fix(markdown): strip garbled running footers

* fix(extractor): reduce garbled OCR false positives

* fix(extractor): tighten garbled text OCR routing

* fix(extractor): decode UTF-16 ToUnicode destinations

* fix(extractor): narrow ToUnicode destination cleanup

* fix(extractor): decode Aptos private ff ligature

.github/workflows/ci.yml

@@ -20,17 +20,7 @@ jobs:
         uses: dtolnay/rust-toolchain@stable
 
       - name: Cache cargo
-        uses: actions/cache@v4
-        with:
-          path: |
-            ~/.cargo/bin/
-            ~/.cargo/registry/index/
-            ~/.cargo/registry/cache/
-            ~/.cargo/git/db/
-            target/
-          key: ${{ runner.os }}-cargo-${{ hashFiles('**/Cargo.lock') }}
-          restore-keys: |
-            ${{ runner.os }}-cargo-
+        uses: Swatinem/rust-cache@v2
 
       - name: Run tests
         run: cargo test --verbose
@@ -61,17 +51,9 @@ jobs:
           components: clippy
 
       - name: Cache cargo
-        uses: actions/cache@v4
+        uses: Swatinem/rust-cache@v2
         with:
-          path: |
-            ~/.cargo/bin/
-            ~/.cargo/registry/index/
-            ~/.cargo/registry/cache/
-            ~/.cargo/git/db/
-            target/
-          key: ${{ runner.os }}-cargo-clippy-${{ hashFiles('**/Cargo.lock') }}
-          restore-keys: |
-            ${{ runner.os }}-cargo-clippy-
+          key: clippy
 
       - name: Run clippy
         run: cargo clippy -- -D warnings
@@ -89,17 +71,9 @@ jobs:
         uses: dtolnay/rust-toolchain@stable
 
       - name: Cache cargo
-        uses: actions/cache@v4
+        uses: Swatinem/rust-cache@v2
         with:
-          path: |
-            ~/.cargo/bin/
-            ~/.cargo/registry/index/
-            ~/.cargo/registry/cache/
-            ~/.cargo/git/db/
-            target/
-          key: ${{ runner.os }}-cargo-build-${{ hashFiles('**/Cargo.lock') }}
-          restore-keys: |
-            ${{ runner.os }}-cargo-build-
+          key: build
 
       - name: Build
         run: cargo build --release --verbose

.github/workflows/publish-crate.yml

@@ -0,0 +1,87 @@
+name: Publish Rust crate
+
+on:
+  push:
+    branches: [main]
+    paths: ['Cargo.toml']
+
+permissions:
+  contents: read
+
+env:
+  CARGO_TERM_COLOR: always
+
+jobs:
+  check-version:
+    name: Check version change
+    runs-on: ubuntu-latest
+    outputs:
+      changed: ${{ steps.check.outputs.changed }}
+      published: ${{ steps.check.outputs.published }}
+      version: ${{ steps.check.outputs.version }}
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 2
+
+      - name: Check if version changed
+        id: check
+        run: |
+          NEW_VERSION=$(python3 -c 'import pathlib, tomllib; print(tomllib.loads(pathlib.Path("Cargo.toml").read_text())["package"]["version"])')
+          OLD_VERSION=$(git show HEAD~1:Cargo.toml | python3 -c 'import sys, tomllib; print(tomllib.loads(sys.stdin.read())["package"]["version"])')
+          echo "old=$OLD_VERSION new=$NEW_VERSION"
+          echo "version=$NEW_VERSION" >> "$GITHUB_OUTPUT"
+
+          if [ "$NEW_VERSION" = "$OLD_VERSION" ]; then
+            echo "changed=false" >> "$GITHUB_OUTPUT"
+            echo "published=false" >> "$GITHUB_OUTPUT"
+            exit 0
+          fi
+
+          echo "changed=true" >> "$GITHUB_OUTPUT"
+
+          HTTP_STATUS=$(curl --silent --show-error --output /tmp/crate-version.json --write-out "%{http_code}" \
+            -H "User-Agent: firecrawl/pdf-inspector publish workflow (https://github.com/firecrawl/pdf-inspector)" \
+            "https://crates.io/api/v1/crates/pdf-inspector/$NEW_VERSION")
+
+          case "$HTTP_STATUS" in
+            200)
+              echo "published=true" >> "$GITHUB_OUTPUT"
+              echo "pdf-inspector v$NEW_VERSION is already published"
+              ;;
+            404)
+              echo "published=false" >> "$GITHUB_OUTPUT"
+              ;;
+            *)
+              cat /tmp/crate-version.json
+              echo "Unexpected crates.io response: $HTTP_STATUS" >&2
+              exit 1
+              ;;
+          esac
+
+  publish:
+    name: Publish to crates.io
+    needs: check-version
+    if: needs.check-version.outputs.changed == 'true' && needs.check-version.outputs.published == 'false'
+    runs-on: ubuntu-latest
+    environment: crates-io
+    permissions:
+      contents: read
+      id-token: write
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Install Rust
+        uses: dtolnay/rust-toolchain@stable
+
+      - name: Verify package
+        run: cargo publish --dry-run
+
+      - name: Authenticate with crates.io
+        id: auth
+        uses: rust-lang/crates-io-auth-action@v1
+
+      - name: Publish crate
+        run: cargo publish
+        env:
+          CARGO_REGISTRY_TOKEN: ${{ steps.auth.outputs.token }}

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "pdf-inspector"
-version = "0.1.0"
+version = "0.1.3"
 edition = "2021"
 autobins = false
 authors = ["Firecrawl Team"]
@@ -17,7 +17,7 @@ crate-type = ["lib", "cdylib"]
 pyo3 = { version = "0.25", features = ["extension-module"], optional = true }
 
 # PDF parsing
-lopdf = { git = "https://github.com/J-F-Liu/lopdf", rev = "7a05512d831415b1f2b1ce522391d6beab8a1284", features = ["rayon"] }
+lopdf = { version = "0.41.0", features = ["rayon"] }
 
 # Error handling
 thiserror = "2.0"

README.md

@@ -1,5 +1,8 @@
 # pdf-inspector
 
+[![Crates.io](https://img.shields.io/crates/v/pdf-inspector.svg)](https://crates.io/crates/pdf-inspector)
+[![npm](https://img.shields.io/npm/v/@firecrawl/pdf-inspector.svg)](https://www.npmjs.com/package/@firecrawl/pdf-inspector)
+
 Fast Rust library for PDF classification and text extraction. Detects whether a PDF is text-based or scanned, extracts text with position awareness, and converts to clean Markdown — all without OCR. Includes bindings for [Python](docs/python.md) and [Node.js](napi/README.md).
 
 Built by [Firecrawl](https://firecrawl.dev) to handle text-based PDFs locally in under 200ms, skipping expensive OCR services for the ~54% of PDFs that don't need them.
@@ -71,9 +74,17 @@ console.log(result.markdown);  // Markdown string or null
 
 ### Rust
 
+Install from [crates.io](https://crates.io/crates/pdf-inspector):
+
+```bash
+cargo add pdf-inspector
+```
+
+Or add it manually:
+
 ```toml
 [dependencies]
-pdf-inspector = { git = "https://github.com/firecrawl/pdf-inspector" }
+pdf-inspector = "0.1"
 ```
 
 ```rust
@@ -91,29 +102,34 @@ if let Some(markdown) = &result.markdown {
 ### CLI
 
 ```bash
+# Install the CLI tools
+cargo install pdf-inspector
+
 # Convert PDF to Markdown
-cargo run --bin pdf2md -- document.pdf
+pdf2md document.pdf
 
 # JSON output (for piping)
-cargo run --bin pdf2md -- document.pdf --json
+pdf2md document.pdf --json
 
 # Raw markdown only (no headers)
-cargo run --bin pdf2md -- document.pdf --raw
+pdf2md document.pdf --raw
 
 # Insert page break markers (<!-- Page N -->)
-cargo run --bin pdf2md -- document.pdf --pages
+pdf2md document.pdf --pages
 
 # Process only specific pages
-cargo run --bin pdf2md -- document.pdf --select-pages 1,3,5-10
+pdf2md document.pdf --select-pages 1,3,5-10
 
 # Detection only (no extraction)
-cargo run --bin detect-pdf -- document.pdf
-cargo run --bin detect-pdf -- document.pdf --json
+detect-pdf document.pdf
+detect-pdf document.pdf --json
 
 # Detection + layout analysis (tables, columns)
-cargo run --bin detect-pdf -- document.pdf --analyze --json
+detect-pdf document.pdf --analyze --json
 ```
 
+From a source checkout, use `cargo run --bin pdf2md -- document.pdf` or `cargo run --bin detect-pdf -- document.pdf` instead.
+
 ## Architecture
 
 ```

SECURITY.md

@@ -0,0 +1,33 @@
+# Security Policy
+
+## Reporting a Vulnerability
+
+If you believe you've found a security vulnerability in pdf-inspector, please
+report it privately so we can fix it before public disclosure.
+
+**Preferred:** Email **help@firecrawl.dev** with:
+
+- A description of the issue and its impact
+- Steps to reproduce (a minimal PDF or input that triggers the bug is ideal)
+- The version or commit hash of pdf-inspector you tested against
+
+**Alternative:** Use GitHub's private vulnerability reporting under the
+[Security tab](https://github.com/firecrawl/pdf-inspector/security/advisories/new).
+
+We'll acknowledge your report in a timely manner and keep you updated on
+remediation progress. Please do not open a public GitHub issue for security
+bugs.
+
+## Scope
+
+In scope:
+- Memory-safety issues (panics, OOB reads, UB) reachable from a crafted PDF
+- Denial-of-service vectors (unbounded allocation, infinite loops) on
+  reasonably-sized inputs
+- Bugs in the `pdf2md` / `detect-pdf` binaries or the `pdf-inspector` crate
+  that affect downstream consumers
+
+Out of scope:
+- Bugs in upstream dependencies (`lopdf`, etc.) — please report those upstream
+- Extraction quality issues (wrong text, missing tables) — open a regular
+  GitHub issue instead

docs/publishing.md

@@ -0,0 +1,21 @@
+# Publishing
+
+The Rust crate is published to [crates.io](https://crates.io/crates/pdf-inspector) with trusted publishing from GitHub Actions. The first release was published manually; future releases publish from `.github/workflows/publish-crate.yml` when a `Cargo.toml` version change lands on `main`.
+
+## crates.io Trusted Publisher
+
+Configure the trusted publisher for the `pdf-inspector` crate with:
+
+- Repository: `firecrawl/pdf-inspector`
+- Workflow: `publish-crate.yml`
+- Environment: `crates-io`
+
+The workflow uses `rust-lang/crates-io-auth-action@v1` to exchange GitHub's OIDC token for a short-lived crates.io token, then passes it to `cargo publish`.
+
+## Release Steps
+
+1. Update `version` in `Cargo.toml`.
+2. Merge the version bump to `main`.
+3. The publish workflow compares the new `Cargo.toml` version with `HEAD~1`, runs `cargo publish --dry-run`, then publishes if that version is not already on crates.io.
+
+If `Cargo.toml` changes without a package version bump, the workflow exits without publishing.

napi/Cargo.lock

@@ -672,8 +672,9 @@ checksum = "5e5032e24019045c762d3c0f28f5b6b8bbf38563a65908389bf7978758920897"
 
 [[package]]
 name = "lopdf"
-version = "0.40.0"
-source = "git+https://github.com/J-F-Liu/lopdf?rev=7a05512d831415b1f2b1ce522391d6beab8a1284#7a05512d831415b1f2b1ce522391d6beab8a1284"
+version = "0.41.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "67513274c50a2b51e5f75d9e682fcf4ab064a8a9c9ae2c3c59309084882bb24d"
 dependencies = [
  "aes",
  "bitflags",
@@ -829,7 +830,7 @@ checksum = "384b8ab6d37215f3c5301a95a4accb5d64aa607f1fcb26a11b5303878451b4fe"
 
 [[package]]
 name = "pdf-inspector"
-version = "0.1.0"
+version = "0.1.3"
 dependencies = [
  "env_logger",
  "log",
@@ -844,7 +845,7 @@ dependencies = [
 
 [[package]]
 name = "pdf-inspector-napi"
-version = "0.2.0"
+version = "0.2.2"
 dependencies = [
  "napi",
  "napi-build",

napi/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "pdf-inspector-napi"
-version = "0.2.0"
+version = "0.2.2"
 edition = "2021"
 
 [lib]

napi/README.md

@@ -37,7 +37,7 @@ console.log(result.confidence)     // 0.875
 
 Extract text within bounding-box regions from a PDF. Designed for hybrid OCR pipelines where a layout model detects regions in rendered page images, and this function extracts text from the PDF structure for text-based pages — skipping GPU OCR.
 
-Each region result includes a `needsOcr` flag that signals unreliable extraction (empty text, GID-encoded fonts, garbage text, encoding issues).
+Each region result includes a `needsOcr` flag that signals unreliable extraction (empty text, GID-encoded fonts, garbage text, encoding issues). When the cause is a suspected garbled text layer, `ocrReason` is set to `"suspected_garbled_text"`.
 
 ```typescript
 import { extractTextInRegions } from '@firecrawl/pdf-inspector'
@@ -84,6 +84,7 @@ interface PageRegionTexts {
 interface RegionText {
   text: string
   needsOcr: boolean         // true when text is unreliable
+  ocrReason?: string        // "suspected_garbled_text" when known
 }
 ```
 

napi/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@firecrawl/pdf-inspector",
-  "version": "1.7.0",
+  "version": "1.9.8",
   "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",

napi/probe-indent.mjs

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

napi/src/lib.rs

@@ -40,6 +40,8 @@ pub struct PdfResult {
     pub processing_time_ms: u32,
     /// 1-indexed page numbers that need OCR.
     pub pages_needing_ocr: Vec<u32>,
+    /// Machine-readable OCR reasons by 1-indexed page.
+    pub ocr_reasons_by_page: Vec<PageOcrReasons>,
     pub title: Option<String>,
     pub confidence: f64,
     pub is_complex_layout: bool,
@@ -48,6 +50,13 @@ pub struct PdfResult {
     pub has_encoding_issues: bool,
 }
 
+/// OCR reasons for a single 1-indexed page.
+#[napi(object)]
+pub struct PageOcrReasons {
+    pub page: u32,
+    pub reasons: Vec<String>,
+}
+
 /// Lightweight PDF classification result.
 #[napi(object)]
 pub struct PdfClassification {
@@ -90,6 +99,8 @@ pub struct RegionText {
     pub text: String,
     /// `true` when the text should not be trusted (empty, GID fonts, garbage, encoding issues).
     pub needs_ocr: bool,
+    /// Machine-readable OCR reason when the cause is known.
+    pub ocr_reason: Option<String>,
 }
 
 /// Extracted text for one page's regions.
@@ -99,6 +110,13 @@ pub struct PageRegionTexts {
     pub regions: Vec<RegionText>,
 }
 
+/// Vector-grid detection result compatible with `extractTablesWithStructure*`.
+#[napi(object)]
+pub struct VectorGridDetectionJs {
+    pub structure_tokens: Vec<String>,
+    pub cell_bboxes: Vec<Vec<f64>>,
+}
+
 // ---------------------------------------------------------------------------
 // Helpers
 // ---------------------------------------------------------------------------
@@ -119,6 +137,7 @@ fn to_napi_result(r: pdf_inspector::PdfProcessResult) -> PdfResult {
         page_count: r.page_count,
         processing_time_ms: r.processing_time_ms as u32,
         pages_needing_ocr: r.pages_needing_ocr,
+        ocr_reasons_by_page: to_napi_page_ocr_reasons(r.ocr_reasons_by_page),
         title: r.title,
         confidence: r.confidence as f64,
         is_complex_layout: r.layout.is_complex,
@@ -128,6 +147,18 @@ fn to_napi_result(r: pdf_inspector::PdfProcessResult) -> PdfResult {
     }
 }
 
+fn to_napi_page_ocr_reasons(
+    reasons: Vec<pdf_inspector::PageOcrReasons>,
+) -> Vec<PageOcrReasons> {
+    reasons
+        .into_iter()
+        .map(|reason| PageOcrReasons {
+            page: reason.page,
+            reasons: reason.reasons,
+        })
+        .collect()
+}
+
 fn convert_item_type(t: &pdf_inspector::types::ItemType) -> (ItemType, Option<String>) {
     match t {
         pdf_inspector::types::ItemType::Text => (ItemType::Text, None),
@@ -317,6 +348,53 @@ pub fn extract_tables_in_regions(
     })
 }
 
+/// Detect a vector ruled-line / rectangle grid inside one page region.
+///
+/// Returns TSR-compatible structure tokens plus crop-pixel cell bboxes, or
+/// `null` when the region does not contain a valid vector grid.
+///
+/// `pageIdx` is 0-indexed. `regionPdfPtBbox` is `[x1,y1,x2,y2]` in PDF
+/// points with top-left origin. `renderDpi` is the DPI of the crop image that
+/// will consume the returned cell bboxes.
+#[napi]
+pub fn detect_vector_grid_in_region(
+    buffer: Buffer,
+    page_idx: u32,
+    region_pdf_pt_bbox: Vec<f64>,
+    render_dpi: f64,
+) -> Result<Option<VectorGridDetectionJs>> {
+    let bytes: Vec<u8> = buffer.to_vec();
+    let region = if region_pdf_pt_bbox.len() == 4 {
+        [
+            region_pdf_pt_bbox[0] as f32,
+            region_pdf_pt_bbox[1] as f32,
+            region_pdf_pt_bbox[2] as f32,
+            region_pdf_pt_bbox[3] as f32,
+        ]
+    } else {
+        [0.0, 0.0, 0.0, 0.0]
+    };
+
+    catch_panic("detect_vector_grid_in_region", move || {
+        let result = pdf_inspector::detect_vector_grid_in_region_mem(
+            &bytes,
+            page_idx,
+            region,
+            render_dpi as f32,
+        )
+        .map_err(|e| to_napi_err(e, "detect_vector_grid_in_region"))?;
+
+        Ok(result.map(|r| VectorGridDetectionJs {
+            structure_tokens: r.structure_tokens,
+            cell_bboxes: r
+                .cell_bboxes
+                .into_iter()
+                .map(|bbox| bbox.into_iter().map(|v| v as f64).collect())
+                .collect(),
+        }))
+    })
+}
+
 /// One cropped table region plus its raw structure-recovery output, for
 /// `extractTablesWithStructure`.
 ///
@@ -422,6 +500,54 @@ pub fn extract_tables_with_structure_cells(
     })
 }
 
+/// One result from `extractTablesWithStructureAuto` — markdown plus a
+/// diagnostic flag identifying which path produced it.
+///
+/// `fallbackReason` is `null` when the TSR-hybrid path produced the
+/// markdown directly. When stage 1's quality check fires (the cells
+/// look like a SLANet detection pathology — phantom rows or multi-row
+/// content in a single cell), the auto path may expand the TSR cells
+/// in-place or run the heuristic table extractor on the same region.
+/// `fallbackReason` carries the diagnostic label (for example
+/// `"multi_row_in_cell_expanded"` or `"phantom_empty_row"`).
+#[napi(object)]
+pub struct TableExtractionResultJs {
+    pub markdown: String,
+    pub fallback_reason: Option<String>,
+}
+
+/// Auto-fallback variant of [`extractTablesWithStructure`].
+///
+/// Runs the TSR-hybrid path, checks the resulting cells for known
+/// SLANet detection pathologies, expands multi-row cells in-place when
+/// possible, and otherwise falls back to the heuristic
+/// `extractTablesInRegions` for inputs where the TSR path looks
+/// compromised.
+///
+/// On clean inputs this returns identical markdown to
+/// `extractTablesWithStructure`; on flagged inputs `fallbackReason` is
+/// set to the recovery path that produced the result.
+#[napi]
+pub fn extract_tables_with_structure_auto(
+    buffer: Buffer,
+    inputs: Vec<TsrTableInputJs>,
+) -> Result<Vec<TableExtractionResultJs>> {
+    let bytes: Vec<u8> = buffer.to_vec();
+    let parsed = parse_tsr_inputs(&inputs);
+
+    catch_panic("extract_tables_with_structure_auto", move || {
+        let result = pdf_inspector::extract_tables_with_structure_auto_mem(&bytes, &parsed)
+            .map_err(|e| to_napi_err(e, "extract_tables_with_structure_auto"))?;
+        Ok(result
+            .into_iter()
+            .map(|r| TableExtractionResultJs {
+                markdown: r.markdown,
+                fallback_reason: r.fallback_reason,
+            })
+            .collect())
+    })
+}
+
 fn parse_tsr_inputs(inputs: &[TsrTableInputJs]) -> Vec<pdf_inspector::TsrTableInput> {
     inputs
         .iter()
@@ -461,6 +587,8 @@ pub struct PageMarkdownResult {
     pub markdown: String,
     /// `true` when text on this page is unreliable.
     pub needs_ocr: bool,
+    /// Machine-readable OCR reason when the cause is known.
+    pub ocr_reason: Option<String>,
 }
 
 /// Combined per-page markdown extraction and layout classification result.
@@ -474,6 +602,8 @@ pub struct PagesExtractionResult {
     pub pages_with_columns: Vec<u32>,
     /// 1-indexed pages that need OCR (scanned/image-based).
     pub pages_needing_ocr: Vec<u32>,
+    /// Machine-readable OCR reasons by 1-indexed page.
+    pub ocr_reasons_by_page: Vec<PageOcrReasons>,
     /// True if any page has tables or columns.
     pub is_complex: bool,
 }
@@ -505,11 +635,13 @@ pub fn extract_pages_markdown(
                     page: r.page,
                     markdown: r.markdown,
                     needs_ocr: r.needs_ocr,
+                    ocr_reason: r.ocr_reason,
                 })
                 .collect(),
             pages_with_tables: result.pages_with_tables,
             pages_with_columns: result.pages_with_columns,
             pages_needing_ocr: result.pages_needing_ocr,
+            ocr_reasons_by_page: to_napi_page_ocr_reasons(result.ocr_reasons_by_page),
             is_complex: result.is_complex,
         })
     })
@@ -546,6 +678,7 @@ fn to_page_region_texts(results: Vec<pdf_inspector::PageRegionResult>) -> Vec<Pa
                 .map(|r| RegionText {
                     text: r.text,
                     needs_ocr: r.needs_ocr,
+                    ocr_reason: r.ocr_reason,
                 })
                 .collect(),
         })

napi/test.mjs

@@ -7,6 +7,7 @@ import {
   extractText,
   extractTextWithPositions,
   extractTextInRegions,
+  detectVectorGridInRegion,
   extractPagesMarkdown,
 } from './index.js';
 
@@ -90,6 +91,17 @@ assert.equal(typeof regionResults[0].regions[0].text, 'string');
 assert.equal(typeof regionResults[0].regions[0].needsOcr, 'boolean');
 console.log('  extractTextInRegions: OK');
 
+// --- detectVectorGridInRegion ---
+console.log('Testing detectVectorGridInRegion...');
+const vectorGrid = detectVectorGridInRegion(fixture, 0, [0, 0, 600, 800], 72);
+assert.ok(vectorGrid === null || typeof vectorGrid === 'object');
+if (vectorGrid) {
+  assert.ok(Array.isArray(vectorGrid.structureTokens));
+  assert.ok(Array.isArray(vectorGrid.cellBboxes));
+  assert.ok(vectorGrid.cellBboxes.every(bbox => Array.isArray(bbox) && bbox.length === 4));
+}
+console.log('  detectVectorGridInRegion: OK');
+
 // --- extractPagesMarkdown ---
 console.log('Testing extractPagesMarkdown...');
 

src/bin/pdf2md.rs

@@ -31,6 +31,22 @@ fn json_escape(s: &str) -> String {
     out
 }
 
+fn format_ocr_reasons_by_page(reasons: &[pdf_inspector::PageOcrReasons]) -> String {
+    reasons
+        .iter()
+        .map(|entry| {
+            let reasons_json = entry
+                .reasons
+                .iter()
+                .map(|reason| format!(r#""{}""#, json_escape(reason)))
+                .collect::<Vec<_>>()
+                .join(",");
+            format!(r#"{{"page":{},"reasons":[{}]}}"#, entry.page, reasons_json)
+        })
+        .collect::<Vec<_>>()
+        .join(",")
+}
+
 /// Parse a page specification like "1,3,5-10,20" into a HashSet of page numbers.
 fn parse_page_spec(spec: &str) -> Result<HashSet<u32>, String> {
     let mut pages = HashSet::new();
@@ -177,12 +193,14 @@ fn main() {
                         .iter()
                         .map(|p| p.to_string())
                         .collect();
+                    let ocr_reasons = format_ocr_reasons_by_page(&result.ocr_reasons_by_page);
                     println!(
-                        r#"{{"pdf_type":"{}","page_count":{},"processing_time_ms":{},"pages_needing_ocr":[{}],"is_complex":{},"pages_with_tables":[{}],"pages_with_columns":[{}],"has_encoding_issues":{}}}"#,
+                        r#"{{"pdf_type":"{}","page_count":{},"processing_time_ms":{},"pages_needing_ocr":[{}],"ocr_reasons_by_page":[{}],"is_complex":{},"pages_with_tables":[{}],"pages_with_columns":[{}],"has_encoding_issues":{}}}"#,
                         pdf_type_str,
                         result.page_count,
                         result.processing_time_ms,
                         ocr_pages.join(","),
+                        ocr_reasons,
                         result.layout.is_complex,
                         table_pages.join(","),
                         col_pages.join(","),
@@ -223,8 +241,9 @@ fn main() {
                     .iter()
                     .map(|p| p.to_string())
                     .collect();
+                let ocr_reasons = format_ocr_reasons_by_page(&result.ocr_reasons_by_page);
                 println!(
-                    r#"{{"pdf_type":"{}","page_count":{},"has_text":{},"processing_time_ms":{},"markdown_length":{},"pages_needing_ocr":[{}],"is_complex":{},"pages_with_tables":[{}],"pages_with_columns":[{}],"has_encoding_issues":{},"markdown":"{}"}}"#,
+                    r#"{{"pdf_type":"{}","page_count":{},"has_text":{},"processing_time_ms":{},"markdown_length":{},"pages_needing_ocr":[{}],"ocr_reasons_by_page":[{}],"is_complex":{},"pages_with_tables":[{}],"pages_with_columns":[{}],"has_encoding_issues":{},"markdown":"{}"}}"#,
                     match result.pdf_type {
                         PdfType::TextBased => "text_based",
                         PdfType::Scanned => "scanned",
@@ -236,6 +255,7 @@ fn main() {
                     result.processing_time_ms,
                     result.markdown.as_ref().map(|m| m.len()).unwrap_or(0),
                     ocr_pages.join(","),
+                    ocr_reasons,
                     result.layout.is_complex,
                     table_pages.join(","),
                     col_pages.join(","),

src/extractor/content_stream.rs

@@ -18,7 +18,7 @@ use super::fonts::{
     get_font_file2_obj_num, get_operand_bytes, CMapDecisionCache,
 };
 use super::xobjects::{extract_form_xobject_text, get_page_xobjects, XObjectType};
-use super::{get_number, multiply_matrices};
+use super::{get_number, image_bbox_from_ctm, multiply_matrices};
 
 /// Strip PDF comments (% to end of line) from content stream bytes.
 ///
@@ -188,7 +188,17 @@ pub(crate) fn extract_page_text_items(
     // Graphics state tracking
     let mut ctm = [1.0f32, 0.0, 0.0, 1.0, 0.0, 0.0]; // Current Transformation Matrix
     let mut text_rendering_mode: i32 = 0; // 0=fill, 1=stroke, 2=fill+stroke, 3=invisible
-    let mut gstate_stack: Vec<([f32; 6], i32, f32, f32)> = Vec::new();
+    #[derive(Clone)]
+    struct SavedGraphicsState {
+        ctm: [f32; 6],
+        text_rendering_mode: i32,
+        char_spacing: f32,
+        word_spacing: f32,
+        text_leading: f32,
+        current_font: String,
+        current_font_size: f32,
+    }
+    let mut gstate_stack: Vec<SavedGraphicsState> = Vec::new();
 
     // Text state tracking
     let mut current_font = String::new();
@@ -227,15 +237,26 @@ pub(crate) fn extract_page_text_items(
         match op.operator.as_str() {
             "q" => {
                 // Save graphics state
-                gstate_stack.push((ctm, text_rendering_mode, char_spacing, word_spacing));
+                gstate_stack.push(SavedGraphicsState {
+                    ctm,
+                    text_rendering_mode,
+                    char_spacing,
+                    word_spacing,
+                    text_leading,
+                    current_font: current_font.clone(),
+                    current_font_size,
+                });
             }
             "Q" => {
                 // Restore graphics state
-                if let Some((saved_ctm, saved_tr, saved_tc, saved_tw)) = gstate_stack.pop() {
-                    ctm = saved_ctm;
-                    text_rendering_mode = saved_tr;
-                    char_spacing = saved_tc;
-                    word_spacing = saved_tw;
+                if let Some(saved) = gstate_stack.pop() {
+                    ctm = saved.ctm;
+                    text_rendering_mode = saved.text_rendering_mode;
+                    char_spacing = saved.char_spacing;
+                    word_spacing = saved.word_spacing;
+                    text_leading = saved.text_leading;
+                    current_font = saved.current_font;
+                    current_font_size = saved.current_font_size;
                 }
             }
             "cm" => {
@@ -385,6 +406,7 @@ pub(crate) fn extract_page_text_items(
                         &font_encodings,
                         &encoding_cache,
                         &mut cmap_decisions,
+                        &font_widths,
                     ) {
                         let combined = multiply_matrices(&text_matrix, &ctm);
                         let rendered_size = effective_font_size(current_font_size, &combined);
@@ -533,6 +555,7 @@ pub(crate) fn extract_page_text_items(
                                     &font_encodings,
                                     &encoding_cache,
                                     &mut cmap_decisions,
+                                    &font_widths,
                                 ) {
                                     current_text.push_str(&text);
                                 }
@@ -620,6 +643,7 @@ pub(crate) fn extract_page_text_items(
                         &font_encodings,
                         &encoding_cache,
                         &mut cmap_decisions,
+                        &font_widths,
                     ) {
                         if !text.trim().is_empty() {
                             let combined = multiply_matrices(&text_matrix, &ctm);
@@ -661,7 +685,29 @@ pub(crate) fn extract_page_text_items(
                         if let Some(xobj_type) = xobjects.get(&xobj_name) {
                             match xobj_type {
                                 XObjectType::Image => {
-                                    // Skip images — text extraction only
+                                    // Emit a positional placeholder for the image
+                                    // so downstream consumers (layout-aware
+                                    // pipelines, figure-OCR routers) can locate
+                                    // raster figures without parsing the PDF
+                                    // again. The text field carries the
+                                    // XObject resource name in the legacy
+                                    // `[Image: Im0]` format that the markdown
+                                    // emitter already recognizes.
+                                    let (x, y, width, height) = image_bbox_from_ctm(&ctm);
+                                    items.push(TextItem {
+                                        text: format!("[Image: {}]", xobj_name),
+                                        x,
+                                        y,
+                                        width,
+                                        height,
+                                        font: String::new(),
+                                        font_size: 0.0,
+                                        page: page_num,
+                                        is_bold: false,
+                                        is_italic: false,
+                                        item_type: ItemType::Image,
+                                        mcid: current_mcid(&marked_content_stack),
+                                    });
                                 }
                                 XObjectType::Form(form_id) => {
                                     // Extract text from Form XObject
@@ -1012,9 +1058,17 @@ pub(crate) fn extract_page_text_items(
     // producing thousands of identical rects that yield a degenerate grid.
     // After dedup, if too few unique clip rects remain we fall through to
     // fill rects (explicitly drawn visible rectangles).
+    //
+    // When fill rects substantially outnumber clip rects, the clips are
+    // typically section-level wrappers and the fills are the actual table
+    // cell backgrounds (e.g. shaded-header tables drawn with `m`/`l`/`h`/`f*`
+    // sequences). In that case, prefer fills.
     if rects.is_empty() {
         dedup_rects(&mut clip_rects);
-        if clip_rects.len() >= 4 {
+        let prefer_fills = !fill_rects.is_empty() && fill_rects.len() >= clip_rects.len() * 3;
+        if prefer_fills {
+            rects = fill_rects;
+        } else if clip_rects.len() >= 4 {
             rects = clip_rects;
         } else if !fill_rects.is_empty() {
             rects = fill_rects;
@@ -1253,6 +1307,91 @@ mod tests {
         assert!(lines.is_empty());
     }
 
+    #[test]
+    fn test_q_restores_current_font_for_text_decoding() {
+        use crate::tounicode::FontCMaps;
+        use lopdf::{dictionary, Object, Stream};
+
+        fn cmap_stream(dst_hex: &str) -> Stream {
+            let cmap = format!(
+                r#"/CIDInit /ProcSet findresource begin
+12 dict begin
+begincmap
+/CIDSystemInfo << /Registry (Adobe) /Ordering (UCS) /Supplement 0 >> def
+/CMapName /Test-UCS def
+/CMapType 2 def
+1 begincodespacerange
+<00> <FF>
+endcodespacerange
+1 beginbfchar
+<41> <{dst_hex}>
+endbfchar
+endcmap
+CMapName currentdict /CMap defineresource pop
+end
+end"#
+            );
+            Stream::new(dictionary! {}, cmap.into_bytes())
+        }
+
+        let mut doc = lopdf::Document::new();
+        let f1_cmap = doc.add_object(Object::Stream(cmap_stream("0058"))); // X
+        let f2_cmap = doc.add_object(Object::Stream(cmap_stream("0059"))); // Y
+        let f1 = doc.add_object(dictionary! {
+            "Type" => "Font",
+            "Subtype" => "Type1",
+            "BaseFont" => "Helvetica",
+            "ToUnicode" => Object::Reference(f1_cmap),
+        });
+        let f2 = doc.add_object(dictionary! {
+            "Type" => "Font",
+            "Subtype" => "Type1",
+            "BaseFont" => "Helvetica",
+            "ToUnicode" => Object::Reference(f2_cmap),
+        });
+
+        let content = b"BT /F1 12 Tf 10 700 Tm <41> Tj ET
+q
+BT /F2 12 Tf 20 700 Tm <41> Tj ET
+Q
+BT 30 700 Tm <41> Tj ET";
+        let content_id = doc.add_object(Object::Stream(Stream::new(
+            dictionary! {},
+            content.to_vec(),
+        )));
+        let page_id = doc.add_object(dictionary! {
+            "Type" => "Page",
+            "Contents" => Object::Reference(content_id),
+            "Resources" => dictionary! {
+                "Font" => dictionary! {
+                    "F1" => Object::Reference(f1),
+                    "F2" => Object::Reference(f2),
+                },
+            },
+            "MediaBox" => vec![0.into(), 0.into(), 612.into(), 792.into()],
+        });
+        let pages_id = doc.add_object(dictionary! {
+            "Type" => "Pages",
+            "Count" => Object::Integer(1),
+            "Kids" => vec![Object::Reference(page_id)],
+        });
+        let catalog_id = doc.add_object(dictionary! {
+            "Type" => "Catalog",
+            "Pages" => Object::Reference(pages_id),
+        });
+        doc.trailer.set("Root", Object::Reference(catalog_id));
+
+        let font_cmaps = FontCMaps::from_doc(&doc);
+        let ((items, _, _), _, _) =
+            extract_page_text_items(&doc, page_id, 1, &font_cmaps, false).unwrap();
+        let text = items
+            .iter()
+            .map(|item| item.text.as_str())
+            .collect::<String>();
+
+        assert_eq!(text, "XYX");
+    }
+
     #[test]
     fn test_strip_pdf_comments() {
         // Basic comment stripping

src/extractor/fonts.rs

@@ -526,6 +526,11 @@ pub(crate) fn parse_font_encoding(
     font_dict: &lopdf::Dictionary,
 ) -> Option<EncodingResult> {
     let encoding_obj = font_dict.get(b"Encoding").ok()?;
+    let base_font_name = font_dict
+        .get(b"BaseFont")
+        .ok()
+        .and_then(|o| o.as_name().ok())
+        .map(|n| String::from_utf8_lossy(n).to_string());
 
     // Encoding can be a name or a dictionary
     match encoding_obj {
@@ -538,12 +543,14 @@ pub(crate) fn parse_font_encoding(
         Object::Reference(obj_ref) => {
             // Reference to encoding dictionary
             if let Ok(enc_dict) = doc.get_dictionary(*obj_ref) {
-                parse_encoding_dictionary(doc, enc_dict)
+                parse_encoding_dictionary(doc, enc_dict, base_font_name.as_deref())
             } else {
                 None
             }
         }
-        Object::Dictionary(enc_dict) => parse_encoding_dictionary(doc, enc_dict),
+        Object::Dictionary(enc_dict) => {
+            parse_encoding_dictionary(doc, enc_dict, base_font_name.as_deref())
+        }
         _ => None,
     }
 }
@@ -562,6 +569,7 @@ pub(crate) struct EncodingResult {
 pub(crate) fn parse_encoding_dictionary(
     doc: &Document,
     enc_dict: &lopdf::Dictionary,
+    base_font_name: Option<&str>,
 ) -> Option<EncodingResult> {
     let differences = enc_dict.get(b"Differences").ok()?;
 
@@ -591,11 +599,9 @@ pub(crate) fn parse_encoding_dictionary(
             Object::Name(name) => {
                 // Map current code to glyph name -> Unicode
                 let glyph_name = String::from_utf8_lossy(&name).to_string();
-                if glyph_name == "fi"
-                    || glyph_name == "fl"
-                    || glyph_name == "ffi"
-                    || glyph_name == "ffl"
-                {
+                let mapped_char = glyph_to_char(&glyph_name)
+                    .or_else(|| private_glyph_to_char(&glyph_name, base_font_name));
+                if mapped_char.is_some_and(is_ligature_char) {
                     debug!(
                         "  Differences: code=0x{:02X} glyph={:?} (ligature)",
                         current_code, glyph_name
@@ -610,7 +616,7 @@ pub(crate) fn parse_encoding_dictionary(
                 {
                     gid_glyph_count += 1;
                 }
-                if let Some(ch) = glyph_to_char(&glyph_name) {
+                if let Some(ch) = mapped_char {
                     encoding_map.insert(current_code, ch);
                 } else {
                     debug!(
@@ -645,6 +651,31 @@ pub(crate) fn parse_encoding_dictionary(
     })
 }
 
+fn private_glyph_to_char(glyph_name: &str, base_font_name: Option<&str>) -> Option<char> {
+    let base_font_name = strip_subset_prefix(base_font_name?);
+
+    // Aptos CFF subsets from Office PDFs can expose the ff ligature as /g431
+    // without a ToUnicode map. Keep this font-scoped because /gNNN names are private.
+    if base_font_name.eq_ignore_ascii_case("Aptos") && glyph_name == "g431" {
+        Some('\u{FB00}')
+    } else {
+        None
+    }
+}
+
+fn strip_subset_prefix(font_name: &str) -> &str {
+    font_name
+        .split_once('+')
+        .map_or(font_name, |(_, stripped)| stripped)
+}
+
+fn is_ligature_char(ch: char) -> bool {
+    matches!(
+        ch,
+        '\u{FB00}' | '\u{FB01}' | '\u{FB02}' | '\u{FB03}' | '\u{FB04}'
+    )
+}
+
 /// Get the CMap lookup key for an Identity-H/V CID font without ToUnicode.
 /// Returns the object number used by `collect_cmaps_from_fonts` to store the CMap:
 /// - FontFile2 or FontFile3 obj_num (for embedded font cmap)
@@ -720,7 +751,13 @@ pub(crate) fn extract_text_from_operand(
     font_encodings: &PageFontEncodings,
     encoding_cache: &HashMap<String, Encoding<'_>>,
     cmap_decisions: &mut CMapDecisionCache,
+    font_widths: &PageFontWidths,
 ) -> Option<String> {
+    let is_type0_cid_font = font_widths
+        .get(current_font)
+        .is_some_and(|info| info.is_cid);
+    let use_cp1252_fallback =
+        should_use_cp1252_single_byte_fallback(base_font_name, is_type0_cid_font);
     let result = (|| -> Option<String> {
         if let Object::String(bytes, _) = obj {
             let mut decode_with_entry = |entry: &crate::tounicode::CMapEntry| -> Option<String> {
@@ -751,9 +788,12 @@ pub(crate) fn extract_text_from_operand(
                                     return Some(ch.to_string());
                                 }
                             }
-                            // 4. Printable ASCII/Latin-1 fallback
+                            // 4. Printable single-byte fallback
                             if b >= 0x20 {
-                                return Some((b as char).to_string());
+                                return Some(
+                                    decode_single_byte_fallback_char(b, use_cp1252_fallback)
+                                        .to_string(),
+                                );
                             }
                             None
                         })
@@ -854,6 +894,13 @@ pub(crate) fn extract_text_from_operand(
             // unmapped. Don't fall through to text-interpretation fallbacks
             // (Latin-1, UTF-16, etc.) which would misinterpret CID bytes as
             // character codes (e.g. CID 0x01A9 → Latin-1 "©").
+            if is_type0_cid_font && bytes.iter().any(|&b| b > 0x7F) {
+                // 2-byte CIDs (Identity-H) are by far the common case; for
+                // an odd byte count we still emit at least one marker so
+                // detection downstream fires.
+                let cid_count = (bytes.len() / 2).max(1);
+                return Some("\u{FFFD}".repeat(cid_count));
+            }
 
             // Try our custom encoding map from Differences arrays.
             // The Differences array overrides specific codes in a base encoding (typically
@@ -869,8 +916,9 @@ pub(crate) fn extract_text_from_operand(
                                 Some(ch)
                             } else if b >= 0x20 {
                                 // Base encoding fallback for printable bytes.
-                                // For codes 0x20-0x7E this matches all standard PDF encodings.
-                                Some(b as char)
+                                // Most PDFs with simple fonts use WinAnsi/PDFDocEncoding
+                                // semantics, not ISO-8859-1 C1 controls.
+                                Some(decode_single_byte_fallback_char(b, use_cp1252_fallback))
                             } else {
                                 None // Skip unmapped control characters
                             }
@@ -926,6 +974,7 @@ pub(crate) fn extract_text_from_operand(
             // Try to decode using cached font encoding from lopdf
             if let Some(encoding) = encoding_cache.get(current_font) {
                 if let Ok(text) = Document::decode_text(encoding, bytes) {
+                    let text = normalize_cp1252_controls(text, use_cp1252_fallback);
                     if text.contains('\u{FFFD}') {
                         debug!(
                             "decode_text produced replacement for font={} bytes_len={}",
@@ -962,13 +1011,119 @@ pub(crate) fn extract_text_from_operand(
                 return Some(symbol_text);
             }
 
-            // Latin-1 fallback
-            Some(bytes.iter().map(|&b| b as char).collect())
+            // Non-CID (Type1 / TrueType / Type3) fonts use single-byte
+            // encodings. In practice the fallback should follow WinAnsi for
+            // 0x80..=0x9F so bytes like 0x92 become smart punctuation instead
+            // of C1 controls that look like CID mojibake.
+            Some(decode_single_byte_fallback(bytes, use_cp1252_fallback))
         } else {
             None
         }
     })();
-    result.map(clean_symbol_pua)
+    result.map(|text| {
+        let text = clean_symbol_pua(text);
+        normalize_cp1252_controls(text, use_cp1252_fallback)
+    })
+}
+
+fn decode_single_byte_fallback(bytes: &[u8], use_cp1252_fallback: bool) -> String {
+    bytes
+        .iter()
+        .map(|&b| decode_single_byte_fallback_char(b, use_cp1252_fallback))
+        .collect()
+}
+
+fn decode_single_byte_fallback_char(byte: u8, use_cp1252_fallback: bool) -> char {
+    if !use_cp1252_fallback {
+        return byte as char;
+    }
+
+    match byte {
+        0x80 => '\u{20AC}',
+        0x82 => '\u{201A}',
+        0x83 => '\u{0192}',
+        0x84 => '\u{201E}',
+        0x85 => '\u{2026}',
+        0x86 => '\u{2020}',
+        0x87 => '\u{2021}',
+        0x88 => '\u{02C6}',
+        0x89 => '\u{2030}',
+        0x8A => '\u{0160}',
+        0x8B => '\u{2039}',
+        0x8C => '\u{0152}',
+        0x8E => '\u{017D}',
+        0x91 => '\u{2018}',
+        0x92 => '\u{2019}',
+        0x93 => '\u{201C}',
+        0x94 => '\u{201D}',
+        0x95 => '\u{2022}',
+        0x96 => '\u{2013}',
+        0x97 => '\u{2014}',
+        0x98 => '\u{02DC}',
+        0x99 => '\u{2122}',
+        0x9A => '\u{0161}',
+        0x9B => '\u{203A}',
+        0x9C => '\u{0153}',
+        0x9E => '\u{017E}',
+        0x9F => '\u{0178}',
+        _ => byte as char,
+    }
+}
+
+fn normalize_cp1252_controls(text: String, use_cp1252_fallback: bool) -> String {
+    if !use_cp1252_fallback {
+        return text;
+    }
+    if !text
+        .chars()
+        .any(|ch| ('\u{0080}'..='\u{009F}').contains(&ch))
+    {
+        return text;
+    }
+
+    text.chars()
+        .map(|ch| {
+            if ('\u{0080}'..='\u{009F}').contains(&ch) {
+                decode_single_byte_fallback_char(ch as u8, true)
+            } else {
+                ch
+            }
+        })
+        .collect()
+}
+
+fn should_use_cp1252_single_byte_fallback(
+    base_font_name: Option<&str>,
+    is_type0_cid_font: bool,
+) -> bool {
+    if is_type0_cid_font {
+        return false;
+    }
+
+    let Some(base_font_name) = base_font_name else {
+        return true;
+    };
+    let font_name = base_font_name
+        .rsplit_once('+')
+        .map_or(base_font_name, |(_, stripped)| stripped)
+        .to_ascii_lowercase();
+
+    // TeX/Computer Modern and math/symbol fonts often place ligatures or
+    // symbols in the C1 byte range. Treating those bytes as Windows-1252 makes
+    // words like "deficiente" become "de…ciente" and "fluid" become "‡uid".
+    let non_cp1252_prefixes = [
+        "cmr", "cmb", "cmmi", "cmsy", "cmex", "cmtt", "cmss", "cmti", "ecrm", "ecbx", "ecti",
+        "tcrm", "tctt", "msam", "msbm", "ttdc",
+    ];
+    if non_cp1252_prefixes
+        .iter()
+        .any(|prefix| font_name.starts_with(prefix))
+    {
+        return false;
+    }
+
+    let non_cp1252_names = ["math", "symbol", "dingbat", "emoji"];
+    !non_cp1252_names.iter().any(|name| font_name.contains(name))
 }
 
 /// Replace PUA characters in the F000-F0FF range with standard Unicode equivalents.
@@ -1090,6 +1245,7 @@ fn score_text(text: &str) -> i32 {
 #[cfg(test)]
 mod tests {
     use super::*;
+    use lopdf::dictionary;
 
     fn make_font_info(widths: &[(u16, u16)], default_width: u16, is_cid: bool) -> FontWidthInfo {
         FontWidthInfo {
@@ -1213,4 +1369,200 @@ mod tests {
         let bad = "###!!!@@@$$$";
         assert!(score_text(good) > score_text(bad));
     }
+
+    fn doc_with_private_differences() -> (Document, lopdf::ObjectId) {
+        let mut doc = Document::with_version("1.7");
+        let encoding_id = doc.add_object(dictionary! {
+            "Differences" => Object::Array(vec![
+                Object::Integer(0x88),
+                Object::Name(b"g431".to_vec()),
+                Object::Name(b"fi".to_vec()),
+                Object::Integer(0xAD),
+                Object::Name(b"fl".to_vec()),
+            ]),
+        });
+
+        (doc, encoding_id)
+    }
+
+    #[test]
+    fn aptos_private_g431_maps_to_ff_ligature() {
+        let (doc, encoding_id) = doc_with_private_differences();
+        let font_dict = dictionary! {
+            "BaseFont" => Object::Name(b"NJEQOD+Aptos".to_vec()),
+            "Encoding" => Object::Reference(encoding_id),
+        };
+
+        let result = parse_font_encoding(&doc, &font_dict).expect("encoding should parse");
+
+        assert_eq!(result.map.get(&0x88u8), Some(&'\u{FB00}'));
+        assert_eq!(result.map.get(&0x89u8), Some(&'\u{FB01}'));
+        assert_eq!(result.map.get(&0xADu8), Some(&'\u{FB02}'));
+    }
+
+    #[test]
+    fn private_g431_does_not_map_for_unrelated_fonts() {
+        let (doc, encoding_id) = doc_with_private_differences();
+        let font_dict = dictionary! {
+            "BaseFont" => Object::Name(b"ABCDEF+OtherFont".to_vec()),
+            "Encoding" => Object::Reference(encoding_id),
+        };
+
+        let result = parse_font_encoding(&doc, &font_dict).expect("encoding should parse");
+
+        assert!(!result.map.contains_key(&0x88u8));
+        assert_eq!(result.map.get(&0x89u8), Some(&'\u{FB01}'));
+        assert_eq!(result.map.get(&0xADu8), Some(&'\u{FB02}'));
+    }
+
+    #[test]
+    fn cid_font_with_unparseable_cmap_does_not_emit_latin1_mojibake() {
+        // Type0/CID font (font_widths reports `is_cid=true`) where the
+        // ToUnicode CMap couldn't be parsed (FontCMaps doesn't have the
+        // obj_num). Bytes are a 2-byte CID stream containing high bytes
+        // that aren't valid UTF-8 — exactly the case in the production
+        // samples (Identity-H text where the ToUnicode CMap was missing
+        // or malformed, scrape_id 019de78c-..., e.g. "Í Ù Z)¿").
+        //
+        // Without the guard, the function falls through to the byte-by-byte
+        // Latin-1 fallback and produces "ÍÙ" (U+00CD U+00D9). The correct
+        // behavior is to emit U+FFFD per CID so downstream
+        // `detect_encoding_issues` flags the page for OCR.
+        let bytes = vec![0xCD_u8, 0xD9, 0xCD, 0xD9];
+        let obj = Object::String(bytes, lopdf::StringFormat::Hexadecimal);
+
+        let font_cmaps = FontCMaps::default();
+        let mut font_tounicode_refs: HashMap<String, u32> = HashMap::new();
+        font_tounicode_refs.insert("F0".to_string(), 999);
+        let inline_cmaps = HashMap::new();
+        let font_encodings: PageFontEncodings = HashMap::new();
+        let encoding_cache: HashMap<String, Encoding<'_>> = HashMap::new();
+        let mut decisions = CMapDecisionCache::new();
+        let mut font_widths: PageFontWidths = HashMap::new();
+        font_widths.insert("F0".to_string(), make_font_info(&[], 1000, true));
+
+        let result = extract_text_from_operand(
+            &obj,
+            "F0",
+            None,
+            &font_cmaps,
+            &font_tounicode_refs,
+            &inline_cmaps,
+            &font_encodings,
+            &encoding_cache,
+            &mut decisions,
+            &font_widths,
+        );
+
+        let text = result.expect("CID font fallback should still emit a marker");
+        assert!(
+            !text.contains('\u{00CD}') && !text.contains('\u{00D9}'),
+            "CID font with unparseable CMap leaked Latin-1 mojibake: {text:?}"
+        );
+        assert!(
+            text.contains('\u{FFFD}'),
+            "CID font with unparseable CMap should emit U+FFFD so detect_encoding_issues fires: {text:?}"
+        );
+    }
+
+    #[test]
+    fn simple_font_single_byte_fallback_passes_high_bytes_through() {
+        // A Type1/TrueType simple font (is_cid=false) with a `/ToUnicode`
+        // reference but no usable CMap and no `/Differences` map.
+        // Per-byte fallback is the canonical interpretation here — these
+        // bytes are character codes, not CIDs. The CID guard must NOT strip
+        // them. Reproduces the false positive that an earlier version of the
+        // guard introduced for fonts in PDFs like pdf-evals/Navigating-
+        // Artificial-Intelligence-..., where bytes like 0xB6 are legitimate
+        // single-byte character codes.
+        let bytes = vec![0x24_u8, 0x47, 0xB6, 0x56]; // "$G¶V"
+        let obj = Object::String(bytes, lopdf::StringFormat::Hexadecimal);
+
+        let font_cmaps = FontCMaps::default();
+        let mut font_tounicode_refs: HashMap<String, u32> = HashMap::new();
+        font_tounicode_refs.insert("F1".to_string(), 999);
+        let inline_cmaps = HashMap::new();
+        let font_encodings: PageFontEncodings = HashMap::new();
+        let encoding_cache: HashMap<String, Encoding<'_>> = HashMap::new();
+        let mut decisions = CMapDecisionCache::new();
+        let mut font_widths: PageFontWidths = HashMap::new();
+        font_widths.insert("F1".to_string(), make_font_info(&[], 1000, false));
+
+        let text = extract_text_from_operand(
+            &obj,
+            "F1",
+            None,
+            &font_cmaps,
+            &font_tounicode_refs,
+            &inline_cmaps,
+            &font_encodings,
+            &encoding_cache,
+            &mut decisions,
+            &font_widths,
+        )
+        .expect("simple font should round-trip Latin-1 bytes");
+        assert_eq!(text, "$G\u{00B6}V");
+        assert!(
+            !text.contains('\u{FFFD}'),
+            "simple font fallback must not stamp FFFD over legitimate bytes: {text:?}"
+        );
+    }
+
+    #[test]
+    fn simple_font_single_byte_fallback_maps_cp1252_punctuation() {
+        let bytes = vec![b'l', 0x92_u8, b'a', b'c', b'a', b'd'];
+        let obj = Object::String(bytes, lopdf::StringFormat::Hexadecimal);
+
+        let font_cmaps = FontCMaps::default();
+        let font_tounicode_refs: HashMap<String, u32> = HashMap::new();
+        let inline_cmaps = HashMap::new();
+        let font_encodings: PageFontEncodings = HashMap::new();
+        let encoding_cache: HashMap<String, Encoding<'_>> = HashMap::new();
+        let mut decisions = CMapDecisionCache::new();
+        let font_widths: PageFontWidths = HashMap::new();
+
+        let text = extract_text_from_operand(
+            &obj,
+            "F1",
+            None,
+            &font_cmaps,
+            &font_tounicode_refs,
+            &inline_cmaps,
+            &font_encodings,
+            &encoding_cache,
+            &mut decisions,
+            &font_widths,
+        )
+        .expect("simple font should decode CP1252 punctuation");
+
+        assert_eq!(text, "l’acad");
+    }
+
+    #[test]
+    fn cached_encoding_decode_normalizes_cp1252_controls() {
+        let text = normalize_cp1252_controls("d\u{92}un \u{96} test".to_string(), true);
+        assert_eq!(text, "d’un – test");
+    }
+
+    #[test]
+    fn tex_font_decode_keeps_c1_ligature_bytes_unmodified() {
+        let text = normalize_cp1252_controls("de\u{85}ciente \u{87}uid".to_string(), false);
+        assert_eq!(text, "de\u{85}ciente \u{87}uid");
+        assert!(!should_use_cp1252_single_byte_fallback(
+            Some("TTdcr10"),
+            false
+        ));
+        assert!(!should_use_cp1252_single_byte_fallback(
+            Some("cmr10"),
+            false
+        ));
+    }
+
+    #[test]
+    fn winansi_text_font_uses_cp1252_fallback() {
+        assert!(should_use_cp1252_single_byte_fallback(
+            Some("BJPQNQ+Times-Roman"),
+            false
+        ));
+    }
 }

src/extractor/layout.rs

@@ -29,8 +29,12 @@ pub(crate) fn detect_columns(
     const MIN_ITEMS_PER_COLUMN: usize = 10;
     const NOISE_FRACTION: f32 = 0.15;
 
-    // Get items for this page
-    let page_items: Vec<&TextItem> = items.iter().filter(|i| i.page == page).collect();
+    // Get items for this page. Strip Image placeholders — an image's left edge
+    // would otherwise count toward the column projection profile.
+    let page_items: Vec<&TextItem> = items
+        .iter()
+        .filter(|i| i.page == page && crate::extractor::is_text_layout_item(i))
+        .collect();
 
     if page_items.is_empty() {
         return vec![];
@@ -1230,11 +1234,7 @@ pub(crate) fn group_into_lines_with_thresholds(
                             ci,
                             item.x,
                             item.y,
-                            if item.text.len() > 60 {
-                                &item.text[..60]
-                            } else {
-                                &item.text
-                            }
+                            super::trace_text_preview(&item.text, 60)
                         );
                     }
                 }

src/extractor/mod.rs

@@ -33,6 +33,13 @@ pub(crate) use layout::ColumnRegion;
 // Public API
 // ---------------------------------------------------------------------------
 
+pub(crate) fn trace_text_preview(text: &str, max_chars: usize) -> &str {
+    match text.char_indices().nth(max_chars) {
+        Some((idx, _)) => &text[..idx],
+        None => text,
+    }
+}
+
 /// Extract text from PDF file as plain string
 pub fn extract_text<P: AsRef<Path>>(path: P) -> Result<String, PdfError> {
     crate::validate_pdf_file(&path)?;
@@ -195,11 +202,7 @@ fn extract_positioned_text_impl(
                     item.width,
                     item.font_size,
                     item.font,
-                    if item.text.len() > 80 {
-                        &item.text[..80]
-                    } else {
-                        &item.text
-                    }
+                    trace_text_preview(&item.text, 80)
                 );
             }
         }
@@ -227,6 +230,69 @@ fn extract_positioned_text_impl(
 // Shared helpers (used by submodules via `super::`)
 // ---------------------------------------------------------------------------
 
+/// Return true when this item should participate in text-layout
+/// heuristics (column detection, table grid detection, line grouping).
+///
+/// Image XObjects emit a positional placeholder via
+/// `extract_text_with_positions` (so layout-aware callers can crop +
+/// caption figures), but their bboxes don't carry text glyphs and would
+/// skew column/row clustering if they reached the heuristics. Hyperlinks
+/// and form fields *do* participate — the existing logic treats them as
+/// text-like and we keep that.
+pub(crate) fn is_text_layout_item(item: &crate::types::TextItem) -> bool {
+    !matches!(item.item_type, crate::types::ItemType::Image)
+}
+
+/// Map a (u, v) point in unit-square coordinates through the 6-element CTM
+/// to page-space. CTM format is `[a, b, c, d, e, f]` per
+/// [`multiply_matrices`].
+fn apply_ctm_point(ctm: &[f32; 6], u: f32, v: f32) -> (f32, f32) {
+    (
+        u * ctm[0] + v * ctm[2] + ctm[4],
+        u * ctm[1] + v * ctm[3] + ctm[5],
+    )
+}
+
+/// Compute the page-space axis-aligned bounding box of an Image XObject
+/// invoked under the given CTM.
+///
+/// Per the PDF spec, an image XObject is always rendered into a unit
+/// square `(0,0)–(1,1)` in its local coordinate system, and the `Do`
+/// operator applies the current CTM to position/scale/rotate that square
+/// onto the page. For the common axis-aligned case (no rotation/shear),
+/// the CTM reduces to `[w, 0, 0, h, x, y]` and the bbox is just
+/// `(x, y, w, h)`. For rotated/sheared images we transform all four
+/// corners and return their axis-aligned bbox so the caller always gets
+/// an upright rectangle.
+///
+/// Coordinates are PDF user space (origin at bottom-left, y-up). Width
+/// and height are non-negative.
+pub(crate) fn image_bbox_from_ctm(ctm: &[f32; 6]) -> (f32, f32, f32, f32) {
+    let corners = [
+        apply_ctm_point(ctm, 0.0, 0.0),
+        apply_ctm_point(ctm, 1.0, 0.0),
+        apply_ctm_point(ctm, 1.0, 1.0),
+        apply_ctm_point(ctm, 0.0, 1.0),
+    ];
+    let (mut x_min, mut x_max) = (corners[0].0, corners[0].0);
+    let (mut y_min, mut y_max) = (corners[0].1, corners[0].1);
+    for (cx, cy) in corners.iter().skip(1) {
+        if *cx < x_min {
+            x_min = *cx;
+        }
+        if *cx > x_max {
+            x_max = *cx;
+        }
+        if *cy < y_min {
+            y_min = *cy;
+        }
+        if *cy > y_max {
+            y_max = *cy;
+        }
+    }
+    (x_min, y_min, x_max - x_min, y_max - y_min)
+}
+
 /// Multiply two 2D transformation matrices
 /// Matrix format: [a, b, c, d, e, f] representing:
 /// | a  b  0 |
@@ -286,6 +352,133 @@ fn effective_merge_width(item: &TextItem) -> f32 {
     }
 }
 
+fn is_standalone_bullet_text(text: &str) -> bool {
+    matches!(text.trim(), "•" | "○" | "●" | "◦")
+}
+
+fn first_text_char(text: &str) -> Option<char> {
+    text.trim_start().chars().next()
+}
+
+fn is_short_alpha_fragment(text: &str) -> bool {
+    let trimmed = text.trim();
+    let char_count = trimmed.chars().count();
+    (1..=4).contains(&char_count) && trimmed.chars().all(char::is_alphabetic)
+}
+
+fn has_phrase_continuation_shape(text: &str) -> bool {
+    let trimmed = text.trim_start();
+    trimmed
+        .chars()
+        .take(24)
+        .any(|ch| ch.is_whitespace() || matches!(ch, '-'))
+}
+
+fn should_preserve_overlapping_stream_order(group: &[&TextItem]) -> bool {
+    if group.len() < 3 {
+        return false;
+    }
+
+    let Some(first) = group.iter().find(|item| !item.text.trim().is_empty()) else {
+        return false;
+    };
+    if group.iter().all(|item| item.mcid.is_none()) {
+        return false;
+    }
+
+    let mut nonempty_count = 0;
+    let mut saw_backtrack = false;
+    let mut nonspace_chars = 0;
+    let mut math_symbol_chars = 0;
+    let mut max_font_size = first.font_size;
+
+    for item in group {
+        if !item.text.trim().is_empty() {
+            nonempty_count += 1;
+        }
+        if (item.font_size - first.font_size).abs() > first.font_size * 0.25 {
+            return false;
+        }
+        max_font_size = max_font_size.max(item.font_size);
+        for ch in item.text.chars().filter(|ch| !ch.is_whitespace()) {
+            nonspace_chars += 1;
+            if matches!(
+                ch,
+                '*' | 'ˆ' | '^' | '=' | '+' | '_' | '[' | ']' | '{' | '}' | '|' | '<' | '>'
+            ) {
+                math_symbol_chars += 1;
+            }
+        }
+    }
+
+    if nonempty_count < 2 {
+        return false;
+    }
+    if nonspace_chars > 0 && math_symbol_chars * 4 > nonspace_chars {
+        return false;
+    }
+
+    let mut sorted_by_x = group.to_vec();
+    sorted_by_x.sort_by(|a, b| a.x.total_cmp(&b.x));
+    let cluster_start = sorted_by_x[0].x;
+    let mut cluster_end = cluster_start + effective_merge_width(sorted_by_x[0]);
+    for item in sorted_by_x.iter().skip(1) {
+        let gap = item.x - cluster_end;
+        if gap > max_font_size * 2.5 {
+            return false;
+        }
+        cluster_end = cluster_end.max(item.x + effective_merge_width(item));
+    }
+    if cluster_end - cluster_start > max_font_size * 36.0 {
+        return false;
+    }
+
+    for index in 0..group.len() - 1 {
+        let previous = group[index];
+        let next = group[index + 1];
+        let font_size = previous.font_size.max(next.font_size);
+        let backtrack_threshold = font_size * 0.25;
+        let previous_start = previous.x;
+        let next_start = next.x;
+        let next_end = next.x + effective_merge_width(next);
+        if next_start < previous_start - backtrack_threshold
+            && next_end > previous_start + backtrack_threshold
+        {
+            let has_near_prefix = group[..=index].iter().rev().take(4).any(|item| {
+                is_short_alpha_fragment(&item.text)
+                    && item.x >= next_start - font_size * 0.5
+                    && item.x <= next_start + font_size * 4.0
+            });
+            let starts_lowercase = first_text_char(&next.text).is_some_and(char::is_lowercase);
+            let phrase_continuation = has_phrase_continuation_shape(&next.text);
+            let has_near_bullet = group[..=index]
+                .iter()
+                .position(|item| {
+                    is_standalone_bullet_text(&item.text) && next_start <= item.x + font_size * 3.0
+                })
+                .is_some_and(|bullet_index| {
+                    if bullet_index >= index {
+                        return false;
+                    }
+                    group[bullet_index + 1..=index]
+                        .iter()
+                        .rev()
+                        .find(|item| !item.text.trim().is_empty())
+                        .is_some_and(|item| {
+                            item.text.trim().chars().count() <= 8
+                                && has_phrase_continuation_shape(&next.text)
+                        })
+                });
+            if (has_near_prefix && starts_lowercase && phrase_continuation) || has_near_bullet {
+                saw_backtrack = true;
+                break;
+            }
+        }
+    }
+
+    saw_backtrack
+}
+
 pub(crate) fn merge_text_items(items: Vec<TextItem>) -> Vec<TextItem> {
     if items.is_empty() {
         return items;
@@ -306,28 +499,32 @@ pub(crate) fn merge_text_items(items: Vec<TextItem>) -> Vec<TextItem> {
         }
     }
 
-    // Sort each group by X position (direction-aware)
-    for (_, _, group) in &mut line_groups {
+    let mut ordered_line_groups: Vec<(u32, f32, Vec<&TextItem>, bool)> = Vec::new();
+
+    // Sort each group by X position (direction-aware), except for lines whose
+    // content stream intentionally backtracks to overlay ActualText fragments.
+    for (page, y, mut group) in line_groups {
         let rtl = is_rtl_text(group.iter().map(|i| &i.text));
+        let preserve_stream_order = !rtl && should_preserve_overlapping_stream_order(&group);
         if rtl {
             group.sort_by(|a, b| b.x.total_cmp(&a.x));
-        } else {
+        } else if !preserve_stream_order {
             group.sort_by(|a, b| a.x.total_cmp(&b.x));
         }
+        ordered_line_groups.push((page, y, group, preserve_stream_order));
     }
 
     // Sort groups by page then Y descending (top of page first)
-    line_groups.sort_by(|a, b| a.0.cmp(&b.0).then_with(|| b.1.total_cmp(&a.1)));
+    ordered_line_groups.sort_by(|a, b| a.0.cmp(&b.0).then_with(|| b.1.total_cmp(&a.1)));
 
     let mut merged = Vec::new();
 
-    for (_, _, group) in &line_groups {
+    for (_, _, group, preserve_stream_order) in &ordered_line_groups {
         let mut i = 0;
         while i < group.len() {
             let first = group[i];
             let mut text = first.text.clone();
             let mut end_x = first.x + effective_merge_width(first);
-            let x_gap_max = first.font_size * 0.5;
 
             let mut j = i + 1;
             while j < group.len() {
@@ -337,10 +534,15 @@ pub(crate) fn merge_text_items(items: Vec<TextItem>) -> Vec<TextItem> {
                     break;
                 }
                 let gap = next.x - end_x;
+                let x_gap_max = if *preserve_stream_order && is_standalone_bullet_text(&text) {
+                    first.font_size * 1.2
+                } else {
+                    first.font_size * 0.5
+                };
                 if gap > x_gap_max {
                     break;
                 }
-                if gap < -first.font_size * 0.5 {
+                if gap < -first.font_size * 0.5 && !preserve_stream_order {
                     break;
                 }
                 // Insert space at word boundaries.
@@ -362,11 +564,19 @@ pub(crate) fn merge_text_items(items: Vec<TextItem>) -> Vec<TextItem> {
                         first.font_size * 0.08
                     }
                 };
-                if gap > threshold {
+                let needs_bullet_space = *preserve_stream_order
+                    && is_standalone_bullet_text(&text)
+                    && !next.text.trim().is_empty();
+                if needs_bullet_space || gap > threshold {
                     text.push(' ');
                 }
                 text.push_str(&next.text);
-                end_x = next.x + effective_merge_width(next);
+                let next_end = next.x + effective_merge_width(next);
+                end_x = if *preserve_stream_order {
+                    end_x.max(next_end)
+                } else {
+                    next_end
+                };
                 j += 1;
             }
 
@@ -511,6 +721,21 @@ mod tests {
         }
     }
 
+    fn with_mcid(mut item: TextItem) -> TextItem {
+        item.mcid = Some(1);
+        item
+    }
+
+    #[test]
+    fn trace_text_preview_truncates_on_char_boundary() {
+        let text = format!("{}{}tail", "a".repeat(79), '\u{FFFD}');
+        let preview = trace_text_preview(&text, 80);
+
+        assert_eq!(preview.chars().count(), 80);
+        assert!(text.is_char_boundary(preview.len()));
+        assert!(preview.ends_with('\u{FFFD}'));
+    }
+
     #[test]
     fn merge_items_no_space_before_period() {
         // Simulate Tc/Tw-adjusted width: "date" width is smaller than the gap
@@ -549,6 +774,83 @@ mod tests {
         assert_eq!(merged[0].text, "hello world");
     }
 
+    #[test]
+    fn merge_items_preserves_stream_order_for_backtracking_heading() {
+        // Some tagged PDFs emit first-letter ActualText fragments, then reset
+        // the text matrix and draw the rest of the word from the line start.
+        let items = vec![
+            with_mcid(make_merge_item("F", 79.4, 4.5)),
+            with_mcid(make_merge_item("r", 83.9, 3.3)),
+            with_mcid(make_merge_item("om tables to data-", 79.4, 89.7)),
+            with_mcid(make_merge_item("", 168.9, 33.9)),
+            with_mcid(make_merge_item("analytics-", 168.9, 75.5)),
+            with_mcid(make_merge_item("ready content", 210.5, 60.8)),
+        ];
+
+        let merged = merge_text_items(items);
+
+        assert_eq!(merged.len(), 1);
+        assert_eq!(
+            merged[0].text,
+            "From tables to data-analytics-ready content"
+        );
+    }
+
+    #[test]
+    fn merge_items_preserves_stream_order_for_reset_word_prefix() {
+        let items = vec![
+            with_mcid(make_merge_item("N", 68.0, 7.0)),
+            with_mcid(make_merge_item("e", 75.1, 4.0)),
+            with_mcid(make_merge_item("w fields created", 68.0, 82.0)),
+        ];
+
+        let merged = merge_text_items(items);
+
+        assert_eq!(merged.len(), 1);
+        assert_eq!(merged[0].text, "New fields created");
+    }
+
+    #[test]
+    fn merge_items_uses_x_order_for_untagged_backtracking_text() {
+        let items = vec![
+            make_merge_item("N", 68.0, 7.0),
+            make_merge_item("e", 75.1, 4.0),
+            make_merge_item("w fields created", 68.2, 82.0),
+        ];
+
+        let merged = merge_text_items(items);
+
+        let texts: Vec<_> = merged.iter().map(|item| item.text.as_str()).collect();
+        assert_eq!(texts, vec!["N", "w fields created", "e"]);
+    }
+
+    #[test]
+    fn merge_items_preserves_bullet_stream_order_with_backtracking() {
+        let items = vec![
+            with_mcid(make_merge_item("•", 79.4, 5.0)),
+            with_mcid(make_merge_item("The MS", 91.0, 32.6)),
+            with_mcid(make_merge_item("A LoS project", 84.4, 70.0)),
+        ];
+
+        let merged = merge_text_items(items);
+
+        assert_eq!(merged.len(), 1);
+        assert_eq!(merged[0].text, "• The MSA LoS project");
+    }
+
+    #[test]
+    fn merge_items_keeps_normal_bullet_gap_limit_without_stream_order() {
+        let items = vec![
+            make_merge_item("•", 79.4, 5.0),
+            make_merge_item("Distant item", 91.0, 60.0),
+        ];
+
+        let merged = merge_text_items(items);
+
+        let texts: Vec<_> = merged.iter().map(|item| item.text.as_str()).collect();
+        assert_eq!(texts, vec!["•", "Distant item"]);
+    }
+
     #[test]
     fn test_group_into_lines() {
         let items = vec![

src/extractor/xobjects.rs

@@ -10,7 +10,7 @@ use super::fonts::{
     build_font_encodings, build_font_widths, compute_string_width_ts, extract_text_from_operand,
     get_font_file2_obj_num, get_operand_bytes, CMapDecisionCache,
 };
-use super::{get_number, multiply_matrices};
+use super::{get_number, image_bbox_from_ctm, multiply_matrices};
 
 const MAX_FORM_XOBJECT_DEPTH: u8 = 5;
 
@@ -262,19 +262,43 @@ fn extract_form_xobject_text_inner(
                 if !op.operands.is_empty() {
                     if let Ok(name) = op.operands[0].as_name() {
                         let xobj_name = String::from_utf8_lossy(name).to_string();
-                        if let Some(XObjectType::Form(nested_id)) = form_xobjects.get(&xobj_name) {
-                            if depth < MAX_FORM_XOBJECT_DEPTH {
-                                let nested_items = extract_form_xobject_text_inner(
-                                    doc,
-                                    *nested_id,
-                                    page_num,
-                                    font_cmaps,
-                                    &ctm,
-                                    cmap_decisions,
-                                    depth + 1,
-                                );
-                                items.extend(nested_items);
+                        match form_xobjects.get(&xobj_name) {
+                            Some(XObjectType::Form(nested_id)) => {
+                                if depth < MAX_FORM_XOBJECT_DEPTH {
+                                    let nested_items = extract_form_xobject_text_inner(
+                                        doc,
+                                        *nested_id,
+                                        page_num,
+                                        font_cmaps,
+                                        &ctm,
+                                        cmap_decisions,
+                                        depth + 1,
+                                    );
+                                    items.extend(nested_items);
+                                }
                             }
+                            Some(XObjectType::Image) => {
+                                // Mirror the top-level Image-XObject emission
+                                // in content_stream.rs so figures embedded
+                                // inside Form XObjects (common in print-to-PDF
+                                // workflows) aren't silently dropped.
+                                let (x, y, width, height) = image_bbox_from_ctm(&ctm);
+                                items.push(TextItem {
+                                    text: format!("[Image: {}]", xobj_name),
+                                    x,
+                                    y,
+                                    width,
+                                    height,
+                                    font: String::new(),
+                                    font_size: 0.0,
+                                    page: page_num,
+                                    is_bold: false,
+                                    is_italic: false,
+                                    item_type: ItemType::Image,
+                                    mcid: None,
+                                });
+                            }
+                            None => {}
                         }
                     }
                 }
@@ -373,6 +397,7 @@ fn extract_form_xobject_text_inner(
                         &font_encodings,
                         &encoding_cache,
                         cmap_decisions,
+                        &font_widths,
                     ) {
                         let combined = multiply_matrices(&text_matrix, &ctm);
                         let rendered_size = effective_font_size(current_font_size, &combined);
@@ -517,6 +542,7 @@ fn extract_form_xobject_text_inner(
                                     &font_encodings,
                                     &encoding_cache,
                                     cmap_decisions,
+                                    &font_widths,
                                 ) {
                                     current_text.push_str(&text);
                                 }

src/markdown/convert.rs

@@ -149,6 +149,79 @@ fn find_isolated_lines(lines: &[TextLine], base_size: f32, para_threshold: f32)
     set
 }
 
+/// Pre-scan body-size all-bold runs that are too long to be headings.
+///
+/// Some academic PDFs use an all-bold abstract/summary paragraph immediately
+/// after the author block.  A line-local bold heading heuristic sees each
+/// wrapped visual line as "standalone" once the first line is misclassified,
+/// producing a stack of `##` headings.  Multi-line body-size bold runs with a
+/// paragraph-sized word count should stay paragraph text.
+fn find_wrapped_bold_paragraph_lines(
+    lines: &[TextLine],
+    base_size: f32,
+    para_threshold: f32,
+) -> HashSet<usize> {
+    let mut set = HashSet::new();
+    let mut i = 0usize;
+
+    while i < lines.len() {
+        if !is_body_size_all_bold_line(&lines[i], base_size) {
+            i += 1;
+            continue;
+        }
+
+        let start = i;
+        let mut end = i;
+        let mut word_count = lines[i].text().split_whitespace().count();
+
+        while end + 1 < lines.len()
+            && is_body_size_all_bold_line(&lines[end + 1], base_size)
+            && is_wrapped_same_style_line(&lines[end], &lines[end + 1], para_threshold)
+        {
+            end += 1;
+            word_count += lines[end].text().split_whitespace().count();
+        }
+
+        let line_count = end - start + 1;
+        if line_count >= 3 && word_count > 20 {
+            for idx in start..=end {
+                set.insert(idx);
+            }
+        }
+
+        i = end + 1;
+    }
+
+    set
+}
+
+fn is_body_size_all_bold_line(line: &TextLine, base_size: f32) -> bool {
+    let Some(first) = line.items.first() else {
+        return false;
+    };
+    first.font_size >= base_size * 0.95
+        && first.font_size < base_size * 1.2
+        && line
+            .items
+            .iter()
+            .all(|item| item.is_bold && (item.font_size - first.font_size).abs() < 0.5)
+}
+
+fn is_wrapped_same_style_line(prev: &TextLine, next: &TextLine, para_threshold: f32) -> bool {
+    if prev.page != next.page {
+        return false;
+    }
+
+    let y_gap = prev.y - next.y;
+    if !(y_gap > 0.0 && y_gap <= para_threshold) {
+        return false;
+    }
+
+    let prev_x = prev.items.first().map(|item| item.x).unwrap_or(0.0);
+    let next_x = next.items.first().map(|item| item.x).unwrap_or(0.0);
+    (prev_x - next_x).abs() <= 40.0
+}
+
 /// Resolve the dominant structure role for a text line by looking up its items' MCIDs.
 ///
 /// Returns the first non-container role found (skipping Document/Part/Sect/Div/NonStruct/Span).
@@ -397,6 +470,8 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
     // between paragraphs at body font size. Inspired by opendataloader's
     // lookahead in HeadingProcessor (prevNode/nextNode context).
     let isolated_lines = find_isolated_lines(&lines, base_size, para_threshold);
+    let wrapped_bold_paragraph_lines =
+        find_wrapped_bold_paragraph_lines(&lines, base_size, para_threshold);
 
     // Detect struct heading levels that are overused (body text mistagged as headings)
     let overused_heading_levels = detect_overused_struct_heading_levels(&lines, struct_roles);
@@ -410,6 +485,7 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
     let mut last_list_x: Option<f32> = None;
     let mut in_code_block = false;
     let mut prev_had_dot_leaders = false;
+    let mut paragraph_in_wrapped_bold_run = false;
     let mut inserted_tables: HashSet<(u32, usize)> = HashSet::new();
     let mut inserted_images: HashSet<(u32, usize)> = HashSet::new();
 
@@ -475,6 +551,7 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
             current_page = line.page;
             prev_y = f32::MAX;
             prev_x = 0.0;
+            paragraph_in_wrapped_bold_run = false;
 
             if options.include_page_numbers {
                 output.push_str(&format!("<!-- Page {} -->\n\n", current_page));
@@ -489,6 +566,7 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
                     if in_paragraph {
                         output.push_str("\n\n");
                         in_paragraph = false;
+                        paragraph_in_wrapped_bold_run = false;
                     }
                     output.push('\n');
                     output.push_str(table_md);
@@ -506,6 +584,7 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
                     if in_paragraph {
                         output.push_str("\n\n");
                         in_paragraph = false;
+                        paragraph_in_wrapped_bold_run = false;
                     }
                     output.push('\n');
                     output.push_str(image_md);
@@ -527,9 +606,18 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
             && y_gap.abs() <= para_threshold
             && (prev_x - line_x).abs() > 50.0
             && prev_y < f32::MAX;
-        if (is_para_break || is_band_switch) && in_paragraph {
+        let line_all_bold = !line.items.is_empty() && line.items.iter().all(|item| item.is_bold);
+        let line_in_wrapped_bold_run = wrapped_bold_paragraph_lines.contains(&line_idx);
+        let is_bold_to_regular_break = in_paragraph
+            && paragraph_in_wrapped_bold_run
+            && !line_in_wrapped_bold_run
+            && !line_all_bold
+            && y_gap > base_size * 1.2
+            && y_gap <= para_threshold;
+        if (is_para_break || is_band_switch || is_bold_to_regular_break) && in_paragraph {
             output.push_str("\n\n");
             in_paragraph = false;
+            paragraph_in_wrapped_bold_run = false;
         }
         // Don't immediately end list on paragraph break
         // Let the continuation check below decide if we're still in a list
@@ -572,6 +660,7 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
             if in_paragraph {
                 output.push_str("\n\n");
                 in_paragraph = false;
+                paragraph_in_wrapped_bold_run = false;
             }
             output.push_str(trimmed);
             output.push_str("\n\n");
@@ -625,6 +714,9 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
                 if !(1..=15).contains(&word_count) {
                     return None;
                 }
+                if wrapped_bold_paragraph_lines.contains(&line_idx) {
+                    return None;
+                }
                 let rarity = font_size_rarity(line_font_size, &font_stats);
                 let all_bold = !line.items.is_empty() && line.items.iter().all(|i| i.is_bold);
                 let standalone = !in_paragraph;
@@ -656,6 +748,7 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
             if in_paragraph {
                 output.push_str("\n\n");
                 in_paragraph = false;
+                paragraph_in_wrapped_bold_run = false;
             }
             let prefix = "#".repeat(level);
             // Use plain text for headers to avoid redundant formatting
@@ -678,6 +771,7 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
             if in_paragraph {
                 output.push_str("\n\n");
                 in_paragraph = false;
+                paragraph_in_wrapped_bold_run = false;
             }
             output.push_str(&format!("- {}", trimmed));
             output.push('\n');
@@ -691,6 +785,7 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
             if in_paragraph {
                 output.push_str("\n\n");
                 in_paragraph = false;
+                paragraph_in_wrapped_bold_run = false;
             }
             let formatted = format_list_item(trimmed);
             output.push_str(&formatted);
@@ -737,6 +832,7 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
             if in_paragraph {
                 output.push_str("\n\n");
                 in_paragraph = false;
+                paragraph_in_wrapped_bold_run = false;
             }
             output.push_str(&format!("> {}\n", trimmed));
             continue;
@@ -747,6 +843,7 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
             if in_paragraph {
                 output.push_str("\n\n");
                 in_paragraph = false;
+                paragraph_in_wrapped_bold_run = false;
             }
             if !in_code_block {
                 output.push_str("```\n");
@@ -767,6 +864,11 @@ pub(super) fn to_markdown_from_lines_with_tables_and_images(
             }
         }
         output.push_str(trimmed);
+        paragraph_in_wrapped_bold_run = if in_paragraph {
+            paragraph_in_wrapped_bold_run || line_in_wrapped_bold_run
+        } else {
+            line_in_wrapped_bold_run
+        };
         in_paragraph = true;
         prev_had_dot_leaders = cur_dot_leaders;
     }
@@ -836,6 +938,8 @@ pub fn to_markdown_from_lines(lines: Vec<TextLine>, options: MarkdownOptions) ->
     let para_threshold = compute_paragraph_threshold(&lines, base_size);
 
     let isolated_lines = find_isolated_lines(&lines, base_size, para_threshold);
+    let wrapped_bold_paragraph_lines =
+        find_wrapped_bold_paragraph_lines(&lines, base_size, para_threshold);
 
     let mut output = String::new();
     let mut current_page = 0u32;
@@ -844,6 +948,7 @@ pub fn to_markdown_from_lines(lines: Vec<TextLine>, options: MarkdownOptions) ->
     let mut in_paragraph = false;
     let mut last_list_x: Option<f32> = None;
     let mut prev_had_dot_leaders = false;
+    let mut paragraph_in_wrapped_bold_run = false;
 
     for (line_idx, line) in lines.iter().enumerate() {
         // Page break
@@ -860,6 +965,7 @@ pub fn to_markdown_from_lines(lines: Vec<TextLine>, options: MarkdownOptions) ->
             in_list = false;
             last_list_x = None;
             prev_had_dot_leaders = false;
+            paragraph_in_wrapped_bold_run = false;
 
             if options.include_page_numbers {
                 output.push_str(&format!("<!-- Page {} -->\n\n", current_page));
@@ -870,9 +976,18 @@ pub fn to_markdown_from_lines(lines: Vec<TextLine>, options: MarkdownOptions) ->
         // (newspaper columns emitted sequentially on the same page).
         let y_gap = prev_y - line.y;
         let is_para_break = y_gap.abs() > para_threshold;
-        if is_para_break && in_paragraph {
+        let line_all_bold = !line.items.is_empty() && line.items.iter().all(|item| item.is_bold);
+        let line_in_wrapped_bold_run = wrapped_bold_paragraph_lines.contains(&line_idx);
+        let is_bold_to_regular_break = in_paragraph
+            && paragraph_in_wrapped_bold_run
+            && !line_in_wrapped_bold_run
+            && !line_all_bold
+            && y_gap > base_size * 1.2
+            && y_gap <= para_threshold;
+        if (is_para_break || is_bold_to_regular_break) && in_paragraph {
             output.push_str("\n\n");
             in_paragraph = false;
+            paragraph_in_wrapped_bold_run = false;
         }
         // Don't immediately end list on paragraph break
         // Let the continuation check below decide if we're still in a list
@@ -896,6 +1011,7 @@ pub fn to_markdown_from_lines(lines: Vec<TextLine>, options: MarkdownOptions) ->
             if in_paragraph {
                 output.push_str("\n\n");
                 in_paragraph = false;
+                paragraph_in_wrapped_bold_run = false;
             }
             output.push_str(trimmed);
             output.push_str("\n\n");
@@ -918,6 +1034,9 @@ pub fn to_markdown_from_lines(lines: Vec<TextLine>, options: MarkdownOptions) ->
                     if !(1..=15).contains(&word_count) {
                         return None;
                     }
+                    if wrapped_bold_paragraph_lines.contains(&line_idx) {
+                        return None;
+                    }
                     let rarity = font_size_rarity(line_font_size, &font_stats);
                     let all_bold = !line.items.is_empty() && line.items.iter().all(|i| i.is_bold);
                     let standalone = !in_paragraph;
@@ -935,6 +1054,7 @@ pub fn to_markdown_from_lines(lines: Vec<TextLine>, options: MarkdownOptions) ->
                 if in_paragraph {
                     output.push_str("\n\n");
                     in_paragraph = false;
+                    paragraph_in_wrapped_bold_run = false;
                 }
                 let prefix = "#".repeat(header_level);
                 // Use plain text for headers to avoid redundant formatting
@@ -949,6 +1069,7 @@ pub fn to_markdown_from_lines(lines: Vec<TextLine>, options: MarkdownOptions) ->
             if in_paragraph {
                 output.push_str("\n\n");
                 in_paragraph = false;
+                paragraph_in_wrapped_bold_run = false;
             }
             let formatted = format_list_item(trimmed);
             output.push_str(&formatted);
@@ -993,6 +1114,7 @@ pub fn to_markdown_from_lines(lines: Vec<TextLine>, options: MarkdownOptions) ->
                 if in_paragraph {
                     output.push_str("\n\n");
                     in_paragraph = false;
+                    paragraph_in_wrapped_bold_run = false;
                 }
                 // Use plain text for code blocks
                 output.push_str(&format!("```\n{}\n```\n", plain_trimmed));
@@ -1010,6 +1132,11 @@ pub fn to_markdown_from_lines(lines: Vec<TextLine>, options: MarkdownOptions) ->
             }
         }
         output.push_str(trimmed);
+        paragraph_in_wrapped_bold_run = if in_paragraph {
+            paragraph_in_wrapped_bold_run || line_in_wrapped_bold_run
+        } else {
+            line_in_wrapped_bold_run
+        };
         in_paragraph = true;
         prev_had_dot_leaders = cur_dot_leaders;
     }
@@ -1356,6 +1483,109 @@ mod tests {
         );
     }
 
+    #[test]
+    fn test_wrapped_bold_abstract_is_not_split_into_headings() {
+        // Regression for arXiv 1107.1353: the opening abstract paragraph is
+        // entirely bold at body size. The first wrapped lines used to become
+        // separate H2 headings, and the following body paragraph was joined to
+        // the bold abstract because the paragraph gap is modest.
+        let make = |text: &str, y: f32, font_size: f32, bold: bool| {
+            let mut item = make_item(text, 1, None);
+            item.y = y;
+            item.font_size = font_size;
+            item.height = font_size;
+            item.is_bold = bold;
+            item
+        };
+
+        let lines = vec![
+            make_line(vec![make(
+                "Quantum Nature of Light Measured With a Single Detector",
+                747.7,
+                25.0,
+                true,
+            )]),
+            make_line(vec![make(
+                "Gesine A. Steudle1*, Stefan Schietinger1, David Höckel1",
+                651.1,
+                11.0,
+                false,
+            )]),
+            make_line(vec![make(
+                "Zwiller2, and Oliver Benson1",
+                638.5,
+                11.0,
+                false,
+            )]),
+            make_line(vec![make(
+                "The introduction of light quanta by Einstein in 1905 triggered strong efforts to",
+                607.5,
+                11.0,
+                true,
+            )]),
+            make_line(vec![make(
+                "demonstrate the quantum properties of light directly, without involving matter",
+                594.8,
+                11.0,
+                true,
+            )]),
+            make_line(vec![make(
+                "quantization. It however took more than seven decades for the quantum granularity",
+                582.2,
+                11.0,
+                true,
+            )]),
+            make_line(vec![make(
+                "of light to be observed in the fluorescence of single atoms. Single atoms emit",
+                569.5,
+                11.0,
+                true,
+            )]),
+            make_line(vec![make(
+                "photons one at a time, this is typically demonstrated with a Hanbury-Brown-Twiss",
+                556.9,
+                11.0,
+                true,
+            )]),
+            make_line(vec![make(
+                "Our work significantly simplifies a widely used photon-correlation technique.",
+                544.2,
+                11.0,
+                true,
+            )]),
+            make_line(vec![make(
+                "A photon is a single excitation of a mode of the electromagnetic field.",
+                528.7,
+                11.0,
+                false,
+            )]),
+        ];
+
+        let md = to_markdown_from_lines_with_tables_and_images(
+            lines,
+            MarkdownOptions::default(),
+            HashMap::new(),
+            HashMap::new(),
+            &std::collections::HashSet::new(),
+            None,
+        );
+
+        assert!(
+            md.contains("# Quantum Nature of Light Measured With a Single Detector"),
+            "title should remain a heading: {md}"
+        );
+        assert!(
+            !md.contains("## The introduction")
+                && !md.contains("## demonstrate")
+                && !md.contains("## quantization"),
+            "bold abstract lines should not become headings: {md}"
+        );
+        assert!(
+            md.contains("technique.**\n\nA photon is a single excitation"),
+            "body paragraph should be separated from bold abstract: {md}"
+        );
+    }
+
     #[test]
     fn test_struct_role_code_multiline_accumulation() {
         let mut line1 = make_item("fn main() {", 1, Some(0));

src/markdown/mod.rs

@@ -422,7 +422,16 @@ impl Default for MarkdownOptions {
             fix_hyphenation: true,
             detect_bold: true,
             detect_italic: true,
-            include_images: true,
+            // `include_images: false` is intentional. The content-stream walker
+            // now emits `ItemType::Image` `TextItem`s for every Image XObject
+            // it encounters (see `extractor/content_stream.rs`). If we rendered
+            // those into markdown by default, every existing caller would
+            // suddenly see `![Image: Im0](image)` placeholders inserted
+            // throughout their output — a silent regression for anyone who
+            // upgrades. Image bboxes are still available via
+            // `extract_text_with_positions` for callers (e.g. layout-aware
+            // pipelines) that want to crop + caption figures themselves.
+            include_images: false,
             include_links: true,
             include_page_numbers: false,
             strip_headers_footers: true,

src/markdown/postprocess.rs

@@ -29,6 +29,7 @@ pub(crate) fn clean_markdown(mut text: String, options: &MarkdownOptions) -> Str
     // text item, which combine with gap-based space insertion to produce
     // double spaces ("Vice  President" instead of "Vice President").
     collapse_consecutive_spaces(&mut text);
+    remove_spaces_before_closing_brackets(&mut text);
 
     // Remove excessive newlines (more than 2 in a row)
     while text.contains("\n\n\n") {
@@ -71,6 +72,20 @@ fn collapse_consecutive_spaces(text: &mut String) {
     *text = result;
 }
 
+/// Remove spaces before closing square brackets.
+/// Unit markers and markdown links occasionally pick up a gap-inserted space
+/// before `]` (e.g. `[kg/m3 ]`), which is cosmetic padding.
+fn remove_spaces_before_closing_brackets(text: &mut String) {
+    let mut result = String::with_capacity(text.len());
+    for ch in text.chars() {
+        if ch == ']' && result.ends_with(' ') {
+            result.pop();
+        }
+        result.push(ch);
+    }
+    *text = result;
+}
+
 /// Collapse dot leaders (runs of 4+ dots) into " ... "
 /// Common in tables of contents: "Introduction...............................1" -> "Introduction ... 1"
 fn collapse_dot_leaders(text: &str) -> String {
@@ -342,6 +357,18 @@ mod tests {
         assert!(result.contains("Chapter 2 ... 20"));
     }
 
+    // --- remove_spaces_before_closing_brackets ---
+
+    #[test]
+    fn test_remove_spaces_before_closing_brackets() {
+        let mut input = "Density [kg/m3 ] and [linked text ](https://example.com)".to_string();
+        remove_spaces_before_closing_brackets(&mut input);
+        assert_eq!(
+            input,
+            "Density [kg/m3] and [linked text](https://example.com)"
+        );
+    }
+
     // --- fix_hyphenation ---
 
     #[test]

src/markdown/preprocess.rs

@@ -208,6 +208,157 @@ fn normalize_for_comparison(s: &str) -> String {
     trimmed.to_string()
 }
 
+/// Compact a comparison key for fuzzy matching of damaged running headers.
+///
+/// Some tagged PDFs emit running footer text with overlapping fragments, so one
+/// page may read "F rom ..." while later pages read "F om r ...". Exact
+/// normalized text still drives candidate discovery; this compact form is only
+/// used when deciding whether a one-off edge line is close enough to an already
+/// repeated candidate.
+fn compact_comparison_key(s: &str) -> String {
+    s.chars()
+        .filter(|c| c.is_ascii_alphanumeric())
+        .map(|c| c.to_ascii_lowercase())
+        .collect()
+}
+
+fn bounded_levenshtein(a: &str, b: &str, max_distance: usize) -> Option<usize> {
+    let a_chars: Vec<char> = a.chars().collect();
+    let b_chars: Vec<char> = b.chars().collect();
+
+    if a_chars.len().abs_diff(b_chars.len()) > max_distance {
+        return None;
+    }
+
+    let mut prev: Vec<usize> = (0..=b_chars.len()).collect();
+    let mut curr = vec![0; b_chars.len() + 1];
+
+    for (i, a_ch) in a_chars.iter().enumerate() {
+        curr[0] = i + 1;
+        let mut row_min = curr[0];
+
+        for (j, b_ch) in b_chars.iter().enumerate() {
+            let cost = usize::from(a_ch != b_ch);
+            curr[j + 1] = (prev[j + 1] + 1).min(curr[j] + 1).min(prev[j] + cost);
+            row_min = row_min.min(curr[j + 1]);
+        }
+
+        if row_min > max_distance {
+            return None;
+        }
+
+        std::mem::swap(&mut prev, &mut curr);
+    }
+
+    let distance = prev[b_chars.len()];
+    (distance <= max_distance).then_some(distance)
+}
+
+fn matches_candidate(
+    normalized: &str,
+    candidates: &HashSet<String>,
+    compact_candidates: &[String],
+) -> bool {
+    if candidates.contains(normalized) {
+        return true;
+    }
+
+    if !has_broken_word_spacing(normalized) {
+        return false;
+    }
+
+    let compact = compact_comparison_key(normalized);
+    if compact.len() < 20 {
+        return false;
+    }
+
+    compact_candidates.iter().any(|candidate| {
+        candidate.len().abs_diff(compact.len()) <= 2
+            && bounded_levenshtein(&compact, candidate, 2).is_some()
+    })
+}
+
+fn ends_with_hyphen(raw: &str) -> bool {
+    matches!(
+        raw.chars().last(),
+        Some('-' | '\u{00ad}' | '\u{2010}' | '\u{2011}' | '\u{2012}' | '\u{2013}')
+    )
+}
+
+fn suspicious_short_token(raw: &str, alpha: &str, contains_equals: bool) -> bool {
+    let len = alpha.chars().count();
+    if len == 0 || len > 2 || ends_with_hyphen(raw) {
+        return false;
+    }
+
+    if contains_equals {
+        let raw_alpha: String = raw.chars().filter(|c| c.is_alphabetic()).collect();
+        if raw_alpha.chars().all(|c| c.is_uppercase()) {
+            return false;
+        }
+    }
+
+    true
+}
+
+fn is_uppercase_heavy(text: &str) -> bool {
+    let mut alpha = 0usize;
+    let mut uppercase = 0usize;
+    let mut lowercase = 0usize;
+
+    for ch in text.chars().filter(|ch| ch.is_alphabetic()) {
+        alpha += 1;
+        if ch.is_uppercase() {
+            uppercase += 1;
+        } else if ch.is_lowercase() {
+            lowercase += 1;
+        }
+    }
+
+    alpha >= 12 && lowercase == 0 && uppercase * 100 / alpha >= 80
+}
+
+fn has_broken_word_spacing(text: &str) -> bool {
+    if is_uppercase_heavy(text) {
+        return false;
+    }
+
+    let contains_equals = text.contains('=');
+    let tokens: Vec<(usize, bool)> = text
+        .split_whitespace()
+        .filter_map(|raw| {
+            let alpha: String = raw
+                .chars()
+                .filter(|c| c.is_alphabetic())
+                .flat_map(|c| c.to_lowercase())
+                .collect();
+            let len = alpha.chars().count();
+            (len > 0).then(|| (len, suspicious_short_token(raw, &alpha, contains_equals)))
+        })
+        .collect();
+
+    if tokens.len() < 4 {
+        return false;
+    }
+
+    let suspicious_tokens = tokens.iter().filter(|(_, suspicious)| *suspicious).count();
+    if suspicious_tokens < 3 {
+        return false;
+    }
+
+    let split_word_windows = tokens
+        .windows(3)
+        .filter(|window| window[0].0 >= 3 && window[1].1 && window[2].0 >= 3)
+        .count();
+    let adjacent_fragments = tokens
+        .windows(2)
+        .filter(|window| window[0].1 && window[1].1)
+        .count();
+
+    suspicious_tokens as f32 / tokens.len() as f32 >= 0.35
+        && (split_word_windows > 0 || adjacent_fragments > 0)
+}
+
 /// Returns true if the line looks like a list item or heading (should not be stripped).
 fn is_structural_line(text: &str) -> bool {
     let t = text.trim_start();
@@ -236,7 +387,9 @@ fn is_decorative_separator(text: &str) -> bool {
 /// Strip lines that repeat on many distinct pages (running headers/footers).
 ///
 /// A line is considered a repeated header/footer if:
-/// 1. Its normalized text appears on `>= max(3, page_count * 30%)` distinct pages
+/// 1. Its normalized text appears on enough distinct pages. The normal threshold
+///    is document-wide; visibly broken/letter-spaced running text can use a
+///    capped chapter-level threshold in long books.
 /// 2. It is at least 10 characters long
 /// 3. It doesn't look like a structural element (heading, list item)
 /// 4. It consistently appears in the top or bottom N distinct Y positions
@@ -253,13 +406,27 @@ fn is_decorative_separator(text: &str) -> bool {
 /// Page numbers are stripped from line text before comparison, so headers like
 /// "Chapter 3 — Page 5" and "Chapter 3 — Page 6" are treated as the same text.
 pub(crate) fn strip_repeated_lines(lines: Vec<TextLine>, page_count: u32) -> Vec<TextLine> {
-    if lines.is_empty() || page_count < 3 {
+    let removal_set = find_repeated_line_indices(&lines, page_count);
+    if removal_set.is_empty() {
         return lines;
     }
 
+    lines
+        .into_iter()
+        .enumerate()
+        .filter(|(idx, _)| !removal_set.contains(idx))
+        .map(|(_, line)| line)
+        .collect()
+}
+
+fn find_repeated_line_indices(lines: &[TextLine], page_count: u32) -> HashSet<usize> {
+    if lines.is_empty() || page_count < 3 {
+        return HashSet::new();
+    }
+
     // Compute Y range per page (min_y, max_y)
     let mut page_y_range: HashMap<u32, (f32, f32)> = HashMap::new();
-    for line in &lines {
+    for line in lines {
         let entry = page_y_range.entry(line.page).or_insert((line.y, line.y));
         if line.y < entry.0 {
             entry.0 = line.y;
@@ -271,7 +438,7 @@ pub(crate) fn strip_repeated_lines(lines: Vec<TextLine>, page_count: u32) -> Vec
 
     // Build sorted Y values per page, so we can check line rank (position from edge)
     let mut page_sorted_ys: HashMap<u32, Vec<f32>> = HashMap::new();
-    for line in &lines {
+    for line in lines {
         page_sorted_ys.entry(line.page).or_default().push(line.y);
     }
     for ys in page_sorted_ys.values_mut() {
@@ -287,23 +454,39 @@ pub(crate) fn strip_repeated_lines(lines: Vec<TextLine>, page_count: u32) -> Vec
     // page margin.
     const EDGE_LINE_COUNT: usize = 5;
 
+    fn y_position_rank(
+        y: f32,
+        page: u32,
+        page_sorted_ys: &HashMap<u32, Vec<f32>>,
+    ) -> Option<(usize, usize)> {
+        let ys = page_sorted_ys.get(&page)?;
+        let pos = ys.iter().position(|&py| (py - y).abs() < 0.1)?;
+        Some((pos, ys.len()))
+    }
+
     /// Returns true if the given Y position is among the first or last N distinct
     /// Y positions on the specified page.
     fn is_y_at_edge(y: f32, page: u32, page_sorted_ys: &HashMap<u32, Vec<f32>>, n: usize) -> bool {
-        let ys = match page_sorted_ys.get(&page) {
-            Some(ys) => ys,
-            None => return false,
+        let Some((pos, len)) = y_position_rank(y, page, page_sorted_ys) else {
+            return false;
         };
-        if ys.len() <= n * 2 {
+        if len <= n * 2 {
             // Page has very few lines — everything is near the edge
             return true;
         }
-        // Check if this Y is among the first or last N
-        let pos = match ys.iter().position(|&py| (py - y).abs() < 0.1) {
-            Some(p) => p,
-            None => return false,
+        pos < n || pos >= len - n
+    }
+
+    fn is_y_at_strict_lower_edge(
+        y: f32,
+        page: u32,
+        page_sorted_ys: &HashMap<u32, Vec<f32>>,
+        n: usize,
+    ) -> bool {
+        let Some((pos, len)) = y_position_rank(y, page, page_sorted_ys) else {
+            return false;
         };
-        pos < n || pos >= ys.len() - n
+        len > n * 2 && pos < n
     }
 
     // Average page span for normalizing Y variance
@@ -327,8 +510,9 @@ pub(crate) fn strip_repeated_lines(lines: Vec<TextLine>, page_count: u32) -> Vec
     // Build frequency maps using normalize_for_comparison.
     // Individual line text -> distinct pages
     let mut freq: HashMap<String, HashSet<u32>> = HashMap::new();
+    let mut bottom_freq: HashMap<String, HashSet<u32>> = HashMap::new();
     let mut y_positions: HashMap<String, Vec<f32>> = HashMap::new();
-    for line in &lines {
+    for line in lines {
         if !is_y_at_edge(line.y, line.page, &page_sorted_ys, EDGE_LINE_COUNT) {
             continue;
         }
@@ -340,6 +524,12 @@ pub(crate) fn strip_repeated_lines(lines: Vec<TextLine>, page_count: u32) -> Vec
         freq.entry(normalized.clone())
             .or_default()
             .insert(line.page);
+        if is_y_at_strict_lower_edge(line.y, line.page, &page_sorted_ys, EDGE_LINE_COUNT) {
+            bottom_freq
+                .entry(normalized.clone())
+                .or_default()
+                .insert(line.page);
+        }
         y_positions.entry(normalized).or_default().push(line.y);
     }
 
@@ -347,6 +537,7 @@ pub(crate) fn strip_repeated_lines(lines: Vec<TextLine>, page_count: u32) -> Vec
     // This catches split column headers where individual fragments don't meet
     // the frequency threshold but the combined row does.
     let mut band_freq: HashMap<String, HashSet<u32>> = HashMap::new();
+    let mut band_bottom_freq: HashMap<String, HashSet<u32>> = HashMap::new();
     let mut band_y_positions: HashMap<String, Vec<f32>> = HashMap::new();
     for (&(page, _), indices) in &y_bands {
         if indices.len() < 2 {
@@ -371,11 +562,36 @@ pub(crate) fn strip_repeated_lines(lines: Vec<TextLine>, page_count: u32) -> Vec
             .entry(normalized.clone())
             .or_default()
             .insert(page);
+        if is_y_at_strict_lower_edge(band_y, page, &page_sorted_ys, EDGE_LINE_COUNT) {
+            band_bottom_freq
+                .entry(normalized.clone())
+                .or_default()
+                .insert(page);
+        }
         band_y_positions.entry(normalized).or_default().push(band_y);
     }
 
-    // Compute threshold
-    let threshold = 3u32.max(page_count * 30 / 100);
+    // Compute thresholds. Keep the conservative document-wide threshold for
+    // clean text, and allow a lower cap only for visibly broken/letter-spaced
+    // running headers in books where each chapter has its own footer/header.
+    let document_threshold = 3u32.max(page_count * 30 / 100);
+    let garbled_chapter_threshold = 3u32.max((page_count * 30 / 100).min(8));
+    let remove_all_bottom_threshold = document_threshold.min(garbled_chapter_threshold);
+    let meets_frequency_threshold =
+        |text: &str, pages: &HashSet<u32>, bottom_pages: &HashMap<String, HashSet<u32>>| -> bool {
+            pages.len() as u32 >= document_threshold
+                || (has_broken_word_spacing(text)
+                    && bottom_pages
+                        .get(text)
+                        .is_some_and(|pages| pages.len() as u32 >= garbled_chapter_threshold))
+        };
+    let should_remove_all_occurrences =
+        |text: &str, bottom_pages: &HashMap<String, HashSet<u32>>| -> bool {
+            has_broken_word_spacing(text)
+                && bottom_pages
+                    .get(text)
+                    .is_some_and(|pages| pages.len() as u32 >= remove_all_bottom_threshold)
+        };
 
     // Check Y-position consistency: headers/footers appear at the same position
     // on every page, table content varies. Require normalized stddev < 5% of
@@ -393,29 +609,61 @@ pub(crate) fn strip_repeated_lines(lines: Vec<TextLine>, page_count: u32) -> Vec
     };
 
     // Identify candidates from individual frequency map
-    let candidates: HashSet<String> = freq
-        .into_iter()
-        .filter(|(text, pages)| {
-            pages.len() as u32 >= threshold
-                && !is_structural_line(text)
-                && has_consistent_y(text, &y_positions)
-        })
-        .map(|(text, _)| text)
+    let mut remove_all_candidates: HashSet<String> = HashSet::new();
+    let mut candidates: HashSet<String> = HashSet::new();
+    for (text, pages) in freq {
+        if meets_frequency_threshold(&text, &pages, &bottom_freq)
+            && !is_structural_line(&text)
+            && has_consistent_y(&text, &y_positions)
+        {
+            if should_remove_all_occurrences(&text, &bottom_freq) {
+                remove_all_candidates.insert(text.clone());
+            }
+            candidates.insert(text);
+        }
+    }
+    let compact_candidates: Vec<String> = candidates
+        .iter()
+        .filter(|text| has_broken_word_spacing(text))
+        .map(|text| compact_comparison_key(text))
+        .filter(|text| text.len() >= 20)
+        .collect();
+    let compact_remove_all_candidates: Vec<String> = remove_all_candidates
+        .iter()
+        .filter(|text| has_broken_word_spacing(text))
+        .map(|text| compact_comparison_key(text))
+        .filter(|text| text.len() >= 20)
         .collect();
 
     // Identify candidates from coalesced band frequency map
-    let band_candidates: HashSet<String> = band_freq
-        .into_iter()
-        .filter(|(text, pages)| {
-            pages.len() as u32 >= threshold
-                && !is_structural_line(text)
-                && has_consistent_y(text, &band_y_positions)
-        })
-        .map(|(text, _)| text)
+    let mut remove_all_band_candidates: HashSet<String> = HashSet::new();
+    let mut band_candidates: HashSet<String> = HashSet::new();
+    for (text, pages) in band_freq {
+        if meets_frequency_threshold(&text, &pages, &band_bottom_freq)
+            && !is_structural_line(&text)
+            && has_consistent_y(&text, &band_y_positions)
+        {
+            if should_remove_all_occurrences(&text, &band_bottom_freq) {
+                remove_all_band_candidates.insert(text.clone());
+            }
+            band_candidates.insert(text);
+        }
+    }
+    let compact_band_candidates: Vec<String> = band_candidates
+        .iter()
+        .filter(|text| has_broken_word_spacing(text))
+        .map(|text| compact_comparison_key(text))
+        .filter(|text| text.len() >= 20)
+        .collect();
+    let compact_remove_all_band_candidates: Vec<String> = remove_all_band_candidates
+        .iter()
+        .filter(|text| has_broken_word_spacing(text))
+        .map(|text| compact_comparison_key(text))
+        .filter(|text| text.len() >= 20)
         .collect();
 
     if candidates.is_empty() && band_candidates.is_empty() {
-        return lines;
+        return HashSet::new();
     }
 
     // Build removal set.
@@ -424,8 +672,10 @@ pub(crate) fn strip_repeated_lines(lines: Vec<TextLine>, page_count: u32) -> Vec
     //   (b) its Y-band's coalesced text matches a band candidate, OR
     //   (c) any sibling in its Y-band was removed (propagation).
     //
-    // The first occurrence (lowest page number) of each repeated header/footer
-    // is kept so that document titles, column headers, etc. appear once.
+    // The first occurrence (lowest page number) of each repeated line is kept
+    // so that document titles, column headers, etc. appear once. Visibly broken
+    // footers proven by repeated lower-edge placement are removed from every
+    // matching edge occurrence, including sparse first pages.
     let mut removal_set: HashSet<usize> = HashSet::new();
 
     // Track which page first shows each candidate (to preserve first occurrence)
@@ -436,7 +686,15 @@ pub(crate) fn strip_repeated_lines(lines: Vec<TextLine>, page_count: u32) -> Vec
         }
         let text = line.text();
         let normalized = normalize_for_comparison(&text);
-        if candidates.contains(&normalized) {
+        if matches_candidate(&normalized, &candidates, &compact_candidates) {
+            if matches_candidate(
+                &normalized,
+                &remove_all_candidates,
+                &compact_remove_all_candidates,
+            ) {
+                removal_set.insert(idx);
+                continue;
+            }
             let first = first_page_individual.entry(normalized).or_insert(line.page);
             if line.page > *first {
                 removal_set.insert(idx);
@@ -465,7 +723,7 @@ pub(crate) fn strip_repeated_lines(lines: Vec<TextLine>, page_count: u32) -> Vec
             .collect::<Vec<_>>()
             .join(" ");
         let normalized = normalize_for_comparison(&coalesced);
-        if band_candidates.contains(&normalized) {
+        if matches_candidate(&normalized, &band_candidates, &compact_band_candidates) {
             let first = first_page_band.entry(normalized).or_insert(page);
             if page < *first {
                 *first = page;
@@ -489,7 +747,17 @@ pub(crate) fn strip_repeated_lines(lines: Vec<TextLine>, page_count: u32) -> Vec
             .collect::<Vec<_>>()
             .join(" ");
         let normalized = normalize_for_comparison(&coalesced);
-        if band_candidates.contains(&normalized) {
+        if matches_candidate(&normalized, &band_candidates, &compact_band_candidates) {
+            if matches_candidate(
+                &normalized,
+                &remove_all_band_candidates,
+                &compact_remove_all_band_candidates,
+            ) {
+                for &idx in &sorted_indices {
+                    removal_set.insert(idx);
+                }
+                continue;
+            }
             let first = first_page_band.get(&normalized).copied().unwrap_or(0);
             if page > first {
                 for &idx in &sorted_indices {
@@ -514,15 +782,10 @@ pub(crate) fn strip_repeated_lines(lines: Vec<TextLine>, page_count: u32) -> Vec
     }
 
     if removal_set.is_empty() {
-        return lines;
+        return HashSet::new();
     }
 
-    lines
-        .into_iter()
-        .enumerate()
-        .filter(|(idx, _)| !removal_set.contains(idx))
-        .map(|(_, line)| line)
-        .collect()
+    removal_set
 }
 
 #[cfg(test)]
@@ -556,6 +819,29 @@ mod tests {
         }
     }
 
+    #[test]
+    fn test_has_broken_word_spacing_detects_split_words() {
+        assert!(has_broken_word_spacing(
+            "F rom p rese rva tion to access a nd be yond"
+        ));
+        assert!(has_broken_word_spacing("Conve rs ing w ith the pas t"));
+        assert!(has_broken_word_spacing("The Na tional Arch ives (U K)"));
+    }
+
+    #[test]
+    fn test_has_broken_word_spacing_ignores_normal_short_words() {
+        assert!(!has_broken_word_spacing(
+            "Reunir talento e empresas é um dos fatores po- sitivos para comunidades de sucesso"
+        ));
+        assert!(!has_broken_word_spacing("Witnessed on behalf of"));
+        assert!(!has_broken_word_spacing(
+            "V = Volume in m3/kg H = Enthalpy in kJ/kg S = Entropy in kJ/kg.K"
+        ));
+        assert!(!has_broken_word_spacing(
+            "TITULAR DEL PODER EJECUTIVO FEDERAL, A TRAVÉS DE LA SECRETARÍA DE ECONOMÍA, A HACER VALER EL PRINCIPIO DE"
+        ));
+    }
+
     #[test]
     fn test_merge_struct_tree_headings() {
         // Two consecutive lines tagged as H2 via struct tree, same font size as body
@@ -683,4 +969,222 @@ mod tests {
             .unwrap();
         assert_eq!(first_header.page, 1, "first occurrence should be on page 1");
     }
+
+    #[test]
+    fn test_strip_repeated_clean_bottom_footers_kept_below_document_threshold() {
+        let mut lines = Vec::new();
+        for page in 1..=8u32 {
+            for row in 0..12u32 {
+                lines.push(make_line(
+                    &format!("unique body content page {page} row {row}"),
+                    9.5,
+                    page,
+                    600.0 - row as f32 * 20.0,
+                    None,
+                ));
+            }
+            lines.push(make_line(
+                &format!("Chapter running footer {}", 90 + page),
+                7.5,
+                page,
+                39.5,
+                None,
+            ));
+        }
+
+        let result = strip_repeated_lines(lines, 200);
+
+        let footer_count = result
+            .iter()
+            .filter(|line| line.text().contains("Chapter running footer"))
+            .count();
+        assert_eq!(
+            footer_count, 8,
+            "clean repeated footer should not use the lower garbled-text threshold"
+        );
+    }
+
+    #[test]
+    fn test_strip_repeated_garbled_bottom_footers_removes_all_occurrences_in_long_doc() {
+        let mut lines = Vec::new();
+        for page in 1..=8u32 {
+            for row in 0..12u32 {
+                lines.push(make_line(
+                    &format!("unique body content page {page} row {row}"),
+                    9.5,
+                    page,
+                    600.0 - row as f32 * 20.0,
+                    None,
+                ));
+            }
+            lines.push(make_line(
+                &format!("M L a t the Na tional Libra ry of N orwa y {}", 90 + page),
+                7.5,
+                page,
+                39.5,
+                None,
+            ));
+        }
+
+        let result = strip_repeated_lines(lines, 200);
+
+        assert!(
+            result
+                .iter()
+                .all(|line| !line.text().contains("Na tional Libra")),
+            "garbled bottom running footer should be removed from every page"
+        );
+        assert!(
+            result
+                .iter()
+                .any(|line| line.text().contains("unique body content page 1 row 0")),
+            "body text should be preserved"
+        );
+    }
+
+    #[test]
+    fn test_strip_repeated_document_wide_garbled_footers_removes_all_occurrences() {
+        let mut lines = Vec::new();
+        for page in 1..=8u32 {
+            for row in 0..12u32 {
+                lines.push(make_line(
+                    &format!("unique body content page {page} row {row}"),
+                    9.5,
+                    page,
+                    600.0 - row as f32 * 20.0,
+                    None,
+                ));
+            }
+            lines.push(make_line(
+                &format!("F rom p rese rva tion to access a nd be yond {}", 90 + page),
+                7.5,
+                page,
+                39.5,
+                None,
+            ));
+        }
+
+        let result = strip_repeated_lines(lines, 8);
+
+        let footer_count = result
+            .iter()
+            .filter(|line| line.text().contains("be yond"))
+            .count();
+        assert_eq!(
+            footer_count, 0,
+            "document-wide garbled footers should be removed from every page"
+        );
+    }
+
+    #[test]
+    fn test_strip_repeated_sparse_uppercase_headers_keep_first_occurrence() {
+        let mut lines = Vec::new();
+        for page in 1..=5u32 {
+            lines.push(make_line(
+                "PROPOSICIÓN CON PUNTO DE ACUERDO POR EL QUE EL SENADO DE LA REPÚBLICA",
+                8.0,
+                page,
+                720.0,
+                None,
+            ));
+            lines.push(make_line(
+                "A TRAVÉS DE LA SECRETARÍA DE ECONOMÍA",
+                8.0,
+                page,
+                704.0,
+                None,
+            ));
+            lines.push(make_line(
+                &format!("unique sparse-page body text {page}"),
+                10.0,
+                page,
+                620.0,
+                None,
+            ));
+        }
+
+        let result = strip_repeated_lines(lines, 5);
+
+        let title_count = result
+            .iter()
+            .filter(|line| line.text().contains("PROPOSICIÓN CON PUNTO"))
+            .count();
+        assert_eq!(
+            title_count, 1,
+            "sparse repeated heading should keep the first occurrence"
+        );
+    }
+
+    #[test]
+    fn test_strip_repeated_bottom_footers_matches_minor_garbling() {
+        let mut lines = Vec::new();
+        for page in 1..=9u32 {
+            for row in 0..12u32 {
+                lines.push(make_line(
+                    &format!("distinct paragraph text page {page} row {row}"),
+                    9.5,
+                    page,
+                    600.0 - row as f32 * 20.0,
+                    None,
+                ));
+            }
+
+            let footer = if page == 1 {
+                "F rom p rese rva tion to access a nd be yond 95"
+            } else {
+                "F om r p rese rva tion to access a nd be yond 97"
+            };
+            lines.push(make_line(footer, 7.5, page, 39.5, None));
+        }
+
+        let result = strip_repeated_lines(lines, 200);
+
+        assert!(
+            result.iter().all(|line| !line.text().contains("be yond")),
+            "fuzzy footer variant should be removed once the repeated form is detected"
+        );
+        assert!(
+            result.iter().any(|line| line
+                .text()
+                .contains("distinct paragraph text page 9 row 11")),
+            "non-footer edge-adjacent body text should be preserved"
+        );
+    }
+
+    #[test]
+    fn test_strip_repeated_bottom_footers_matches_sparse_first_page_variant() {
+        let mut lines = Vec::new();
+        for page in 1..=9u32 {
+            let body_rows = if page == 1 { 3 } else { 12 };
+            for row in 0..body_rows {
+                lines.push(make_line(
+                    &format!("distinct paragraph text page {page} row {row}"),
+                    9.5,
+                    page,
+                    600.0 - row as f32 * 20.0,
+                    None,
+                ));
+            }
+
+            let footer = if page == 1 {
+                "F rom p rese rva tion to access a nd be yond 95"
+            } else {
+                "F om r p rese rva tion to access a nd be yond 97"
+            };
+            lines.push(make_line(footer, 7.5, page, 39.5, None));
+        }
+
+        let result = strip_repeated_lines(lines, 200);
+
+        assert!(
+            result.iter().all(|line| !line.text().contains("be yond")),
+            "sparse first page variant should be removed once later lower-edge footers prove the candidate"
+        );
+        assert!(
+            result
+                .iter()
+                .any(|line| line.text().contains("distinct paragraph text page 1 row 0")),
+            "sparse first page body text should be preserved"
+        );
+    }
 }

src/python.rs

@@ -30,6 +30,9 @@ pub struct PyPdfResult {
     /// 1-indexed page numbers that need OCR.
     #[pyo3(get)]
     pub pages_needing_ocr: Vec<u32>,
+    /// Machine-readable OCR reasons by 1-indexed page.
+    #[pyo3(get)]
+    pub ocr_reasons_by_page: Vec<PyPageOcrReasons>,
     /// Title from PDF metadata.
     #[pyo3(get)]
     pub title: Option<String>,
@@ -60,6 +63,28 @@ impl PyPdfResult {
     }
 }
 
+/// OCR reasons for a single 1-indexed page.
+#[pyclass(name = "PageOcrReasons")]
+#[derive(Clone)]
+pub struct PyPageOcrReasons {
+    /// 1-indexed page number.
+    #[pyo3(get)]
+    pub page: u32,
+    /// Machine-readable OCR reason identifiers.
+    #[pyo3(get)]
+    pub reasons: Vec<String>,
+}
+
+#[pymethods]
+impl PyPageOcrReasons {
+    fn __repr__(&self) -> String {
+        format!(
+            "PageOcrReasons(page={}, reasons={:?})",
+            self.page, self.reasons
+        )
+    }
+}
+
 // ---------------------------------------------------------------------------
 // Classification wrapper (lightweight)
 // ---------------------------------------------------------------------------
@@ -106,6 +131,9 @@ pub struct PyRegionText {
     /// True when the text should not be trusted (empty, GID fonts, garbage, encoding issues).
     #[pyo3(get)]
     pub needs_ocr: bool,
+    /// Machine-readable OCR reason when the cause is known.
+    #[pyo3(get)]
+    pub ocr_reason: Option<String>,
 }
 
 #[pymethods]
@@ -160,6 +188,9 @@ pub struct PyPageMarkdown {
     /// encoding issues, garbage text, or empty extraction).
     #[pyo3(get)]
     pub needs_ocr: bool,
+    /// Machine-readable OCR reason when the cause is known.
+    #[pyo3(get)]
+    pub ocr_reason: Option<String>,
 }
 
 #[pymethods]
@@ -190,6 +221,9 @@ pub struct PyPagesExtractionResult {
     /// 1-indexed pages that need OCR (scanned/image-based or unreliable text).
     #[pyo3(get)]
     pub pages_needing_ocr: Vec<u32>,
+    /// Machine-readable OCR reasons by 1-indexed page.
+    #[pyo3(get)]
+    pub ocr_reasons_by_page: Vec<PyPageOcrReasons>,
     /// True if any page has tables or columns.
     #[pyo3(get)]
     pub is_complex: bool,
@@ -268,6 +302,7 @@ fn to_py_result(r: crate::PdfProcessResult) -> PyPdfResult {
         page_count: r.page_count,
         processing_time_ms: r.processing_time_ms,
         pages_needing_ocr: r.pages_needing_ocr,
+        ocr_reasons_by_page: to_py_page_ocr_reasons(r.ocr_reasons_by_page),
         title: r.title,
         confidence: r.confidence,
         is_complex_layout: r.layout.is_complex,
@@ -277,6 +312,16 @@ fn to_py_result(r: crate::PdfProcessResult) -> PyPdfResult {
     }
 }
 
+fn to_py_page_ocr_reasons(reasons: Vec<crate::PageOcrReasons>) -> Vec<PyPageOcrReasons> {
+    reasons
+        .into_iter()
+        .map(|reason| PyPageOcrReasons {
+            page: reason.page,
+            reasons: reason.reasons,
+        })
+        .collect()
+}
+
 fn to_py_err(e: crate::PdfError) -> PyErr {
     PyValueError::new_err(e.to_string())
 }
@@ -350,11 +395,13 @@ fn to_py_pages_result(r: crate::PagesExtractionResult) -> PyPagesExtractionResul
                 page: p.page,
                 markdown: p.markdown,
                 needs_ocr: p.needs_ocr,
+                ocr_reason: p.ocr_reason,
             })
             .collect(),
         pages_with_tables: r.pages_with_tables,
         pages_with_columns: r.pages_with_columns,
         pages_needing_ocr: r.pages_needing_ocr,
+        ocr_reasons_by_page: to_py_page_ocr_reasons(r.ocr_reasons_by_page),
         is_complex: r.is_complex,
     }
 }
@@ -370,6 +417,7 @@ fn convert_region_results(results: Vec<crate::PageRegionResult>) -> Vec<PyPageRe
                 .map(|r| PyRegionText {
                     text: r.text,
                     needs_ocr: r.needs_ocr,
+                    ocr_reason: r.ocr_reason,
                 })
                 .collect(),
         })
@@ -563,6 +611,7 @@ fn extract_pages_markdown_bytes(
 #[pymodule]
 fn pdf_inspector(m: &Bound<'_, PyModule>) -> PyResult<()> {
     m.add_class::<PyPdfResult>()?;
+    m.add_class::<PyPageOcrReasons>()?;
     m.add_class::<PyPdfClassification>()?;
     m.add_class::<PyTextItem>()?;
     m.add_class::<PyRegionText>()?;

src/tables/detect_lines.rs

@@ -4,11 +4,74 @@
 //! gridlines.  Many IRS forms and government PDFs use these instead of
 //! `re` (rectangle) operators.
 
+use std::collections::HashSet;
+
 use crate::tables::Table;
 use crate::types::{PdfLine, TextItem};
 
 use super::detect_rects::{assign_items_to_grid, snap_edges};
 
+/// Derive column edges from the x-endpoints of horizontal-rule
+/// segments when no vertical lines were drawn.
+///
+/// Catalog and archival-finding-aid tables are commonly drawn with
+/// per-row horizontal rules broken into N segments (one segment per
+/// cell), with no vertical dividers at all. The segment break points
+/// (e.g. `[50, 127], [127, 485], [485, 562]` per row) implicitly
+/// encode the column boundaries.
+///
+/// Returns column edges if ≥3 distinct x-positions each show up as a
+/// segment endpoint on ≥50% of the unique horizontal-line rows.
+/// Returns `None` otherwise — decorative rules with varying widths
+/// shouldn't be mistaken for a table.
+fn derive_columns_from_horizontal_segments(horizontals: &[(f32, f32, f32)]) -> Option<Vec<f32>> {
+    if horizontals.len() < 3 {
+        return None;
+    }
+
+    let mut endpoints: Vec<f32> = Vec::with_capacity(horizontals.len() * 2);
+    for &(_, x_min, x_max) in horizontals {
+        endpoints.push(x_min);
+        endpoints.push(x_max);
+    }
+    let clusters = snap_edges(&endpoints, 5.0);
+    if clusters.len() < 3 {
+        return None;
+    }
+
+    // Bucket y-values to count unique rows. Tolerance ~0.1pt (×10
+    // rounding) tolerates the snap_edges 3pt clustering used later
+    // for row edges.
+    let unique_rows: HashSet<i32> = horizontals
+        .iter()
+        .map(|&(y, _, _)| (y * 10.0).round() as i32)
+        .collect();
+    if unique_rows.len() < 2 {
+        return None;
+    }
+    let min_rows = (unique_rows.len() as f32 * 0.5).ceil() as usize;
+
+    let qualifying: Vec<f32> = clusters
+        .iter()
+        .copied()
+        .filter(|&cluster_x| {
+            let rows_touched: HashSet<i32> = horizontals
+                .iter()
+                .filter(|&&(_, x_min, x_max)| {
+                    (x_min - cluster_x).abs() < 5.0 || (x_max - cluster_x).abs() < 5.0
+                })
+                .map(|&(y, _, _)| (y * 10.0).round() as i32)
+                .collect();
+            rows_touched.len() >= min_rows
+        })
+        .collect();
+
+    if qualifying.len() < 3 {
+        return None;
+    }
+    Some(qualifying)
+}
+
 /// Detect tables from line segments on a given page.
 ///
 /// Lines are classified as horizontal or vertical, snapped into grid edges,
@@ -52,25 +115,50 @@ pub fn detect_tables_from_lines(items: &[TextItem], lines: &[PdfLine], page: u32
         // Diagonal lines are ignored
     }
 
-    if horizontals.len() < 3 || verticals.len() < 2 {
+    if horizontals.len() < 3 {
         return Vec::new();
     }
 
+    // If no/very-few vertical lines are drawn, try to derive column edges
+    // from the x-endpoints of the horizontal-rule segments. Catalog and
+    // archival-finding-aid layouts commonly draw each row's horizontal
+    // rule as N segments (one per cell), with no vertical dividers at
+    // all — the segment break points encode the column boundaries.
+    let implicit_col_edges: Option<Vec<f32>> = if verticals.len() < 2 {
+        derive_columns_from_horizontal_segments(&horizontals)
+    } else {
+        None
+    };
+    if verticals.len() < 2 && implicit_col_edges.is_none() {
+        return Vec::new();
+    }
+    let cols_from_segments = implicit_col_edges.is_some();
+
     log::debug!(
-        "detect_lines p{}: {} horiz, {} vert lines (of {} total on page)",
+        "detect_lines p{}: {} horiz, {} vert lines (of {} total on page){}",
         page,
         horizontals.len(),
         verticals.len(),
-        page_lines.len()
+        page_lines.len(),
+        if cols_from_segments {
+            " — columns from horizontal segments"
+        } else {
+            ""
+        }
     );
 
     // Snap Y-values of horizontal lines → row edges
     let h_ys: Vec<f32> = horizontals.iter().map(|(y, _, _)| *y).collect();
     let row_edges = snap_edges(&h_ys, 3.0);
 
-    // Snap X-values of vertical lines → column edges
-    let v_xs: Vec<f32> = verticals.iter().map(|(x, _, _)| *x).collect();
-    let col_edges = snap_edges(&v_xs, 3.0);
+    // Column edges from drawn verticals when present, else from the
+    // horizontal-segment endpoints derived above.
+    let col_edges = if let Some(c) = implicit_col_edges {
+        c
+    } else {
+        let v_xs: Vec<f32> = verticals.iter().map(|(x, _, _)| *x).collect();
+        snap_edges(&v_xs, 3.0)
+    };
 
     log::debug!(
         "detect_lines p{}: {} row edges, {} col edges after snap",
@@ -110,15 +198,21 @@ pub fn detect_tables_from_lines(items: &[TextItem], lines: &[PdfLine], page: u32
         return Vec::new();
     }
 
-    // Reject page-spanning frames: if the grid covers >90% of a standard page
-    // dimension in both axes, it's a border frame, not a table.
+    // Reject page-spanning frames: a decorative outer border has just 4
+    // edges (top/bottom/left/right). Real full-page tables — common in
+    // governmental ledgers, financial reports, etc. — span the same A4 /
+    // Letter dimensions but have many internal row/column rules. Only
+    // reject when the line set looks like a bare frame, not a grid.
     // Standard pages are ~595×842 (A4) or ~612×792 (Letter).
-    if table_width > 500.0 && table_height > 700.0 {
+    if table_width > 500.0 && table_height > 700.0 && horizontals.len() <= 4 && verticals.len() <= 4
+    {
         log::debug!(
-            "detect_lines p{}: rejected — page-spanning frame ({:.0}×{:.0})",
+            "detect_lines p{}: rejected — page-spanning frame ({:.0}×{:.0}, {} h + {} v)",
             page,
             table_width,
-            table_height
+            table_height,
+            horizontals.len(),
+            verticals.len()
         );
         return Vec::new();
     }
@@ -146,24 +240,33 @@ pub fn detect_tables_from_lines(items: &[TextItem], lines: &[PdfLine], page: u32
 
     // Validate vertical lines: at least 2 should span a meaningful height.
     // Full spanning (>30%) is ideal, but accept many shorter lines (>10%)
-    // for tables with partial column separators.
-    let spanning_v = verticals
-        .iter()
-        .filter(|(_, y_min, y_max)| (y_max - y_min) > table_height * 0.3)
-        .count();
-    let partial_v = verticals
-        .iter()
-        .filter(|(_, y_min, y_max)| (y_max - y_min) > table_height * 0.10)
-        .count();
-    if spanning_v < 2 && partial_v < 4 {
-        log::debug!(
-            "detect_lines p{}: rejected — {} spanning + {} partial V lines",
-            page,
-            spanning_v,
-            partial_v
-        );
-        return Vec::new();
-    }
+    // for tables with partial column separators. Skipped entirely when
+    // columns came from horizontal-segment endpoints — there are no
+    // vertical lines to validate against, and the segment-endpoint
+    // consistency check in `derive_columns_from_horizontal_segments`
+    // is the equivalent guard.
+    let spanning_v = if cols_from_segments {
+        0
+    } else {
+        let s = verticals
+            .iter()
+            .filter(|(_, y_min, y_max)| (y_max - y_min) > table_height * 0.3)
+            .count();
+        let p = verticals
+            .iter()
+            .filter(|(_, y_min, y_max)| (y_max - y_min) > table_height * 0.10)
+            .count();
+        if s < 2 && p < 4 {
+            log::debug!(
+                "detect_lines p{}: rejected — {} spanning + {} partial V lines",
+                page,
+                s,
+                p
+            );
+            return Vec::new();
+        }
+        s
+    };
 
     // Row edges need to be in descending order (top of page = higher Y first)
     let mut row_edges_desc = row_edges;
@@ -410,6 +513,135 @@ mod tests {
         assert!(tables.is_empty());
     }
 
+    #[test]
+    fn test_horizontal_segments_only_implicit_columns_accepted() {
+        // Catalog/finding-aid pattern: each row's horizontal rule is
+        // drawn as 3 segments at consistent x-endpoints (50, 127, 485,
+        // 562), with no vertical lines anywhere. The segment break
+        // points must be inferred as column edges.
+        let mut lines = Vec::new();
+        // Slightly uneven row spacing so the chart-gridline rejector
+        // (CV < 0.02) doesn't fire.
+        let row_ys = [80.0_f32, 145.0, 215.0, 280.0, 350.0, 415.0, 485.0];
+        for &y in &row_ys {
+            lines.push(make_hline(y, 50.0, 127.0, 1));
+            lines.push(make_hline(y, 127.0, 485.0, 1));
+            lines.push(make_hline(y, 485.0, 562.0, 1));
+        }
+        // Populate every cell so capture / density checks pass.
+        let mut items = Vec::new();
+        for w in row_ys.windows(2) {
+            let row_y = (w[0] + w[1]) / 2.0;
+            items.push(make_item("id", 80.0, row_y, 1));
+            items.push(make_item("description here", 200.0, row_y, 1));
+            items.push(make_item("date", 510.0, row_y, 1));
+        }
+        let tables = detect_tables_from_lines(&items, &lines, 1);
+        assert_eq!(
+            tables.len(),
+            1,
+            "horizontal-segment-only grid should be accepted"
+        );
+        let t = &tables[0];
+        assert!(
+            t.cells.len() >= 4,
+            "expected ≥4 rows, got {}",
+            t.cells.len()
+        );
+        assert_eq!(t.cells[0].len(), 3, "expected 3 columns");
+    }
+
+    #[test]
+    fn test_horizontal_segments_with_inconsistent_endpoints_rejected() {
+        // Decorative rules of varying widths shouldn't be detected as a
+        // table — each line has its own x-endpoints, no consistent
+        // column boundary survives the 50%-of-rows threshold.
+        let lines = vec![
+            make_hline(100.0, 50.0, 150.0, 1),
+            make_hline(200.0, 50.0, 220.0, 1),
+            make_hline(300.0, 50.0, 310.0, 1),
+            make_hline(400.0, 50.0, 470.0, 1),
+        ];
+        let items = vec![
+            make_item("decorative", 100.0, 150.0, 1),
+            make_item("text", 100.0, 250.0, 1),
+        ];
+        let tables = detect_tables_from_lines(&items, &lines, 1);
+        assert!(
+            tables.is_empty(),
+            "varying-width decorative rules should not be detected"
+        );
+    }
+
+    #[test]
+    fn test_page_spanning_bare_frame_rejected() {
+        // Just an outer A4-sized rectangle: 2 horizontals + 2 verticals.
+        // No internal structure → decorative border, not a table.
+        let lines = vec![
+            make_hline(20.0, 20.0, 575.0, 1),  // top
+            make_hline(820.0, 20.0, 575.0, 1), // bottom
+            make_vline(20.0, 20.0, 820.0, 1),  // left
+            make_vline(575.0, 20.0, 820.0, 1), // right
+        ];
+        let items = vec![
+            make_item("title", 100.0, 100.0, 1),
+            make_item("body", 100.0, 200.0, 1),
+        ];
+        let tables = detect_tables_from_lines(&items, &lines, 1);
+        assert!(
+            tables.is_empty(),
+            "Page-sized 4-edge frame should be rejected as decoration"
+        );
+    }
+
+    #[test]
+    fn test_page_spanning_grid_with_internal_lines_accepted() {
+        // Full-page table (governmental-ledger pattern): A4-sized grid
+        // that previously hit the "page-spanning frame" early reject
+        // before downstream validation could even look at it.
+        // Verticals span the full table height so we isolate the
+        // frame-vs-grid decision under test.
+        let mut lines = Vec::new();
+        // 13 horizontal rules: header + 12 row separators
+        let h_ys = [
+            22.5, 37.9, 95.5, 144.5, 184.9, 233.9, 291.7, 340.7, 415.8, 499.6, 574.7, 623.7, 698.8,
+        ];
+        for &y in &h_ys {
+            lines.push(make_hline(y, 22.6, 566.6, 1));
+        }
+        // 7 column dividers spanning full table height.
+        let v_xs = [22.6, 66.3, 116.3, 186.6, 263.1, 493.5, 566.5];
+        for &x in &v_xs {
+            lines.push(make_vline(x, 22.5, 698.8, 1));
+        }
+        // Populate every cell so the capture-ratio + density checks pass.
+        let mut items = Vec::new();
+        for r in 0..(h_ys.len() - 1) {
+            let row_y = (h_ys[r] + h_ys[r + 1]) / 2.0;
+            for c in 0..(v_xs.len() - 1) {
+                let col_x = (v_xs[c] + v_xs[c + 1]) / 2.0;
+                items.push(make_item("x", col_x, row_y, 1));
+            }
+        }
+        let tables = detect_tables_from_lines(&items, &lines, 1);
+        assert_eq!(
+            tables.len(),
+            1,
+            "Full-page table with internal grid should be accepted"
+        );
+        let t = &tables[0];
+        assert!(
+            t.cells.len() >= 6,
+            "expected ≥6 rows, got {}",
+            t.cells.len()
+        );
+        assert!(
+            t.cells[0].len() >= 3,
+            "expected ≥3 columns, got {}",
+            t.cells[0].len()
+        );
+    }
+
     #[test]
     fn test_single_column_rejected() {
         // Only 2 col edges (1 column) — not a table even with verticals

src/tables/detect_rects.rs

@@ -231,6 +231,15 @@ pub fn detect_tables_from_rects(
     rects: &[PdfRect],
     page: u32,
 ) -> (Vec<Table>, Vec<RectHintRegion>) {
+    // Strip Image placeholders before column/row clustering — an image's bbox
+    // would otherwise show up as a spurious column edge. See `is_text_layout_item`.
+    let items_owned: Vec<TextItem> = items
+        .iter()
+        .filter(|i| crate::extractor::is_text_layout_item(i))
+        .cloned()
+        .collect();
+    let items = items_owned.as_slice();
+
     // Filter rects on this page; normalize negative widths/heights; skip tiny rects.
     let mut page_rects: Vec<(f32, f32, f32, f32)> = Vec::new(); // (x, y, w, h) normalized
     for r in rects {
@@ -285,7 +294,11 @@ pub fn detect_tables_from_rects(
         //
         // Only remove when the container is a similarly-sized cell (height
         // ratio < 4×), NOT when the container is a table-wide background
-        // that dwarfs the sub-rect.
+        // that dwarfs the sub-rect.  Origin-anchored page-background rects
+        // also disqualify as containers — they normally exceed the 4× ratio,
+        // but when the sub-rect is itself a tall table-frame the ratio can
+        // fall under the gate, and dropping the frame collapses cluster
+        // adjacency between adjacent column-cell groups.
         //
         // Skip this O(n²) dedup when there are too many rects — pages with
         // thousands of vector-drawing rects won't benefit from cell dedup.
@@ -295,9 +308,11 @@ pub fn detect_tables_from_rects(
             page_rects.retain(|&(ax, ay, aw, ah)| {
                 let tol = 2.0;
                 !snapshot.iter().any(|&(bx, by, bw, bh)| {
+                    let container_is_page_bg = bx < 5.0 && by < 5.0;
                     // b must strictly contain a (b is larger in area)
                     bw * bh > aw * ah * 1.2
                         && bh < ah * 4.0 // container must be similarly sized, not a table background
+                        && !container_is_page_bg
                         && bx <= ax + tol
                         && (bx + bw) >= (ax + aw) - tol
                         && by <= ay + tol
@@ -1388,13 +1403,15 @@ fn detect_row_stripe_table(
         .max()
         .unwrap_or(0);
     // Allow longer cells for multi-column tables (descriptions in one column
-    // are common).  Single-column or 2-column "tables" with giant cells are
-    // almost always layout backgrounds.
+    // are common). Narrow grids with giant cells are usually layout
+    // backgrounds — but only when the row count is also small. A 4+-row
+    // key/value table with one descriptive column reads as a real table
+    // on every other gate, so don't reject it on cell length alone.
     let max_allowed = if num_cols >= 3 { 2000 } else { 500 };
-    if max_cell_len > max_allowed {
+    if max_cell_len > max_allowed && non_empty_rows < 4 {
         debug!(
-            "  row-stripe rejected: max cell length {} > {} (layout background)",
-            max_cell_len, max_allowed
+            "  row-stripe rejected: max cell length {} > {} (layout background, {} rows)",
+            max_cell_len, max_allowed, non_empty_rows
         );
         return None;
     }
@@ -1587,25 +1604,111 @@ fn detect_row_stripe_table_from_cell_rects(
         return None;
     }
 
-    // Derive columns from text X-position clustering
+    // Derive columns from text X-position clustering, but prefer rect
+    // X-edges when they already provide a tighter scaffold.  Some PDFs draw
+    // only the row-index cells in the body plus a full header row; that is
+    // not dense enough for `try_build_grid`, but the header rects still define
+    // the real columns.  Text starts inside wide cells can otherwise split the
+    // table into spurious sub-columns.
     let columns = cluster_x_positions(&page_items, 15.0);
-    if columns.len() < 2 {
+    let text_col_edges = if columns.len() >= 2 {
+        let mut edges: Vec<f32> = Vec::with_capacity(columns.len() + 1);
+        let min_x = page_items.iter().map(|(_, i)| i.x).reduce(f32::min)?;
+        edges.push(min_x - 5.0);
+        for pair in columns.windows(2) {
+            edges.push((pair[0] + pair[1]) / 2.0);
+        }
+        let max_x_right = page_items
+            .iter()
+            .map(|(_, i)| i.x + i.width)
+            .reduce(f32::max)?;
+        edges.push(max_x_right + 5.0);
+        Some(edges)
+    } else {
+        None
+    };
+
+    let rect_col_edges = {
+        let mut x_vals = Vec::with_capacity(content_rects.len() * 2);
+        for &&(x, _, w, _) in &content_rects {
+            x_vals.push(x);
+            x_vals.push(x + w);
+        }
+        let mut edges = snap_edges(&x_vals, 6.0);
+        edges.sort_by(|a, b| a.total_cmp(b));
+        if (3..=26).contains(&edges.len()) {
+            Some(edges)
+        } else {
+            None
+        }
+    };
+
+    // For wired-grid tables whose header text is centered/right-aligned but
+    // whose data is left-aligned, cluster_x_positions can drop the header-only
+    // x-cluster in its singleton-filter pass and merge adjacent data clusters
+    // when the gap is below threshold, losing a column. Rect borders are
+    // ground truth in that case — but only when each rect column actually
+    // holds text. Decorative or background rects (prose laid out in a frame,
+    // cell-fill rects with extra borders) can produce more rect-derived
+    // columns than the text supports; preferring rects there would split a
+    // logical column into spurious sub-columns.
+    let rect_cols_match_text = match (&rect_col_edges, &text_col_edges) {
+        (Some(rect_edges), _) if rect_edges.len() >= 4 => {
+            let num_rect_cols = rect_edges.len() - 1;
+            let mut col_item_counts = vec![0usize; num_rect_cols];
+            for (_, item) in &page_items {
+                let cx = item.x + item.width / 2.0;
+                for c in 0..num_rect_cols {
+                    if cx >= rect_edges[c] - 2.0 && cx <= rect_edges[c + 1] + 2.0 {
+                        col_item_counts[c] += 1;
+                        break;
+                    }
+                }
+            }
+            // Require every rect column to hold multiple text items. A rect
+            // column with no (or only one) item is decorative or the rect grid
+            // is detecting a spurious column the data does not need; in those
+            // cases the old text-cluster preference is the safer fallback.
+            col_item_counts.iter().all(|&n| n >= 2)
+        }
+        _ => false,
+    };
+
+    let (col_edges, columns_from_text) = match (rect_col_edges, text_col_edges) {
+        (Some(rect_edges), text_edges_opt) if rect_cols_match_text => {
+            debug!(
+                "  cell-rect using {} rect-derived columns (text clusters: {}; rect cols well-distributed)",
+                rect_edges.len() - 1,
+                text_edges_opt
+                    .as_ref()
+                    .map(|e| (e.len() - 1) as i32)
+                    .unwrap_or(-1)
+            );
+            (rect_edges, false)
+        }
+        (Some(rect_edges), Some(text_edges)) if rect_edges.len() <= text_edges.len() => {
+            debug!(
+                "  cell-rect using {} rect-derived columns over {} text clusters",
+                rect_edges.len() - 1,
+                text_edges.len() - 1
+            );
+            (rect_edges, false)
+        }
+        (_, Some(text_edges)) => (text_edges, true),
+        (Some(rect_edges), None) => (rect_edges, false),
+        (None, None) => {
+            debug!(
+                "  cell-rect rejected: only {} columns from text clustering",
+                columns.len()
+            );
+            return None;
+        }
+    };
+
+    if col_edges.len() < 3 {
         return None;
     }
 
-    // Build column edges
-    let mut col_edges: Vec<f32> = Vec::with_capacity(columns.len() + 1);
-    let min_x = page_items.iter().map(|(_, i)| i.x).reduce(f32::min)?;
-    col_edges.push(min_x - 5.0);
-    for pair in columns.windows(2) {
-        col_edges.push((pair[0] + pair[1]) / 2.0);
-    }
-    let max_x_right = page_items
-        .iter()
-        .map(|(_, i)| i.x + i.width)
-        .reduce(f32::max)?;
-    col_edges.push(max_x_right + 5.0);
-
     let num_cols = col_edges.len() - 1;
     let num_rows = row_edges.len() - 1;
 
@@ -1617,12 +1720,25 @@ fn detect_row_stripe_table_from_cell_rects(
         page_items.len()
     );
 
-    let (cells, item_indices) = assign_items_to_grid(items, &col_edges, &row_edges, page);
+    let (mut cells, item_indices) = assign_items_to_grid(items, &col_edges, &row_edges, page);
 
     if item_indices.is_empty() {
         return None;
     }
 
+    let mut row_edges = row_edges;
+    let (collapsed_cells, collapsed_row_edges, collapsed_rows) =
+        collapse_multiline_description_rows(cells, row_edges, &col_edges);
+    let has_wrapped_description_rows = collapsed_rows > 0;
+    cells = collapsed_cells;
+    row_edges = collapsed_row_edges;
+    if collapsed_rows > 0 {
+        debug!(
+            "  cell-rect collapsed {} wrapped description rows",
+            collapsed_rows
+        );
+    }
+
     // Validate: >=2 non-empty rows, >=25% density
     let non_empty_rows = cells
         .iter()
@@ -1636,6 +1752,7 @@ fn detect_row_stripe_table_from_cell_rects(
         return None;
     }
 
+    let num_rows = cells.len();
     let total_cells = (num_cols * num_rows) as f32;
     let non_empty_cells = cells
         .iter()
@@ -1655,17 +1772,21 @@ fn detect_row_stripe_table_from_cell_rects(
         return None;
     }
 
-    // Reject tables with paragraph-length cells (layout backgrounds, not tables)
+    // Reject tables with paragraph-length cells — typically layout
+    // backgrounds (sidebars, banners) where a single big rectangle
+    // contains a wall of prose.  Spare multi-row key/value tables where
+    // the value column is a multi-bullet description: those pass every
+    // other gate and shouldn't get killed on cell length alone.
     let max_cell_len = cells
         .iter()
         .flat_map(|row| row.iter())
         .map(|c| c.len())
         .max()
         .unwrap_or(0);
-    if max_cell_len > 500 {
+    if max_cell_len > 500 && non_empty_rows < 4 {
         debug!(
-            "  cell-rect rejected: max cell length {} > 500",
-            max_cell_len
+            "  cell-rect rejected: max cell length {} > 500 ({} rows, layout background)",
+            max_cell_len, non_empty_rows
         );
         return None;
     }
@@ -1681,13 +1802,36 @@ fn detect_row_stripe_table_from_cell_rects(
 
     // Reject "tables" that are actually prose in a framed region.
     // Columns here come from text X-position clustering; when prose wraps
-    // inside a bounding-box rect (e.g. chat-transcript figures) the
-    // word-boundary gaps cluster into many spurious columns, and the
-    // resulting cells hold sentence fragments riddled with common English
-    // function words. Count cells with any such word and reject when
-    // 20%+ of non-empty cells match — real tabular data (labels, units,
-    // numbers) rarely contains these words.
-    if num_cols >= 4 {
+    // inside a bounding-box rect (e.g. chat-transcript figures, two-column
+    // legal-text blocks in forms) the word-boundary gaps cluster into
+    // spurious columns, and the resulting cells hold sentence fragments
+    // riddled with common English function words.
+    //
+    // Apply at any column count >= 2. The 2-col case is the bite — a
+    // paragraph wrapped into 2 justified columns produces the same
+    // surface signal as a real "label / value" table in the
+    // well-distributed-cols check (both cols populated), so we need a
+    // content-based signal to tell them apart.
+    //
+    // Layered checks combine after the 20%-of-cells prose-word
+    // trigger fires:
+    //   (a) Long-cell content: prose-in-a-frame averages ~70-100 chars
+    //       per non-empty cell (sentence fragments); real data tables
+    //       are typically <30 chars, occasionally up to ~55 for
+    //       descriptive 4-col tables. The 65-char threshold cleanly
+    //       separates them on observed fixtures (accessory_building
+    //       prose=74 chars, upstage data=53, greencomp=20). This
+    //       overrides the well-distributed relaxation — long cells
+    //       are the strongest prose signal even when both cols are
+    //       populated.
+    //   (b) Two-column text-only scaffold: when both columns were inferred
+    //       from text starts rather than rect edges, prose fragments can look
+    //       perfectly balanced. Require rect evidence for this relaxed shape.
+    //   (c) Well-distributed columns: ≥75% of cols hold ≥2 non-empty
+    //       cells. Catches the prose-paragraph-as-many-cols shape
+    //       while admitting real "label / value / description /
+    //       benefit"-style tables.
+    if num_cols >= 2 {
         const PROSE_WORDS: &[&str] = &[
             "a", "an", "the", "of", "to", "is", "was", "are", "were", "be", "been", "in", "on",
             "at", "with", "for", "by", "as", "and", "or", "but", "this", "that", "these", "those",
@@ -1697,6 +1841,7 @@ fn detect_row_stripe_table_from_cell_rects(
         ];
         let mut prose_cells = 0usize;
         let mut counted = 0usize;
+        let mut total_chars = 0usize;
         for row in &cells {
             for cell in row {
                 let t = cell.trim();
@@ -1704,6 +1849,7 @@ fn detect_row_stripe_table_from_cell_rects(
                     continue;
                 }
                 counted += 1;
+                total_chars += t.chars().count();
                 let lower = t.to_ascii_lowercase();
                 let has_prose_word = lower
                     .split(|c: char| !c.is_ascii_alphabetic() && c != '\'')
@@ -1714,11 +1860,65 @@ fn detect_row_stripe_table_from_cell_rects(
             }
         }
         if counted > 0 && prose_cells * 5 >= counted {
+            // (a) Long-cell content: overrides the well-distributed
+            // relaxation. The 2-col prose-in-a-frame case populates
+            // both cols (passes well-distributed) but every cell
+            // holds a sentence fragment, so mean cell length is the
+            // discriminator.
+            const PROSE_MEAN_CHAR_THRESHOLD: usize = 65;
+            let mean_chars = total_chars / counted;
+            if mean_chars > PROSE_MEAN_CHAR_THRESHOLD && !has_wrapped_description_rows {
+                debug!(
+                    "  cell-rect rejected: prose-in-frame, mean non-empty cell {} chars > {} (prose words {}/{})",
+                    mean_chars, PROSE_MEAN_CHAR_THRESHOLD, prose_cells, counted
+                );
+                return None;
+            } else if mean_chars > PROSE_MEAN_CHAR_THRESHOLD {
+                debug!(
+                    "  cell-rect prose check relaxed: wrapped description rows, mean {} chars (prose words {}/{})",
+                    mean_chars, prose_cells, counted
+                );
+            }
+
+            // (b) Two text-derived columns are not enough vector evidence once
+            // the content looks prose-like. Real 2-col rect tables still pass
+            // when the column scaffold comes from drawn cell geometry.
+            if columns_from_text && num_cols == 2 {
+                debug!(
+                    "  cell-rect rejected: prose-in-frame with text-derived 2-col scaffold (mean {} chars, prose words {}/{})",
+                    mean_chars, prose_cells, counted
+                );
+                return None;
+            }
+
+            // (c) Well-distributed columns.
+            let filled_cols = (0..num_cols)
+                .filter(|&c| {
+                    cells
+                        .iter()
+                        .filter(|row| {
+                            !row.get(c)
+                                .map(String::as_str)
+                                .unwrap_or("")
+                                .trim()
+                                .is_empty()
+                        })
+                        .count()
+                        >= 2
+                })
+                .count();
+            let well_distributed = filled_cols * 4 >= num_cols * 3;
+            if !well_distributed {
+                debug!(
+                    "  cell-rect rejected: {}/{} cells contain prose function words — likely prose ({}/{} cols filled, mean {} chars)",
+                    prose_cells, counted, filled_cols, num_cols, mean_chars
+                );
+                return None;
+            }
             debug!(
-                "  cell-rect rejected: {}/{} cells contain prose function words — likely prose",
-                prose_cells, counted
+                "  cell-rect prose check relaxed: {}/{} cols filled, mean {} chars — table-with-description-col",
+                filled_cols, num_cols, mean_chars
             );
-            return None;
         }
     }
 
@@ -1739,6 +1939,132 @@ fn detect_row_stripe_table_from_cell_rects(
     Some(Table::new(column_centers, row_centers, cells, item_indices))
 }
 
+/// Merge wrapped description-line bands back into their visual data rows.
+///
+/// Some Word/PDF exports draw enough rectangle geometry to prove a table exists
+/// but expose Y bands per wrapped text line instead of per cell row. In the
+/// common mapping-table shape, a narrow row-label column precedes one wide
+/// description column, and wrapped continuation bands have content only in that
+/// wide column. Merge only that high-confidence shape so framed prose still
+/// falls through the existing prose guards.
+fn collapse_multiline_description_rows(
+    cells: Vec<Vec<String>>,
+    row_edges: Vec<f32>,
+    col_edges: &[f32],
+) -> (Vec<Vec<String>>, Vec<f32>, usize) {
+    let num_rows = cells.len();
+    let num_cols = col_edges.len().saturating_sub(1);
+    if num_rows < 3 || num_cols < 3 || row_edges.len() != num_rows + 1 {
+        return (cells, row_edges, 0);
+    }
+
+    let table_width = col_edges[num_cols] - col_edges[0];
+    if table_width <= 0.0 {
+        return (cells, row_edges, 0);
+    }
+
+    let Some((description_col, description_width)) = (0..num_cols)
+        .map(|c| (c, col_edges[c + 1] - col_edges[c]))
+        .max_by(|a, b| a.1.total_cmp(&b.1))
+    else {
+        return (cells, row_edges, 0);
+    };
+
+    // Require a preceding row-label column. Without it (e.g. a prose frame
+    // split into text-start columns), "one populated wide column" is not enough
+    // evidence to find visual row starts safely.
+    if description_col == 0 || description_width < table_width * 0.35 {
+        return (cells, row_edges, 0);
+    }
+
+    let row_has_left_label = |row: &[String]| {
+        row.iter()
+            .take(description_col)
+            .any(|cell| !cell.trim().is_empty())
+    };
+    let labeled_rows = cells.iter().filter(|row| row_has_left_label(row)).count();
+    if labeled_rows < 2 {
+        return (cells, row_edges, 0);
+    }
+
+    let mut merged_rows = 0usize;
+    let mut wrapped_description_rows = 0usize;
+    let mut new_cells: Vec<Vec<String>> = Vec::with_capacity(num_rows);
+    let mut new_edges = Vec::with_capacity(row_edges.len());
+    new_edges.push(row_edges[0]);
+
+    for (row_idx, row) in cells.into_iter().enumerate() {
+        let desc_text = row
+            .get(description_col)
+            .map(String::as_str)
+            .unwrap_or("")
+            .trim();
+        let left_label = row_has_left_label(&row);
+        let non_desc_non_empty = row
+            .iter()
+            .enumerate()
+            .filter(|(col, cell)| *col != description_col && !cell.trim().is_empty())
+            .count();
+
+        // Wrapped continuation bands contain only description-column text.
+        // The preceding label/marker column is empty because the visual row's
+        // label cell spans the whole wrapped block.
+        let is_description_continuation = row_idx > 0
+            && !desc_text.is_empty()
+            && !left_label
+            && non_desc_non_empty == 0
+            && !new_cells.is_empty();
+
+        // Header cells are often split as "Controls" / "Version" in the first
+        // column while the other header labels sit on the first band.
+        let only_first_col = row
+            .iter()
+            .enumerate()
+            .all(|(col, cell)| col == 0 || cell.trim().is_empty());
+        let is_header_continuation = row_idx > 0
+            && only_first_col
+            && row
+                .first()
+                .is_some_and(|cell| !cell.trim().is_empty() && cell.chars().count() <= 24)
+            && !new_cells.is_empty()
+            && new_cells
+                .last()
+                .is_some_and(|prev| prev.iter().filter(|c| !c.trim().is_empty()).count() >= 2);
+
+        if is_description_continuation || is_header_continuation {
+            if let Some(prev) = new_cells.last_mut() {
+                for (col, cell) in row.iter().enumerate() {
+                    let text = cell.trim();
+                    if text.is_empty() {
+                        continue;
+                    }
+                    if !prev[col].trim().is_empty() {
+                        prev[col].push(' ');
+                    }
+                    prev[col].push_str(text);
+                }
+            }
+            merged_rows += 1;
+            if is_description_continuation {
+                wrapped_description_rows += 1;
+            }
+        } else {
+            if !new_cells.is_empty() {
+                new_edges.push(row_edges[row_idx]);
+            }
+            new_cells.push(row);
+        }
+    }
+
+    new_edges.push(*row_edges.last().unwrap());
+
+    if merged_rows == 0 || new_cells.len() < 2 || new_edges.len() != new_cells.len() + 1 {
+        return (new_cells, row_edges, 0);
+    }
+
+    (new_cells, new_edges, wrapped_description_rows)
+}
+
 /// Detect a table by merging all cluster rects into one group.
 ///
 /// This handles clip-path PDFs where each column's cell rects form a separate
@@ -1874,18 +2200,20 @@ fn detect_merged_cluster_table(
         return None;
     }
 
-    // Reject if any cell has excessive text — layout background rects produce
-    // "cells" containing paragraphs, not short data-table values.
+    // Reject if any cell has excessive text — layout background rects
+    // produce "cells" containing paragraphs, not short data-table values.
+    // Multi-row key/value tables can legitimately have one column of
+    // long descriptive text, so only reject narrow-row layouts here.
     let max_cell_len = cells
         .iter()
         .flat_map(|row| row.iter())
         .map(|c| c.len())
         .max()
         .unwrap_or(0);
-    if max_cell_len > 500 {
+    if max_cell_len > 500 && non_empty_rows < 4 {
         debug!(
-            "  merged-cluster rejected: max cell length {} > 500 (layout background)",
-            max_cell_len
+            "  merged-cluster rejected: max cell length {} > 500 ({} rows, layout background)",
+            max_cell_len, non_empty_rows
         );
         return None;
     }
@@ -2315,6 +2643,46 @@ mod tests {
         );
     }
 
+    #[test]
+    fn test_row_stripe_accepts_multi_row_key_value_long_cells() {
+        // Multi-row 2-column key/value table where one value cell holds
+        // a paragraph (>500 chars).  The old `max_cell_len > 500` check
+        // rejected this shape as a "layout background"; with the
+        // multi-row guard, it should be accepted.
+        let mut rects = Vec::new();
+        let row_h = 25.0_f32;
+        let y_top = 700.0_f32;
+        for i in 0..8 {
+            let y = y_top - (i as f32) * row_h;
+            rects.push((40.0, y, 510.0, row_h));
+        }
+        let mut items = Vec::new();
+        for i in 0..8 {
+            let row_center_y = y_top - (i as f32) * row_h + row_h / 2.0;
+            // Left column: short label
+            items.push(make_item(&format!("Field {}", i), 45.0, row_center_y, 10.0));
+            // Right column: short value, except the last row which is a paragraph
+            let value = if i == 7 {
+                "X".repeat(800)
+            } else {
+                "value".to_string()
+            };
+            items.push(make_item(&value, 300.0, row_center_y, 10.0));
+        }
+        let result = detect_row_stripe_table(&items, &rects, 1);
+        assert!(
+            result.is_some(),
+            "multi-row key/value table with one long cell should be accepted"
+        );
+        let t = result.unwrap();
+        assert!(
+            t.cells.len() >= 4,
+            "expected ≥4 rows, got {}",
+            t.cells.len()
+        );
+        assert_eq!(t.cells[0].len(), 2, "expected 2 columns");
+    }
+
     // --- propagate_merged_cells ---
 
     #[test]
@@ -2929,6 +3297,232 @@ mod tests {
         // If tables were detected, that's also acceptable
     }
 
+    #[test]
+    fn text_derived_two_col_prose_is_not_cell_rect_table() {
+        let page = 1;
+        let mut rects = Vec::new();
+        for row in 0..8 {
+            rects.push(PdfRect {
+                x: 50.0,
+                y: 100.0 + row as f32 * 20.0,
+                width: 180.0,
+                height: 18.0,
+                page,
+            });
+        }
+
+        let mut items = Vec::new();
+        let left = [
+            "the annual plan was revised",
+            "and the team noted changes",
+            "this section explains limits",
+            "with additional notes below",
+            "the policy was reviewed",
+            "and results are summarized",
+            "this appendix describes scope",
+            "with examples for reference",
+        ];
+        let right = [
+            "for each area in the review",
+            "as part of the assessment",
+            "that were applied in context",
+            "to support the conclusion",
+            "for use by the committee",
+            "as shown in the narrative",
+            "that remain under discussion",
+            "to clarify the method",
+        ];
+        for row in 0..8 {
+            let y = 104.0 + row as f32 * 20.0;
+            let mut left_item = make_item(left[row], 60.0, y, 9.0);
+            left_item.width = 50.0;
+            items.push(left_item);
+            let mut right_item = make_item(right[row], 150.0, y, 9.0);
+            right_item.width = 50.0;
+            items.push(right_item);
+        }
+
+        let (tables, _hints) = detect_tables_from_rects(&items, &rects, page);
+        assert!(
+            tables.is_empty(),
+            "text-derived two-column prose must not be accepted as a rect table; got {:?}",
+            tables
+                .iter()
+                .map(|t| (t.rows.len(), t.columns.len()))
+                .collect::<Vec<_>>()
+        );
+    }
+
+    #[test]
+    fn multiline_indented_description_rows_collapse_to_visual_rows() {
+        let page = 1;
+        let col_edges = [0.0, 60.0, 420.0, 460.0, 500.0, 540.0];
+        let row_edges = [
+            340.0, 320.0, 300.0, 270.0, 250.0, 230.0, 200.0, 180.0, 160.0,
+        ];
+
+        let mut rects = Vec::new();
+        for row in 0..row_edges.len() - 1 {
+            let y_top = row_edges[row];
+            let y_bot = row_edges[row + 1];
+            for col in 0..col_edges.len() - 1 {
+                rects.push((
+                    col_edges[col],
+                    y_bot,
+                    col_edges[col + 1] - col_edges[col],
+                    y_top - y_bot,
+                ));
+            }
+        }
+
+        let mut items = vec![
+            make_item("Controls", 8.0, 330.0, 9.0),
+            make_item("Control", 70.0, 330.0, 9.0),
+            make_item("IG 1", 428.0, 330.0, 9.0),
+            make_item("IG 2", 468.0, 330.0, 9.0),
+            make_item("IG 3", 508.0, 330.0, 9.0),
+            make_item("Version", 8.0, 310.0, 9.0),
+            make_item("v8", 20.0, 285.0, 9.0),
+            make_item(
+                "4.5 Implement and Manage a Firewall on End-User Devices",
+                70.0,
+                285.0,
+                9.0,
+            ),
+            make_item("*", 438.0, 285.0, 9.0),
+            make_item("*", 478.0, 285.0, 9.0),
+            make_item("*", 518.0, 285.0, 9.0),
+            make_item("v7", 20.0, 215.0, 9.0),
+            make_item(
+                "9.4 Apply Host-based Firewalls or Port-Filtering",
+                70.0,
+                215.0,
+                9.0,
+            ),
+            make_item("*", 478.0, 215.0, 9.0),
+            make_item("*", 518.0, 215.0, 9.0),
+        ];
+        items.push(make_item(
+            "Implement and manage a host-based firewall or port-filtering tool",
+            84.0,
+            260.0,
+            8.0,
+        ));
+        items.push(make_item(
+            "on end-user devices with a default-deny rule",
+            84.0,
+            240.0,
+            8.0,
+        ));
+        items.push(make_item(
+            "Apply host-based firewalls or port filtering tools on end systems",
+            84.0,
+            190.0,
+            8.0,
+        ));
+        items.push(make_item(
+            "and deny unauthorized network communication",
+            84.0,
+            170.0,
+            8.0,
+        ));
+
+        let table = detect_row_stripe_table_from_cell_rects(&items, &rects, page)
+            .expect("expected multiline description table");
+        assert_eq!(table.columns.len(), 5);
+        assert_eq!(
+            table.rows.len(),
+            3,
+            "wrapped lines should collapse to header plus two data rows"
+        );
+        assert_eq!(table.cells[0][0], "Controls Version");
+        assert!(table.cells[1][1].contains("host-based firewall"));
+        assert!(table.cells[1][1].contains("default-deny rule"));
+        assert!(table.cells[2][1].contains("deny unauthorized"));
+    }
+
+    /// Wire-bordered 4-column table whose header text is centered/right-aligned
+    /// inside each cell while the data is left-aligned: cluster_x_positions
+    /// merges adjacent columns (data Item→EAN gap is below threshold) and
+    /// drops the header-only x-clusters in the filter pass, leaving only 3
+    /// text-derived columns. Rect borders are 4 columns of ground truth.
+    /// Before the fix the cell-rect path preferred text edges when they were
+    /// the smaller set — losing a column. After the fix, 3+ rect columns
+    /// always win.
+    #[test]
+    fn wired_header_data_misaligned_keeps_all_columns_from_rects() {
+        let page = 1;
+        // 4 cols: Item | EAN | Nombre | Cant
+        let col_xs = [380.0_f32, 410.0, 470.0, 660.0, 700.0];
+        // Header + 9 data rows at 15pt tall each (y descending).
+        let row_ys: Vec<f32> = (0..=10).map(|r| 400.0 - 15.0 * r as f32).collect();
+
+        let mut rects: Vec<(f32, f32, f32, f32)> = Vec::new();
+        for r in 0..10 {
+            let y_top = row_ys[r];
+            let y_bot = row_ys[r + 1];
+            for c in 0..4 {
+                rects.push((col_xs[c], y_bot, col_xs[c + 1] - col_xs[c], y_top - y_bot));
+            }
+        }
+
+        let mut items: Vec<TextItem> = Vec::new();
+        // Header row (y ≈ 392.5): headers sit further to the right than data
+        // because they are centered/right-aligned in the cells.
+        items.push(make_item("Item", 389.0, 392.5, 9.0));
+        items.push(make_item("EAN", 432.0, 392.5, 9.0));
+        items.push(make_item("Nombre", 552.0, 392.5, 9.0));
+        items.push(make_item("Cant", 672.0, 392.5, 9.0));
+
+        let names = [
+            "Arnes Frontal",
+            "Arnes Motor",
+            "Arnes Piso",
+            "Arnes Techo",
+            "Arnes Puerta",
+            "Arnes Tablero",
+            "Arnes Trasero",
+            "Arnes Lateral",
+            "Arnes Sensor",
+        ];
+        for r in 0..9 {
+            let y = 377.5 - 15.0 * r as f32;
+            items.push(make_item(&(r + 1).to_string(), 396.0, y, 9.0));
+            items.push(make_item("7701023403016", 410.0, y, 9.0));
+            items.push(make_item(names[r], 480.0, y, 9.0));
+            items.push(make_item("1", 680.0, y, 9.0));
+        }
+
+        let table = detect_row_stripe_table_from_cell_rects(&items, &rects, page)
+            .expect("wired 4-column table with header/data x-misalignment must detect");
+        assert_eq!(
+            table.columns.len(),
+            4,
+            "expected 4 columns from rect borders; cells: {:?}",
+            table.cells
+        );
+        for c in 0..4 {
+            let any_populated = table.cells.iter().any(|row| !row[c].trim().is_empty());
+            assert!(
+                any_populated,
+                "column {} empty across all rows; cells: {:?}",
+                c, table.cells
+            );
+        }
+        // Header row populated in all 4 cells.
+        let header = &table.cells[0];
+        assert_eq!(header[0].trim(), "Item");
+        assert_eq!(header[1].trim(), "EAN");
+        assert_eq!(header[2].trim(), "Nombre");
+        assert_eq!(header[3].trim(), "Cant");
+        // First data row: Item="1", EAN, name, count="1" — no Item↔EAN merge.
+        let data1 = &table.cells[1];
+        assert_eq!(data1[0].trim(), "1");
+        assert_eq!(data1[1].trim(), "7701023403016");
+        assert!(data1[2].trim().contains("Arnes"));
+        assert_eq!(data1[3].trim(), "1");
+    }
+
     #[test]
     fn failed_cluster_no_hint_without_items() {
         // Rects with no text items inside → no failed-cluster hint generated.

src/tables/format.rs

@@ -160,6 +160,81 @@ fn starts_with_uppercase_word(cell: &str) -> bool {
         .is_some_and(|c| c.is_uppercase())
 }
 
+fn starts_with_uppercase_alpha(cell: &str) -> bool {
+    cell.chars()
+        .find(|c| c.is_alphabetic())
+        .is_some_and(|c| c.is_uppercase())
+}
+
+fn starts_with_lowercase_alpha(cell: &str) -> bool {
+    cell.chars()
+        .find(|c| c.is_alphabetic())
+        .is_some_and(|c| c.is_lowercase())
+}
+
+fn starts_with_numbered_label(cell: &str) -> bool {
+    let trimmed = cell.trim_start();
+    let digit_count = trimmed.chars().take_while(|c| c.is_ascii_digit()).count();
+
+    digit_count > 0
+        && digit_count <= 3
+        && trimmed
+            .chars()
+            .nth(digit_count)
+            .is_some_and(|c| matches!(c, '.' | ')' | '-' | ':'))
+}
+
+fn alpha_word_count(cell: &str) -> usize {
+    cell.split_whitespace()
+        .filter(|word| word.chars().any(|c| c.is_alphabetic()))
+        .count()
+}
+
+fn looks_like_compact_entry_label(cell: &str) -> bool {
+    let trimmed = cell.trim();
+    if trimmed.len() < 3 || trimmed.len() > 80 {
+        return false;
+    }
+
+    if !starts_with_uppercase_alpha(trimmed) && !starts_with_numbered_label(trimmed) {
+        return false;
+    }
+
+    if trimmed.ends_with(['.', ',', ';', ':']) {
+        return false;
+    }
+
+    let words = alpha_word_count(trimmed);
+    (1..=6).contains(&words)
+}
+
+fn looks_like_plain_section_label(cell: &str) -> bool {
+    let trimmed = cell.trim();
+    if trimmed.len() < 4 || trimmed.len() > 40 {
+        return false;
+    }
+    if trimmed.ends_with(['.', ',', ';', ':']) || trimmed.contains(|ch: char| ch.is_ascii_digit()) {
+        return false;
+    }
+    if trimmed.len() <= 4 && trimmed.chars().all(|ch| !ch.is_lowercase()) {
+        return false;
+    }
+    trimmed
+        .chars()
+        .all(|ch| ch.is_alphabetic() || ch.is_whitespace() || matches!(ch, '&' | '/' | '-'))
+        && starts_with_uppercase_alpha(trimmed)
+        && (1..=4).contains(&alpha_word_count(trimmed))
+}
+
+fn ends_like_incomplete_phrase(cell: &str) -> bool {
+    let lower = cell.trim_end().to_ascii_lowercase();
+    lower.ends_with(" and")
+        || lower.ends_with(" or")
+        || lower.ends_with(',')
+        || lower.ends_with('-')
+        || lower.ends_with('/')
+}
+
 /// 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();
@@ -185,6 +260,9 @@ fn clean_table_cells(cells: &[Vec<String>]) -> (Vec<Vec<String>>, Vec<String>) {
             continue;
         }
 
+        let num_cols = row.len();
+        let filled_cells = row.iter().filter(|c| !c.trim().is_empty()).count();
+
         // Check if this is a continuation row (first column is empty but others have content).
         // A row with only 1 short non-empty cell (besides the first) is more likely a
         // section sub-header (e.g. "JAN", "FEB") than overflow text — don't merge it.
@@ -222,31 +300,76 @@ fn clean_table_cells(cells: &[Vec<String>]) -> (Vec<Vec<String>>, Vec<String>) {
             .iter()
             .filter(|cell| starts_with_uppercase_word(cell))
             .count();
+        let first_non_empty_col = row.iter().position(|c| !c.trim().is_empty());
+        let first_non_empty_cell = first_non_empty_col
+            .and_then(|idx| row.get(idx))
+            .map(|c| c.trim())
+            .unwrap_or("");
+        let title_like_later_cells = first_non_empty_col
+            .map(|idx| {
+                row.iter()
+                    .skip(idx + 1)
+                    .map(|c| c.trim())
+                    .filter(|c| !c.is_empty() && starts_with_uppercase_alpha(c))
+                    .count()
+            })
+            .unwrap_or(0);
+        let prev_first_cell_empty = cleaned
+            .last()
+            .and_then(|r| r.first())
+            .is_some_and(|c| c.trim().is_empty());
+        let prev_first_cell = cleaned
+            .last()
+            .and_then(|r| r.first())
+            .map(|c| c.trim())
+            .unwrap_or("");
+        let header_filled = cleaned
+            .first()
+            .map(|r| r.iter().filter(|c| !c.trim().is_empty()).count())
+            .unwrap_or(num_cols);
         let looks_like_spanning_first_column_row = first_cell.is_empty()
             && row.len() >= 4
             && non_first_cells.len() == row.len().saturating_sub(1)
             && uppercase_leading_cells >= non_first_cells.len().saturating_sub(1);
+        // Hierarchical tables often use a row-spanned first column: sub-rows
+        // leave column 0 blank, then start a compact title-like label in
+        // column 1.  Wrapped continuations in the existing fixtures start
+        // mid-sentence/lowercase ("continued text here", "with 3.5%...") or
+        // carry lowercase fragments in the later cells, so keep those mergeable.
+        let looks_like_hierarchical_subrow = first_cell.is_empty()
+            && row.len() >= 3
+            && first_non_empty_col == Some(1)
+            && looks_like_compact_entry_label(first_non_empty_cell)
+            && ((non_first_cells.len() >= 2 && title_like_later_cells > 0)
+                || (non_first_cells.len() == 1
+                    && prev_first_cell_empty
+                    && alpha_word_count(first_non_empty_cell) >= 2));
+        let looks_like_new_first_column_entry = !first_cell.is_empty()
+            && (starts_with_numbered_label(first_cell) || starts_with_uppercase_alpha(first_cell))
+            && filled_cells >= 2
+            && non_first_cells
+                .iter()
+                .any(|cell| looks_like_compact_entry_label(cell));
+        let looks_like_section_label_row = !first_cell.is_empty()
+            && filled_cells == 1
+            && header_filled >= 3
+            && looks_like_plain_section_label(first_cell);
         // Classic continuation: first cell empty, content in other cells
         let is_classic_continuation = first_cell.is_empty()
             && !non_first_cells.is_empty()
             && !is_short_subheader
             && !looks_like_data_row
             && !looks_like_spanning_first_column_row
+            && !looks_like_hierarchical_subrow
             && cleaned.len() > 1;
 
         // Wrapped-cell continuation: row has fewer filled cells than the header
         // row, suggesting it's overflow text from the previous row's cells.
         // Only trigger when the previous row has significantly more filled cells.
-        let num_cols = row.len();
-        let filled_cells = row.iter().filter(|c| !c.trim().is_empty()).count();
         let prev_filled = cleaned
             .last()
             .map(|r| r.iter().filter(|c| !c.trim().is_empty()).count())
             .unwrap_or(0);
-        let header_filled = cleaned
-            .first()
-            .map(|r| r.iter().filter(|c| !c.trim().is_empty()).count())
-            .unwrap_or(num_cols);
         // Merge when the row has significantly fewer filled cells than header.
         // For wide tables (5+ cols), require ≤50% of header cells.
         // For narrow tables (2-4 cols), require fewer than header cells.
@@ -257,11 +380,18 @@ fn clean_table_cells(cells: &[Vec<String>]) -> (Vec<Vec<String>>, Vec<String>) {
         } else {
             header_filled.saturating_sub(1)
         };
+        let continues_wrapped_first_column_label = !first_cell.is_empty()
+            && starts_with_lowercase_alpha(first_cell)
+            && ends_like_incomplete_phrase(prev_first_cell);
         let is_wrapped_continuation = cleaned.len() > 1
             && filled_cells <= max_filled_for_merge
-            && prev_filled > filled_cells
+            && (prev_filled > filled_cells
+                || (continues_wrapped_first_column_label && prev_filled >= filled_cells))
             && !looks_like_data_row
             && !looks_like_spanning_first_column_row
+            && !looks_like_hierarchical_subrow
+            && !looks_like_new_first_column_entry
+            && !looks_like_section_label_row
             && !is_short_subheader;
 
         let is_continuation = is_classic_continuation || is_wrapped_continuation;
@@ -407,6 +537,44 @@ mod tests {
         assert!(cleaned[1][1].contains("continued text here"));
     }
 
+    #[test]
+    fn test_clean_table_cells_first_column_section_label_not_merged() {
+        let cells = vec![
+            vec![
+                "Properties".into(),
+                "Conditions".into(),
+                "Method".into(),
+                "Typical values".into(),
+                "Units".into(),
+            ],
+            vec![
+                "Melt Flow Rate".into(),
+                "230 C/2.16 kg".into(),
+                "ASTM D1238".into(),
+                "3.0".into(),
+                "g/10 min".into(),
+            ],
+            vec![
+                "Mechanical".into(),
+                "".into(),
+                "".into(),
+                "".into(),
+                "".into(),
+            ],
+            vec![
+                "Tensile Stress at Yield".into(),
+                "50 mm/min".into(),
+                "ASTM D638".into(),
+                "31".into(),
+                "MPa".into(),
+            ],
+        ];
+        let (cleaned, _) = clean_table_cells(&cells);
+
+        assert_eq!(cleaned.len(), 4);
+        assert_eq!(cleaned[2][0], "Mechanical");
+    }
+
     #[test]
     fn test_clean_table_cells_short_subheader_not_merged() {
         let cells = vec![
@@ -459,6 +627,100 @@ mod tests {
         assert_eq!(cleaned[2][1], "Uncertainty around other copies");
     }
 
+    #[test]
+    fn test_clean_table_cells_numbered_hierarchy_rows_not_overmerged() {
+        let cells = vec![
+            vec![
+                "Group".into(),
+                "Task".into(),
+                "Detail".into(),
+                "Benefit".into(),
+            ],
+            vec![
+                "1. Group alpha".into(),
+                "Task setup and".into(),
+                "Begin setup".into(),
+                "Faster start".into(),
+            ],
+            vec![
+                "".into(),
+                "management".into(),
+                "recommended profile".into(),
+                "with saved defaults".into(),
+            ],
+            vec![
+                "2. Group beta and".into(),
+                "Storage setup".into(),
+                "Provides upload tools".into(),
+                "".into(),
+            ],
+            vec![
+                "fine-tuning".into(),
+                "".into(),
+                "for filtered inputs".into(),
+                "service".into(),
+            ],
+            vec![
+                "".into(),
+                "Label workspace".into(),
+                "Creates review sets".into(),
+                "Lets teams review".into(),
+            ],
+            vec![
+                "".into(),
+                "Model training".into(),
+                "".into(),
+                "Supports custom model".into(),
+            ],
+        ];
+        let (cleaned, _) = clean_table_cells(&cells);
+
+        assert_eq!(cleaned.len(), 5);
+        assert_eq!(cleaned[1][0], "1. Group alpha");
+        assert_eq!(cleaned[1][1], "Task setup and management");
+        assert_eq!(cleaned[2][0], "2. Group beta and fine-tuning");
+        assert_eq!(cleaned[2][1], "Storage setup");
+        assert_eq!(cleaned[3][1], "Label workspace");
+        assert_eq!(cleaned[4][1], "Model training");
+    }
+
+    #[test]
+    fn test_clean_table_cells_partial_hierarchical_subrow_not_merged() {
+        let cells = vec![
+            vec![
+                "Group".into(),
+                "Task".into(),
+                "Detail".into(),
+                "Benefit".into(),
+            ],
+            vec![
+                "Group A".into(),
+                "Alpha task".into(),
+                "Initial detail".into(),
+                "Initial benefit".into(),
+            ],
+            vec![
+                "".into(),
+                "Beta task".into(),
+                "Parallel detail".into(),
+                "".into(),
+            ],
+            vec![
+                "".into(),
+                "second line".into(),
+                "additional detail".into(),
+                "".into(),
+            ],
+        ];
+        let (cleaned, _) = clean_table_cells(&cells);
+
+        assert_eq!(cleaned.len(), 3);
+        assert_eq!(cleaned[1][1], "Alpha task");
+        assert_eq!(cleaned[2][0], "");
+        assert_eq!(cleaned[2][1], "Beta task second line");
+        assert_eq!(cleaned[2][2], "Parallel detail additional detail");
+    }
+
     #[test]
     fn test_clean_table_cells_full_width_continuation_row_still_merges_when_lowercase() {
         let cells = vec![

src/tounicode.rs

@@ -329,7 +329,7 @@ impl ToUnicodeCMap {
                 if let (Some(start), Some(end), Some(base)) = (
                     parse_hex_u16(&start_hex),
                     parse_hex_u16(&end_hex),
-                    parse_hex_u32(&base_hex),
+                    hex_to_unicode_scalar(&base_hex),
                 ) {
                     self.ranges.push((start, end, base));
                 }
@@ -575,32 +575,86 @@ fn parse_hex_u16(hex: &str) -> Option<u16> {
     u16::from_str_radix(hex.trim(), 16).ok()
 }
 
-/// Parse a hex string to u32
-fn parse_hex_u32(hex: &str) -> Option<u32> {
-    u32::from_str_radix(hex.trim(), 16).ok()
-}
-
-/// Convert a hex string to a Unicode string
-/// Handles both 2-byte (BMP) and 4-byte (supplementary) codepoints
+/// Convert a ToUnicode destination hex string to Unicode.
+///
+/// PDF ToUnicode destinations are UTF-16BE strings. Supplementary-plane
+/// characters are encoded as surrogate pairs, so treating each 4-hex chunk as
+/// a scalar drops emoji like D83CDF1F.
 fn hex_to_unicode_string(hex: &str) -> Option<String> {
-    let hex = hex.trim();
-    let mut result = String::new();
-
-    // Process 4 hex digits at a time
-    let mut i = 0;
-    while i + 4 <= hex.len() {
-        if let Ok(cp) = u32::from_str_radix(&hex[i..i + 4], 16) {
-            if let Some(c) = char::from_u32(cp) {
-                result.push(c);
-            }
-        }
-        i += 4;
+    let hex: String = hex.chars().filter(|ch| !ch.is_ascii_whitespace()).collect();
+    if hex.is_empty() || !hex.len().is_multiple_of(2) {
+        return None;
     }
 
-    if result.is_empty() {
-        None
+    let bytes: Option<Vec<u8>> = (0..hex.len())
+        .step_by(2)
+        .map(|i| u8::from_str_radix(&hex[i..i + 2], 16).ok())
+        .collect();
+    let bytes = bytes?;
+
+    if bytes.len().is_multiple_of(2) {
+        let units: Vec<u16> = bytes
+            .chunks_exact(2)
+            .map(|chunk| u16::from_be_bytes([chunk[0], chunk[1]]))
+            .collect();
+        if let Ok(result) = String::from_utf16(&units) {
+            if !result.is_empty() {
+                return Some(normalize_tounicode_destination(result));
+            }
+        }
+    }
+
+    // Be permissive for non-standard one-byte destinations.
+    if bytes.len() == 1 {
+        let ch = bytes[0] as char;
+        if !ch.is_control() || ch == '\t' || ch == '\n' {
+            return Some(ch.to_string());
+        }
+    }
+
+    None
+}
+
+fn normalize_tounicode_destination(text: String) -> String {
+    let is_multi_char = text.chars().nth(1).is_some();
+
+    // Some malformed producer CMaps put a list of alternative whitespace or
+    // hyphen codepoints into one destination. Keep ordinary multi-character
+    // mappings intact unless that malformed signature is present.
+    if is_multi_char
+        && text.chars().all(char::is_whitespace)
+        && text.chars().any(|ch| matches!(ch, '\t' | '\n' | '\r'))
+    {
+        return if text.contains('\t') {
+            "\t".to_string()
+        } else {
+            " ".to_string()
+        };
+    }
+
+    if is_multi_char
+        && text.contains('\u{00ad}')
+        && text.chars().all(|ch| {
+            matches!(
+                ch,
+                '-' | '\u{00ad}' | '\u{2010}' | '\u{2011}' | '\u{2012}' | '\u{2013}' | '\u{2212}'
+            )
+        })
+    {
+        return "-".to_string();
+    }
+
+    text
+}
+
+fn hex_to_unicode_scalar(hex: &str) -> Option<u32> {
+    let text = hex_to_unicode_string(hex)?;
+    let mut chars = text.chars();
+    let ch = chars.next()?;
+    if chars.next().is_none() {
+        Some(ch as u32)
     } else {
-        Some(result)
+        None
     }
 }
 
@@ -2607,6 +2661,97 @@ endbfrange
         assert_eq!(cmap.lookup(0x0005), Some("C".to_string()));
     }
 
+    #[test]
+    fn test_parse_bfchar_surrogate_pair_emoji() {
+        let cmap_content = r#"
+1 begincodespacerange
+<00> <FF>
+endcodespacerange
+2 beginbfchar
+<16> <D83CDF1F>
+<9D> <D83CDFAD>
+endbfchar
+"#;
+        let cmap = ToUnicodeCMap::parse(cmap_content.as_bytes()).unwrap();
+
+        assert_eq!(cmap.code_byte_length, 1);
+        assert_eq!(cmap.lookup(0x16), Some("🌟".to_string()));
+        assert_eq!(cmap.lookup(0x9D), Some("🎭".to_string()));
+    }
+
+    #[test]
+    fn test_parse_bfrange_surrogate_pair_base() {
+        let cmap_content = r#"
+1 begincodespacerange
+<00> <FF>
+endcodespacerange
+1 beginbfrange
+<C8> <C9> <D83CDFD8>
+endbfrange
+"#;
+        let cmap = ToUnicodeCMap::parse(cmap_content.as_bytes()).unwrap();
+
+        assert_eq!(cmap.code_byte_length, 1);
+        assert_eq!(cmap.lookup(0xC8), Some("🏘".to_string()));
+        assert_eq!(cmap.lookup(0xC9), Some("🏙".to_string()));
+    }
+
+    #[test]
+    fn test_parse_bfrange_preserves_single_hyphen_like_base() {
+        let cmap_content = r#"
+1 begincodespacerange
+<00> <FF>
+endcodespacerange
+1 beginbfrange
+<21> <22> <2013>
+endbfrange
+"#;
+        let cmap = ToUnicodeCMap::parse(cmap_content.as_bytes()).unwrap();
+
+        assert_eq!(cmap.lookup(0x21), Some("–".to_string()));
+        assert_eq!(cmap.lookup(0x22), Some("—".to_string()));
+    }
+
+    #[test]
+    fn test_parse_spaced_destination_hex_without_control_noise() {
+        let cmap_content = r#"
+1 begincodespacerange
+<00> <FF>
+endcodespacerange
+3 beginbfchar
+<21> < 0009 000d 0020 00a0 >
+<22> < 002d 00ad 2010 >
+<23> <00a0>
+endbfchar
+"#;
+        let cmap = ToUnicodeCMap::parse(cmap_content.as_bytes()).unwrap();
+
+        assert_eq!(cmap.lookup(0x21), Some("\t".to_string()));
+        assert_eq!(cmap.lookup(0x22), Some("-".to_string()));
+        assert_eq!(cmap.lookup(0x23), Some("\u{00a0}".to_string()));
+    }
+
+    #[test]
+    fn test_parse_preserves_valid_multi_character_destinations() {
+        let cmap_content = r#"
+1 begincodespacerange
+<00> <FF>
+endcodespacerange
+4 beginbfchar
+<21> <002d002d>
+<22> <20132013>
+<23> <002000a0>
+<24> <00660069>
+endbfchar
+"#;
+        let cmap = ToUnicodeCMap::parse(cmap_content.as_bytes()).unwrap();
+
+        assert_eq!(cmap.lookup(0x21), Some("--".to_string()));
+        assert_eq!(cmap.lookup(0x22), Some("––".to_string()));
+        assert_eq!(cmap.lookup(0x23), Some(" \u{00a0}".to_string()));
+        assert_eq!(cmap.lookup(0x24), Some("fi".to_string()));
+    }
+
     #[test]
     fn test_remap_to_sequential() {
         // Simulate a broken CMap where GIDs are from pre-subsetting:

tests/snapshots/td9264.md

@@ -54,9 +54,14 @@ company other than a life insurance company shall make a return on Form 1120PC.
 
 annual statement (or a pro forma annual statement), including the underwriting and investment exhibit for the year covered by such return.
 
-||(3) Foreign insurance companies. The provisions of paragraphs (c)(1) and|
-|---|---|
-||(c)(2) of this section concerning the returns and statements of insurance companies subject to tax under section 801 or section 831 also apply to foreign insurance companies subject to tax under those sections, except that the copy of the annual statement required to be submitted with the return shall, in the case of a foreign insurance company that is not required to file an annual statement, be a copy of the pro forma annual statement relating to the United States business of such company. (4) Exception for insurance companies filing their Federal income tax returns electronically. If an insurance company described in paragraph (c)(1), (c)(2), or (c)(3) of this section files its Federal income tax return electronically, it should not include on or with such return its annual statement (or pro forma annual statement), or any portion thereof. Such statement must be available at all times for inspection by authorized Internal Revenue Service officers or employees and retained for so long as such statements may be material in the administration of any internal revenue law. See §1.6001-1(e). (5) Definition. For purposes of this section, the term annual statement means the annual statement, the form of which is approved by the National Association of Insurance Commissioners (NAIC), which is filed by an insurance company for the year with the insurance departments of States, Territories, and the District of|
+(3) Foreign insurance companies. The provisions of paragraphs (c)(1) and
+(c)(2) of this section concerning the returns and statements of insurance companies subject to tax under section 801 or section 831 also apply to foreign insurance companies subject to tax under those sections, except that the copy of the annual statement required to be submitted with the return shall, in the case of a foreign insurance company that is not required to file an annual statement, be a copy of the pro forma annual statement relating to the United States business of such company.
+(4) Exception for insurance companies filing their Federal income tax returns
+electronically. If an insurance company described in paragraph (c)(1), (c)(2), or
+
+(c)(3) of this section files its Federal income tax return electronically, it should not include on or with such return its annual statement (or pro forma annual statement), or any portion thereof. Such statement must be available at all times for inspection by authorized Internal Revenue Service officers or employees and retained for so long as such statements may be material in the administration of any internal revenue law. See §1.6001-1(e).
+(5) Definition. For purposes of this section, the term annual statement means
+the annual statement, the form of which is approved by the National Association of Insurance Commissioners (NAIC), which is filed by an insurance company for the year with the insurance departments of States, Territories, and the District of
 
 Columbia. The term annual statement also includes a pro forma annual statement if the insurance company is not required to file the NAIC annual statement.
 
@@ -201,3 +206,4 @@ CFR part or section where Current OMB identified or described control No.
 Deputy Commissioner for Services and Enforcement.
 
 Approved: May 19, 2006 Eric Solomon Acting Deputy Assistant Secretary of the Treasury (Tax Policy).
+