docs: note multi-row expansion scope

Clarify that the row-band cap intentionally keeps v1 focused on common small row-loss cases while larger compressions continue to use heuristic fallback.

Made-with: Cursor

napi/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@firecrawl/pdf-inspector",
-  "version": "1.5.0",
+  "version": "1.8.0",
   "description": "Fast PDF classification and text extraction. Detect text-based vs scanned PDFs, extract text by region with quality checks. Native Rust performance via napi-rs.",
   "main": "index.js",
   "types": "index.d.ts",

napi/src/lib.rs

@@ -99,6 +99,13 @@ 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
 // ---------------------------------------------------------------------------
@@ -317,6 +324,236 @@ 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`.
+///
+/// `structureTokens` and `cellBboxes` are typically produced by an external
+/// table-structure recognition model (e.g. SLANet on PaddleOCR) running on
+/// a rendered crop of the page. pdf-inspector uses the structure to lay out
+/// the cells and pulls the cell text from the native PDF — no OCR involved.
+#[napi(object)]
+pub struct TsrTableInputJs {
+    /// 0-indexed page number where the crop was taken from.
+    pub page: u32,
+    /// Crop bbox on the page, `[x1, y1, x2, y2]` in PDF points with
+    /// top-left origin.
+    pub crop_pdf_pt_bbox: Vec<f64>,
+    /// DPI the crop image was rendered at (e.g. `200.0`).
+    pub render_dpi: f64,
+    /// Raw structure tokens emitted by the TSR model, in document order.
+    pub structure_tokens: Vec<String>,
+    /// One bbox per cell (in document order). May be 4-element
+    /// `[x1,y1,x2,y2]` or 8-element 4-corner polygon, in crop image-pixel
+    /// space.
+    pub cell_bboxes: Vec<Vec<f64>>,
+}
+
+/// Extract markdown tables using externally-supplied structure recovery.
+///
+/// For each input, pairs structure tokens with cell bboxes (rowspan/colspan
+/// aware), converts each cell bbox from crop image-pixels into page PDF
+/// points, pulls the cell's text from the native PDF, and emits a markdown
+/// pipe-table.
+///
+/// Returns one markdown string per input, in input order.
+#[napi]
+pub fn extract_tables_with_structure(
+    buffer: Buffer,
+    inputs: Vec<TsrTableInputJs>,
+) -> Result<Vec<String>> {
+    let bytes: Vec<u8> = buffer.to_vec();
+    let parsed = parse_tsr_inputs(&inputs);
+
+    catch_panic("extract_tables_with_structure", move || {
+        pdf_inspector::extract_tables_with_structure_mem(&bytes, &parsed)
+            .map_err(|e| to_napi_err(e, "extract_tables_with_structure"))
+    })
+}
+
+/// One resolved cell from `extractTablesWithStructureCells`.
+#[napi(object)]
+pub struct StructuredCellJs {
+    /// 0-indexed grid row.
+    pub row: u32,
+    /// 0-indexed grid column.
+    pub col: u32,
+    /// 1 for a normal cell.
+    pub rowspan: u32,
+    /// 1 for a normal cell.
+    pub colspan: u32,
+    /// `true` when the cell is a `<th>` or sits inside `<thead>`.
+    pub is_header: bool,
+    /// Text extracted from the native PDF for this cell (may be empty).
+    pub text: String,
+    /// Axis-aligned bbox `[x1, y1, x2, y2]` in page PDF-points, top-left
+    /// origin. Useful for debug overlays or per-cell post-processing.
+    pub page_pt_bbox: Vec<f64>,
+}
+
+/// Extract structured cells using externally-supplied structure recovery.
+///
+/// Lower-level sibling of [`extractTablesWithStructure`]: instead of
+/// rendering markdown, returns the resolved cells (row, col, rowspan,
+/// colspan, isHeader, text, pagePtBbox) so callers can drive their own
+/// rendering, debug overlays, or per-cell post-processing.
+///
+/// Returns one `Array<StructuredCellJs>` per input, in input order.
+#[napi]
+pub fn extract_tables_with_structure_cells(
+    buffer: Buffer,
+    inputs: Vec<TsrTableInputJs>,
+) -> Result<Vec<Vec<StructuredCellJs>>> {
+    let bytes: Vec<u8> = buffer.to_vec();
+    let parsed = parse_tsr_inputs(&inputs);
+
+    catch_panic("extract_tables_with_structure_cells", move || {
+        let result = pdf_inspector::extract_tables_with_structure_cells_mem(&bytes, &parsed)
+            .map_err(|e| to_napi_err(e, "extract_tables_with_structure_cells"))?;
+        Ok(result
+            .into_iter()
+            .map(|cells| {
+                cells
+                    .into_iter()
+                    .map(|c| StructuredCellJs {
+                        row: c.row as u32,
+                        col: c.col as u32,
+                        rowspan: c.rowspan as u32,
+                        colspan: c.colspan as u32,
+                        is_header: c.is_header,
+                        text: c.text,
+                        page_pt_bbox: c.page_pt_bbox.iter().map(|v| *v as f64).collect(),
+                    })
+                    .collect()
+            })
+            .collect())
+    })
+}
+
+/// One result from `extractTablesWithStructureAuto` — markdown plus a
+/// diagnostic flag identifying which path produced it.
+///
+/// `fallbackReason` is `null` when the TSR-hybrid path produced the
+/// markdown directly. When stage 1's quality check fires (the cells
+/// look like a SLANet detection pathology — phantom rows or multi-row
+/// content in a single cell), the auto path may expand the TSR cells
+/// in-place or run the heuristic table extractor on the same region.
+/// `fallbackReason` carries the diagnostic label (for example
+/// `"multi_row_in_cell_expanded"` or `"phantom_empty_row"`).
+#[napi(object)]
+pub struct TableExtractionResultJs {
+    pub markdown: String,
+    pub fallback_reason: Option<String>,
+}
+
+/// Auto-fallback variant of [`extractTablesWithStructure`].
+///
+/// Runs the TSR-hybrid path, checks the resulting cells for known
+/// SLANet detection pathologies, expands multi-row cells in-place when
+/// possible, and otherwise falls back to the heuristic
+/// `extractTablesInRegions` for inputs where the TSR path looks
+/// compromised.
+///
+/// On clean inputs this returns identical markdown to
+/// `extractTablesWithStructure`; on flagged inputs `fallbackReason` is
+/// set to the recovery path that produced the result.
+#[napi]
+pub fn extract_tables_with_structure_auto(
+    buffer: Buffer,
+    inputs: Vec<TsrTableInputJs>,
+) -> Result<Vec<TableExtractionResultJs>> {
+    let bytes: Vec<u8> = buffer.to_vec();
+    let parsed = parse_tsr_inputs(&inputs);
+
+    catch_panic("extract_tables_with_structure_auto", move || {
+        let result = pdf_inspector::extract_tables_with_structure_auto_mem(&bytes, &parsed)
+            .map_err(|e| to_napi_err(e, "extract_tables_with_structure_auto"))?;
+        Ok(result
+            .into_iter()
+            .map(|r| TableExtractionResultJs {
+                markdown: r.markdown,
+                fallback_reason: r.fallback_reason,
+            })
+            .collect())
+    })
+}
+
+fn parse_tsr_inputs(inputs: &[TsrTableInputJs]) -> Vec<pdf_inspector::TsrTableInput> {
+    inputs
+        .iter()
+        .map(|i| {
+            let crop = if i.crop_pdf_pt_bbox.len() == 4 {
+                [
+                    i.crop_pdf_pt_bbox[0] as f32,
+                    i.crop_pdf_pt_bbox[1] as f32,
+                    i.crop_pdf_pt_bbox[2] as f32,
+                    i.crop_pdf_pt_bbox[3] as f32,
+                ]
+            } else {
+                [0.0, 0.0, 0.0, 0.0]
+            };
+            let cell_bboxes: Vec<Vec<f32>> = i
+                .cell_bboxes
+                .iter()
+                .map(|bb| bb.iter().map(|v| *v as f32).collect())
+                .collect();
+            pdf_inspector::TsrTableInput {
+                page: i.page,
+                crop_pdf_pt_bbox: crop,
+                render_dpi: i.render_dpi as f32,
+                structure_tokens: i.structure_tokens.clone(),
+                cell_bboxes,
+            }
+        })
+        .collect()
+}
+
 /// Per-page markdown extraction result.
 #[napi(object)]
 pub struct PageMarkdownResult {
@@ -360,9 +597,8 @@ pub fn extract_pages_markdown(
 ) -> Result<PagesExtractionResult> {
     let bytes: Vec<u8> = buffer.to_vec();
     catch_panic("extract_pages_markdown", move || {
-        let result =
-            pdf_inspector::extract_pages_markdown_mem(&bytes, pages.as_deref())
-                .map_err(|e| to_napi_err(e, "extract_pages_markdown"))?;
+        let result = pdf_inspector::extract_pages_markdown_mem(&bytes, pages.as_deref())
+            .map_err(|e| to_napi_err(e, "extract_pages_markdown"))?;
         Ok(PagesExtractionResult {
             pages: result
                 .pages

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/detect_pdf.rs

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

src/detector.rs

@@ -97,26 +97,9 @@ pub fn detect_pdf_type_with_config<P: AsRef<Path>>(
 ) -> Result<PdfTypeResult, PdfError> {
     crate::validate_pdf_file(&path)?;
 
-    // First, load metadata only (fast operation)
-    let metadata = match Document::load_metadata(&path) {
-        Ok(m) => m,
-        Err(ref e) if crate::is_encrypted_lopdf_error(e) => {
-            Document::load_metadata_with_password(&path, "")?
-        }
-        Err(e) => return Err(e.into()),
-    };
+    let (doc, page_count) = crate::load_document_from_path(&path)?;
 
-    // Then load the full document for content inspection
-    // We use filtered loading to skip heavy objects we don't need
-    let doc = match Document::load(&path) {
-        Ok(d) => d,
-        Err(ref e) if crate::is_encrypted_lopdf_error(e) => {
-            Document::load_with_password(&path, "")?
-        }
-        Err(e) => return Err(e.into()),
-    };
-
-    detect_from_document(&doc, metadata.page_count, &config)
+    detect_from_document(&doc, page_count, &config)
 }
 
 /// Detect PDF type from memory buffer
@@ -131,25 +114,64 @@ pub fn detect_pdf_type_mem_with_config(
 ) -> Result<PdfTypeResult, PdfError> {
     crate::validate_pdf_bytes(buffer)?;
 
-    // Load metadata first (fast)
-    let metadata = match Document::load_metadata_mem(buffer) {
-        Ok(m) => m,
-        Err(ref e) if crate::is_encrypted_lopdf_error(e) => {
-            Document::load_metadata_mem_with_password(buffer, "")?
-        }
-        Err(e) => return Err(e.into()),
-    };
+    let (doc, page_count) = crate::load_document_from_mem(buffer)?;
 
-    // Load document for inspection
-    let doc = match Document::load_mem(buffer) {
-        Ok(d) => d,
-        Err(ref e) if crate::is_encrypted_lopdf_error(e) => {
-            Document::load_mem_with_options(buffer, lopdf::LoadOptions::with_password(""))?
-        }
-        Err(e) => return Err(e.into()),
-    };
+    detect_from_document(&doc, page_count, &config)
+}
 
-    detect_from_document(&doc, metadata.page_count, &config)
+/// Heuristic page-count fallback for malformed PDFs that cannot be parsed.
+///
+/// This scans raw bytes for page dictionaries (`/Type /Page`) while excluding
+/// the page tree node (`/Type /Pages`). It is intended as a low-confidence hint
+/// for diagnostics; parsed page-tree counts remain authoritative.
+pub fn estimate_page_count_from_bytes(buffer: &[u8]) -> u32 {
+    let mut count = 0u32;
+    let mut pos = 0usize;
+
+    while let Some(rel_idx) = find_bytes(&buffer[pos..], b"/Type") {
+        let mut value_pos = pos + rel_idx + b"/Type".len();
+        value_pos = skip_pdf_whitespace(buffer, value_pos);
+
+        if buffer.get(value_pos) == Some(&b'/') {
+            let name_start = value_pos + 1;
+            let name_end = name_start + b"Page".len();
+            if name_end <= buffer.len()
+                && &buffer[name_start..name_end] == b"Page"
+                && buffer
+                    .get(name_end)
+                    .is_none_or(|b| is_pdf_name_delimiter(*b))
+            {
+                count += 1;
+            }
+        }
+
+        pos += rel_idx + b"/Type".len();
+    }
+
+    count
+}
+
+fn find_bytes(haystack: &[u8], needle: &[u8]) -> Option<usize> {
+    haystack.windows(needle.len()).position(|w| w == needle)
+}
+
+fn skip_pdf_whitespace(buffer: &[u8], mut pos: usize) -> usize {
+    while pos < buffer.len() && is_pdf_whitespace(buffer[pos]) {
+        pos += 1;
+    }
+    pos
+}
+
+fn is_pdf_whitespace(byte: u8) -> bool {
+    matches!(byte, b'\0' | b'\t' | b'\n' | 0x0C | b'\r' | b' ')
+}
+
+fn is_pdf_name_delimiter(byte: u8) -> bool {
+    is_pdf_whitespace(byte)
+        || matches!(
+            byte,
+            b'(' | b')' | b'<' | b'>' | b'[' | b']' | b'{' | b'}' | b'/' | b'%'
+        )
 }
 
 /// Detection logic on a pre-loaded document.

src/extractor/mod.rs

@@ -36,26 +36,14 @@ pub(crate) use layout::ColumnRegion;
 /// Extract text from PDF file as plain string
 pub fn extract_text<P: AsRef<Path>>(path: P) -> Result<String, PdfError> {
     crate::validate_pdf_file(&path)?;
-    let doc = match Document::load(&path) {
-        Ok(d) => d,
-        Err(ref e) if crate::is_encrypted_lopdf_error(e) => {
-            Document::load_with_password(&path, "")?
-        }
-        Err(e) => return Err(e.into()),
-    };
+    let (doc, _) = crate::load_document_from_path(&path)?;
     extract_text_from_doc(&doc)
 }
 
 /// Extract text from PDF memory buffer
 pub fn extract_text_mem(buffer: &[u8]) -> Result<String, PdfError> {
     crate::validate_pdf_bytes(buffer)?;
-    let doc = match Document::load_mem(buffer) {
-        Ok(d) => d,
-        Err(ref e) if crate::is_encrypted_lopdf_error(e) => {
-            Document::load_mem_with_options(buffer, lopdf::LoadOptions::with_password(""))?
-        }
-        Err(e) => return Err(e.into()),
-    };
+    let (doc, _) = crate::load_document_from_mem(buffer)?;
     extract_text_from_doc(&doc)
 }
 
@@ -91,13 +79,7 @@ pub(crate) fn extract_text_with_positions_and_rects<P: AsRef<Path>>(
     page_filter: Option<&HashSet<u32>>,
 ) -> Result<PageExtraction, PdfError> {
     crate::validate_pdf_file(&path)?;
-    let doc = match Document::load(&path) {
-        Ok(d) => d,
-        Err(ref e) if crate::is_encrypted_lopdf_error(e) => {
-            Document::load_with_password(&path, "")?
-        }
-        Err(e) => return Err(e.into()),
-    };
+    let (doc, _) = crate::load_document_from_path(&path)?;
     let font_cmaps = FontCMaps::from_doc(&doc);
     let (extraction, _thresholds, _gid_pages) =
         extract_positioned_text_from_doc(&doc, &font_cmaps, page_filter)?;
@@ -124,13 +106,7 @@ pub(crate) fn extract_text_with_positions_mem_and_rects(
     page_filter: Option<&HashSet<u32>>,
 ) -> Result<PageExtraction, PdfError> {
     crate::validate_pdf_bytes(buffer)?;
-    let doc = match Document::load_mem(buffer) {
-        Ok(d) => d,
-        Err(ref e) if crate::is_encrypted_lopdf_error(e) => {
-            Document::load_mem_with_options(buffer, lopdf::LoadOptions::with_password(""))?
-        }
-        Err(e) => return Err(e.into()),
-    };
+    let (doc, _) = crate::load_document_from_mem(buffer)?;
     let font_cmaps = FontCMaps::from_doc(&doc);
     let (extraction, _thresholds, _gid_pages) =
         extract_positioned_text_from_doc(&doc, &font_cmaps, page_filter)?;

src/tables/mod.rs

@@ -9,6 +9,7 @@ mod detect_struct;
 mod financial;
 mod format;
 mod grid;
+pub mod structured;
 
 pub use detect_heuristic::detect_tables;
 pub(crate) use detect_heuristic::is_table_of_contents;
@@ -17,6 +18,7 @@ pub(crate) use detect_rects::cluster_rects;
 pub use detect_rects::{detect_tables_from_rects, RectHintRegion};
 pub use detect_struct::detect_tables_from_struct_tree;
 pub use format::table_to_markdown;
+pub use structured::{cells_to_markdown, StructuredCell};
 
 use crate::types::TextItem;
 

src/tables/structured.rs

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