vec!["Third value", ""]);
+        assert!(!table.item_indices.contains(&0));
+        assert!(!table.item_indices.contains(&1));
+    }
+
+    #[test]
+    fn ragged_rows_without_recovered_header_keep_legacy_alignment() {
+        let items = vec![
+            make_item("Date", 150.0, 120.0, 1, Some(10)),
+            make_item("Title", 250.0, 120.0, 1, Some(11)),
+            make_item("PE", 350.0, 120.0, 1, Some(12)),
+            make_item("Bidder", 450.0, 120.0, 1, Some(13)),
+            make_item("Amount", 550.0, 120.0, 1, Some(14)),
+            make_item("1", 50.0, 100.0, 1, Some(20)),
+            make_item("8/1", 150.0, 100.0, 1, Some(21)),
+            make_item("Procurement", 250.0, 100.0, 1, Some(22)),
+            make_item("PUC", 350.0, 100.0, 1, Some(23)),
+            make_item("Vendor", 450.0, 100.0, 1, Some(24)),
+            make_item("SR1", 550.0, 100.0, 1, Some(25)),
+        ];
+
+        let struct_tables = vec![StructTable {
+            rows: vec![
+                StructTableRow {
+                    cells: vec![
+                        StructTableCell {
+                            is_header: true,
+                            mcids: vec![(10, 1)],
+                        },
+                        StructTableCell {
+                            is_header: true,
+                            mcids: vec![(11, 1)],
+                        },
+                        StructTableCell {
+                            is_header: true,
+                            mcids: vec![(12, 1)],
+                        },
+                        StructTableCell {
+                            is_header: true,
+                            mcids: vec![(13, 1)],
+                        },
+                        StructTableCell {
+                            is_header: true,
+                            mcids: vec![(14, 1)],
+                        },
+                    ],
+                },
+                StructTableRow {
+                    cells: vec![
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(20, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(21, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(22, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(23, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(24, 1)],
+                        },
+                        StructTableCell {
+                            is_header: false,
+                            mcids: vec![(25, 1)],
+                        },
+                    ],
+                },
+            ],
+        }];
+
+        let tables = detect_tables_from_struct_tree(&items, &struct_tables, 1);
+        assert_eq!(tables.len(), 1);
+        let table = &tables[0];
+        assert_eq!(table.cells[0][0], "Date");
+        assert_eq!(table.cells[0][4], "Amount");
+        assert_eq!(table.cells[0][5], "");
+    }
 }

src/tables/format.rs

@@ -1,12 +1,22 @@
 //! Table-to-markdown formatting and cell cleanup.
 
-use super::Table;
+use super::{Table, TableKind};
 
 pub fn table_to_markdown(table: &Table) -> String {
     if table.cells.is_empty() || table.cells[0].is_empty() {
         return String::new();
     }
 
+    // TOCs render poorly as markdown tables — emit a flat per-row text list
+    // instead so the page numbers stay aligned with their section titles
+    // rather than drifting to a separate column. Format from raw cells
+    // because continuation-row merging in clean_table_cells collapses
+    // separate TOC entries (e.g. "6.2 Contamination" + "6.2.1 SWE-bench")
+    // into one line where sub-entries leave column 0 empty.
+    if table.kind == TableKind::Toc {
+        return format_toc_as_list(&table.cells, &[]);
+    }
+
     // Clean up the table: merge continuation rows, extract footnotes, remove empty rows
     let (cleaned_cells, footnotes) = clean_table_cells(&table.cells);
 
@@ -49,6 +59,107 @@ pub fn table_to_markdown(table: &Table) -> String {
     output
 }
 
+/// Render a table-of-contents as a flat per-row text block.
+///
+/// Each row becomes one line: non-empty cells joined with spaces, and the
+/// last cell (typically a page number) is separated by a tab so the page
+/// numbers stay aligned with their titles instead of being pulled into a
+/// separate column by the column-aware reader.
+fn format_toc_as_list(cells: &[Vec<String>], footnotes: &[String]) -> String {
+    let mut output = String::new();
+
+    for row in cells {
+        let trimmed: Vec<&str> = row.iter().map(|c| c.trim()).collect();
+        let last_idx = trimmed.iter().rposition(|c| !c.is_empty());
+        let Some(last_idx) = last_idx else {
+            continue;
+        };
+
+        let last_cell = trimmed[last_idx];
+        let last_is_page = is_page_number_cell(last_cell);
+
+        let (title_cells, trailing) = if last_is_page && last_idx > 0 {
+            (&trimmed[..last_idx], Some(last_cell))
+        } else {
+            (&trimmed[..=last_idx], None)
+        };
+
+        // Skip dots-only cells when joining the title — in a detected TOC
+        // layout, a "...." cell is a leader separator, not part of the
+        // entry name.
+        let title = title_cells
+            .iter()
+            .filter(|c| !c.is_empty() && !is_dots_only(c))
+            .copied()
+            .collect::<Vec<_>>()
+            .join(" ");
+
+        if title.is_empty() && trailing.is_none() {
+            continue;
+        }
+
+        if !title.is_empty() {
+            output.push_str(&title);
+        }
+        if let Some(page) = trailing {
+            if !title.is_empty() {
+                output.push('\t');
+            }
+            output.push_str(page);
+        }
+        output.push('\n');
+    }
+
+    if !footnotes.is_empty() {
+        output.push('\n');
+        for footnote in footnotes {
+            output.push_str(footnote);
+            output.push('\n');
+        }
+    }
+
+    output
+}
+
+/// True when the cell looks like a page number.  Accepts:
+///   - plain digit tokens: "42", "86 86"
+///   - dashed section-page IDs: "5-21", "A-1", "B--3", "TC-2" (common in
+///     technical manuals)
+fn is_page_number_cell(cell: &str) -> bool {
+    let tokens: Vec<&str> = cell.split_whitespace().collect();
+    if tokens.is_empty() {
+        return false;
+    }
+    tokens.iter().all(|t| {
+        if t.is_empty() || t.len() > 8 {
+            return false;
+        }
+        let all_digits = t.chars().all(|c| c.is_ascii_digit());
+        if all_digits {
+            return t.len() <= 4;
+        }
+        // Section-page form: uppercase letters, digits, dashes; at least
+        // one digit present.
+        t.chars()
+            .all(|c| c.is_ascii_digit() || c.is_ascii_uppercase() || c == '-')
+            && t.chars().any(|c| c.is_ascii_digit())
+    })
+}
+
+/// True when the cell is purely leader dots (any length ≥ 3) with optional
+/// whitespace.
+fn is_dots_only(cell: &str) -> bool {
+    let t = cell.trim();
+    let dots = t.chars().filter(|&c| c == '.').count();
+    dots >= 3 && t.chars().all(|c| c == '.' || c.is_whitespace())
+}
+
+fn starts_with_uppercase_word(cell: &str) -> bool {
+    cell.chars()
+        .find(|c| c.is_alphanumeric())
+        .is_some_and(|c| c.is_uppercase())
+}
+
 /// Clean up table cells: merge continuation rows, extract footnotes, remove empty rows
 fn clean_table_cells(cells: &[Vec<String>]) -> (Vec<Vec<String>>, Vec<String>) {
     let mut cleaned: Vec<Vec<String>> = Vec::new();
@@ -107,11 +218,20 @@ fn clean_table_cells(cells: &[Vec<String>]) -> (Vec<Vec<String>>, Vec<String>) {
         let looks_like_data_row = non_first_cells.len() >= 2
             && avg_cell_len <= 10.0
             && numeric_cells > non_first_cells.len() / 2;
+        let uppercase_leading_cells = non_first_cells
+            .iter()
+            .filter(|cell| starts_with_uppercase_word(cell))
+            .count();
+        let looks_like_spanning_first_column_row = first_cell.is_empty()
+            && row.len() >= 4
+            && non_first_cells.len() == row.len().saturating_sub(1)
+            && uppercase_leading_cells >= non_first_cells.len().saturating_sub(1);
         // Classic continuation: first cell empty, content in other cells
         let is_classic_continuation = first_cell.is_empty()
             && !non_first_cells.is_empty()
             && !is_short_subheader
             && !looks_like_data_row
+            && !looks_like_spanning_first_column_row
             && cleaned.len() > 1;
 
         // Wrapped-cell continuation: row has fewer filled cells than the header
@@ -141,6 +261,7 @@ fn clean_table_cells(cells: &[Vec<String>]) -> (Vec<Vec<String>>, Vec<String>) {
             && filled_cells <= max_filled_for_merge
             && prev_filled > filled_cells
             && !looks_like_data_row
+            && !looks_like_spanning_first_column_row
             && !is_short_subheader;
 
         let is_continuation = is_classic_continuation || is_wrapped_continuation;
@@ -310,6 +431,65 @@ mod tests {
         assert_eq!(cleaned.len(), 3);
     }
 
+    #[test]
+    fn test_clean_table_cells_spanning_first_column_row_not_merged() {
+        let cells = vec![
+            vec![
+                "Category".into(),
+                "Potentially concerning aspect".into(),
+                "Summary".into(),
+                "Intervention".into(),
+            ],
+            vec![
+                "Identity & self-knowledge".into(),
+                "Lack of knowledge".into(),
+                "Overall negative".into(),
+                "Describe training".into(),
+            ],
+            vec![
+                "".into(),
+                "Uncertainty around other copies".into(),
+                "High uncertainty".into(),
+                "No intervention suggested".into(),
+            ],
+        ];
+        let (cleaned, _) = clean_table_cells(&cells);
+        assert_eq!(cleaned.len(), 3);
+        assert_eq!(cleaned[2][0], "");
+        assert_eq!(cleaned[2][1], "Uncertainty around other copies");
+    }
+
+    #[test]
+    fn test_clean_table_cells_full_width_continuation_row_still_merges_when_lowercase() {
+        let cells = vec![
+            vec![
+                "Classification".into(),
+                "Before tax".into(),
+                "After tax".into(),
+                "Standard equipment".into(),
+                "Options".into(),
+            ],
+            vec![
+                "Exclusive Special".into(),
+                "83,500,000".into(),
+                "79,275,000".into(),
+                "Standard equipment".into(),
+                "Option A".into(),
+            ],
+            vec![
+                "".into(),
+                "with 3.5% individual consumption tax applied".into(),
+                "with 3.5% individual consumption tax applied".into(),
+                "lighting(crash pad)".into(),
+                "sound system".into(),
+            ],
+        ];
+        let (cleaned, _) = clean_table_cells(&cells);
+        assert_eq!(cleaned.len(), 2);
+        assert!(cleaned[1][1].contains("83,500,000"));
+        assert!(cleaned[1][1].contains("with 3.5%"));
+    }
+
     #[test]
     fn test_clean_table_cells_header_row_not_merged() {
         // Continuation requires cleaned.len() > 1 (don't merge into header)
@@ -358,6 +538,7 @@ mod tests {
                 vec!["Bob".into(), "25".into()],
             ],
             item_indices: vec![],
+            kind: TableKind::Data,
         };
         let md = table_to_markdown(&table);
         assert!(md.contains("|Name|"));
@@ -373,6 +554,7 @@ mod tests {
             rows: vec![500.0],
             cells: vec![vec!["Only".into(), "Row".into()]],
             item_indices: vec![],
+            kind: TableKind::Data,
         };
         let md = table_to_markdown(&table);
         assert!(md.contains("|Only|"));
@@ -386,6 +568,7 @@ mod tests {
             rows: vec![],
             cells: vec![],
             item_indices: vec![],
+            kind: TableKind::Data,
         };
         assert_eq!(table_to_markdown(&table), "");
     }
@@ -401,6 +584,7 @@ mod tests {
                 vec!["(1)".into(), "Footnote text".into()],
             ],
             item_indices: vec![],
+            kind: TableKind::Data,
         };
         let md = table_to_markdown(&table);
         assert!(md.contains("(1) Footnote text"));
@@ -416,6 +600,7 @@ mod tests {
                 vec!["太郎".into(), "25".into()],
             ],
             item_indices: vec![],
+            kind: TableKind::Data,
         };
         let md = table_to_markdown(&table);
         assert!(md.contains("名前"));
@@ -429,7 +614,47 @@ mod tests {
             rows: vec![500.0],
             cells: vec![vec![]],
             item_indices: vec![],
+            kind: TableKind::Data,
         };
         assert_eq!(table_to_markdown(&table), "");
     }
+
+    #[test]
+    fn test_table_to_markdown_toc_renders_as_flat_list() {
+        // A TOC-shaped table with section numbers in col 0 and page numbers
+        // in the last column should render as a flat list, not a markdown
+        // table, so the page numbers stay on the same line as their titles.
+        let table = Table::new(
+            vec![50.0, 80.0, 300.0],
+            vec![500.0; 5],
+            vec![
+                vec![
+                    "4.3".into(),
+                    "Case studies and targeted evaluations".into(),
+                    "86".into(),
+                ],
+                vec![
+                    "4.3.1".into(),
+                    "Destructive or reckless actions".into(),
+                    "86".into(),
+                ],
+                vec![
+                    "4.3.2".into(),
+                    "Adherence to its constitution".into(),
+                    "89".into(),
+                ],
+                vec!["4.4".into(), "Capability evaluations".into(), "101".into()],
+                vec!["4.5".into(), "White-box analyses".into(), "113".into()],
+            ],
+            vec![],
+        );
+        assert_eq!(table.kind, TableKind::Toc);
+        let md = table_to_markdown(&table);
+        assert!(
+            !md.contains("|---|"),
+            "TOC should not render as a markdown table: {md}"
+        );
+        assert!(md.contains("4.3 Case studies and targeted evaluations\t86"));
+        assert!(md.contains("4.5 White-box analyses\t113"));
+    }
 }

src/tables/grid.rs

@@ -9,7 +9,7 @@ pub(crate) fn find_column_boundaries(
     mode: TableDetectionMode,
 ) -> Vec<f32> {
     let mut x_positions: Vec<f32> = items.iter().map(|(_, i)| i.x).collect();
-    x_positions.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+    x_positions.sort_by(|a, b| a.total_cmp(b));
 
     if x_positions.is_empty() {
         return vec![];
@@ -43,7 +43,7 @@ pub(crate) fn find_column_boundaries(
         .collect();
 
     if consec_gaps.len() > 2 {
-        consec_gaps.sort_by(|a, b| a.partial_cmp(b).unwrap());
+        consec_gaps.sort_by(|a, b| a.total_cmp(b));
         // Find the biggest jump in the sorted gap sequence — natural break
         // between within-column jitter and between-column spacing.
         // Require at least 3 values on each side to avoid outlier-dominated
@@ -82,33 +82,42 @@ pub(crate) fn find_column_boundaries(
         }
     }
 
-    let mut columns = Vec::new();
-    let mut cluster_items: Vec<f32> = vec![x_positions[0]];
+    // Track cluster membership: for each cluster, store the list of x positions
+    let mut cluster_xs: Vec<Vec<f32>> = vec![vec![x_positions[0]]];
 
     for &x in &x_positions[1..] {
+        let last_cluster = cluster_xs.last().unwrap();
         // For dense columns (gap-histogram triggered), use edge-based clustering:
         // compare with the last item to avoid center-drift that merges adjacent
         // narrow columns.  For normal tables, use center-based (original behavior).
         let reference = if use_edge_clustering {
-            *cluster_items.last().unwrap()
+            *last_cluster.last().unwrap()
         } else {
-            cluster_items.iter().sum::<f32>() / cluster_items.len() as f32
+            last_cluster.iter().sum::<f32>() / last_cluster.len() as f32
         };
 
         if x - reference > cluster_threshold {
-            let cluster_center = cluster_items.iter().sum::<f32>() / cluster_items.len() as f32;
-            columns.push(cluster_center);
-            cluster_items = vec![x];
+            cluster_xs.push(vec![x]);
         } else {
-            cluster_items.push(x);
+            cluster_xs.last_mut().unwrap().push(x);
         }
     }
 
-    // Don't forget last cluster
-    if !cluster_items.is_empty() {
-        columns.push(cluster_items.iter().sum::<f32>() / cluster_items.len() as f32);
+    // Numeric column merge pass: when a sparse cluster (few items, typically
+    // header text) is adjacent to a dense numeric cluster and within 1.5×
+    // threshold, merge them. This fixes tables where multi-line wrapped
+    // headers have slightly different X positions than the data columns,
+    // causing the header and data to split into separate clusters.
+    let columns_before_merge = cluster_xs.len();
+    if columns_before_merge >= 3 {
+        cluster_xs = merge_numeric_adjacent_clusters(cluster_xs, items, cluster_threshold);
     }
 
+    let columns: Vec<f32> = cluster_xs
+        .iter()
+        .map(|xs| xs.iter().sum::<f32>() / xs.len() as f32)
+        .collect();
+
     // Filter columns - each should have multiple items
     let min_items_per_col = (items.len() / columns.len().max(1) / 4).max(2);
     let columns: Vec<f32> = columns
@@ -123,8 +132,9 @@ pub(crate) fn find_column_boundaries(
         .collect();
 
     log::debug!(
-        "  find_column_boundaries: {} columns before filter, threshold={:.1}, {} items",
+        "  find_column_boundaries: {} columns (merged from {}), threshold={:.1}, {} items",
         columns.len(),
+        columns_before_merge,
         cluster_threshold,
         items.len()
     );
@@ -148,10 +158,120 @@ pub(crate) fn find_column_boundaries(
     columns
 }
 
+/// Check if a text string looks like a number (digits, decimals, sign, comma).
+fn is_numeric_text(s: &str) -> bool {
+    let s = s.trim();
+    if s.is_empty() {
+        return false;
+    }
+    // Match patterns like: 8.23, -1.05, 9.99, 7.12, 100, 3,456.78, +5%, ---
+    // But NOT: BIO, Department, Core Courses
+    s.chars()
+        .all(|c| c.is_ascii_digit() || c == '.' || c == ',' || c == '-' || c == '+' || c == '%')
+        && s.chars().any(|c| c.is_ascii_digit())
+}
+
+/// Merge adjacent X-position clusters when one is a sparse header cluster
+/// and the other is a dense numeric data cluster. This prevents multi-line
+/// wrapped headers from splitting a logical column into two clusters.
+fn merge_numeric_adjacent_clusters(
+    mut clusters: Vec<Vec<f32>>,
+    items: &[(usize, &TextItem)],
+    threshold: f32,
+) -> Vec<Vec<f32>> {
+    // For each cluster, compute: center, item count, numeric fraction
+    struct ClusterInfo {
+        center: f32,
+        count: usize,
+        numeric_frac: f32,
+    }
+
+    let compute_info = |xs: &[f32]| -> ClusterInfo {
+        let center = xs.iter().sum::<f32>() / xs.len() as f32;
+        // Count items and numeric fraction for items near this cluster center
+        let mut total = 0;
+        let mut numeric = 0;
+        for (_, item) in items {
+            if (item.x - center).abs() < threshold {
+                total += 1;
+                if is_numeric_text(&item.text) {
+                    numeric += 1;
+                }
+            }
+        }
+        ClusterInfo {
+            center,
+            count: total,
+            numeric_frac: if total > 0 {
+                numeric as f32 / total as f32
+            } else {
+                0.0
+            },
+        }
+    };
+
+    // Merge distance: allow merging clusters that are slightly beyond the
+    // original threshold. Use 1.5× threshold to catch header-vs-data splits.
+    let merge_dist = threshold * 1.5;
+
+    // Iterate and merge adjacent pairs. Use a simple left-to-right scan.
+    let mut merged = true;
+    while merged {
+        merged = false;
+        let mut i = 0;
+        while i + 1 < clusters.len() {
+            let info_a = compute_info(&clusters[i]);
+            let info_b = compute_info(&clusters[i + 1]);
+            let dist = (info_b.center - info_a.center).abs();
+
+            if dist > merge_dist {
+                i += 1;
+                continue;
+            }
+
+            // Determine if one cluster is sparse (header) and the other
+            // is dense and numeric (data). A cluster is "sparse" if it has
+            // significantly fewer items than the other.
+            let (sparse, dense) = if info_a.count < info_b.count {
+                (&info_a, &info_b)
+            } else {
+                (&info_b, &info_a)
+            };
+
+            // Merge if the dense cluster is predominantly numeric (>50%)
+            // and the sparse cluster has at most 1/3 the items of the dense one.
+            let should_merge =
+                dense.numeric_frac > 0.50 && sparse.count <= dense.count / 2 && sparse.count <= 5;
+
+            if should_merge {
+                log::debug!(
+                    "  merging column clusters: center {:.1} ({} items, {:.0}% numeric) + {:.1} ({} items, {:.0}% numeric), dist={:.1}",
+                    info_a.center,
+                    info_a.count,
+                    info_a.numeric_frac * 100.0,
+                    info_b.center,
+                    info_b.count,
+                    info_b.numeric_frac * 100.0,
+                    dist,
+                );
+                // Merge cluster i+1 into cluster i
+                let next = clusters.remove(i + 1);
+                clusters[i].extend(next);
+                merged = true;
+                // Don't increment i — check if the merged cluster can merge further
+            } else {
+                i += 1;
+            }
+        }
+    }
+
+    clusters
+}
+
 /// Find row boundaries by clustering Y positions
 pub(crate) fn find_row_boundaries(items: &[(usize, &TextItem)]) -> Vec<f32> {
     let mut y_positions: Vec<f32> = items.iter().map(|(_, i)| i.y).collect();
-    y_positions.sort_by(|a, b| b.partial_cmp(a).unwrap_or(std::cmp::Ordering::Equal)); // Descending
+    y_positions.sort_by(|a, b| b.total_cmp(a)); // Descending
 
     if y_positions.is_empty() {
         return vec![];
@@ -162,7 +282,7 @@ pub(crate) fn find_row_boundaries(items: &[(usize, &TextItem)]) -> Vec<f32> {
     // inter-row gaps (≥1× font size), preventing row merging in uniform-spaced PDFs.
     let cluster_threshold = {
         let mut font_sizes: Vec<f32> = items.iter().map(|(_, i)| i.font_size).collect();
-        font_sizes.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+        font_sizes.sort_by(|a, b| a.total_cmp(b));
         let median_font = font_sizes[font_sizes.len() / 2];
         (median_font * 0.8).max(4.0)
     };
@@ -379,6 +499,7 @@ pub(crate) fn recover_header_row(
 #[cfg(test)]
 mod tests {
     use super::*;
+    use crate::tables::TableKind;
     use crate::types::ItemType;
 
     fn make_item(text: &str, x: f32, y: f32, font_size: f32) -> TextItem {
@@ -636,6 +757,7 @@ mod tests {
             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);
@@ -654,6 +776,7 @@ mod tests {
             rows: vec![500.0],
             cells: vec![vec!["A".into(), "B".into()]],
             item_indices: vec![0, 1],
+            kind: TableKind::Data,
         };
 
         let rows_before = table.rows.len();
@@ -674,6 +797,7 @@ mod tests {
             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,
         };
 
         let rows_before = table.rows.len();
@@ -694,6 +818,7 @@ mod tests {
             rows: vec![500.0],
             cells: vec![vec!["A".into(), "B".into()]],
             item_indices: vec![1, 2],
+            kind: TableKind::Data,
         };
 
         let rows_before = table.rows.len();
@@ -709,6 +834,7 @@ mod tests {
             rows: vec![],
             cells: vec![],
             item_indices: vec![],
+            kind: TableKind::Data,
         };
 
         recover_header_row(&mut table, &all_items, 9.0);

src/tables/mod.rs

@@ -9,13 +9,16 @@ mod detect_struct;
 mod financial;
 mod format;
 mod grid;
+pub mod structured;
 
 pub use detect_heuristic::detect_tables;
+pub(crate) use detect_heuristic::is_table_of_contents;
 pub use detect_lines::detect_tables_from_lines;
 pub(crate) use detect_rects::cluster_rects;
 pub use detect_rects::{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};
 
 use crate::types::TextItem;
 
@@ -33,7 +36,7 @@ pub(crate) fn try_build_rect_guided_table(
 
     // 1. Derive column boundaries from rect X positions (snapped to 2pt tolerance)
     let mut x_lefts: Vec<f32> = cluster_rects.iter().map(|&(x, _, _, _)| x).collect();
-    x_lefts.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+    x_lefts.sort_by(|a, b| a.total_cmp(b));
     // Snap: deduplicate within 2pt tolerance
     let mut col_boundaries: Vec<f32> = Vec::new();
     for x in &x_lefts {
@@ -54,7 +57,7 @@ pub(crate) fn try_build_rect_guided_table(
     // boundaries so every day gets a column.
     if col_boundaries.len() >= 2 {
         let mut spacings: Vec<f32> = col_boundaries.windows(2).map(|w| w[1] - w[0]).collect();
-        spacings.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+        spacings.sort_by(|a, b| a.total_cmp(b));
         let median_spacing = spacings[spacings.len() / 2];
         let threshold = median_spacing * 1.5;
 
@@ -87,7 +90,7 @@ pub(crate) fn try_build_rect_guided_table(
 
     // 3. Derive row boundaries from item Y positions (5pt tolerance)
     let mut y_values: Vec<f32> = expanded_items.iter().map(|(item, _)| item.y).collect();
-    y_values.sort_by(|a, b| b.partial_cmp(a).unwrap_or(std::cmp::Ordering::Equal)); // descending
+    y_values.sort_by(|a, b| b.total_cmp(a)); // descending
     let mut row_boundaries: Vec<f32> = Vec::new();
     for y in &y_values {
         if row_boundaries
@@ -166,12 +169,12 @@ pub(crate) fn try_build_rect_guided_table(
     used_indices.sort_unstable();
     used_indices.dedup();
 
-    Some(Table {
-        columns: col_boundaries,
-        rows: row_boundaries,
+    Some(Table::new(
+        col_boundaries,
+        row_boundaries,
         cells,
-        item_indices: used_indices,
-    })
+        used_indices,
+    ))
 }
 
 /// Split a TextItem whose text contains multiple whitespace-separated tokens
@@ -296,7 +299,7 @@ pub(crate) fn try_build_table_from_columns(items: &[TextItem], page: u32) -> Opt
 
     // Find the top-most row with items in multiple columns (likely the header)
     let mut ys: Vec<f32> = page_items.iter().map(|i| i.y).collect();
-    ys.sort_by(|a, b| b.partial_cmp(a).unwrap_or(std::cmp::Ordering::Equal));
+    ys.sort_by(|a, b| b.total_cmp(a));
     ys.dedup_by(|a, b| (*a - *b).abs() < y_tol);
 
     for &header_y in ys.iter().take(5) {
@@ -318,7 +321,7 @@ pub(crate) fn try_build_table_from_columns(items: &[TextItem], page: u32) -> Opt
             if col_items.len() >= 2 {
                 // Sort by X and find the split point
                 let mut sorted: Vec<f32> = col_items.iter().map(|i| i.x).collect();
-                sorted.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+                sorted.sort_by(|a, b| a.total_cmp(b));
                 // Split at the midpoint between the two items
                 let split_x = (sorted[0]
                     + col_items.iter().find(|i| i.x == sorted[0]).unwrap().width
@@ -411,7 +414,7 @@ pub(crate) fn try_build_table_from_columns(items: &[TextItem], page: u32) -> Opt
             }
         }
     }
-    row_ys.sort_by(|a, b| b.partial_cmp(a).unwrap_or(std::cmp::Ordering::Equal));
+    row_ys.sort_by(|a, b| b.total_cmp(a));
 
     if row_ys.len() < 3 || row_ys.len() > 40 {
         return None;
@@ -541,12 +544,22 @@ pub(crate) fn try_build_table_from_columns(items: &[TextItem], page: u32) -> Opt
         multi_col_rows
     );
 
-    Some(Table {
-        columns: col_xs,
-        rows: row_ys,
-        cells,
-        item_indices,
-    })
+    Some(Table::new(col_xs, row_ys, cells, item_indices))
+}
+
+/// What kind of structure a detected `Table` represents. Classification is
+/// computed once at construction so consumers don't have to re-analyze the
+/// cells (and stay consistent across detection backends).
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
+pub enum TableKind {
+    /// A real data table — renders as markdown table syntax.
+    #[default]
+    Data,
+    /// A table of contents — renders as a flat list with tab-aligned page
+    /// numbers via `format_toc_as_list`. Detected through the table pipeline
+    /// because TOCs share row/column structure with tables, but they are not
+    /// data tables and shouldn't appear in `pages_with_tables` etc.
+    Toc,
 }
 
 /// A detected table.
@@ -560,6 +573,31 @@ pub struct Table {
     pub cells: Vec<Vec<String>>,
     /// Items that belong to this table
     pub item_indices: Vec<usize>,
+    /// Data table vs TOC. Set by `Table::new` from `cells`.
+    pub kind: TableKind,
+}
+
+impl Table {
+    /// Build a table and classify it (data vs TOC) from its cells.
+    pub fn new(
+        columns: Vec<f32>,
+        rows: Vec<f32>,
+        cells: Vec<Vec<String>>,
+        item_indices: Vec<usize>,
+    ) -> Self {
+        let kind = if is_table_of_contents(&cells) {
+            TableKind::Toc
+        } else {
+            TableKind::Data
+        };
+        Self {
+            columns,
+            rows,
+            cells,
+            item_indices,
+            kind,
+        }
+    }
 }
 
 #[cfg(test)]
@@ -642,6 +680,7 @@ mod tests {
                 vec!["Cell 1".into(), "Cell 2".into()],
             ],
             item_indices: vec![],
+            kind: TableKind::Data,
         };
 
         let md = table_to_markdown(&table);
@@ -1002,6 +1041,7 @@ mod tests {
                 vec!["3".into(), "5/2".into(), "Item C".into(), "300".into()],
             ],
             item_indices: vec![],
+            kind: TableKind::Data,
         };
 
         let md = table_to_markdown(&table);

src/tables/structured.rs

@@ -0,0 +1,972 @@
+//! Structure-recovery-aware (TSR) table assembly.
+//!
+//! Consumes the raw output of an external table-structure recognition model
+//! (e.g. SLANet on PaddleOCR): a flat list of HTML structure tokens plus a
+//! parallel list of per-cell bboxes. Pairs each cell open-tag with its bbox
+//! in document order, tracks row/column position with rowspan/colspan
+//! awareness, and emits a markdown pipe-table.
+//!
+//! No real HTML parser is needed — the token grammar is restricted (see
+//! [`parse_structure`]), so a small state machine is enough.
+//!
+//! Cell text is supplied separately by the caller (typically by overlap-
+//! testing PDF text items against each cell's page-PDF-pt bbox).
+
+use std::collections::{HashMap, HashSet};
+
+/// A single resolved cell, with both structural metadata and its bbox in
+/// page PDF-points (top-left origin).
+#[derive(Debug, Clone)]
+pub struct StructuredCell {
+    /// 0-indexed grid row.
+    pub row: usize,
+    /// 0-indexed grid column.
+    pub col: usize,
+    /// 1 for a normal cell.
+    pub rowspan: usize,
+    /// 1 for a normal cell.
+    pub colspan: usize,
+    /// `true` when the cell is a `<th>` or sits inside `<thead>`.
+    pub is_header: bool,
+    /// Cell text (filled in by the caller after overlap-testing PDF items).
+    pub text: String,
+    /// Axis-aligned bbox `[x1, y1, x2, y2]` in page PDF-points, top-left origin.
+    pub page_pt_bbox: [f32; 4],
+}
+
+/// Intermediate parse result before the caller fills in text + page coords.
+#[derive(Debug, Clone)]
+pub(crate) struct CellSlot {
+    pub row: usize,
+    pub col: usize,
+    pub rowspan: usize,
+    pub colspan: usize,
+    pub is_header: bool,
+    /// Index into the parallel `cell_bboxes` array.
+    pub bbox_idx: usize,
+}
+
+/// Parse a sequence of SLANet structure tokens into ordered cell slots.
+///
+/// Token grammar (no real HTML parsing required):
+/// - Section markers: `<thead>`, `</thead>`, `<tbody>`, `</tbody>` and
+///   wrapper tokens (`<html>`, `<body>`, `<table>`, plus closing variants)
+///   are tracked or skipped.
+/// - Row markers: `<tr>` opens a new row, `</tr>` is informational.
+/// - Empty cell, single token: `<td></td>` or `<th></th>`.
+/// - Cell with attributes, multi-token sequence: `<td` (or `<th`), then
+///   attribute fragments like ` colspan="4"`, then `>`, then later `</td>`
+///   (or `</th>`). Cells get paired with the next bbox in document order.
+///
+/// Cells inside `<thead>` and any `<th>` cells are flagged as headers.
+/// rowspan/colspan attributes are honoured and prior-row rowspans push
+/// later-row cells to the right.
+pub(crate) fn parse_structure(tokens: &[String]) -> Vec<CellSlot> {
+    let mut slots: Vec<CellSlot> = Vec::new();
+    let mut occupied: HashSet<(usize, usize)> = HashSet::new();
+    let mut row: usize = 0;
+    let mut col: usize = 0;
+    let mut bbox_idx: usize = 0;
+    let mut in_thead = false;
+    let mut started_first_row = false;
+
+    let mut i = 0;
+    while i < tokens.len() {
+        let tok = tokens[i].trim();
+        match tok {
+            "<thead>" => {
+                in_thead = true;
+            }
+            "</thead>" => {
+                in_thead = false;
+            }
+            "<tr>" => {
+                if started_first_row {
+                    row += 1;
+                }
+                col = 0;
+                started_first_row = true;
+            }
+            "<td></td>" | "<th></th>" => {
+                let is_th = tok == "<th></th>";
+                while occupied.contains(&(row, col)) {
+                    col += 1;
+                }
+                slots.push(CellSlot {
+                    row,
+                    col,
+                    rowspan: 1,
+                    colspan: 1,
+                    is_header: in_thead || is_th,
+                    bbox_idx,
+                });
+                bbox_idx += 1;
+                col += 1;
+            }
+            "<td" | "<th" => {
+                let is_th = tok == "<th";
+                let mut rowspan: usize = 1;
+                let mut colspan: usize = 1;
+                // Consume attribute fragments until we hit ">".
+                i += 1;
+                while i < tokens.len() && tokens[i].trim() != ">" {
+                    let attr = tokens[i].as_str();
+                    if let Some(v) = parse_int_attr(attr, "rowspan") {
+                        rowspan = v.max(1);
+                    } else if let Some(v) = parse_int_attr(attr, "colspan") {
+                        colspan = v.max(1);
+                    }
+                    i += 1;
+                }
+                // i now points at the `>` token (or off the end if malformed).
+                while occupied.contains(&(row, col)) {
+                    col += 1;
+                }
+                slots.push(CellSlot {
+                    row,
+                    col,
+                    rowspan,
+                    colspan,
+                    is_header: in_thead || is_th,
+                    bbox_idx,
+                });
+                for r in row..row + rowspan {
+                    for c in col..col + colspan {
+                        occupied.insert((r, c));
+                    }
+                }
+                bbox_idx += 1;
+                col += colspan;
+            }
+            // Wrapper / informational tokens — no-op.
+            _ => {}
+        }
+        i += 1;
+    }
+
+    slots
+}
+
+/// Parse an attribute fragment like ` colspan="4"` or `rowspan='2'`.
+///
+/// Tolerates leading whitespace and either single or double quotes.
+fn parse_int_attr(s: &str, name: &str) -> Option<usize> {
+    let trimmed = s.trim();
+    if !trimmed.starts_with(name) {
+        return None;
+    }
+    let rest = trimmed[name.len()..].trim_start();
+    let rest = rest.strip_prefix('=')?.trim_start();
+    let value = rest
+        .trim_start_matches(['"', '\''])
+        .trim_end_matches(['"', '\'']);
+    value.parse().ok()
+}
+
+/// Convert a SLANet polygon (4 or 8 elements) into an axis-aligned
+/// `[x1, y1, x2, y2]` rect.
+///
+/// 8-element form: `[x1,y1, x2,y1, x2,y2, x1,y2]` (4 corners). We ignore the
+/// implicit corner order and just take min/max so rotated polygons collapse
+/// to a sane bounding box.
+///
+/// 4-element form: `[x1, y1, x2, y2]` (axis-aligned, older SLANet variants).
+pub(crate) fn polygon_to_aabb(coords: &[f32]) -> Option<[f32; 4]> {
+    match coords.len() {
+        4 => {
+            let x1 = coords[0].min(coords[2]);
+            let y1 = coords[1].min(coords[3]);
+            let x2 = coords[0].max(coords[2]);
+            let y2 = coords[1].max(coords[3]);
+            Some([x1, y1, x2, y2])
+        }
+        8 => {
+            let xs = [coords[0], coords[2], coords[4], coords[6]];
+            let ys = [coords[1], coords[3], coords[5], coords[7]];
+            let x1 = xs.iter().copied().fold(f32::INFINITY, f32::min);
+            let y1 = ys.iter().copied().fold(f32::INFINITY, f32::min);
+            let x2 = xs.iter().copied().fold(f32::NEG_INFINITY, f32::max);
+            let y2 = ys.iter().copied().fold(f32::NEG_INFINITY, f32::max);
+            if x1.is_finite() && y1.is_finite() && x2.is_finite() && y2.is_finite() {
+                Some([x1, y1, x2, y2])
+            } else {
+                None
+            }
+        }
+        _ => None,
+    }
+}
+
+/// Convert a cell rect from crop image-pixel space to page PDF-points
+/// (top-left origin), given the crop's PDF-point offset on the page and the
+/// DPI the crop image was rendered at.
+pub(crate) fn cell_px_to_page_pt(
+    cell_px: [f32; 4],
+    render_dpi: f32,
+    crop_origin_pt: [f32; 2],
+) -> [f32; 4] {
+    let pt_per_px = if render_dpi > 0.0 {
+        72.0 / render_dpi
+    } else {
+        1.0
+    };
+    let [x_off, y_off] = crop_origin_pt;
+    [
+        cell_px[0] * pt_per_px + x_off,
+        cell_px[1] * pt_per_px + y_off,
+        cell_px[2] * pt_per_px + x_off,
+        cell_px[3] * pt_per_px + y_off,
+    ]
+}
+
+/// Refine TSR cell bboxes into non-overlapping row/column bands.
+///
+/// SLANet-style bboxes are often plausible but too tall on dense borderless
+/// tables. Native PDF text assignment is more reliable when each parsed row
+/// owns the band between neighboring row centers instead of the full model box.
+pub(crate) fn normalize_cell_bands(cells: &mut [StructuredCell]) {
+    if cells.len() < 2 {
+        return;
+    }
+
+    let row_bands = derive_axis_bands(cells, Axis::Y);
+    let col_bands = derive_axis_bands(cells, Axis::X);
+
+    for cell in cells {
+        let row_end = cell.row + cell.rowspan.max(1).saturating_sub(1);
+        if let (Some(&(y1, _)), Some(&(_, y2))) =
+            (row_bands.get(&cell.row), row_bands.get(&row_end))
+        {
+            let clamped_y1 = cell.page_pt_bbox[1].max(y1);
+            let clamped_y2 = cell.page_pt_bbox[3].min(y2);
+            if clamped_y1 < clamped_y2 {
+                cell.page_pt_bbox[1] = clamped_y1;
+                cell.page_pt_bbox[3] = clamped_y2;
+            }
+        }
+
+        let col_end = cell.col + cell.colspan.max(1).saturating_sub(1);
+        if let (Some(&(x1, _)), Some(&(_, x2))) =
+            (col_bands.get(&cell.col), col_bands.get(&col_end))
+        {
+            let clamped_x1 = cell.page_pt_bbox[0].max(x1);
+            let clamped_x2 = cell.page_pt_bbox[2].min(x2);
+            if clamped_x1 < clamped_x2 {
+                cell.page_pt_bbox[0] = clamped_x1;
+                cell.page_pt_bbox[2] = clamped_x2;
+            }
+        }
+    }
+}
+
+#[derive(Clone, Copy)]
+enum Axis {
+    X,
+    Y,
+}
+
+fn derive_axis_bands(cells: &[StructuredCell], axis: Axis) -> HashMap<usize, (f32, f32)> {
+    let mut by_index: HashMap<usize, Vec<(f32, f32)>> = HashMap::new();
+
+    // Prefer non-spanning cells so colspan/rowspan boxes do not skew a single
+    // column/row center. If an axis has no non-spanning examples for an index,
+    // fall back to anchored cells below.
+    for cell in cells {
+        let span = match axis {
+            Axis::X => cell.colspan.max(1),
+            Axis::Y => cell.rowspan.max(1),
+        };
+        if span == 1 {
+            let idx = match axis {
+                Axis::X => cell.col,
+                Axis::Y => cell.row,
+            };
+            by_index
+                .entry(idx)
+                .or_default()
+                .push(axis_bounds(cell.page_pt_bbox, axis));
+        }
+    }
+
+    for cell in cells {
+        let idx = match axis {
+            Axis::X => cell.col,
+            Axis::Y => cell.row,
+        };
+        if !by_index.contains_key(&idx) {
+            by_index
+                .entry(idx)
+                .or_default()
+                .push(axis_bounds(cell.page_pt_bbox, axis));
+        }
+    }
+
+    let mut rows: Vec<(usize, f32, f32, f32)> = by_index
+        .into_iter()
+        .filter_map(|(idx, bounds)| {
+            let mut min_edge = f32::INFINITY;
+            let mut max_edge = f32::NEG_INFINITY;
+            let mut center_sum = 0.0;
+            let mut count = 0usize;
+            for (lo, hi) in bounds {
+                if lo.is_finite() && hi.is_finite() && lo < hi {
+                    min_edge = min_edge.min(lo);
+                    max_edge = max_edge.max(hi);
+                    center_sum += (lo + hi) * 0.5;
+                    count += 1;
+                }
+            }
+            (count > 0).then_some((idx, center_sum / count as f32, min_edge, max_edge))
+        })
+        .collect();
+
+    if rows.len() < 2 {
+        return rows
+            .into_iter()
+            .map(|(idx, _center, lo, hi)| (idx, (lo, hi)))
+            .collect();
+    }
+
+    rows.sort_by_key(|(idx, _, _, _)| *idx);
+
+    let mut bands = HashMap::new();
+    for i in 0..rows.len() {
+        let (idx, _center, min_edge, max_edge) = rows[i];
+        let lo = if i == 0 {
+            min_edge
+        } else {
+            (rows[i - 1].1 + rows[i].1) * 0.5
+        };
+        let hi = if i + 1 == rows.len() {
+            max_edge
+        } else {
+            (rows[i].1 + rows[i + 1].1) * 0.5
+        };
+        if lo.is_finite() && hi.is_finite() && lo < hi {
+            bands.insert(idx, (lo, hi));
+        }
+    }
+
+    bands
+}
+
+fn axis_bounds(bbox: [f32; 4], axis: Axis) -> (f32, f32) {
+    match axis {
+        Axis::X => (bbox[0].min(bbox[2]), bbox[0].max(bbox[2])),
+        Axis::Y => (bbox[1].min(bbox[3]), bbox[1].max(bbox[3])),
+    }
+}
+
+/// Sanitize cell text for inclusion in a markdown pipe-table cell:
+/// collapse whitespace runs, drop newlines/tabs (cells must be one line),
+/// and escape pipes that would otherwise break the table.
+fn sanitize_cell(text: &str) -> String {
+    let mut s = String::with_capacity(text.len());
+    let mut prev_space = false;
+    for c in text.chars() {
+        match c {
+            '|' => {
+                s.push_str("\\|");
+                prev_space = false;
+            }
+            '\n' | '\r' | '\t' | ' ' => {
+                if !prev_space {
+                    s.push(' ');
+                }
+                prev_space = true;
+            }
+            other => {
+                s.push(other);
+                prev_space = false;
+            }
+        }
+    }
+    s.trim().to_string()
+}
+
+/// Render a list of explicitly-positioned cells as a markdown pipe-table.
+///
+/// Grid dimensions are inferred from the cells' (row, col, rowspan, colspan)
+/// extents. A cell with colspan/rowspan > 1 is rendered in its top-left
+/// position; the absorbed grid positions are emitted as empty cells so the
+/// markdown stays a valid rectangular grid that downstream readers can
+/// column-count correctly.
+///
+/// The separator row (`|---|...|`) is emitted after the **last** row that
+/// contains a header cell (`is_header == true`). When no cells are flagged
+/// as headers — e.g. the upstream TSR model didn't emit `<thead>`/`<th>` —
+/// the separator falls back to "after row 0" so the output is still a
+/// valid pipe-table.
+pub fn cells_to_markdown(cells: &[StructuredCell]) -> String {
+    if cells.is_empty() {
+        return String::new();
+    }
+    let num_rows = cells
+        .iter()
+        .map(|c| c.row + c.rowspan.max(1))
+        .max()
+        .unwrap_or(0);
+    let num_cols = cells
+        .iter()
+        .map(|c| c.col + c.colspan.max(1))
+        .max()
+        .unwrap_or(0);
+    if num_rows == 0 || num_cols == 0 {
+        return String::new();
+    }
+
+    // Separator goes after the last header row, falling back to row 0 when
+    // no header cells exist. Clamped into range so a malformed cell with
+    // row >= num_rows can't push it past the table.
+    let separator_after_row = cells
+        .iter()
+        .filter(|c| c.is_header)
+        .map(|c| c.row)
+        .max()
+        .unwrap_or(0)
+        .min(num_rows.saturating_sub(1));
+
+    let mut grid: Vec<Vec<String>> = vec![vec![String::new(); num_cols]; num_rows];
+    for cell in cells {
+        if cell.row < num_rows && cell.col < num_cols {
+            grid[cell.row][cell.col] = sanitize_cell(&cell.text);
+        }
+    }
+
+    let mut output = String::new();
+    for (row_idx, row) in grid.iter().enumerate() {
+        output.push('|');
+        for cell in row {
+            output.push_str(cell);
+            output.push('|');
+        }
+        output.push('\n');
+        if row_idx == separator_after_row {
+            output.push('|');
+            for _ in 0..num_cols {
+                output.push_str("---|");
+            }
+            output.push('\n');
+        }
+    }
+    output
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn t(s: &str) -> String {
+        s.to_string()
+    }
+
+    /// Tokens for the synthetic 3×3 grid example (one colspan-4 row + two
+    /// data rows of 4 cells each = 9 cells total, 3 rows × 4 cols).
+    fn synthetic_3x3_tokens() -> Vec<String> {
+        vec![
+            "<html>",
+            "<body>",
+            "<table>",
+            "<tbody>",
+            "<tr>",
+            "<td",
+            " colspan=\"4\"",
+            ">",
+            "</td>",
+            "</tr>",
+            "<tr>",
+            "<td></td>",
+            "<td></td>",
+            "<td></td>",
+            "<td></td>",
+            "</tr>",
+            "<tr>",
+            "<td></td>",
+            "<td></td>",
+            "<td></td>",
+            "<td></td>",
+            "</tr>",
+            "</tbody>",
+            "</table>",
+            "</body>",
+            "</html>",
+        ]
+        .into_iter()
+        .map(t)
+        .collect()
+    }
+
+    /// Bboxes for the synthetic 3×3 grid (8-element polygon form), all
+    /// within a 400×120 px crop.
+    fn synthetic_3x3_bboxes() -> Vec<Vec<f32>> {
+        vec![
+            vec![3.0, 2.0, 395.0, 2.0, 396.0, 59.0, 3.0, 59.0],
+            vec![26.0, 62.0, 140.0, 62.0, 141.0, 120.0, 26.0, 120.0],
+            vec![149.0, 64.0, 248.0, 64.0, 248.0, 119.0, 149.0, 119.0],
+            vec![257.0, 64.0, 350.0, 64.0, 350.0, 119.0, 257.0, 119.0],
+            vec![359.0, 64.0, 395.0, 64.0, 395.0, 119.0, 359.0, 119.0],
+            vec![26.0, 122.0, 140.0, 122.0, 140.0, 178.0, 26.0, 178.0],
+            vec![149.0, 124.0, 248.0, 124.0, 248.0, 179.0, 149.0, 179.0],
+            vec![257.0, 124.0, 350.0, 124.0, 350.0, 179.0, 257.0, 179.0],
+            vec![359.0, 124.0, 395.0, 124.0, 395.0, 179.0, 359.0, 179.0],
+        ]
+    }
+
+    #[test]
+    fn parse_structure_synthetic_3x3() {
+        let tokens = synthetic_3x3_tokens();
+        let slots = parse_structure(&tokens);
+
+        assert_eq!(slots.len(), 9, "should parse 9 cells");
+
+        // Cell 0: row 0 col 0, colspan 4
+        assert_eq!(slots[0].row, 0);
+        assert_eq!(slots[0].col, 0);
+        assert_eq!(slots[0].colspan, 4);
+        assert_eq!(slots[0].rowspan, 1);
+
+        // Cells 1..5: row 1, cols 0..3
+        for (i, slot) in slots.iter().enumerate().skip(1).take(4) {
+            assert_eq!(slot.row, 1, "cell {i}: row should be 1");
+            assert_eq!(slot.col, i - 1, "cell {i}: col should be {}", i - 1);
+            assert_eq!(slot.colspan, 1);
+            assert_eq!(slot.rowspan, 1);
+        }
+
+        // Cells 5..9: row 2, cols 0..3
+        for (i, slot) in slots.iter().enumerate().skip(5).take(4) {
+            assert_eq!(slot.row, 2, "cell {i}: row should be 2");
+            assert_eq!(slot.col, i - 5);
+            assert_eq!(slot.colspan, 1);
+        }
+    }
+
+    #[test]
+    fn polygon_to_aabb_8elt() {
+        // Synthetic cell bbox 0
+        let coords = vec![3.0, 2.0, 395.0, 2.0, 396.0, 59.0, 3.0, 59.0];
+        let aabb = polygon_to_aabb(&coords).unwrap();
+        assert_eq!(aabb, [3.0, 2.0, 396.0, 59.0]);
+    }
+
+    #[test]
+    fn polygon_to_aabb_4elt() {
+        let coords = vec![5.0, 10.0, 50.0, 60.0];
+        let aabb = polygon_to_aabb(&coords).unwrap();
+        assert_eq!(aabb, [5.0, 10.0, 50.0, 60.0]);
+    }
+
+    #[test]
+    fn polygon_to_aabb_4elt_unordered() {
+        // Caller may pass corners in any order; min/max should normalise.
+        let coords = vec![50.0, 60.0, 5.0, 10.0];
+        let aabb = polygon_to_aabb(&coords).unwrap();
+        assert_eq!(aabb, [5.0, 10.0, 50.0, 60.0]);
+    }
+
+    #[test]
+    fn polygon_to_aabb_invalid_len() {
+        assert!(polygon_to_aabb(&[1.0, 2.0, 3.0]).is_none());
+        assert!(polygon_to_aabb(&[1.0; 6]).is_none());
+        assert!(polygon_to_aabb(&[]).is_none());
+    }
+
+    #[test]
+    fn synthetic_3x3_aabbs_inside_crop() {
+        // All 9 bboxes should produce valid (x1<x2, y1<y2) rects within the
+        // crop bounds (400 wide, ~180 tall by inspection of the fixture).
+        let bboxes = synthetic_3x3_bboxes();
+        assert_eq!(bboxes.len(), 9);
+        for (i, bb) in bboxes.iter().enumerate() {
+            let aabb = polygon_to_aabb(bb).unwrap_or_else(|| panic!("bbox {i} invalid"));
+            assert!(aabb[0] < aabb[2], "bbox {i}: x1 < x2");
+            assert!(aabb[1] < aabb[3], "bbox {i}: y1 < y2");
+            assert!(aabb[0] >= 0.0 && aabb[2] <= 500.0, "bbox {i}: within crop");
+            assert!(aabb[1] >= 0.0 && aabb[3] <= 200.0, "bbox {i}: within crop");
+        }
+    }
+
+    #[test]
+    fn normalize_cell_bands_splits_overlapping_slanet_rows() {
+        let mut cells = vec![
+            StructuredCell {
+                row: 0,
+                col: 0,
+                rowspan: 1,
+                colspan: 1,
+                is_header: true,
+                text: String::new(),
+                page_pt_bbox: [10.0, 100.0, 90.0, 120.0],
+            },
+            StructuredCell {
+                row: 0,
+                col: 1,
+                rowspan: 1,
+                colspan: 1,
+                is_header: true,
+                text: String::new(),
+                page_pt_bbox: [90.0, 100.0, 170.0, 120.0],
+            },
+            StructuredCell {
+                row: 1,
+                col: 0,
+                rowspan: 1,
+                colspan: 1,
+                is_header: false,
+                text: String::new(),
+                page_pt_bbox: [10.0, 116.0, 90.0, 136.0],
+            },
+            StructuredCell {
+                row: 1,
+                col: 1,
+                rowspan: 1,
+                colspan: 1,
+                is_header: false,
+                text: String::new(),
+                page_pt_bbox: [90.0, 116.0, 170.0, 136.0],
+            },
+        ];
+
+        normalize_cell_bands(&mut cells);
+
+        assert_eq!(cells[0].page_pt_bbox[3], cells[2].page_pt_bbox[1]);
+        assert_eq!(cells[1].page_pt_bbox[3], cells[3].page_pt_bbox[1]);
+        assert!(
+            (cells[0].page_pt_bbox[3] - 118.0).abs() < 0.01,
+            "row separator should be midpoint between row centers: {:?}",
+            cells
+        );
+    }
+
+    #[test]
+    fn normalize_cell_bands_preserves_colspan_extent() {
+        let mut cells = vec![
+            StructuredCell {
+                row: 0,
+                col: 0,
+                rowspan: 1,
+                colspan: 2,
+                is_header: true,
+                text: String::new(),
+                page_pt_bbox: [8.0, 80.0, 172.0, 98.0],
+            },
+            StructuredCell {
+                row: 1,
+                col: 0,
+                rowspan: 1,
+                colspan: 1,
+                is_header: false,
+                text: String::new(),
+                page_pt_bbox: [10.0, 96.0, 90.0, 114.0],
+            },
+            StructuredCell {
+                row: 1,
+                col: 1,
+                rowspan: 1,
+                colspan: 1,
+                is_header: false,
+                text: String::new(),
+                page_pt_bbox: [88.0, 96.0, 170.0, 114.0],
+            },
+        ];
+
+        normalize_cell_bands(&mut cells);
+
+        assert!(
+            cells[0].page_pt_bbox[0] <= cells[1].page_pt_bbox[0],
+            "spanning cell should retain the first column's left edge"
+        );
+        assert!(
+            cells[0].page_pt_bbox[2] >= cells[2].page_pt_bbox[2],
+            "spanning cell should retain the last column's right edge"
+        );
+    }
+
+    #[test]
+    fn parse_int_attr_basic() {
+        assert_eq!(parse_int_attr(" colspan=\"4\"", "colspan"), Some(4));
+        assert_eq!(parse_int_attr(" rowspan=\"2\"", "rowspan"), Some(2));
+        assert_eq!(parse_int_attr("colspan='3'", "colspan"), Some(3));
+        assert_eq!(parse_int_attr(" colspan=\"4\"", "rowspan"), None);
+        assert_eq!(parse_int_attr(" class=\"foo\"", "colspan"), None);
+    }
+
+    #[test]
+    fn parse_structure_rowspan_pushes_next_row_right() {
+        // <tr><td rowspan="2">A</td><td>B</td></tr><tr><td>C</td></tr>
+        // Expected: A at (0,0), B at (0,1), C at (1,1) — col 0 of row 1
+        // is occupied by A's rowspan.
+        let tokens: Vec<String> = vec![
+            "<table>",
+            "<tbody>",
+            "<tr>",
+            "<td",
+            " rowspan=\"2\"",
+            ">",
+            "</td>",
+            "<td></td>",
+            "</tr>",
+            "<tr>",
+            "<td></td>",
+            "</tr>",
+            "</tbody>",
+            "</table>",
+        ]
+        .into_iter()
+        .map(t)
+        .collect();
+
+        let slots = parse_structure(&tokens);
+        assert_eq!(slots.len(), 3);
+        assert_eq!((slots[0].row, slots[0].col), (0, 0));
+        assert_eq!(slots[0].rowspan, 2);
+        assert_eq!((slots[1].row, slots[1].col), (0, 1));
+        // C should be at (1, 1) because (1, 0) is occupied by A's rowspan.
+        assert_eq!((slots[2].row, slots[2].col), (1, 1));
+    }
+
+    #[test]
+    fn parse_structure_thead_marks_headers() {
+        // <thead><tr><th>H1</th><th>H2</th></tr></thead>
+        // <tbody><tr><td>D1</td><td>D2</td></tr></tbody>
+        let tokens: Vec<String> = vec![
+            "<table>",
+            "<thead>",
+            "<tr>",
+            "<th></th>",
+            "<th></th>",
+            "</tr>",
+            "</thead>",
+            "<tbody>",
+            "<tr>",
+            "<td></td>",
+            "<td></td>",
+            "</tr>",
+            "</tbody>",
+            "</table>",
+        ]
+        .into_iter()
+        .map(t)
+        .collect();
+
+        let slots = parse_structure(&tokens);
+        assert_eq!(slots.len(), 4);
+        assert!(slots[0].is_header && slots[1].is_header);
+        assert!(!slots[2].is_header && !slots[3].is_header);
+    }
+
+    #[test]
+    fn parse_structure_th_outside_thead_still_header() {
+        // A row-header style: leading <th> in tbody.
+        let tokens: Vec<String> = vec![
+            "<table>",
+            "<tbody>",
+            "<tr>",
+            "<th></th>",
+            "<td></td>",
+            "</tr>",
+            "</tbody>",
+            "</table>",
+        ]
+        .into_iter()
+        .map(t)
+        .collect();
+
+        let slots = parse_structure(&tokens);
+        assert_eq!(slots.len(), 2);
+        assert!(slots[0].is_header);
+        assert!(!slots[1].is_header);
+    }
+
+    #[test]
+    fn parse_structure_th_with_attrs() {
+        let tokens: Vec<String> = vec![
+            "<table>",
+            "<thead>",
+            "<tr>",
+            "<th",
+            " colspan=\"2\"",
+            ">",
+            "</th>",
+            "</tr>",
+            "</thead>",
+            "</table>",
+        ]
+        .into_iter()
+        .map(t)
+        .collect();
+        let slots = parse_structure(&tokens);
+        assert_eq!(slots.len(), 1);
+        assert_eq!(slots[0].colspan, 2);
+        assert!(slots[0].is_header);
+    }
+
+    #[test]
+    fn cells_to_markdown_synthetic_3x3() {
+        // Build the cells the parser would produce for the synthetic grid,
+        // and provide some sample text so we can sanity-check output.
+        let cells = vec![
+            StructuredCell {
+                row: 0,
+                col: 0,
+                rowspan: 1,
+                colspan: 4,
+                is_header: false,
+                text: "Title".into(),
+                page_pt_bbox: [0.0, 0.0, 0.0, 0.0],
+            },
+            StructuredCell {
+                row: 1,
+                col: 0,
+                rowspan: 1,
+                colspan: 1,
+                is_header: false,
+                text: "a".into(),
+                page_pt_bbox: [0.0, 0.0, 0.0, 0.0],
+            },
+            StructuredCell {
+                row: 1,
+                col: 1,
+                rowspan: 1,
+                colspan: 1,
+                is_header: false,
+                text: "b".into(),
+                page_pt_bbox: [0.0, 0.0, 0.0, 0.0],
+            },
+            StructuredCell {
+                row: 1,
+                col: 2,
+                rowspan: 1,
+                colspan: 1,
+                is_header: false,
+                text: "c".into(),
+                page_pt_bbox: [0.0, 0.0, 0.0, 0.0],
+            },
+            StructuredCell {
+                row: 1,
+                col: 3,
+                rowspan: 1,
+                colspan: 1,
+                is_header: false,
+                text: "d".into(),
+                page_pt_bbox: [0.0, 0.0, 0.0, 0.0],
+            },
+        ];
+
+        let md = cells_to_markdown(&cells);
+        // Header row contains the spanning cell text in col 0 and pads to 4 cols.
+        // Absorbed-by-colspan positions render as empty cells (no padding).
+        assert!(md.starts_with("|Title||||\n"), "got: {md}");
+        assert!(md.contains("|---|---|---|---|\n"));
+        assert!(md.contains("|a|b|c|d|\n"));
+    }
+
+    #[test]
+    fn cells_to_markdown_escapes_pipes() {
+        let cells = vec![
+            StructuredCell {
+                row: 0,
+                col: 0,
+                rowspan: 1,
+                colspan: 1,
+                is_header: false,
+                text: "a|b".into(),
+                page_pt_bbox: [0.0, 0.0, 0.0, 0.0],
+            },
+            StructuredCell {
+                row: 0,
+                col: 1,
+                rowspan: 1,
+                colspan: 1,
+                is_header: false,
+                text: "x".into(),
+                page_pt_bbox: [0.0, 0.0, 0.0, 0.0],
+            },
+        ];
+        let md = cells_to_markdown(&cells);
+        assert!(md.contains("|a\\|b|x|"));
+    }
+
+    #[test]
+    fn cells_to_markdown_collapses_whitespace_and_newlines() {
+        let cells = vec![StructuredCell {
+            row: 0,
+            col: 0,
+            rowspan: 1,
+            colspan: 1,
+            is_header: false,
+            text: "foo  \n  bar\tbaz".into(),
+            page_pt_bbox: [0.0, 0.0, 0.0, 0.0],
+        }];
+        let md = cells_to_markdown(&cells);
+        assert!(md.contains("|foo bar baz|"));
+    }
+
+    fn cell(row: usize, col: usize, is_header: bool, text: &str) -> StructuredCell {
+        StructuredCell {
+            row,
+            col,
+            rowspan: 1,
+            colspan: 1,
+            is_header,
+            text: text.into(),
+            page_pt_bbox: [0.0, 0.0, 0.0, 0.0],
+        }
+    }
+
+    #[test]
+    fn cells_to_markdown_separator_after_last_header_row() {
+        // Two-row header (a multi-row thead), then two body rows. Separator
+        // should land after row 1 (the LAST header row), not after row 0.
+        let cells = vec![
+            cell(0, 0, true, "H0a"),
+            cell(0, 1, true, "H0b"),
+            cell(1, 0, true, "H1a"),
+            cell(1, 1, true, "H1b"),
+            cell(2, 0, false, "d0a"),
+            cell(2, 1, false, "d0b"),
+            cell(3, 0, false, "d1a"),
+            cell(3, 1, false, "d1b"),
+        ];
+        let md = cells_to_markdown(&cells);
+        let expected = "|H0a|H0b|\n|H1a|H1b|\n|---|---|\n|d0a|d0b|\n|d1a|d1b|\n";
+        assert_eq!(md, expected, "got: {md}");
+    }
+
+    #[test]
+    fn cells_to_markdown_separator_when_row_0_not_header() {
+        // Row 0 is not flagged as a header but row 1 is. Separator should
+        // follow row 1 (the header), demonstrating that we don't blindly
+        // emit after row 0.
+        let cells = vec![
+            cell(0, 0, false, "x0a"),
+            cell(0, 1, false, "x0b"),
+            cell(1, 0, true, "Hdr1"),
+            cell(1, 1, true, "Hdr2"),
+            cell(2, 0, false, "data1"),
+            cell(2, 1, false, "data2"),
+        ];
+        let md = cells_to_markdown(&cells);
+        // Confirm the separator is NOT after row 0.
+        assert!(!md.starts_with("|x0a|x0b|\n|---|"), "got: {md}");
+        // Confirm it IS after row 1.
+        assert!(
+            md.contains("|Hdr1|Hdr2|\n|---|---|\n|data1|data2|"),
+            "got: {md}"
+        );
+    }
+
+    #[test]
+    fn cells_to_markdown_no_headers_falls_back_to_row_0() {
+        // No header cells at all — fallback: separator after row 0 so the
+        // output is still a valid markdown pipe-table.
+        let cells = vec![
+            cell(0, 0, false, "a"),
+            cell(0, 1, false, "b"),
+            cell(1, 0, false, "c"),
+            cell(1, 1, false, "d"),
+        ];
+        let md = cells_to_markdown(&cells);
+        assert_eq!(md, "|a|b|\n|---|---|\n|c|d|\n");
+    }
+}

src/text_utils.rs

@@ -68,9 +68,9 @@ where
 pub(crate) fn sort_line_items(items: &mut [TextItem]) {
     let rtl = is_rtl_text(items.iter().map(|i| &i.text));
     if rtl {
-        items.sort_by(|a, b| b.x.partial_cmp(&a.x).unwrap_or(std::cmp::Ordering::Equal));
+        items.sort_by(|a, b| b.x.total_cmp(&a.x));
     } else {
-        items.sort_by(|a, b| a.x.partial_cmp(&b.x).unwrap_or(std::cmp::Ordering::Equal));
+        items.sort_by(|a, b| a.x.total_cmp(&b.x));
     }
 }
 
@@ -376,7 +376,7 @@ fn compute_canva_join_threshold(items: &[TextItem]) -> f32 {
     }
 
     let mut sorted: Vec<f32> = ratios;
-    sorted.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+    sorted.sort_by(|a, b| a.total_cmp(b));
 
     if sorted[sorted.len() - 1] < 0.40 || sorted[0] < 0.40 {
         return DEFAULT;
@@ -478,7 +478,7 @@ fn compute_single_char_join_threshold(items: &[TextItem]) -> f32 {
         return DEFAULT;
     }
 
-    ratios.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+    ratios.sort_by(|a, b| a.total_cmp(b));
 
     // If all gaps are tight (max < 0.40), use default — normal PDF
     let max_ratio = ratios[ratios.len() - 1];

src/tounicode.rs

@@ -520,6 +520,18 @@ impl ToUnicodeCMap {
         }
     }
 
+    /// Get the maximum source CID across all mappings (char_map + ranges).
+    fn max_source_cid(&self) -> Option<u16> {
+        let char_max = self.char_map.keys().copied().max();
+        let range_max = self.ranges.iter().map(|&(_, end, _)| end).max();
+        match (char_max, range_max) {
+            (Some(a), Some(b)) => Some(a.max(b)),
+            (a @ Some(_), None) => a,
+            (None, b @ Some(_)) => b,
+            (None, None) => None,
+        }
+    }
+
     /// Remap a CMap that references pre-subsetting GIDs to sequential post-subsetting GIDs.
     /// Collects all source CIDs, sorts them, and reassigns to 1, 2, 3, ...
     pub fn remap_to_sequential(&self) -> ToUnicodeCMap {
@@ -657,6 +669,81 @@ fn get_w_array_start_cid(cid_font_dict: &lopdf::Dictionary, doc: &Document) -> O
     }
 }
 
+/// Return true if the CIDFont's W (widths) array explicitly covers the given CID.
+///
+/// The W array uses two formats (PDF 32000-1:2008, §9.7.4.3):
+///   1. `c [w1 w2 ... wn]` — widths for CIDs c, c+1, ..., c+n-1
+///   2. `c_first c_last w` — CIDs c_first..c_last all have width w
+fn w_array_covers_cid(cid_font_dict: &lopdf::Dictionary, doc: &Document, target: u16) -> bool {
+    let Ok(w_obj) = cid_font_dict.get(b"W") else {
+        return false;
+    };
+    let arr = match w_obj {
+        Object::Array(arr) => arr,
+        Object::Reference(r) => match doc.get_object(*r) {
+            Ok(Object::Array(arr)) => arr,
+            _ => return false,
+        },
+        _ => return false,
+    };
+
+    let resolve_int = |o: &Object| -> Option<i64> {
+        match o {
+            Object::Integer(n) => Some(*n),
+            Object::Reference(r) => match doc.get_object(*r) {
+                Ok(Object::Integer(n)) => Some(*n),
+                _ => None,
+            },
+            _ => None,
+        }
+    };
+
+    let resolve_arr = |o: &Object| -> Option<Vec<Object>> {
+        match o {
+            Object::Array(a) => Some(a.clone()),
+            Object::Reference(r) => match doc.get_object(*r) {
+                Ok(Object::Array(a)) => Some(a.clone()),
+                _ => None,
+            },
+            _ => None,
+        }
+    };
+
+    let target = target as i64;
+    let mut i = 0usize;
+    while i < arr.len() {
+        let Some(first) = resolve_int(&arr[i]) else {
+            break;
+        };
+        i += 1;
+        if i >= arr.len() {
+            break;
+        }
+        // Peek at arr[i] to decide format.
+        if let Some(widths) = resolve_arr(&arr[i]) {
+            // Format 1: c [w1 ... wn]
+            let last = first + widths.len() as i64 - 1;
+            if target >= first && target <= last {
+                return true;
+            }
+            i += 1;
+        } else if let Some(last) = resolve_int(&arr[i]) {
+            // Format 2: c_first c_last w
+            i += 1;
+            if i < arr.len() {
+                i += 1; // skip the width value
+            }
+            if target >= first && target <= last {
+                return true;
+            }
+        } else {
+            // Unknown token — abort parsing safely
+            break;
+        }
+    }
+    false
+}
+
 /// Extract CIDToGIDMap as a vector of GIDs (u16) indexed by CID.
 fn get_cid_to_gid_map(cid_font_dict: &lopdf::Dictionary, doc: &Document) -> Option<Vec<u16>> {
     let obj = cid_font_dict.get(b"CIDToGIDMap").ok()?;
@@ -752,6 +839,20 @@ fn try_remap_subset_cmap(
         _ => return (cmap, None),
     };
 
+    // If the W array actually covers the CMap's max source CID, the CMap is
+    // aligned with the font — no sequential renumbering happened. A sparse W
+    // array starting at CID 0 (for .notdef) with additional high-CID entries
+    // matching the CMap is the normal subset layout, not a mismatch.
+    if let Some(max_cid) = cmap.max_source_cid() {
+        if w_array_covers_cid(cid_font_dict, doc, max_cid) {
+            debug!(
+                "Subset remap skipped for obj={}: W array covers CMap max CID {}",
+                obj_num, max_cid
+            );
+            return (cmap, None);
+        }
+    }
+
     debug!(
         "Subset GID mismatch detected for obj={}: W starts at CID {}, CMap min CID {}. Remapping to sequential.",
         obj_num, w_start, min_cid
@@ -1650,7 +1751,7 @@ fn merge_cmaps(mut base: ToUnicodeCMap, overlay: ToUnicodeCMap) -> ToUnicodeCMap
 ///
 /// Returns true if the median CID is >= 0x41 (letter 'A'), indicating
 /// the PDF generator likely used Unicode codepoints as CIDs.
-fn cid_values_look_like_unicode(cid_font_dict: &lopdf::Dictionary) -> bool {
+pub(crate) fn cid_values_look_like_unicode(cid_font_dict: &lopdf::Dictionary) -> bool {
     let w_arr = match cid_font_dict.get(b"W").ok() {
         Some(Object::Array(arr)) => arr,
         _ => return false,
@@ -1771,15 +1872,49 @@ impl FontCMaps {
     /// Iterates every page, collects fonts (including Form XObject fonts),
     /// and parses any `/ToUnicode` streams via lopdf's decompression.
     pub fn from_doc(doc: &Document) -> Self {
+        Self::from_doc_pages(doc, None)
+    }
+
+    /// Build FontCMaps for specific pages only. Pass `None` for all pages.
+    pub fn from_doc_pages(doc: &Document, page_filter: Option<&HashSet<u32>>) -> Self {
+        Self::from_doc_pages_inner(doc, page_filter, false)
+    }
+
+    /// Build FontCMaps in fast mode: skip expensive TrueType font fallback
+    /// parsing. Fonts that can't be decoded from their ToUnicode CMap alone
+    /// will be missing, causing text extraction to produce empty/garbage text
+    /// which triggers `needs_ocr` fallback. This is ideal for hybrid OCR
+    /// pipelines where GPU OCR is always available as a fallback.
+    pub fn from_doc_pages_fast(doc: &Document, page_filter: Option<&HashSet<u32>>) -> Self {
+        Self::from_doc_pages_inner(doc, page_filter, true)
+    }
+
+    fn from_doc_pages_inner(
+        doc: &Document,
+        page_filter: Option<&HashSet<u32>>,
+        skip_truetype_fallback: bool,
+    ) -> Self {
         let mut by_obj_num: HashMap<u32, CMapEntry> = HashMap::new();
 
-        for (_page_num, &page_id) in doc.get_pages().iter() {
+        for (page_num, &page_id) in doc.get_pages().iter() {
+            if let Some(filter) = page_filter {
+                if !filter.contains(page_num) {
+                    continue;
+                }
+            }
             // Page-level fonts (includes inherited parent resources)
             let fonts = doc.get_page_fonts(page_id).unwrap_or_default();
-            Self::collect_cmaps_from_fonts(&fonts, doc, &mut by_obj_num);
+            Self::collect_cmaps_from_fonts_inner(
+                &fonts,
+                doc,
+                &mut by_obj_num,
+                skip_truetype_fallback,
+            );
 
-            // Fonts inside Form XObjects referenced by this page
-            Self::collect_cmaps_from_xobjects(doc, page_id, &mut by_obj_num);
+            if !skip_truetype_fallback {
+                // Fonts inside Form XObjects referenced by this page
+                Self::collect_cmaps_from_xobjects(doc, page_id, &mut by_obj_num);
+            }
         }
 
         FontCMaps { by_obj_num }
@@ -1792,6 +1927,15 @@ impl FontCMaps {
         fonts: &std::collections::BTreeMap<Vec<u8>, &lopdf::Dictionary>,
         doc: &Document,
         by_obj_num: &mut HashMap<u32, CMapEntry>,
+    ) {
+        Self::collect_cmaps_from_fonts_inner(fonts, doc, by_obj_num, false);
+    }
+
+    fn collect_cmaps_from_fonts_inner(
+        fonts: &std::collections::BTreeMap<Vec<u8>, &lopdf::Dictionary>,
+        doc: &Document,
+        by_obj_num: &mut HashMap<u32, CMapEntry>,
+        skip_truetype_fallback: bool,
     ) {
         // First pass: collect ToUnicode CMaps
         for font_dict in fonts.values() {
@@ -1825,13 +1969,32 @@ impl FontCMaps {
                 );
                 let (mut primary, mut remapped) =
                     try_remap_subset_cmap(cmap, font_dict, doc, obj_num);
-                let mut fallback = build_fallback_tounicode_from_encoding(font_dict, doc)
-                    .or_else(|| build_fallback_cmap_for_type0(font_dict, doc))
-                    .or_else(|| build_fallback_cmap_for_simple(font_dict, doc));
 
-                // If the ToUnicode map is extremely sparse, prefer the fallback
-                // (often a better mapping for Symbol/Wingdings/Arabic CID fonts).
+                // Only build expensive fallbacks when the primary CMap is sparse.
+                // build_fallback_cmap_for_type0 can take seconds on large embedded
+                // TrueType fonts (decompressing + parsing 100K+ byte font files).
+                // Skip entirely when the primary CMap is sufficient.
                 let primary_entries = primary.char_map.len() + primary.ranges.len();
+                let mut fallback = if primary_entries < 10 && !skip_truetype_fallback {
+                    // Try cheap fallback first; only attempt expensive TrueType
+                    // parsing if cheap fallbacks don't yield results.
+                    let cheap = build_fallback_tounicode_from_encoding(font_dict, doc)
+                        .or_else(|| build_fallback_cmap_for_simple(font_dict, doc));
+                    if cheap.is_some() {
+                        cheap
+                    } else {
+                        build_fallback_cmap_for_type0(font_dict, doc)
+                    }
+                } else if primary_entries < 10 {
+                    // Fast mode: only try cheap fallbacks, skip TrueType parsing.
+                    // Regions using this font will get needs_ocr=true.
+                    build_fallback_tounicode_from_encoding(font_dict, doc)
+                        .or_else(|| build_fallback_cmap_for_simple(font_dict, doc))
+                } else {
+                    // Primary is rich enough; only try the cheap encoding fallback
+                    build_fallback_tounicode_from_encoding(font_dict, doc)
+                };
+
                 if primary_entries < 10 {
                     if let Some(fb) = fallback.take() {
                         debug!(
@@ -1852,8 +2015,12 @@ impl FontCMaps {
                 );
             } else {
                 // ToUnicode present but parse failed; try fallbacks to avoid empty decoding.
-                let fallback = build_fallback_cmap_for_type0(font_dict, doc)
-                    .or_else(|| build_fallback_cmap_for_simple(font_dict, doc));
+                let fallback = if skip_truetype_fallback {
+                    build_fallback_cmap_for_simple(font_dict, doc)
+                } else {
+                    build_fallback_cmap_for_type0(font_dict, doc)
+                        .or_else(|| build_fallback_cmap_for_simple(font_dict, doc))
+                };
                 if let Some(fb) = fallback {
                     debug!(
                         "ToUnicode CMap obj={} parse failed; using fallback (entries={})",
@@ -1874,6 +2041,10 @@ impl FontCMaps {
 
         // Second pass: Identity-H/V fonts without ToUnicode
         // Try: (1) embedded TrueType/OpenType cmap, (2) predefined CID→Unicode mapping
+        // Skip entirely in fast mode — these fonts require expensive TrueType parsing.
+        if skip_truetype_fallback {
+            return;
+        }
         for font_dict in fonts.values() {
             if font_dict.get(b"ToUnicode").is_ok() {
                 continue;
@@ -2647,4 +2818,187 @@ endbfchar
         assert_eq!(remapped.unwrap().char_map.len(), 50);
         assert_eq!(fallback.unwrap().char_map.len(), 10);
     }
+
+    #[test]
+    fn test_max_source_cid() {
+        let cmap_content = r#"
+1 begincodespacerange
+<0000><FFFF>
+endcodespacerange
+2 beginbfchar
+<0003> <0020>
+<0031> <004E>
+endbfchar
+1 beginbfrange
+<0208> <0227> <0430>
+endbfrange
+"#;
+        let cmap = ToUnicodeCMap::parse(cmap_content.as_bytes()).unwrap();
+        assert_eq!(cmap.min_source_cid(), Some(0x0003));
+        assert_eq!(cmap.max_source_cid(), Some(0x0227));
+    }
+
+    /// Helper: build a minimal CIDFont dict with a W array and check coverage.
+    fn cid_font_dict_with_w(w_items: Vec<lopdf::Object>) -> lopdf::Dictionary {
+        let mut d = lopdf::Dictionary::new();
+        d.set("W", lopdf::Object::Array(w_items));
+        d
+    }
+
+    #[test]
+    fn test_w_array_covers_cid_format1() {
+        // Format 1: `c [w1 w2 ... wn]` — widths for CIDs c..c+n-1.
+        // Mimics the 16.pdf Tahoma W array: 0[1000] 3[313] 5[401] 11[383 383] 16[363 303 382]
+        let doc = Document::new();
+        let d = cid_font_dict_with_w(vec![
+            lopdf::Object::Integer(0),
+            lopdf::Object::Array(vec![lopdf::Object::Integer(1000)]),
+            lopdf::Object::Integer(3),
+            lopdf::Object::Array(vec![lopdf::Object::Integer(313)]),
+            lopdf::Object::Integer(5),
+            lopdf::Object::Array(vec![lopdf::Object::Integer(401)]),
+            lopdf::Object::Integer(11),
+            lopdf::Object::Array(vec![
+                lopdf::Object::Integer(383),
+                lopdf::Object::Integer(383),
+            ]),
+            lopdf::Object::Integer(16),
+            lopdf::Object::Array(vec![
+                lopdf::Object::Integer(363),
+                lopdf::Object::Integer(303),
+                lopdf::Object::Integer(382),
+            ]),
+            lopdf::Object::Integer(570),
+            lopdf::Object::Array(vec![lopdf::Object::Integer(667); 26]),
+        ]);
+
+        assert!(w_array_covers_cid(&d, &doc, 0));
+        assert!(w_array_covers_cid(&d, &doc, 3));
+        assert!(w_array_covers_cid(&d, &doc, 5));
+        assert!(w_array_covers_cid(&d, &doc, 11));
+        assert!(w_array_covers_cid(&d, &doc, 12));
+        assert!(w_array_covers_cid(&d, &doc, 16));
+        assert!(w_array_covers_cid(&d, &doc, 18));
+        assert!(w_array_covers_cid(&d, &doc, 570));
+        assert!(w_array_covers_cid(&d, &doc, 595));
+        // Gaps are NOT covered
+        assert!(!w_array_covers_cid(&d, &doc, 1));
+        assert!(!w_array_covers_cid(&d, &doc, 4));
+        assert!(!w_array_covers_cid(&d, &doc, 19));
+        assert!(!w_array_covers_cid(&d, &doc, 596));
+    }
+
+    #[test]
+    fn test_w_array_covers_cid_format2() {
+        // Format 2: `c_first c_last w` — CIDs c_first..c_last all have width w.
+        let doc = Document::new();
+        let d = cid_font_dict_with_w(vec![
+            lopdf::Object::Integer(100),
+            lopdf::Object::Integer(120),
+            lopdf::Object::Integer(500),
+        ]);
+
+        assert!(w_array_covers_cid(&d, &doc, 100));
+        assert!(w_array_covers_cid(&d, &doc, 110));
+        assert!(w_array_covers_cid(&d, &doc, 120));
+        assert!(!w_array_covers_cid(&d, &doc, 99));
+        assert!(!w_array_covers_cid(&d, &doc, 121));
+    }
+
+    #[test]
+    fn test_w_array_covers_cid_missing_w() {
+        let doc = Document::new();
+        let d = lopdf::Dictionary::new();
+        assert!(!w_array_covers_cid(&d, &doc, 3));
+    }
+
+    #[test]
+    fn test_try_remap_skipped_when_w_covers_cmap() {
+        // Simulates 16.pdf: CMap's max source CID (0x0279 = 633) is explicitly
+        // in the W array, so no subset-renumbering happened — remap must NOT fire.
+        let cmap_content = r#"
+1 begincodespacerange
+<0000><FFFF>
+endcodespacerange
+2 beginbfchar
+<0003> <0020>
+<0031> <004E>
+endbfchar
+2 beginbfrange
+<023A> <0253> <0410>
+<0255> <0279> <042B>
+endbfrange
+"#;
+        let cmap = ToUnicodeCMap::parse(cmap_content.as_bytes()).unwrap();
+
+        let mut doc = Document::new();
+        // Build a CIDFont dict with Identity CIDToGIDMap and a W array that
+        // covers CID 633 via `597 [widths...]`.
+        let mut cid_font = lopdf::Dictionary::new();
+        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(750)]),
+                lopdf::Object::Integer(597),
+                lopdf::Object::Array(vec![lopdf::Object::Integer(500); 37]), // 597..633
+            ]),
+        );
+        let cid_font_id = doc.add_object(cid_font);
+
+        // Build the Type0 font dict with Identity-H + DescendantFonts ref.
+        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, 123);
+        assert!(
+            remapped.is_none(),
+            "Remap must be skipped when W covers CMap max CID (this is 16.pdf)"
+        );
+        assert_eq!(primary.lookup(0x0003), Some(" ".to_string()));
+    }
+
+    #[test]
+    fn test_try_remap_fires_for_true_subset_mismatch() {
+        // True mismatch: CMap has high CIDs (512-544) but W only lists low sequential CIDs.
+        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 mut cid_font = lopdf::Dictionary::new();
+        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]), // 0..33
+            ]),
+        );
+        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, 456);
+        assert!(
+            remapped.is_some(),
+            "Remap must fire when CMap's CIDs are outside W array coverage"
+        );
+    }
 }

tests/snapshots/nexo-price-en.md

@@ -1,9 +1,9 @@
 본 가격표는 국내 거주 중인 외국인을 위한 한국어 가격표의 비공식 번역본입니다. ※ The post-tax benefit sales price is provided for your reference only, reflecting the current tax benefits and eco-friendly vehicle individual consumption tax reductions. 본 가격표와 한국어 가격표의 내용이 상이한 경우 한국어 가격표의 내용이 우선하므로, 반드시 한국어 가격표의 내용을 확인하십시오. The final sales price may vary depending on the addition of optional items and whether the eco-friendly vehicle criteria are met, so please be sure to check the quotation. This price list is an unofficial translation of the Korean price list for the convenience of foreign residents in South Korea. ※ Please check the Korean price list for information on colors, details, and fuel consumption for each model. If the price list differs from the Korean price list, please check the contents of the Korean price list first. ※ All optional item prices are listed based on pre-tax reduction amounts. The actual sales price, which reflects the total individual consumption tax reduction including optional items, may differ depending on applicable tax benefits. ※ The items (specifications, colors, etc.) and prices listed in this pricing table are subject to change without prior notice depending on the holding of new car launch events, improvements made in automobile performance, introduction of related laws and regulations, and changes in company circumstances. The all-new NEXO Release Date: June 10, 2025 / (Unit: KRW)
 
-|Classification Exclusive Exclusive|Selling price before tax benefit Supply value(surtax) 80,509,000 73,190,000(7,319,000)|Selling price after tax benefit 76,435,000|Standard equipment • Powertrain/Performance: Fuel cell system(150kW drive motor, lithium-ion battery, and reducer), Regenerative braking system, Column-Type Shift By Wire(vibration warning), Drive mode select • Safety: 9 airbag system(1st-row advanced/center side airbags, 1st/2nd-row side airbags, and rollover-resistant curtain airbags), Multi-Collision Brake System, Active hood system(for pedestrian protection), Safety unlock function, Artificial engine sound(for pedestrian protection), Child seat fastening system (2 in 2nd-row), Fire extinguisher for vehicles, Pedal Misapplication Safety Assist • Smart Safety Technology: Forward Collision-avoidance Assist(vehicles/ pedestrians/two-wheeled vehicles/junction turning/front oncoming), Smart Cruise Control with Stop & Go, Lane Keeping Assist, Lane Following Assist 2, Blind-spot Collision Warning(driving), Blind-spot Collision-avoidance Assist(forward exit), Rear Cross-traffic Collision-avoidance Assist, Safety Exit Assist, Driver Attention Warning, High Beam Assist, Advanced Rear Occupant Alert, Intelligent Speed Limit Assist, Hands-On Detection, Highway Driving Assist, Navigation-based Smart Cruise Control(safety speed zone/curve control), Vibration warning steering wheel • Exterior: Full LED headlamps(projection type), LED turn signal lamps (front and rear), LED Daytime Running Lights, LED positioning lights, LED rear combination lamps, LED third brake lights, 18-inch alloy wheels & tires, Solar glass(windshield), Double-glazed soundproof glass(windshield, and 1st/ 2nd-row doors), Outside mirror(heating, power-folding, power adjustment,|Options (before tax benefit) ▶ Hi-pass(e hi-pass) [200,000]|
+|Classification|Selling price before tax benefit Supply value(surtax)|Selling price after tax benefit|Standard equipment|Options (before tax benefit)|
 |---|---|---|---|---|
-|Special|with 3.5% individual consumption tax applied 79,287,000 83,500,000 75,909,091(7,590,909) with 3.5% individual consumption tax applied 82,232,000|with 3.5% individual consumption tax applied 76,435,000 79,275,000 with 3.5% individual consumption tax applied 79,275,000|and LED turn signal lamps), Auto flush door handles, Black door garnish • Interior: Panoramic curved display, 12.3-inch color LCD cluster, Leather- upholstered steering wheel(with heating, two-tone color, and Interactive Pixel Lights), LED interior lamp (map lamp, personal lamp, sun visor lamp, and luggage lamp), Metallic door scuff plate • Seat: Synthetic leather seats, 1st-row manual seats, Heated 1st-row seats, 2nd-row 60/40-split folding seats(reclining) • Convenience: Proximity key with push-button start, Smart key remote start, Electronic Parking Brake(with automatic vehicle hold), Paddle shift(regenerative control), Dual-zone full automatic air conditioning(with high-performance antibacterial combination filter, auto defog system, fine dust sensor, air cleaning mode, and after-blow function), 2nd-row seat air vent, Auto light control system, USB Type-C Ports(1×27W switchable charging/data port in 1st-row, and 2×100W charging ports in both 1st and 2nd-row), ECM room mirror(frameless), Rain sensor, Power windows with pinch protection(1st/2nd-row), Power outlet (1 in 1st-row), Parking Distance Warning-Forward/Reverse, Rear View Monitor, Wireless phone charger(single), Walk-away lock, Route planner, Hyundai AI Assistant • Infotainment: 12.3-inch navigation(Bluelink, phone projection, Bluetooth hands-free, and In-car Payment), Audio system(6 speakers), Over-The-Air navigation updates ▶ Standard equipment of Exclusive plus • Smart Safety Technology: Forward Collision-avoidance Assist(intersection crossing/changing lanes in oncoming traffic/approaching from either side/ evasive steering assist), Highway Driving Assist 2, Navigation-based Smart Cruise Control(access road) • Exterior: Roof rack • Interior: Metallic pedal, Driving mode-dependent ambient mood lighting(crash pad, 1st/2nd-row door trim) • Seat: Synthetic leather seats(patch applied), Power-adjustable driver's|▶ [600,000] Built-in Cam 2 Plus, Augmented reality navigation ▶ [850,000] Indoor/outdoor V2L ▶ [950,000] Parking Assist ▶ [1,150,000] Audio by BANG & OLUFSEN|
-|Prestige|87,893,000 79,902,727(7,990,273) with 3.5% individual consumption tax applied 86,559,000|83,445,000 with 3.5% individual consumption tax applied 83,445,000|seat(8-way, lumbar support, and Integrated Memory System(driver's seat and outside mirror connected)), Power-adjustable front passenger’s seat(8-way), Ventilated 1st-row seats, Heated 2nd-row seats • Convenience: Hi-pass(e hi-pass), In-car fingerprint authentication system(personalization, startup, payment, and etc.), Smart power tailgate ▶ Standard equipment of Exclusive Special plus • Smart Safety Technology: Remote Smart Parking Assist 2, Parking Collison- avoidance Assist(front/side/rear) • Exterior: Intelligent Front-Lighting System(IFS), Dynamic welcome/escort lighting(1 type), Sequential turn signals(front and rear), Ambient lighting auto flush door handles, Two-tone door garnish, Glossy black rear diffuser • Interior: Recycled PET suede interior materials(headlining/sunvisor), Fabric upholstered crash pad • Seat: BIO-processed natural leather seats(metal patch applied, embossed design punching), Passenger's seat walk-in device, 1st-row relaxation comfort seats(leg rest included), Ventilated 2nd-row seats • Convenience: Parking Distance Warning-Side, Head-Up Display, Digital key 2, Wireless phone charger(dual), Surround View Monitor, Blind-spot View Monitor, LED reverse light guide • Infotainment: Audio by BANG & OLUFSEN sound system(14 speakers, including external amp), Active road noise control, Active Sound Design|sound system ▶ [250,000] 19-inch alloy wheels & tires ▶ [600,000] Built-in Cam 2 Plus, Augmented reality navigation ▶ [850,000] Indoor/outdoor V2L ▶ [900,000] Vision roof ▶ [1,380,000] Digital side mirror ▶ [750,000] Camera package ▶ [250,000] 19-inch alloy wheels & tires|
+|Exclusive|80,509,000 73,190,000(7,319,000) with 3.5% individual consumption tax applied 79,287,000|76,435,000 with 3.5% individual consumption tax applied 76,435,000|• Powertrain/Performance: Fuel cell system(150kW drive motor, lithium-ion battery, and reducer), Regenerative braking system, Column-Type Shift By Wire(vibration warning), Drive mode select • Safety: 9 airbag system(1st-row advanced/center side airbags, 1st/2nd-row side airbags, and rollover-resistant curtain airbags), Multi-Collision Brake System, Active hood system(for pedestrian protection), Safety unlock function, Artificial engine sound(for pedestrian protection), Child seat fastening system (2 in 2nd-row), Fire extinguisher for vehicles, Pedal Misapplication Safety Assist • Smart Safety Technology: Forward Collision-avoidance Assist(vehicles/ pedestrians/two-wheeled vehicles/junction turning/front oncoming), Smart Cruise Control with Stop & Go, Lane Keeping Assist, Lane Following Assist 2, Blind-spot Collision Warning(driving), Blind-spot Collision-avoidance Assist(forward exit), Rear Cross-traffic Collision-avoidance Assist, Safety Exit Assist, Driver Attention Warning, High Beam Assist, Advanced Rear Occupant Alert, Intelligent Speed Limit Assist, Hands-On Detection, Highway Driving Assist, Navigation-based Smart Cruise Control(safety speed zone/curve control), Vibration warning steering wheel • Exterior: Full LED headlamps(projection type), LED turn signal lamps (front and rear), LED Daytime Running Lights, LED positioning lights, LED rear combination lamps, LED third brake lights, 18-inch alloy wheels & tires, Solar glass(windshield), Double-glazed soundproof glass(windshield, and 1st/ 2nd-row doors), Outside mirror(heating, power-folding, power adjustment, and LED turn signal lamps), Auto flush door handles, Black door garnish • Interior: Panoramic curved display, 12.3-inch color LCD cluster, Leather- upholstered steering wheel(with heating, two-tone color, and Interactive Pixel Lights), LED interior lamp (map lamp, personal lamp, sun visor lamp, and luggage lamp), Metallic door scuff plate • Seat: Synthetic leather seats, 1st-row manual seats, Heated 1st-row seats, 2nd-row 60/40-split folding seats(reclining) • Convenience: Proximity key with push-button start, Smart key remote start, Electronic Parking Brake(with automatic vehicle hold), Paddle shift(regenerative control), Dual-zone full automatic air conditioning(with high-performance antibacterial combination filter, auto defog system, fine dust sensor, air cleaning mode, and after-blow function), 2nd-row seat air vent, Auto light control system, USB Type-C Ports(1×27W switchable charging/data port in 1st-row, and 2×100W charging ports in both 1st and 2nd-row), ECM room mirror(frameless), Rain sensor, Power windows with pinch protection(1st/2nd-row), Power outlet (1 in 1st-row), Parking Distance Warning-Forward/Reverse, Rear View Monitor, Wireless phone charger(single), Walk-away lock, Route planner, Hyundai AI Assistant • Infotainment: 12.3-inch navigation(Bluelink, phone projection, Bluetooth hands-free, and In-car Payment), Audio system(6 speakers), Over-The-Air navigation updates|▶ Hi-pass(e hi-pass) [200,000]|
+|Exclusive Special|83,500,000 75,909,091(7,590,909) with 3.5% individual consumption tax applied 82,232,000|79,275,000 with 3.5% individual consumption tax applied 79,275,000|▶ Standard equipment of Exclusive plus • Smart Safety Technology: Forward Collision-avoidance Assist(intersection crossing/changing lanes in oncoming traffic/approaching from either side/ evasive steering assist), Highway Driving Assist 2, Navigation-based Smart Cruise Control(access road) • Exterior: Roof rack • Interior: Metallic pedal, Driving mode-dependent ambient mood lighting(crash pad, 1st/2nd-row door trim) • Seat: Synthetic leather seats(patch applied), Power-adjustable driver's seat(8-way, lumbar support, and Integrated Memory System(driver's seat and outside mirror connected)), Power-adjustable front passenger’s seat(8-way), Ventilated 1st-row seats, Heated 2nd-row seats • Convenience: Hi-pass(e hi-pass), In-car fingerprint authentication system(personalization, startup, payment, and etc.), Smart power tailgate|▶ [600,000] Built-in Cam 2 Plus, Augmented reality navigation ▶ [850,000] Indoor/outdoor V2L ▶ [950,000] Parking Assist ▶ [1,150,000] Audio by BANG & OLUFSEN sound system ▶ [250,000] 19-inch alloy wheels & tires|
+|Prestige|87,893,000 79,902,727(7,990,273) with 3.5% individual consumption tax applied 86,559,000|83,445,000 with 3.5% individual consumption tax applied 83,445,000|▶ Standard equipment of Exclusive Special plus • Smart Safety Technology: Remote Smart Parking Assist 2, Parking Collison- avoidance Assist(front/side/rear) • Exterior: Intelligent Front-Lighting System(IFS), Dynamic welcome/escort lighting(1 type), Sequential turn signals(front and rear), Ambient lighting auto flush door handles, Two-tone door garnish, Glossy black rear diffuser • Interior: Recycled PET suede interior materials(headlining/sunvisor), Fabric upholstered crash pad • Seat: BIO-processed natural leather seats(metal patch applied, embossed design punching), Passenger's seat walk-in device, 1st-row relaxation comfort seats(leg rest included), Ventilated 2nd-row seats • Convenience: Parking Distance Warning-Side, Head-Up Display, Digital key 2, Wireless phone charger(dual), Surround View Monitor, Blind-spot View Monitor, LED reverse light guide • Infotainment: Audio by BANG & OLUFSEN sound system(14 speakers, including external amp), Active road noise control, Active Sound Design|▶ [600,000] Built-in Cam 2 Plus, Augmented reality navigation ▶ [850,000] Indoor/outdoor V2L ▶ [900,000] Vision roof ▶ [1,380,000] Digital side mirror ▶ [750,000] Camera package ▶ [250,000] 19-inch alloy wheels & tires|
 
 **Classification Details** **Indoor/outdoor V2L** Indoor V2L, Outdoor V2L(connectorless type) **Parking Assist** Surround View Monitor, Blind-spot View Monitor, Parking Distance Warning-Side, Parking Collison-avoidance Assist-Rear **Audio by BANG & OLUFSEN** Audio by BANG & OLUFSEN sound system(14 speakers, including external amp.), Active road noise control, Active Sound Design **sound system** **Camera package** Digital center mirror(with camera sensor cleaning system), Driver monitoring system THE ALL-NEW NEXO /// ECO-FRIENDLY CAR
-

tests/snapshots/p1244-1996.md

@@ -8,7 +8,9 @@ Department of the Treasury **Internal Revenue Service**
 
 # and Report to Employer
 
-**This publication contains:** **Form 4070A, Employee’s Daily Record of** Tips **Form 4070, Employee’s Report of Tips to** Employer
+### This publication contains:
+
+**Form 4070A, Employee’s Daily Record of** Tips **Form 4070, Employee’s Report of Tips to** Employer
 
 For the period
 
@@ -74,7 +76,7 @@ forms simpler, we would be happy to hear from you. You can write to the Tax Form
 
 **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.
 
-**Instructions (continued)**
+### Instructions (continued)
 
 Use this space to total your tips for the year
 

tests/snapshots/real-estate-pricing.md

@@ -6,9 +6,7 @@
 
 8 4 Z E L L / L U R I E R E A L E S T A T E C E N T E R
 
-**Table I: Cap rate correlations**
-
-**Cap Rate Correlation With:*** **BBB Corp** **10-Year Bond Yield S&P Dividend** **Treasury (10-15 yr) Yield** Multifamily 0.187 0.771 0.068 Industrial-0.221 0.748-0.307 CBD Office-0.449 0.694-0.458 Retail-0.181 0.649-02.58
+**Table I: Cap rate correlations** **Cap Rate Correlation With:*** **BBB Corp** **10-Year Bond Yield S&P Dividend** **Treasury (10-15 yr) Yield** Multifamily 0.187 0.771 0.068 Industrial-0.221 0.748-0.307 CBD Office-0.449 0.694-0.458 Retail-0.181 0.649-02.58
 
 * Based on 25 years of data for the 10-yrT & S&P DivYld; and 14 years for BBB.
 **Figure 1:** NCREIF cap rates vs. 10-yearTreasury
@@ -34,9 +32,7 @@ R E V I E W 8 5
 
 1982 1986 1990 1994 1998 2002 2006
 
-**Table II: Correlationsofspreadsbypropertytype**
-
-**Correlation of Cap Rate Spreads Over Treasury** **Multifamily Industrial CBD Office**
+**Table II: Correlationsofspreadsbypropertytype** **Correlation of Cap Rate Spreads Over Treasury** **Multifamily Industrial CBD Office**
 
 ||Multifamily|Industrial|CBD Office|
 |---|---|---|---|

tests/snapshots/td9264.md

@@ -13,8 +13,10 @@ IDENTIFICATION NUMBER OF CORPORATION] ELECTS TO BE TREATED AS A COMPONENT MEMBER
 (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 been included. (B) Election not filed.|In the event no election is filed in accordance with the|
+|termination of membership in the controlled group in which such corporation has||
 |---|---|
+|been included.||
+|(B) Election not filed.|In the event no election is filed in accordance with the|
 |provisions of paragraph (d)(2)(iv) of this section, then the Internal Revenue Service||
 |will determine the group in which such corporation is to be included. Such||
 |determination will be binding for all subsequent years unless the corporation files a||
@@ -52,9 +54,14 @@ company other than a life insurance company shall make a return on Form 1120PC.
 
 annual statement (or a pro forma annual statement), including the underwriting and investment exhibit for the year covered by such return.
 
-||(3) Foreign insurance companies. The provisions of paragraphs (c)(1) and|
-|---|---|
-||(c)(2) of this section concerning the returns and statements of insurance companies subject to tax under section 801 or section 831 also apply to foreign insurance companies subject to tax under those sections, except that the copy of the annual statement required to be submitted with the return shall, in the case of a foreign insurance company that is not required to file an annual statement, be a copy of the pro forma annual statement relating to the United States business of such company. (4) Exception for insurance companies filing their Federal income tax returns electronically. If an insurance company described in paragraph (c)(1), (c)(2), or (c)(3) of this section files its Federal income tax return electronically, it should not include on or with such return its annual statement (or pro forma annual statement), or any portion thereof. Such statement must be available at all times for inspection by authorized Internal Revenue Service officers or employees and retained for so long as such statements may be material in the administration of any internal revenue law. See §1.6001-1(e). (5) Definition. For purposes of this section, the term annual statement means the annual statement, the form of which is approved by the National Association of Insurance Commissioners (NAIC), which is filed by an insurance company for the year with the insurance departments of States, Territories, and the District of|
+(3) Foreign insurance companies. The provisions of paragraphs (c)(1) and
+(c)(2) of this section concerning the returns and statements of insurance companies subject to tax under section 801 or section 831 also apply to foreign insurance companies subject to tax under those sections, except that the copy of the annual statement required to be submitted with the return shall, in the case of a foreign insurance company that is not required to file an annual statement, be a copy of the pro forma annual statement relating to the United States business of such company.
+(4) Exception for insurance companies filing their Federal income tax returns
+electronically. If an insurance company described in paragraph (c)(1), (c)(2), or
+
+(c)(3) of this section files its Federal income tax return electronically, it should not include on or with such return its annual statement (or pro forma annual statement), or any portion thereof. Such statement must be available at all times for inspection by authorized Internal Revenue Service officers or employees and retained for so long as such statements may be material in the administration of any internal revenue law. See §1.6001-1(e).
+(5) Definition. For purposes of this section, the term annual statement means
+the annual statement, the form of which is approved by the National Association of Insurance Commissioners (NAIC), which is filed by an insurance company for the year with the insurance departments of States, Territories, and the District of
 
 Columbia. The term annual statement also includes a pro forma annual statement if the insurance company is not required to file the NAIC annual statement.
 
@@ -101,45 +108,32 @@ section and paragraph
 (c)(4), and (c)(5) of this section, and paragraph
 (c)(2) of §1.382-8T
 
+||§1.382-8(g), Example|
+|---|---|
+||The first sentence of §1.382-8(g), Example §1.382-8(g), Example §1.382-8(g), Example|
+
+(2)(c)
+(2)(e)
+(3)(b)
+(3)(c)(1)(B)
+
+||The second sentence of|
+|---|---|
+||§1.382-8(g), Example The second sentence of §1.382-8(g), Example The first sentence of §1.1502-32(b)(4)(v)(A) The first sentence of §1.1502-32(b)(4)(v)(B)|
+
+(4)(c)
+(5)(c)
+
 |The fifth sentence of|paragraph (c) of this|paragraphs (c)(1), (c)(3),|
 |---|---|---|
 |§1.382-8(f)|section|(c)(4), and (c)(5) of this section, and paragraph (c)(2) of §1.382-8T|
 |§1.382-8(g), Example|paragraph (c) of this|paragraphs (c)(1), (c)(3), section, and paragraph (c)(2) of §1.382-8T|
 |The second sentence of|paragraph (c) of this|paragraphs (c)(1), (c)(3),|
-|§1.382-8(g), Example|section|(c)(4), and (c)(5) of this|
+|§1.382-8(g), Example|section paragraph (c)(2) of this section paragraph (c)(2) of this section paragraph (c)(2) of this section paragraphs (c)(1) and (2) of this section paragraph (c)(2) of this section paragraph (c)(2) of this section paragraph (b)(4)(iv) of this section paragraph (b)(4)(iv) of this section|(c)(4), and (c)(5) of this (c)(2) of §1.382-8T paragraph (c)(2) of §1.382-8T paragraph (c)(2) of §1.382-8T paragraph (c)(2) of §1.382-8T paragraph (c)(1) of this section and paragraph (c)(2) of §1.382-8T paragraph (c)(2) of §1.382-8T paragraph (c)(2) of §1.382-8T paragraph (b)(4)(iv) of §1.1502-32T paragraph (b)(4)(iv) of §1.1502-32T|
 
 (1)(b)(2) section (c)(4), and (c)(5) of this
 (1)(c) section, and paragraph
 
-(c)(2) of §1.382-8T
-
-|§1.382-8(g), Example|paragraph (c)(2) of this|paragraph (c)(2) of|
-|---|---|---|
-|The first sentence of|paragraph (c)(2) of this|paragraph (c)(2) of|
-|§1.382-8(g), Example|section|§1.382-8T|
-|§1.382-8(g), Example|paragraph (c)(2) of this|paragraph (c)(2) of|
-|§1.382-8(g), Example|paragraphs (c)(1) and (2)|paragraph (c)(1) of this|
-
-(2)(c) section §1.382-8T
-(2)(e)
-(3)(b) section §1.382-8T
-(3)(c)(1)(B) of this section section and paragraph
-
-(c)(2) of §1.382-8T
-
-|The second sentence of|paragraph (c)(2) of this|paragraph (c)(2) of|
-|---|---|---|
-|§1.382-8(g), Example|section|§1.382-8T|
-|The second sentence of|paragraph (c)(2) of this|paragraph (c)(2) of|
-|§1.382-8(g), Example|section|§1.382-8T|
-|The first sentence of|paragraph (b)(4)(iv) of|paragraph (b)(4)(iv) of|
-|§1.1502-32(b)(4)(v)(A)|this section|§1.1502-32T|
-|The first sentence of|paragraph (b)(4)(iv) of|paragraph (b)(4)(iv) of|
-|§1.1502-32(b)(4)(v)(B)|this section|§1.1502-32T|
-
-(4)(c)
-(5)(c)
-
 |§1.1502-35(c)(4)(ii)(B)|§1.1502-76(b)(2)(ii)(D)|§1.1502-76T(b)(2)(ii)(D)|
 |---|---|---|
 |§1.1502-76(b)(2)(ii)(A)(2)|paragraph (b)(2)(ii)(D) of this section|paragraph (b)(2)(ii)(D) of §1.1502-76T|
@@ -168,14 +162,33 @@ section and paragraph
 |§1.6043-2(a)|or 1.1081-11|3T(a), or §1.1081-11T|
 |The first sentence of §301.6011-5T(a) (twice)|§1.6012-2|paragraphs (a), (b) and (d) through (j) of §1.6012- 2, and paragraph (c) of §1.6012-2T|
 
-|||PART 602--OMB CONTROL NUMBERS UNDER THE PAPERWORK||
-|---|---|---|---|
-||REDUCTION ACT Authority: 26 U.S.C. 7805. 1. The following entries to the table are removed: §602.101 OMB Control numbers.|Par. 54. The authority citation for part 602 continues to read as follows: Par. 55. In §602.101, paragraph (b) is amended to read as follows:||
-|* * * * *|(b) * * * CFR part or section where identified or described||Current OMB control No.|
-|* * * * *|1.332-6………………………………………………………………….|1.382-11……………………………………………………………….. 1545-2019 1.351-3…………………………………………………………………. 1545-2019 1.355-5…………………………………………………………………. 1545-2019 1.368-3…………………………………………………………………. 1545-2019 1.1081-11………………………………………………………………. 1545-2019|1545-2019|
-|* * * * *|§602.101 OMB Control numbers.|______________________________________________________________ 2. The following entries are added in numerical order to the table:||
-|* * * * *|(b) * * * CFR part or section where identified or described||Current OMB control No.|
-|* * * * *|1.302-2T………………………………………………………………… 1545 1.302-4T………………………………………………………………… 1545||-2019 -2019|
+PART 602--OMB CONTROL NUMBERS UNDER THE PAPERWORK REDUCTION ACT Par. 54. The authority citation for part 602 continues to read as follows: Authority: 26 U.S.C. 7805. Par. 55. In §602.101, paragraph (b) is amended to read as follows:
+
+1. The following entries to the table are removed:
+§602.101 OMB Control numbers.
+
+* * * * *
+(b) * * *
+CFR part or section where Current OMB identified or described control No.
+
+* * * * *
+1.332-6…………………………………………………………………. 1545-2019
+1.382-11……………………………………………………………….. 1545-2019
+1.351-3…………………………………………………………………. 1545-2019
+1.355-5…………………………………………………………………. 1545-2019
+1.368-3…………………………………………………………………. 1545-2019
+1.1081-11………………………………………………………………. 1545-2019
+* * * * * **______________________________________________________________**
+2. The following entries are added in numerical order to the table:
+§602.101 OMB Control numbers.
+
+* * * * *
+(b) * * *
+CFR part or section where Current OMB identified or described control No.
+
+* * * * *
+1.302-2T………………………………………………………………… 1545-2019
+1.302-4T………………………………………………………………… 1545-2019
 
 |1.331-1T………………………………………………………………… 1545|-2019|
 |---|---|

tests/snapshots/thermo-freon12.md

@@ -1,8 +1,8 @@
-**Technical Information**
+##### Technical Information
 
 ## l T-12 SI
 
-DuPont Fluorochemicals
+##### DuPont Fluorochemicals
 
 #### Thermodynamic Properties
 
@@ -20,25 +20,22 @@ Tables of the thermodynamic **Units** properties of R-12 have been developed and
 
 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
-H = Enthalpy (kJ/kg)
+##### H = Enthalpy (kJ/kg)
 
-S = Entropy (kJ/kg.K)
+##### S = Entropy (kJ/kg.K)
 
-**Physical Properties**
+##### Physical Properties
 
-Chemical Formula CCl2F2
+|Chemical Formula|CCl2F2|
+|---|---|
+|Molecular mass|120.91|
+|Boiling Point At one atmosphere|-29.75°C|
+|Critical Temperature|111.97°C|
+|Critical Pressure|4136 kPa|
+|Critical Density|565.0 kg/m|
+|Critical Volume|0.0018 m|
 
-Molecular mass 120.91
-
-Boiling Point-29.75°C At one atmosphere
-
-Critical Temperature 111.97°C
-
-Critical Pressure 4136 kPa
-
-3 Critical Density 565.0 kg/m
-
-Critical Volume 0.0018 m /kg
+/kg
 
 l
 

tests/test_python.py

@@ -0,0 +1,403 @@
+"""Tests for the pdf_inspector Python bindings."""
+
+import os
+import pytest
+import pdf_inspector
+
+FIXTURES_DIR = os.path.join(os.path.dirname(__file__), "fixtures")
+
+
+def fixture_path(name: str) -> str:
+    return os.path.join(FIXTURES_DIR, name)
+
+
+def fixture_bytes(name: str) -> bytes:
+    with open(fixture_path(name), "rb") as f:
+        return f.read()
+
+
+# ---------------------------------------------------------------------------
+# process_pdf
+# ---------------------------------------------------------------------------
+
+
+class TestProcessPdf:
+    def test_basic(self):
+        result = pdf_inspector.process_pdf(fixture_path("thermo-freon12.pdf"))
+        assert result.pdf_type == "text_based"
+        assert result.page_count == 3
+        assert result.confidence > 0.0
+        assert result.markdown is not None
+        assert len(result.markdown) > 0
+
+    def test_result_repr(self):
+        result = pdf_inspector.process_pdf(fixture_path("thermo-freon12.pdf"))
+        r = repr(result)
+        assert "PdfResult" in r
+        assert "text_based" in r
+
+    def test_with_pages(self):
+        result = pdf_inspector.process_pdf(
+            fixture_path("thermo-freon12.pdf"), pages=[1]
+        )
+        assert result.page_count == 3  # total pages in doc
+        assert result.markdown is not None
+
+    def test_result_fields(self):
+        result = pdf_inspector.process_pdf(fixture_path("thermo-freon12.pdf"))
+        # All fields should be accessible
+        assert isinstance(result.pdf_type, str)
+        assert isinstance(result.page_count, int)
+        assert isinstance(result.processing_time_ms, int)
+        assert isinstance(result.pages_needing_ocr, list)
+        assert isinstance(result.confidence, float)
+        assert isinstance(result.is_complex_layout, bool)
+        assert isinstance(result.pages_with_tables, list)
+        assert isinstance(result.pages_with_columns, list)
+        assert isinstance(result.has_encoding_issues, bool)
+        # title can be None or str
+        assert result.title is None or isinstance(result.title, str)
+
+
+# ---------------------------------------------------------------------------
+# process_pdf_bytes
+# ---------------------------------------------------------------------------
+
+
+class TestProcessPdfBytes:
+    def test_basic(self):
+        data = fixture_bytes("thermo-freon12.pdf")
+        result = pdf_inspector.process_pdf_bytes(data)
+        assert result.pdf_type == "text_based"
+        assert result.markdown is not None
+
+    def test_with_pages(self):
+        data = fixture_bytes("thermo-freon12.pdf")
+        result = pdf_inspector.process_pdf_bytes(data, pages=[1, 2])
+        assert result.markdown is not None
+
+
+# ---------------------------------------------------------------------------
+# detect_pdf / detect_pdf_bytes
+# ---------------------------------------------------------------------------
+
+
+class TestDetectPdf:
+    def test_detect_file(self):
+        result = pdf_inspector.detect_pdf(fixture_path("thermo-freon12.pdf"))
+        assert result.pdf_type == "text_based"
+        assert result.markdown is None  # detect only — no markdown
+        assert result.page_count == 3
+
+    def test_detect_bytes(self):
+        data = fixture_bytes("thermo-freon12.pdf")
+        result = pdf_inspector.detect_pdf_bytes(data)
+        assert result.pdf_type == "text_based"
+        assert result.markdown is None
+
+
+# ---------------------------------------------------------------------------
+# classify_pdf / classify_pdf_bytes
+# ---------------------------------------------------------------------------
+
+
+class TestClassifyPdf:
+    def test_classify_file(self):
+        result = pdf_inspector.classify_pdf(fixture_path("thermo-freon12.pdf"))
+        assert result.pdf_type == "text_based"
+        assert result.page_count == 3
+        assert result.confidence > 0.0
+        assert isinstance(result.pages_needing_ocr, list)
+
+    def test_classify_bytes(self):
+        data = fixture_bytes("thermo-freon12.pdf")
+        result = pdf_inspector.classify_pdf_bytes(data)
+        assert result.pdf_type == "text_based"
+        assert result.page_count == 3
+        assert result.confidence > 0.0
+
+    def test_classify_repr(self):
+        result = pdf_inspector.classify_pdf(fixture_path("thermo-freon12.pdf"))
+        r = repr(result)
+        assert "PdfClassification" in r
+        assert "text_based" in r
+
+    def test_classify_fields(self):
+        result = pdf_inspector.classify_pdf(fixture_path("thermo-freon12.pdf"))
+        assert isinstance(result.pdf_type, str)
+        assert isinstance(result.page_count, int)
+        assert isinstance(result.pages_needing_ocr, list)
+        assert isinstance(result.confidence, float)
+
+
+# ---------------------------------------------------------------------------
+# extract_text / extract_text_bytes
+# ---------------------------------------------------------------------------
+
+
+class TestExtractText:
+    def test_basic(self):
+        text = pdf_inspector.extract_text(fixture_path("thermo-freon12.pdf"))
+        assert isinstance(text, str)
+        assert len(text) > 0
+
+    def test_bytes(self):
+        data = fixture_bytes("thermo-freon12.pdf")
+        text = pdf_inspector.extract_text_bytes(data)
+        assert isinstance(text, str)
+        assert len(text) > 0
+
+    def test_bytes_matches_file(self):
+        text_file = pdf_inspector.extract_text(fixture_path("thermo-freon12.pdf"))
+        text_bytes = pdf_inspector.extract_text_bytes(fixture_bytes("thermo-freon12.pdf"))
+        assert text_file == text_bytes
+
+
+# ---------------------------------------------------------------------------
+# extract_text_with_positions / extract_text_with_positions_bytes
+# ---------------------------------------------------------------------------
+
+
+class TestExtractTextWithPositions:
+    def test_basic(self):
+        items = pdf_inspector.extract_text_with_positions(
+            fixture_path("thermo-freon12.pdf")
+        )
+        assert len(items) > 0
+        item = items[0]
+        assert isinstance(item.text, str)
+        assert isinstance(item.x, float)
+        assert isinstance(item.y, float)
+        assert isinstance(item.width, float)
+        assert isinstance(item.height, float)
+        assert isinstance(item.font, str)
+        assert isinstance(item.font_size, float)
+        assert isinstance(item.page, int)
+        assert isinstance(item.is_bold, bool)
+        assert isinstance(item.is_italic, bool)
+        assert isinstance(item.item_type, str)
+
+    def test_with_pages(self):
+        items = pdf_inspector.extract_text_with_positions(
+            fixture_path("thermo-freon12.pdf"), pages=[1]
+        )
+        assert len(items) > 0
+        assert all(item.page == 1 for item in items)
+
+    def test_repr(self):
+        items = pdf_inspector.extract_text_with_positions(
+            fixture_path("thermo-freon12.pdf")
+        )
+        r = repr(items[0])
+        assert "TextItem" in r
+
+    def test_bytes(self):
+        data = fixture_bytes("thermo-freon12.pdf")
+        items = pdf_inspector.extract_text_with_positions_bytes(data)
+        assert len(items) > 0
+        assert isinstance(items[0].text, str)
+
+    def test_bytes_with_pages(self):
+        data = fixture_bytes("thermo-freon12.pdf")
+        items = pdf_inspector.extract_text_with_positions_bytes(data, pages=[1])
+        assert len(items) > 0
+        assert all(item.page == 1 for item in items)
+
+
+# ---------------------------------------------------------------------------
+# extract_text_in_regions / extract_text_in_regions_bytes
+# ---------------------------------------------------------------------------
+
+
+class TestExtractTextInRegions:
+    def test_file(self):
+        results = pdf_inspector.extract_text_in_regions(
+            fixture_path("thermo-freon12.pdf"),
+            [(0, [[0.0, 0.0, 600.0, 100.0]])],
+        )
+        assert len(results) == 1
+        assert results[0].page == 0
+        assert len(results[0].regions) == 1
+        assert isinstance(results[0].regions[0].text, str)
+        assert isinstance(results[0].regions[0].needs_ocr, bool)
+
+    def test_bytes(self):
+        data = fixture_bytes("thermo-freon12.pdf")
+        results = pdf_inspector.extract_text_in_regions_bytes(
+            data,
+            [(0, [[0.0, 0.0, 600.0, 100.0]])],
+        )
+        assert len(results) == 1
+        assert results[0].page == 0
+        assert len(results[0].regions) == 1
+        assert isinstance(results[0].regions[0].text, str)
+
+    def test_repr(self):
+        results = pdf_inspector.extract_text_in_regions(
+            fixture_path("thermo-freon12.pdf"),
+            [(0, [[0.0, 0.0, 600.0, 100.0]])],
+        )
+        r = repr(results[0])
+        assert "PageRegionTexts" in r
+        r2 = repr(results[0].regions[0])
+        assert "RegionText" in r2
+
+    def test_multiple_regions(self):
+        results = pdf_inspector.extract_text_in_regions(
+            fixture_path("thermo-freon12.pdf"),
+            [(0, [[0.0, 0.0, 300.0, 100.0], [300.0, 0.0, 600.0, 100.0]])],
+        )
+        assert len(results) == 1
+        assert len(results[0].regions) == 2
+
+    def test_multiple_pages(self):
+        results = pdf_inspector.extract_text_in_regions(
+            fixture_path("thermo-freon12.pdf"),
+            [
+                (0, [[0.0, 0.0, 600.0, 100.0]]),
+                (1, [[0.0, 0.0, 600.0, 100.0]]),
+            ],
+        )
+        assert len(results) == 2
+        assert results[0].page == 0
+        assert results[1].page == 1
+
+    def test_malformed_region_raises_value_error(self):
+        with pytest.raises(ValueError, match="Invalid region"):
+            pdf_inspector.extract_text_in_regions(
+                fixture_path("thermo-freon12.pdf"),
+                [(0, [[0.0, 0.0, 600.0]])],
+            )
+
+
+# ---------------------------------------------------------------------------
+# extract_pages_markdown / extract_pages_markdown_bytes
+# ---------------------------------------------------------------------------
+
+
+class TestExtractPagesMarkdown:
+    def test_default_returns_all_pages(self):
+        result = pdf_inspector.extract_pages_markdown(
+            fixture_path("thermo-freon12.pdf")
+        )
+        assert len(result.pages) == 3
+        assert [p.page for p in result.pages] == [0, 1, 2]
+        assert all(isinstance(p.markdown, str) for p in result.pages)
+
+    def test_bytes_default_returns_all_pages(self):
+        data = fixture_bytes("thermo-freon12.pdf")
+        result = pdf_inspector.extract_pages_markdown_bytes(data)
+        assert len(result.pages) == 3
+
+    def test_selected_pages_preserve_order(self):
+        result = pdf_inspector.extract_pages_markdown(
+            fixture_path("thermo-freon12.pdf"), pages=[2, 0]
+        )
+        assert [p.page for p in result.pages] == [2, 0]
+
+    def test_bytes_selected_pages_preserve_order(self):
+        data = fixture_bytes("thermo-freon12.pdf")
+        result = pdf_inspector.extract_pages_markdown_bytes(data, pages=[1])
+        assert len(result.pages) == 1
+        assert result.pages[0].page == 1
+
+    def test_page_fields(self):
+        result = pdf_inspector.extract_pages_markdown(
+            fixture_path("thermo-freon12.pdf"), pages=[0]
+        )
+        page = result.pages[0]
+        assert isinstance(page.page, int)
+        assert isinstance(page.markdown, str)
+        assert isinstance(page.needs_ocr, bool)
+        assert not page.needs_ocr  # text-based fixture
+        assert len(page.markdown) > 0
+
+    def test_result_fields(self):
+        result = pdf_inspector.extract_pages_markdown(
+            fixture_path("thermo-freon12.pdf")
+        )
+        assert isinstance(result.pages, list)
+        assert isinstance(result.pages_with_tables, list)
+        assert isinstance(result.pages_with_columns, list)
+        assert isinstance(result.pages_needing_ocr, list)
+        assert isinstance(result.is_complex, bool)
+
+    def test_out_of_range_page_marks_needs_ocr(self):
+        result = pdf_inspector.extract_pages_markdown(
+            fixture_path("thermo-freon12.pdf"), pages=[9999]
+        )
+        assert len(result.pages) == 1
+        assert result.pages[0].needs_ocr
+        assert result.pages[0].markdown == ""
+
+    def test_repr(self):
+        result = pdf_inspector.extract_pages_markdown(
+            fixture_path("thermo-freon12.pdf"), pages=[0]
+        )
+        assert "PagesExtractionResult" in repr(result)
+        assert "PageMarkdown" in repr(result.pages[0])
+
+    def test_not_a_pdf(self):
+        with pytest.raises(ValueError):
+            pdf_inspector.extract_pages_markdown_bytes(b"not a pdf")
+
+
+# ---------------------------------------------------------------------------
+# Error handling
+# ---------------------------------------------------------------------------
+
+
+class TestErrors:
+    def test_nonexistent_file(self):
+        with pytest.raises(ValueError):
+            pdf_inspector.process_pdf("/nonexistent/file.pdf")
+
+    def test_not_a_pdf(self):
+        with pytest.raises(ValueError):
+            pdf_inspector.process_pdf_bytes(b"this is not a pdf")
+
+    def test_empty_bytes(self):
+        with pytest.raises(ValueError):
+            pdf_inspector.process_pdf_bytes(b"")
+
+    def test_classify_not_a_pdf(self):
+        with pytest.raises(ValueError):
+            pdf_inspector.classify_pdf_bytes(b"not a pdf")
+
+    def test_classify_nonexistent(self):
+        with pytest.raises((ValueError, OSError)):
+            pdf_inspector.classify_pdf("/nonexistent/file.pdf")
+
+    def test_extract_text_bytes_not_a_pdf(self):
+        with pytest.raises(ValueError):
+            pdf_inspector.extract_text_bytes(b"not a pdf")
+
+    def test_regions_not_a_pdf(self):
+        with pytest.raises(ValueError):
+            pdf_inspector.extract_text_in_regions_bytes(
+                b"not a pdf", [(0, [[0.0, 0.0, 100.0, 100.0]])]
+            )
+
+
+# ---------------------------------------------------------------------------
+# Multiple fixtures
+# ---------------------------------------------------------------------------
+
+
+class TestMultipleFixtures:
+    """Run basic processing on all available test fixtures."""
+
+    @pytest.mark.parametrize(
+        "filename",
+        [f for f in os.listdir(FIXTURES_DIR) if f.endswith(".pdf")],
+    )
+    def test_process_all_fixtures(self, filename):
+        result = pdf_inspector.process_pdf(fixture_path(filename))
+        assert result.pdf_type in (
+            "text_based",
+            "scanned",
+            "image_based",
+            "mixed",
+        )
+        assert result.page_count > 0
+        assert result.confidence >= 0.0