fix: split interleaved tiny-text blocks at the same Y

Slide-deck PDFs often stack several PDF text blocks at the same Y in a
tiny (<3pt) copy-paste layer — typically a decorative row of glyphs
(e.g. a palette legend) plus a readable paragraph such as a disclaimer.
Before this change, our line grouper fused them by Y and then sorted
the combined items by X, which character-interleaved the two streams
into gibberish like "r1e2s1pe8ct3i5ve0o6w1n1e1r4s" for the string
"respective owners".

The fix runs in merge_text_items, while items are still in content-
stream order. For tiny fonts (<3pt on both sides) we also split a
Y-bucket when the incoming item's X is a sharp backtrack from the
bucket's current right edge — the signature of a new `BT ... Tm`
block resetting the text position. Normal-sized text (≥3pt) keeps the
existing behaviour so tables and other valid layouts don't get
over-split.

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
 
@@ -68,7 +97,7 @@ jobs:
 
   publish:
     name: Publish to npm
-    needs: build
+    needs: [check-version, build]
     runs-on: ubuntu-latest
     permissions:
       contents: read

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)

README.md

@@ -55,12 +55,12 @@ print(result.markdown)   # Markdown string or None
 ### Node.js
 
 ```bash
-npm install @firecrawl/pdf-inspector-js
+npm install @firecrawl/pdf-inspector
 ```
 
 ```javascript
 import { readFileSync } from 'fs';
-import { processPdf, classifyPdf } from '@firecrawl/pdf-inspector-js';
+import { processPdf, classifyPdf } from '@firecrawl/pdf-inspector';
 
 const result = processPdf(readFileSync('document.pdf'));
 console.log(result.pdfType);   // "TextBased", "Scanned", "ImageBased", "Mixed"

napi/README.md

@@ -1,4 +1,4 @@
-# firecrawl-pdf-inspector
+# PDF Inspector
 
 Fast PDF classification and region-based text extraction for Node.js/Bun. Native Rust performance via [napi-rs](https://napi.rs).
 
@@ -7,9 +7,9 @@ Built by [Firecrawl](https://firecrawl.dev) for hybrid OCR pipelines — extract
 ## Install
 
 ```bash
-npm install firecrawl-pdf-inspector
+npm install @firecrawl/pdf-inspector
 # or
-bun add firecrawl-pdf-inspector
+bun add @firecrawl/pdf-inspector
 ```
 
 Prebuilt binaries included for **linux-x64** and **macOS ARM64**. No Rust toolchain needed.
@@ -21,7 +21,7 @@ Prebuilt binaries included for **linux-x64** and **macOS ARM64**. No Rust toolch
 Classify a PDF as TextBased, Scanned, Mixed, or ImageBased (~10-50ms). Returns which pages need OCR.
 
 ```typescript
-import { classifyPdf } from 'firecrawl-pdf-inspector'
+import { classifyPdf } from '@firecrawl/pdf-inspector'
 import { readFileSync } from 'fs'
 
 const pdf = readFileSync('document.pdf')
@@ -40,7 +40,7 @@ Extract text within bounding-box regions from a PDF. Designed for hybrid OCR pip
 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'
+import { extractTextInRegions } from '@firecrawl/pdf-inspector'
 
 const result = extractTextInRegions(pdf, [
   {

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/package.json

@@ -1,9 +1,12 @@
 {
-  "name": "firecrawl-pdf-inspector",
-  "version": "0.4.3",
+  "name": "@firecrawl/pdf-inspector",
+  "version": "1.2.0",
   "description": "Fast PDF classification and text extraction. Detect text-based vs scanned PDFs, extract text by region with quality checks. Native Rust performance via napi-rs.",
   "main": "index.js",
   "types": "index.d.ts",
+  "bin": {
+    "pdf-inspector": "bin/pdf-inspector.mjs"
+  },
   "license": "MIT",
   "keywords": [
     "pdf",
@@ -20,6 +23,7 @@
     "index.js",
     "index.d.ts",
     "*.node",
+    "bin/",
     "README.md"
   ],
   "repository": {
@@ -34,7 +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/src/lib.rs

@@ -5,6 +5,28 @@ 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
 // ---------------------------------------------------------------------------
@@ -12,7 +34,7 @@ use std::panic;
 /// Full PDF processing result with markdown and metadata.
 #[napi(object)]
 pub struct PdfResult {
-    pub pdf_type: String,
+    pub pdf_type: PdfType,
     pub markdown: Option<String>,
     pub page_count: u32,
     pub processing_time_ms: u32,
@@ -29,7 +51,7 @@ pub struct PdfResult {
 /// Lightweight PDF classification result.
 #[napi(object)]
 pub struct PdfClassification {
-    pub pdf_type: String,
+    pub pdf_type: PdfType,
     pub page_count: u32,
     /// 0-indexed page numbers that need OCR.
     pub pages_needing_ocr: Vec<u32>,
@@ -49,7 +71,9 @@ pub struct TextItem {
     pub page: u32,
     pub is_bold: bool,
     pub is_italic: bool,
-    pub item_type: String,
+    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).
@@ -79,18 +103,18 @@ pub struct PageRegionTexts {
 // Helpers
 // ---------------------------------------------------------------------------
 
-fn pdf_type_string(t: pdf_inspector::PdfType) -> String {
+fn convert_pdf_type(t: pdf_inspector::PdfType) -> PdfType {
     match t {
-        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(),
+        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: pdf_type_string(r.pdf_type),
+        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,
@@ -104,12 +128,12 @@ fn to_napi_result(r: pdf_inspector::PdfProcessResult) -> PdfResult {
     }
 }
 
-fn item_type_string(t: &pdf_inspector::types::ItemType) -> String {
+fn convert_item_type(t: &pdf_inspector::types::ItemType) -> (ItemType, Option<String>) {
     match t {
-        pdf_inspector::types::ItemType::Text => "text".into(),
-        pdf_inspector::types::ItemType::Image => "image".into(),
-        pdf_inspector::types::ItemType::Link(url) => format!("link:{url}"),
-        pdf_inspector::types::ItemType::FormField => "form_field".into(),
+        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),
     }
 }
 
@@ -181,7 +205,7 @@ pub fn classify_pdf(buffer: Buffer) -> Result<PdfClassification> {
         let result =
             pdf_inspector::classify_pdf_mem(&bytes).map_err(|e| to_napi_err(e, "classify_pdf"))?;
         Ok(PdfClassification {
-            pdf_type: pdf_type_string(result.pdf_type),
+            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,
@@ -222,18 +246,22 @@ pub fn extract_text_with_positions(
 
         Ok(items
             .into_iter()
-            .map(|item| 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: item_type_string(&item.item_type),
+            .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())
     })
@@ -289,6 +317,65 @@ pub fn extract_tables_in_regions(
     })
 }
 
+/// 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 specific 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.
+#[napi]
+pub fn extract_pages_markdown(
+    buffer: Buffer,
+    pages: 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)
+            .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()

src/extractor/content_stream.rs

@@ -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];
@@ -425,6 +433,10 @@ 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 {
@@ -700,6 +712,7 @@ 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 +720,13 @@ 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;

src/extractor/mod.rs

@@ -315,21 +315,79 @@ pub(crate) fn merge_text_items(items: Vec<TextItem>) -> Vec<TextItem> {
         return items;
     }
 
-    // Group items by (page, Y position) with 5pt tolerance
+    // Group items by (page, Y position) with 5pt tolerance, walking in
+    // stream order.  A new group also starts when the incoming item's X
+    // would be a significant backtrack — signature of a new PDF text
+    // block (`BT ... Tm`) starting at the left of the same Y band.
+    // Without this split, slide-deck PDFs that stack several copy-paste
+    // layers at the same tiny-text Y (e.g. a color-palette row and a
+    // disclaimer body) later sort-by-X into character-interleaved
+    // gibberish.
     let y_tolerance = 5.0;
-    let mut line_groups: Vec<(u32, f32, Vec<&TextItem>)> = Vec::new();
+    struct LineGroup<'a> {
+        page: u32,
+        y: f32,
+        end_x: f32,
+        font_size: f32,
+        items: Vec<&'a TextItem>,
+    }
+    let mut line_groups: Vec<LineGroup> = Vec::new();
 
     for item in &items {
-        let found = line_groups
-            .iter_mut()
-            .find(|(pg, y, _)| *pg == item.page && (item.y - *y).abs() < y_tolerance);
-        if let Some((_, _, group)) = found {
-            group.push(item);
-        } else {
-            line_groups.push((item.page, item.y, vec![item]));
+        // Find the most recently touched matching (page, y) bucket that
+        // the item can legitimately continue (no sharp X backtrack).
+        // Scanning from the end means the last-updated bucket wins, so a
+        // new back-tracked block reliably opens a fresh group instead of
+        // re-joining the old one.
+        let item_right_limit = item.x + effective_merge_width(item);
+        let mut target: Option<usize> = None;
+        for (i, g) in line_groups.iter().enumerate().rev() {
+            if g.page != item.page {
+                continue;
+            }
+            if (g.y - item.y).abs() >= y_tolerance {
+                continue;
+            }
+            // Detect a sharp X backtrack that can only be a new PDF text
+            // block (`BT ... Tm`) starting at the same Y band.  Only
+            // trigger for very small fonts: in slide-deck PDFs, multiple
+            // text blocks get stacked at the same Y in a tiny copy-paste
+            // layer, and sort-by-X later interleaves them.  Normal-sized
+            // fonts (≥ 3pt) stay on the original behaviour — splitting
+            // there tends to break tables and other valid layouts.
+            if item.font_size < 3.0 && g.font_size < 3.0 {
+                let backtrack = g.end_x - item.x;
+                let min_backtrack = (item.font_size * 5.0).max(5.0);
+                if backtrack > min_backtrack {
+                    continue;
+                }
+            }
+            target = Some(i);
+            break;
+        }
+        match target {
+            Some(i) => {
+                line_groups[i].end_x = line_groups[i].end_x.max(item_right_limit);
+                line_groups[i].items.push(item);
+            }
+            None => {
+                line_groups.push(LineGroup {
+                    page: item.page,
+                    y: item.y,
+                    end_x: item_right_limit,
+                    font_size: item.font_size,
+                    items: vec![item],
+                });
+            }
         }
     }
 
+    // Re-shape into the tuple layout the rest of this function expects.
+    let mut line_groups: Vec<(u32, f32, Vec<&TextItem>)> = line_groups
+        .into_iter()
+        .map(|g| (g.page, g.y, g.items))
+        .collect();
+
     // Sort each group by X position (direction-aware)
     for (_, _, group) in &mut line_groups {
         let rtl = is_rtl_text(group.iter().map(|i| &i.text));
@@ -573,6 +631,112 @@ mod tests {
         assert_eq!(merged[0].text, "hello world");
     }
 
+    fn make_tiny_item(text: &str, x: f32, width: f32) -> TextItem {
+        TextItem {
+            text: text.into(),
+            x,
+            y: 700.0,
+            width,
+            height: 1.3,
+            font: "F1".into(),
+            font_size: 1.3, // tiny copy-paste / accessibility layer font
+            page: 1,
+            is_bold: false,
+            is_italic: false,
+            item_type: ItemType::Text,
+            mcid: None,
+        }
+    }
+
+    #[test]
+    fn merge_items_splits_tiny_interleaved_text_blocks_at_same_y() {
+        // Slide-deck copy-paste layers render several `BT` blocks into a
+        // tiny (<3pt) text layer all at the same Y.  Before the fix,
+        // grouping by Y alone would fuse them into one bucket, and the
+        // subsequent sort-by-X would character-interleave them:
+        // "A O B S G T..." instead of "Alpha...Omega...".
+        //
+        // The sharp X backtrack between blocks must open a new group so
+        // block-A items stay before block-B items in the output, even
+        // after their shared Y bucket would be X-sorted.
+        let items = vec![
+            make_tiny_item("Alpha", 10.0, 2.0),
+            make_tiny_item("Gamma", 40.0, 2.0),
+            // Block B: Tm resets X back to the left.  With the fix, this
+            // opens a new group; without, X-sort would put "Omega"
+            // between "Alpha" and "Gamma" (interleaving).
+            make_tiny_item("Omega", 20.0, 2.0),
+            make_tiny_item("Theta", 50.0, 2.0),
+        ];
+        let merged = merge_text_items(items);
+        let texts: Vec<String> = merged.iter().map(|m| m.text.clone()).collect();
+        // The critical property is that "Alpha" + "Gamma" appear
+        // together in stream order before "Omega" + "Theta" — not
+        // interleaved.  Individual merges within each block depend on
+        // gap thresholds that don't matter for this invariant.
+        let alpha_idx = texts.iter().position(|t| t.contains("Alpha")).unwrap();
+        let gamma_idx = texts.iter().position(|t| t.contains("Gamma")).unwrap();
+        let omega_idx = texts.iter().position(|t| t.contains("Omega")).unwrap();
+        let theta_idx = texts.iter().position(|t| t.contains("Theta")).unwrap();
+        assert!(
+            alpha_idx < omega_idx && gamma_idx < omega_idx,
+            "block-A items (Alpha, Gamma) must precede block-B items (Omega, Theta); got {texts:?}"
+        );
+        assert!(
+            omega_idx < theta_idx,
+            "Omega must come before Theta within block B; got {texts:?}"
+        );
+    }
+
+    #[test]
+    fn merge_items_does_not_split_normal_font_table_rows() {
+        // Normal body/table text at 12pt must not be split by the
+        // X-backtrack rule — that rule applies only to tiny (<3pt)
+        // copy-paste layers.  A single 12pt group here means the rule
+        // didn't fire: the items all end up X-sorted together, so
+        // "extra" at X=50 sits between "Row1-ColA" (10) and "Row1-ColB"
+        // (100) rather than staying at the end.
+        let items = vec![
+            make_merge_item("Row1-ColA", 10.0, 50.0),
+            make_merge_item("Row1-ColB", 100.0, 50.0),
+            make_merge_item("Row1-ColC", 200.0, 50.0),
+            make_merge_item("extra", 50.0, 30.0),
+        ];
+        let merged = merge_text_items(items);
+        let texts: Vec<String> = merged.iter().map(|m| m.text.clone()).collect();
+        // Find positions of the non-merged anchors.
+        let a_idx = texts
+            .iter()
+            .position(|t| t.contains("Row1-ColA"))
+            .expect("ColA in output");
+        let extra_idx = texts
+            .iter()
+            .position(|t| t.contains("extra"))
+            .expect("extra in output");
+        let b_idx = texts
+            .iter()
+            .position(|t| t.contains("Row1-ColB"))
+            .expect("ColB in output");
+        assert!(
+            a_idx < extra_idx && extra_idx < b_idx,
+            "12pt items must be X-sorted together (Row1-ColA, extra, Row1-ColB); got {texts:?}"
+        );
+    }
+
+    #[test]
+    fn merge_items_tolerates_small_kerning_overlap() {
+        // Small negative gaps (ligature/kerning) must NOT split a word.
+        // A ~1pt overlap between "Wo" and "rld" is well within normal
+        // kerning and should stay in one group.
+        let items = vec![
+            make_merge_item("Wo", 100.0, 20.0),  // end = 120
+            make_merge_item("rld", 119.0, 25.0), // 1pt overlap: kerning, not a reset
+        ];
+        let merged = merge_text_items(items);
+        assert_eq!(merged.len(), 1);
+        assert_eq!(merged[0].text, "World");
+    }
+
     #[test]
     fn test_group_into_lines() {
         let items = vec![

src/lib.rs

@@ -1,3 +1,9 @@
+// Rust 1.95 introduced collapsible_match for `if` inside match arms.
+// The content-stream parsers use this pattern extensively (match on operator
+// name, then check `in_text_block && !op.operands.is_empty()`). Collapsing
+// these into match guards would hurt readability. Allow crate-wide.
+#![allow(clippy::collapsible_match)]
+
 //! Smart PDF detection and text extraction using lopdf
 //!
 //! # Quick start
@@ -299,6 +305,149 @@ pub fn classify_pdf_mem(buffer: &[u8]) -> Result<PdfClassification, PdfError> {
     })
 }
 
+// =========================================================================
+// Per-page markdown extraction
+// =========================================================================
+
+/// Per-page markdown extraction result.
+#[derive(Debug)]
+pub struct PageMarkdown {
+    /// 0-indexed page number.
+    pub page: u32,
+    /// Formatted markdown for this page.
+    pub markdown: String,
+    /// `true` when text on this page is unreliable (GID-encoded fonts,
+    /// encoding issues, garbage text, or empty extraction).
+    pub needs_ocr: bool,
+}
+
+/// Combined per-page markdown extraction and layout classification result.
+#[derive(Debug)]
+pub struct PagesExtractionResult {
+    /// Per-page markdown results.
+    pub pages: Vec<PageMarkdown>,
+    /// 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 specific pages of a PDF, with layout
+/// classification metadata.
+///
+/// Unlike [`process_pdf_mem`] which returns one concatenated markdown string,
+/// this returns per-page markdown so callers can mix direct extraction
+/// (for simple text pages) with GPU OCR (for complex/scanned pages).
+///
+/// Font statistics are computed from the full document so header
+/// detection thresholds are consistent regardless of which pages are
+/// requested. Per-page `needs_ocr` is set when the page has GID-encoded
+/// fonts, encoding issues, or garbage text.
+///
+/// Layout complexity (tables, columns) is computed from the full document
+/// at near-zero cost since the items/rects/lines are already in memory.
+pub fn extract_pages_markdown_mem(
+    buffer: &[u8],
+    pages: &[u32],
+) -> Result<PagesExtractionResult, PdfError> {
+    validate_pdf_bytes(buffer)?;
+    let (doc, page_count) = load_document_from_mem(buffer)?;
+    let font_cmaps = FontCMaps::from_doc(&doc);
+
+    // Extract ALL pages to get accurate, document-wide font stats.
+    let ((all_items, all_rects, all_lines), page_thresholds, gid_pages) =
+        extractor::extract_positioned_text_from_doc(&doc, &font_cmaps, None)?;
+
+    // Compute layout complexity from full document (near-zero cost).
+    let complexity = compute_layout_complexity(&all_items, &all_rects, &all_lines);
+
+    // Compute font stats from full document (cross-page consistency).
+    let font_stats = markdown::analysis::calculate_font_stats_from_items(&all_items);
+
+    let mut results = Vec::with_capacity(pages.len());
+    let mut pages_needing_ocr = Vec::new();
+
+    for &page_0idx in pages {
+        // Out-of-range pages → empty + needs_ocr
+        if page_0idx >= page_count {
+            pages_needing_ocr.push(page_0idx + 1);
+            results.push(PageMarkdown {
+                page: page_0idx,
+                markdown: String::new(),
+                needs_ocr: true,
+            });
+            continue;
+        }
+
+        let page_1idx = page_0idx + 1;
+
+        // Filter items/rects for this page only
+        let page_items: Vec<TextItem> = all_items
+            .iter()
+            .filter(|i| i.page == page_1idx)
+            .cloned()
+            .collect();
+
+        let page_rects: Vec<PdfRect> = all_rects
+            .iter()
+            .filter(|r| r.page == page_1idx)
+            .cloned()
+            .collect();
+
+        let has_gid = gid_pages.contains(&page_1idx);
+
+        // Build markdown with document-wide font stats
+        let options = MarkdownOptions {
+            base_font_size: Some(font_stats.most_common_size),
+            include_page_numbers: false,
+            strip_headers_footers: false,
+            ..MarkdownOptions::default()
+        };
+
+        let md = markdown::to_markdown_from_items_with_rects_and_lines(
+            page_items,
+            options,
+            &page_rects,
+            &[],
+            &page_thresholds,
+            None,
+            &[],
+        );
+
+        let needs_ocr = md.trim().is_empty()
+            || has_gid
+            || is_garbage_text(&md)
+            || is_cid_garbage(&md)
+            || detect_encoding_issues(&md);
+
+        if needs_ocr {
+            pages_needing_ocr.push(page_1idx);
+        }
+
+        results.push(PageMarkdown {
+            page: page_0idx,
+            markdown: if needs_ocr { String::new() } else { md },
+            needs_ocr,
+        });
+    }
+
+    Ok(PagesExtractionResult {
+        pages: results,
+        pages_with_tables: complexity.pages_with_tables,
+        pages_with_columns: complexity.pages_with_columns,
+        pages_needing_ocr,
+        is_complex: complexity.is_complex,
+    })
+}
+
+// =========================================================================
+// Region-based text extraction (for hybrid OCR pipelines)
+// =========================================================================
+
 /// Result for a single region's text extraction.
 #[derive(Debug)]
 pub struct RegionText {

src/tables/detect_heuristic.rs

@@ -653,15 +653,20 @@ 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
+    if is_paragraph_content(&cells) {
+        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;
     }
 
@@ -898,77 +903,247 @@ 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(super) 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.
@@ -1144,6 +1319,33 @@ 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;
@@ -1398,4 +1600,227 @@ mod tests {
             "data table with dot-leader labels should not be rejected as TOC"
         );
     }
+
+    #[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/format.rs

@@ -1,5 +1,6 @@
 //! Table-to-markdown formatting and cell cleanup.
 
+use super::detect_heuristic::is_table_of_contents;
 use super::Table;
 
 pub fn table_to_markdown(table: &Table) -> String {
@@ -7,6 +8,22 @@ pub fn table_to_markdown(table: &Table) -> String {
         return String::new();
     }
 
+    // Detect TOC on the raw cells: clean_table_cells merges rows in ways
+    // that can make genuine data tables superficially resemble a TOC
+    // (short numeric cells, few columns) — but the raw detection here
+    // preserves the original multi-column structure and only matches the
+    // true TOC pattern.
+    //
+    // Tables of contents 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: continuation-row merging collapses separate
+    // TOC entries (e.g. "6.2 Contamination" + "6.2.1 SWE-bench") into a
+    // single line because sub-entries leave column 0 empty.
+    if is_table_of_contents(&table.cells) {
+        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 +66,101 @@ 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())
+}
+
 /// 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();
@@ -432,4 +544,42 @@ mod tests {
         };
         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 {
+            columns: vec![50.0, 80.0, 300.0],
+            rows: vec![500.0; 5],
+            cells: 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()],
+            ],
+            item_indices: vec![],
+        };
+        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

@@ -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()
     );
@@ -148,6 +158,116 @@ 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();

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
@@ -2717,4 +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/integration_tests.rs

@@ -4,9 +4,10 @@ use pdf_inspector::detector::{DetectionConfig, ScanStrategy};
 use pdf_inspector::extractor::group_into_lines;
 use pdf_inspector::types::TextLine;
 use pdf_inspector::{
-    detect_pdf_type, extract_tables_in_regions_mem, extract_text, extract_text_in_regions_mem,
-    extract_text_with_positions, process_pdf_mem, process_pdf_with_options, to_markdown,
-    MarkdownOptions, PdfError, PdfOptions, PdfType, TextItem,
+    detect_pdf_type, extract_pages_markdown_mem, extract_tables_in_regions_mem, extract_text,
+    extract_text_in_regions_mem, extract_text_with_positions, process_pdf_mem,
+    process_pdf_with_options, to_markdown, MarkdownOptions, PdfError, PdfOptions, PdfType,
+    TextItem,
 };
 use std::collections::HashSet;
 
@@ -1436,3 +1437,196 @@ fn test_extract_tables_in_regions_nonexistent_page() {
     assert!(region.needs_ocr);
     assert!(region.text.is_empty());
 }
+
+#[test]
+fn test_bits_pilani_page4_table_detection() {
+    // Page 4 (0-indexed 3) has a table with multi-line wrapped headers and
+    // numeric data columns. The heuristic detector previously failed because:
+    // 1. Header items at different X positions than data created extra column
+    //    clusters (6 cols instead of 4)
+    // 2. Spanning super-header row ("First Degree | First Degree") produced
+    //    duplicate header cells that looks_like_partial_table_ex rejected
+    let buf = std::fs::read("tests/fixtures/bits_pilani_feedback.pdf").unwrap();
+    let results =
+        extract_tables_in_regions_mem(&buf, &[(3, vec![[0.0, 0.0, 612.0, 792.0]])]).unwrap();
+    assert_eq!(results.len(), 1);
+    let region = &results[0].regions[0];
+    assert!(
+        !region.needs_ocr,
+        "Page 4 table should be detected, got needs_ocr=true"
+    );
+    assert!(
+        region.text.contains("BIO"),
+        "Should contain department name BIO"
+    );
+    assert!(region.text.contains("8.23"), "Should contain numeric data");
+}
+
+#[test]
+fn test_bits_pilani_page8_table_detection() {
+    // Page 8 (0-indexed 7) has a numbered-row table that already worked.
+    // Verify it still works after changes.
+    let buf = std::fs::read("tests/fixtures/bits_pilani_feedback.pdf").unwrap();
+    let results =
+        extract_tables_in_regions_mem(&buf, &[(7, vec![[0.0, 0.0, 612.0, 792.0]])]).unwrap();
+    assert_eq!(results.len(), 1);
+    let region = &results[0].regions[0];
+    assert!(!region.needs_ocr, "Page 8 table should still be detected");
+}
+
+// =========================================================================
+// extract_pages_markdown_mem tests
+// =========================================================================
+
+#[test]
+fn test_extract_pages_markdown_basic() {
+    let buf = std::fs::read("tests/fixtures/nexo-price-en.pdf").unwrap();
+    let result = extract_pages_markdown_mem(&buf, &[0, 1]).unwrap();
+
+    assert_eq!(result.pages.len(), 2);
+    assert_eq!(result.pages[0].page, 0);
+    assert_eq!(result.pages[1].page, 1);
+    // Text-based PDF should produce non-empty markdown
+    assert!(!result.pages[0].markdown.is_empty());
+    assert!(!result.pages[0].needs_ocr);
+}
+
+#[test]
+fn test_extract_pages_markdown_page_ordering() {
+    let buf = std::fs::read("tests/fixtures/nexo-price-en.pdf").unwrap();
+    // Request pages in non-sequential order
+    let result = extract_pages_markdown_mem(&buf, &[1, 0]).unwrap();
+
+    assert_eq!(result.pages.len(), 2);
+    // Results should match input order, not document order
+    assert_eq!(result.pages[0].page, 1);
+    assert_eq!(result.pages[1].page, 0);
+}
+
+#[test]
+fn test_extract_pages_markdown_out_of_range() {
+    let buf = std::fs::read("tests/fixtures/nexo-price-en.pdf").unwrap();
+    let result = extract_pages_markdown_mem(&buf, &[9999]).unwrap();
+
+    assert_eq!(result.pages.len(), 1);
+    assert_eq!(result.pages[0].page, 9999);
+    assert!(result.pages[0].markdown.is_empty());
+    assert!(result.pages[0].needs_ocr);
+    assert!(result.pages_needing_ocr.contains(&10000)); // 1-indexed
+}
+
+#[test]
+fn test_extract_pages_markdown_empty_pages_list() {
+    let buf = std::fs::read("tests/fixtures/nexo-price-en.pdf").unwrap();
+    let result = extract_pages_markdown_mem(&buf, &[]).unwrap();
+    assert!(result.pages.is_empty());
+}
+
+#[test]
+fn test_extract_pages_markdown_single_page() {
+    let buf = std::fs::read("tests/fixtures/nexo-price-en.pdf").unwrap();
+    let result = extract_pages_markdown_mem(&buf, &[0]).unwrap();
+
+    assert_eq!(result.pages.len(), 1);
+    assert_eq!(result.pages[0].page, 0);
+    assert!(!result.pages[0].markdown.is_empty());
+    assert!(!result.pages[0].needs_ocr);
+}
+
+#[test]
+fn test_extract_pages_markdown_invalid_buffer() {
+    let result = extract_pages_markdown_mem(b"not a pdf", &[0]);
+    assert!(result.is_err());
+}
+
+#[test]
+fn test_extract_pages_markdown_gid_pages_need_ocr() {
+    // shinagawa_identity_h.pdf has GID-encoded fonts
+    let buf = std::fs::read("tests/fixtures/shinagawa_identity_h.pdf").unwrap();
+    let result = extract_pages_markdown_mem(&buf, &[0]).unwrap();
+
+    assert_eq!(result.pages.len(), 1);
+    assert!(result.pages[0].needs_ocr);
+    assert!(result.pages_needing_ocr.contains(&1)); // 1-indexed
+}
+
+#[test]
+fn test_extract_pages_markdown_classification_with_tables() {
+    // nexo-price-en.pdf is known to have tables
+    let buf = std::fs::read("tests/fixtures/nexo-price-en.pdf").unwrap();
+    let page_count = process_pdf_mem(&buf).unwrap().page_count;
+    let page_indices: Vec<u32> = (0..page_count).collect();
+    let result = extract_pages_markdown_mem(&buf, &page_indices).unwrap();
+
+    assert!(
+        !result.pages_with_tables.is_empty(),
+        "nexo-price-en.pdf should have pages with tables"
+    );
+    assert!(result.is_complex);
+}
+
+#[test]
+fn test_extract_pages_markdown_simple_pdf_no_complexity() {
+    // bare_name_struct.pdf is a simple document with a heading and code block
+    let buf = std::fs::read("tests/fixtures/bare_name_struct.pdf").unwrap();
+    let result = extract_pages_markdown_mem(&buf, &[0]).unwrap();
+
+    assert!(result.pages_with_tables.is_empty());
+    assert!(result.pages_with_columns.is_empty());
+    assert!(!result.is_complex);
+}
+
+#[test]
+fn test_extract_pages_markdown_classification_matches_process_pdf() {
+    let buf = std::fs::read("tests/fixtures/nexo-price-en.pdf").unwrap();
+    let full = process_pdf_mem(&buf).unwrap();
+    let page_count = full.page_count;
+    let page_indices: Vec<u32> = (0..page_count).collect();
+    let result = extract_pages_markdown_mem(&buf, &page_indices).unwrap();
+
+    assert_eq!(
+        result.pages_with_tables, full.layout.pages_with_tables,
+        "pages_with_tables should match process_pdf"
+    );
+    assert_eq!(
+        result.pages_with_columns, full.layout.pages_with_columns,
+        "pages_with_columns should match process_pdf"
+    );
+}
+
+#[test]
+fn test_extract_pages_markdown_consistency_with_process_pdf() {
+    let buf = std::fs::read("tests/fixtures/nexo-price-en.pdf").unwrap();
+
+    // Get full process_pdf output
+    let full = process_pdf_mem(&buf).unwrap();
+    let full_md = full.markdown.unwrap_or_default();
+
+    // Get per-page output for all pages
+    let page_count = full.page_count;
+    let page_indices: Vec<u32> = (0..page_count).collect();
+    let result = extract_pages_markdown_mem(&buf, &page_indices).unwrap();
+
+    // Concatenated per-page markdown should contain substantial overlap with
+    // the full output (exact match not expected due to header/footer stripping
+    // and cross-page paragraph merging differences)
+    let concat: String = result
+        .pages
+        .iter()
+        .map(|p| p.markdown.as_str())
+        .collect::<Vec<_>>()
+        .join("\n");
+
+    // Both should be non-empty for a text-based PDF
+    assert!(!full_md.is_empty());
+    assert!(!concat.is_empty());
+
+    // The per-page version should contain at least 50% of the full content's
+    // length (accounting for header/footer stripping differences)
+    assert!(
+        concat.len() * 2 >= full_md.len(),
+        "per-page concat ({} chars) is too short vs full ({} chars)",
+        concat.len(),
+        full_md.len()
+    );
+}