fix(pdf-inspector): recover key-value region tables

* emit ItemType::Image bboxes for Image XObjects (was: silently dropped)

Background. ItemType::Image, MarkdownOptions::include_images, and the
markdown emitter's image-collection path have all been in the tree
for a while, but no producer ever populated them — content_stream.rs
explicitly `// Skip images — text extraction only` at the Do
operator, and the nested Form-XObject walker in xobjects.rs only
matched XObjectType::Form, silently dropping Image entries. The
declared types were dead code.

This PR lights them up. At every Do that resolves to an Image
XObject (both top-level and nested inside Form XObjects), we now
compute the page-space bbox from the current CTM via a new
`image_bbox_from_ctm` helper — handling both axis-aligned and
rotated/sheared placements via 4-corner AABB — and emit a TextItem
with `item_type: ItemType::Image` and the legacy `[Image: <name>]`
text payload that the markdown emitter already knows how to render.

Callers can now find raster figures via `extract_text_with_positions`
(and the `_mem` variant, newly re-exported at the crate root) without
needing to re-parse the PDF or run a vision/layout model. The intended
consumer is layout-aware text pipelines that want to crop figures and
caption them out-of-band.

Two backstops to avoid silent breakage for existing callers:

  1. `MarkdownOptions::include_images` default flipped `true → false`.
     If it stayed at `true`, every existing user of
     `extract_pages_markdown` would suddenly see `![Image: Im0](image)`
     placeholders inserted throughout their output the moment they
     upgraded. Image data is still available structurally via
     `extract_text_with_positions`; rendering it into markdown is now
     an opt-in. New regression test asserts `extract_pages_markdown`
     output is unchanged for the image-bearing fixture.

  2. Image items now also skip the layout heuristics
     (`detect_columns`, `detect_tables_from_rects`) via a new
     `is_text_layout_item` predicate. Without this filter, an image's
     left edge would land in the column-projection profile and skew
     table column detection — surfaced by
     `vector_grid_tests::upstage_key_functions_four_cols` going from 4
     detected columns to 5 in CI before the filter was added.

Re-exporting `extract_text_with_positions_mem` at the crate root —
strictly additive; mirrors how `extract_pages_markdown_mem` is already
available there.

Tests:

  - test_extract_text_with_positions_emits_image_bboxes — minimal PDF
    with one 200×100 image at (50, 600); asserts one Image item with
    correct bbox + page + text.
  - test_image_xobject_bbox_handles_rotated_ctm — 90° rotated image
    via shear-component CTM; asserts AABB is correct (handles non-
    axis-aligned placements via 4-corner clamp).
  - test_image_emission_does_not_change_default_markdown — asserts no
    `Image:` token leaks into default markdown output, regression
    guard for the include_images flip.
  - test_markdown_options_default_has_include_images_false — explicit
    sentinel so anyone flipping it back catches it in CI.

* Bump version from 1.8.15 to 1.9.0

napi/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@firecrawl/pdf-inspector",
-  "version": "1.9.0",
+  "version": "1.9.3",
   "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",

src/lib.rs

@@ -823,7 +823,10 @@ pub fn extract_tables_in_regions_mem(
                     // symmetrically low under font-decode failure — this
                     // guard breaks that symmetry by comparing against
                     // bbox area, which is independent of extraction.
-                    if region_text_density_too_low(region_text_chars, region_area) {
+                    if source != TableCandidateSource::KeyValue
+                        && region_text_density_too_low(region_text_chars, region_area)
+                        && !markdown_table_body_is_dense(&md)
+                    {
                         return None;
                     }
                     let shape = markdown_table_shape(&md);
@@ -834,7 +837,11 @@ pub fn extract_tables_in_regions_mem(
                         Some(TableCandidateIssue::LineRowUndercount)
                     } else if wide_table_sparse_prefix_undercount(&md) {
                         Some(TableCandidateIssue::SparseWideUndercount)
-                    } else if text_cluster_column_undercount(&matched, shape) {
+                    } else if !matches!(
+                        source,
+                        TableCandidateSource::Line | TableCandidateSource::KeyValue
+                    ) && text_cluster_column_undercount(&matched, shape)
+                    {
                         Some(TableCandidateIssue::TextColumnUndercount)
                     } else if prose_grid_fragment_needs_ocr(&md) {
                         Some(TableCandidateIssue::ProseGridFragment)
@@ -877,6 +884,16 @@ pub fn extract_tables_in_regions_mem(
             {
                 candidates.push(candidate);
             }
+            if let Some(table) = tables::try_build_table_from_columns(&matched, page_1idx) {
+                if let Some(candidate) = evaluate(TableCandidateSource::Column, &table) {
+                    candidates.push(candidate);
+                }
+            }
+            if let Some(table) = tables::try_build_key_value_table_from_rows(&matched, page_1idx) {
+                if let Some(candidate) = evaluate(TableCandidateSource::KeyValue, &table) {
+                    candidates.push(candidate);
+                }
+            }
 
             match select_table_candidate(&candidates) {
                 Some(candidate) => page_results.push(RegionText {
@@ -3748,6 +3765,8 @@ enum TableCandidateSource {
     Rect,
     Line,
     Heuristic,
+    Column,
+    KeyValue,
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
@@ -3782,8 +3801,12 @@ fn select_table_candidate(candidates: &[TableCandidate]) -> Option<&TableCandida
     // serving a tidy-looking fragment.
     if first.issue == Some(TableCandidateIssue::LineRowUndercount) {
         return candidates.iter().find(|candidate| {
-            candidate.source == TableCandidateSource::Heuristic
-                && candidate.issue.is_none()
+            matches!(
+                candidate.source,
+                TableCandidateSource::Heuristic
+                    | TableCandidateSource::Column
+                    | TableCandidateSource::KeyValue
+            ) && candidate.issue.is_none()
                 && candidate.shape.cols * 10 >= first.shape.cols * 13
         });
     }
@@ -3801,14 +3824,31 @@ fn select_table_candidate(candidates: &[TableCandidate]) -> Option<&TableCandida
         TableCandidateSource::Rect | TableCandidateSource::Line
     ) {
         if let Some(heuristic) = candidates.iter().find(|candidate| {
-            candidate.source == TableCandidateSource::Heuristic
-                && candidate.issue.is_none()
+            matches!(
+                candidate.source,
+                TableCandidateSource::Heuristic
+                    | TableCandidateSource::Column
+                    | TableCandidateSource::KeyValue
+            ) && candidate.issue.is_none()
                 && heuristic_substantially_better(candidate.shape, accepted.shape)
         }) {
             accepted = heuristic;
         }
     }
 
+    if accepted.source == TableCandidateSource::Heuristic {
+        if let Some(layout_candidate) = candidates.iter().find(|candidate| {
+            matches!(
+                candidate.source,
+                TableCandidateSource::Column | TableCandidateSource::KeyValue
+            ) && candidate.issue.is_none()
+                && candidate.shape.cols >= accepted.shape.cols
+                && candidate.shape.rows > accepted.shape.rows
+        }) {
+            accepted = layout_candidate;
+        }
+    }
+
     Some(accepted)
 }
 
@@ -4253,12 +4293,21 @@ fn looks_like_partial_table_ex(markdown: &str, layout_assisted: bool) -> bool {
     }
 
     // Failure mode 2: header has empty cells in a multi-column table.
-    // When layout-assisted, allow up to 1 empty header cell (common in
-    // tables with merged/spanning header cells that we can't represent).
-    let empty_count = header_cells.iter().filter(|c| c.is_empty()).count();
+    // When layout-assisted, tolerate merged/spanning header gaps if the
+    // body is dense. Region bboxes from a layout model often start at a
+    // visual table whose header cannot be represented faithfully in a
+    // flat pipe table, while the body rows are still complete enough to use.
+    let header_empty_indices: Vec<usize> = header_cells
+        .iter()
+        .enumerate()
+        .filter_map(|(idx, cell)| cell.is_empty().then_some(idx))
+        .collect();
+    let empty_count = header_empty_indices.len();
     if layout_assisted {
-        // Reject only if >1 empty header cell (2+ means serious boundary issue)
-        if n_cols >= 3 && empty_count >= 2 {
+        if n_cols >= 3
+            && empty_count >= 2
+            && !layout_assisted_empty_header_has_dense_body(markdown, n_cols)
+        {
             return true;
         }
     } else if n_cols >= 3 && empty_count >= 1 {
@@ -4296,7 +4345,16 @@ fn looks_like_partial_table_ex(markdown: &str, layout_assisted: bool) -> bool {
             // (totals, subtotals) are common.
             let threshold = if layout_assisted { 2 } else { 3 };
             if n_cols >= 3 && empty_data * threshold >= n_cols {
-                return true;
+                let sparse_row_shares_header_spacer = layout_assisted
+                    && data_inner.iter().enumerate().any(|(idx, cell)| {
+                        cell.trim().is_empty() && header_empty_indices.contains(&idx)
+                    })
+                    && layout_assisted_empty_header_has_dense_body(markdown, n_cols);
+                let sparse_row_is_section_label = layout_assisted
+                    && layout_assisted_sparse_section_row_is_ok(data_inner, markdown, n_cols);
+                if !sparse_row_shares_header_spacer && !sparse_row_is_section_label {
+                    return true;
+                }
             }
         }
     }
@@ -4354,6 +4412,88 @@ fn looks_like_partial_table_ex(markdown: &str, layout_assisted: bool) -> bool {
     false
 }
 
+fn layout_assisted_empty_header_has_dense_body(markdown: &str, n_cols: usize) -> bool {
+    let rows = markdown_pipe_rows(markdown);
+    let data_rows: Vec<&Vec<&str>> = rows
+        .iter()
+        .skip(1)
+        .filter(|row| row.iter().any(|cell| !cell.trim().is_empty()))
+        .collect();
+    if data_rows.len() < 2 || n_cols < 3 {
+        return false;
+    }
+
+    let total_cells = data_rows.len() * n_cols;
+    let mut filled_cells = 0usize;
+    let mut rows_with_multiple_cells = 0usize;
+    let mut max_filled_in_row = 0usize;
+    for row in &data_rows {
+        let filled = row.iter().filter(|cell| !cell.trim().is_empty()).count();
+        filled_cells += filled;
+        max_filled_in_row = max_filled_in_row.max(filled);
+        if filled >= 2 {
+            rows_with_multiple_cells += 1;
+        }
+    }
+
+    // Dense enough to be a useful extraction despite lossy merged headers.
+    // The row-count gate avoids accepting a single tidy row under a broken
+    // header, and the density gate keeps sparse fragments on the OCR path.
+    rows_with_multiple_cells * 2 >= data_rows.len()
+        && max_filled_in_row >= n_cols.min(3)
+        && filled_cells * 100 >= total_cells * 45
+}
+
+fn layout_assisted_sparse_section_row_is_ok(row: &[&str], markdown: &str, n_cols: usize) -> bool {
+    let labels: Vec<&str> = row
+        .iter()
+        .map(|cell| cell.trim())
+        .filter(|cell| !cell.is_empty())
+        .collect();
+    if labels.len() != 1 {
+        return false;
+    }
+    let label = labels[0];
+    if label.len() > 40 || !label.chars().any(|ch| ch.is_alphabetic()) {
+        return false;
+    }
+    if label.ends_with('.') || label.ends_with('!') || label.ends_with('?') || label.ends_with(':')
+    {
+        return false;
+    }
+    layout_assisted_empty_header_has_dense_body(markdown, n_cols)
+}
+
+fn markdown_table_body_is_dense(markdown: &str) -> bool {
+    let rows = markdown_pipe_rows(markdown);
+    let data_rows: Vec<&Vec<&str>> = rows
+        .iter()
+        .skip(1)
+        .filter(|row| row.iter().any(|cell| !cell.trim().is_empty()))
+        .collect();
+    if data_rows.len() < 3 {
+        return false;
+    }
+
+    let cols = rows.iter().map(|row| row.len()).max().unwrap_or_default();
+    if cols < 3 {
+        return false;
+    }
+
+    let mut filled_cells = 0usize;
+    let mut rows_with_multiple_cells = 0usize;
+    for row in &data_rows {
+        let filled = row.iter().filter(|cell| !cell.trim().is_empty()).count();
+        filled_cells += filled;
+        if filled >= cols.min(3) {
+            rows_with_multiple_cells += 1;
+        }
+    }
+
+    let total_cells = data_rows.len() * cols;
+    rows_with_multiple_cells * 2 >= data_rows.len() && filled_cells * 100 >= total_cells * 45
+}
+
 /// Original strict validation (no layout assistance). Used by tests and
 /// full-page extraction paths that don't have layout model assistance.
 #[cfg(test)]
@@ -4705,6 +4845,16 @@ mod table_candidate_selection_tests {
         assert_eq!(selected.source, TableCandidateSource::Heuristic);
     }
 
+    #[test]
+    fn prefers_clean_column_fallback_when_it_recovers_more_rows() {
+        let candidates = vec![
+            candidate(TableCandidateSource::Heuristic, 6, 5, None),
+            candidate(TableCandidateSource::Column, 7, 5, None),
+        ];
+        let selected = select_table_candidate(&candidates).unwrap();
+        assert_eq!(selected.source, TableCandidateSource::Column);
+    }
+
     #[test]
     fn line_candidate_collapsing_captured_y_clusters_is_suspicious() {
         let long = "value value value value value value value value value value value value";
@@ -4809,7 +4959,9 @@ mod table_candidate_selection_tests {
 
 #[cfg(test)]
 mod looks_like_partial_table_tests {
-    use super::{looks_like_partial_table, looks_like_partial_table_ex};
+    use super::{
+        looks_like_partial_table, looks_like_partial_table_ex, markdown_table_body_is_dense,
+    };
 
     #[test]
     fn good_table_passes() {
@@ -4944,11 +5096,29 @@ mod looks_like_partial_table_tests {
 
     #[test]
     fn two_empty_headers_still_rejected_when_layout_assisted() {
-        // 2+ empty headers is still bad even with layout assistance.
+        // A single tidy row is not enough evidence to trust a badly gapped header.
         let md = "|A|||D|\n|---|---|---|---|\n|x|y|z|w|";
         assert!(
             looks_like_partial_table_ex(md, true),
-            "2 empty headers rejected even layout-assisted"
+            "2 empty headers with only one body row are rejected even layout-assisted"
+        );
+    }
+
+    #[test]
+    fn dense_body_with_empty_merged_header_passes_when_layout_assisted() {
+        let md = "|Year||Unadjusted Basis|||\n\
+                  |---|---|---|---|---|\n\
+                  |1|.1667|$100,000|$16,670|$16,670|\n\
+                  |2|.3333|$100,000|$33,330|$50,000|\n\
+                  |3|.3333|$100,000|$33,330|$88,330|\n\
+                  |4|.1667|$100,000|$16,670|$100,000|";
+        assert!(
+            looks_like_partial_table(md),
+            "strict mode still rejects merged-header gaps"
+        );
+        assert!(
+            !looks_like_partial_table_ex(md, true),
+            "layout-assisted should trust a dense body under a merged header"
         );
     }
 
@@ -4973,6 +5143,40 @@ mod looks_like_partial_table_tests {
         );
     }
 
+    #[test]
+    fn sparse_first_row_with_header_spacer_passes_when_layout_assisted() {
+        let md = "|Properties|Instruction||Training Datasets Alignment|\n\
+                  |---|---|---|---|\n\
+                  ||Alpaca-GPT4 OpenOrca Synth. Math-Instruct||Orca DPO Pairs Ultrafeedback Cleaned|\n\
+                  |Total # Samples|52K 2.91M 126K||12.9K 60.8K 126K|";
+        assert!(
+            looks_like_partial_table(md),
+            "strict mode rejects the sparse first row"
+        );
+        assert!(
+            !looks_like_partial_table_ex(md, true),
+            "layout-assisted should allow sparse rows that share a header spacer column"
+        );
+    }
+
+    #[test]
+    fn sparse_section_row_passes_when_layout_assisted_body_is_dense() {
+        let md = "|Properties|Conditions|Method|Typical values|Units|\n\
+                  |---|---|---|---|---|\n\
+                  |Rheology|||||\n\
+                  |Melt Flow Rate|230 C/2.16 kg|ASTM D1238|3.0|g/10 min|\n\
+                  |Tensile Stress at Yield|50 mm/min|ASTM D638|31|MPa|\n\
+                  |Elongation at Yield|50 mm/min|ASTM D638|8|%|";
+        assert!(
+            looks_like_partial_table(md),
+            "strict mode rejects the sparse first row"
+        );
+        assert!(
+            !looks_like_partial_table_ex(md, true),
+            "layout-assisted should allow a short section label above dense table rows"
+        );
+    }
+
     #[test]
     fn paragraph_still_rejected_when_layout_assisted() {
         // Paragraph detection is not relaxed — it's a genuine extraction issue.
@@ -5026,6 +5230,27 @@ mod looks_like_partial_table_tests {
             "duplicate headers rejected even layout-assisted"
         );
     }
+
+    #[test]
+    fn dense_numeric_table_body_is_structurally_trusted() {
+        let md = "|Year|3-Year|5-Year|7-Year|\n\
+                  |---|---|---|---|\n\
+                  |1|33.0%|20.00%|14.29%|\n\
+                  |2|44.45%|32.00%|24.49%|\n\
+                  |3|14.81%|19.20%|17.49%|\n\
+                  |4|7.41%|11.52%|12.49%|";
+        assert!(markdown_table_body_is_dense(md));
+    }
+
+    #[test]
+    fn sparse_markdown_fragment_is_not_structurally_trusted() {
+        let md = "|A|B|C|D|\n\
+                  |---|---|---|---|\n\
+                  |x||||\n\
+                  |||y||\n\
+                  ||||z|";
+        assert!(!markdown_table_body_is_dense(md));
+    }
 }
 
 /// Analyse extracted items and rects for layout complexity.

src/tables/format.rs

@@ -208,6 +208,24 @@ fn looks_like_compact_entry_label(cell: &str) -> bool {
     (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")
@@ -305,6 +323,10 @@ fn clean_table_cells(cells: &[Vec<String>]) -> (Vec<Vec<String>>, Vec<String>) {
             .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)
@@ -328,6 +350,10 @@ fn clean_table_cells(cells: &[Vec<String>]) -> (Vec<Vec<String>>, Vec<String>) {
             && 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()
@@ -344,10 +370,6 @@ fn clean_table_cells(cells: &[Vec<String>]) -> (Vec<Vec<String>>, Vec<String>) {
             .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.
@@ -369,6 +391,7 @@ fn clean_table_cells(cells: &[Vec<String>]) -> (Vec<Vec<String>>, Vec<String>) {
             && !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;
@@ -514,6 +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![

src/tables/mod.rs

@@ -460,6 +460,7 @@ pub(crate) fn try_build_table_from_columns(items: &[TextItem], page: u32) -> Opt
             item_indices.push(item_idx);
         }
     }
+    merge_superscript_marker_rows(&mut row_ys, &mut cells);
 
     // Validate: need reasonable fill rate
     let total_cells = row_ys.len() * columns.len();
@@ -547,6 +548,536 @@ pub(crate) fn try_build_table_from_columns(items: &[TextItem], page: u32) -> Opt
     Some(Table::new(col_xs, row_ys, cells, item_indices))
 }
 
+/// Build a region-scoped two-column key/value table from text baselines.
+///
+/// This intentionally lives outside the full-page heuristic detector. Layout
+/// callers already supplied a table-shaped bbox, and some real table regions
+/// are plain product/spec forms with only two visual columns. The main column
+/// fallback starts at four columns to avoid newspaper/prose false positives;
+/// this path keeps tighter key/value-specific guards instead.
+pub(crate) fn try_build_key_value_table_from_rows(items: &[TextItem], page: u32) -> Option<Table> {
+    let page_items: Vec<RowItem> = items
+        .iter()
+        .enumerate()
+        .filter(|(_, item)| item.page == page && !item.text.trim().is_empty())
+        .map(|(idx, item)| RowItem {
+            index: idx,
+            item: item.clone(),
+        })
+        .collect();
+
+    if page_items.len() < 4 {
+        return None;
+    }
+
+    let median_font_size = median_f32(page_items.iter().map(|ri| ri.item.font_size).collect())
+        .unwrap_or(10.0)
+        .max(1.0);
+    let y_tol = (median_font_size * 0.75).clamp(4.0, 9.0);
+    let rows = group_key_value_visual_rows(page_items, y_tol);
+    if rows.len() < 2 || rows.len() > 80 {
+        return None;
+    }
+
+    let split_x = infer_key_value_split_x(&rows, median_font_size)?;
+    let mut kv_rows: Vec<KeyValueRow> = Vec::new();
+    let mut paired_rows = 0usize;
+    let mut section_rows = 0usize;
+    let mut left_label_like = 0usize;
+    let mut left_starts = Vec::new();
+    let mut right_starts = Vec::new();
+
+    for row in &rows {
+        let mut left_items = Vec::new();
+        let mut right_items = Vec::new();
+        for item in &row.items {
+            if item.item.x < split_x {
+                left_items.push(item);
+            } else {
+                right_items.push(item);
+            }
+        }
+
+        let left = join_row_item_text(&left_items);
+        let right = join_row_item_text(&right_items);
+        if left.is_empty() && right.is_empty() {
+            continue;
+        }
+
+        let mut item_indices: Vec<usize> = row.items.iter().map(|ri| ri.index).collect();
+        item_indices.sort_unstable();
+        item_indices.dedup();
+
+        if !left.is_empty() && !right.is_empty() {
+            paired_rows += 1;
+            if looks_like_key_value_label(&left) {
+                left_label_like += 1;
+            }
+            if let Some(x) = left_items.first().map(|ri| ri.item.x) {
+                left_starts.push(x);
+            }
+            if let Some(x) = right_items.first().map(|ri| ri.item.x) {
+                right_starts.push(x);
+            }
+        } else if !left.is_empty() {
+            section_rows += 1;
+        }
+
+        kv_rows.push(KeyValueRow {
+            y: row.y,
+            left,
+            right,
+            item_indices,
+        });
+    }
+
+    if kv_rows.len() < 2 || paired_rows < 2 {
+        return None;
+    }
+
+    let header_inferred = key_value_first_pair_is_header(&kv_rows);
+    let data_pairs = if header_inferred {
+        paired_rows.saturating_sub(1)
+    } else {
+        paired_rows
+    };
+    if data_pairs < 1 {
+        return None;
+    }
+
+    if section_rows > paired_rows * 2 + 2 {
+        return None;
+    }
+
+    let label_rows_for_score = if header_inferred {
+        paired_rows.saturating_sub(1)
+    } else {
+        paired_rows
+    };
+    let label_like_for_score = if header_inferred && !kv_rows.is_empty() {
+        left_label_like.saturating_sub(1)
+    } else {
+        left_label_like
+    };
+    if label_rows_for_score >= 2 && label_like_for_score * 2 < label_rows_for_score {
+        return None;
+    }
+
+    let left_x = median_f32(left_starts).unwrap_or_else(|| {
+        rows.iter()
+            .flat_map(|row| row.items.iter().map(|ri| ri.item.x))
+            .fold(f32::INFINITY, f32::min)
+    });
+    let right_x = median_f32(right_starts).unwrap_or(split_x);
+    if !left_x.is_finite() || !right_x.is_finite() || right_x - left_x < 40.0 {
+        return None;
+    }
+    let right_cluster_count = significant_side_x_clusters(&rows, split_x, false);
+    let marker_rows = marker_matrix_value_rows(&kv_rows);
+    if (right_cluster_count >= 5 && paired_rows >= 3)
+        || (right_cluster_count >= 3 && marker_rows >= 3 && marker_rows * 2 >= paired_rows)
+    {
+        return None;
+    }
+
+    if key_value_rows_look_like_prose(&kv_rows, header_inferred) {
+        return None;
+    }
+
+    let mut table_rows = Vec::new();
+    let mut cells = Vec::new();
+    let mut item_indices = Vec::new();
+
+    let mut start_idx = 0usize;
+    if header_inferred {
+        let header = &kv_rows[0];
+        table_rows.push(header.y);
+        cells.push(vec![header.left.clone(), header.right.clone()]);
+        item_indices.extend(header.item_indices.iter().copied());
+        start_idx = 1;
+    } else {
+        table_rows.push(kv_rows.first().map(|row| row.y + y_tol).unwrap_or(0.0));
+        cells.push(vec!["Field".to_string(), "Value".to_string()]);
+    }
+
+    for row in kv_rows.iter().skip(start_idx) {
+        if !row.left.is_empty() && !row.right.is_empty() {
+            table_rows.push(row.y);
+            cells.push(vec![row.left.clone(), row.right.clone()]);
+            item_indices.extend(row.item_indices.iter().copied());
+        } else if !row.left.is_empty() {
+            table_rows.push(row.y);
+            cells.push(vec!["Section".to_string(), row.left.clone()]);
+            item_indices.extend(row.item_indices.iter().copied());
+        } else if !row.right.is_empty() {
+            if let Some(last) = cells.last_mut() {
+                if let Some(value) = last.get_mut(1) {
+                    if !value.trim().is_empty() {
+                        value.push(' ');
+                    }
+                    value.push_str(&row.right);
+                    item_indices.extend(row.item_indices.iter().copied());
+                }
+            }
+        }
+    }
+
+    if cells.len() < 2 {
+        return None;
+    }
+
+    item_indices.sort_unstable();
+    item_indices.dedup();
+
+    log::debug!(
+        "key-value table: {} rows, pairs={}, sections={}, split_x={:.1}",
+        cells.len(),
+        paired_rows,
+        section_rows,
+        split_x
+    );
+
+    Some(Table::new(
+        vec![left_x, right_x],
+        table_rows,
+        cells,
+        item_indices,
+    ))
+}
+
+#[derive(Debug, Clone)]
+struct RowItem {
+    index: usize,
+    item: TextItem,
+}
+
+#[derive(Debug, Clone)]
+struct VisualRow {
+    y: f32,
+    items: Vec<RowItem>,
+}
+
+#[derive(Debug, Clone)]
+struct KeyValueRow {
+    y: f32,
+    left: String,
+    right: String,
+    item_indices: Vec<usize>,
+}
+
+fn group_key_value_visual_rows(mut items: Vec<RowItem>, y_tol: f32) -> Vec<VisualRow> {
+    items.sort_by(|a, b| {
+        b.item
+            .y
+            .total_cmp(&a.item.y)
+            .then_with(|| a.item.x.total_cmp(&b.item.x))
+    });
+
+    let mut rows: Vec<VisualRow> = Vec::new();
+    for row_item in items {
+        if let Some(row) = rows
+            .iter_mut()
+            .find(|row| (row.y - row_item.item.y).abs() <= y_tol)
+        {
+            let len = row.items.len() as f32;
+            row.y = (row.y * len + row_item.item.y) / (len + 1.0);
+            row.items.push(row_item);
+            continue;
+        }
+
+        rows.push(VisualRow {
+            y: row_item.item.y,
+            items: vec![row_item],
+        });
+    }
+
+    for row in &mut rows {
+        row.items.sort_by(|a, b| a.item.x.total_cmp(&b.item.x));
+    }
+    rows.sort_by(|a, b| b.y.total_cmp(&a.y));
+    rows
+}
+
+fn infer_key_value_split_x(rows: &[VisualRow], median_font_size: f32) -> Option<f32> {
+    let min_gap = (median_font_size * 2.0).max(24.0);
+    let mut splits = Vec::new();
+
+    for row in rows {
+        if row.items.len() < 2 {
+            continue;
+        }
+
+        let mut best_gap = 0.0f32;
+        let mut best_split = None;
+        for pair in row.items.windows(2) {
+            let left = &pair[0].item;
+            let right = &pair[1].item;
+            let left_right = left.x + left.width.max(0.0);
+            let gap = right.x - left_right;
+            if gap > best_gap {
+                best_gap = gap;
+                best_split = Some(left_right + gap / 2.0);
+            }
+        }
+
+        if best_gap >= min_gap {
+            if let Some(split) = best_split {
+                splits.push(split);
+            }
+        }
+    }
+
+    if splits.len() < 2 {
+        return None;
+    }
+
+    median_f32(splits)
+}
+
+fn join_row_item_text(items: &[&RowItem]) -> String {
+    let mut parts = Vec::new();
+    for item in items {
+        let trimmed = item.item.text.trim();
+        if !trimmed.is_empty() {
+            parts.push(trimmed);
+        }
+    }
+    normalize_cell_text(&parts.join(" "))
+}
+
+fn normalize_cell_text(text: &str) -> String {
+    text.split_whitespace().collect::<Vec<_>>().join(" ")
+}
+
+fn key_value_first_pair_is_header(rows: &[KeyValueRow]) -> bool {
+    let Some(first) = rows.first() else {
+        return false;
+    };
+    if first.left.is_empty() || first.right.is_empty() {
+        return false;
+    }
+    if !looks_like_key_value_header_cell(&first.left)
+        || !looks_like_key_value_header_cell(&first.right)
+    {
+        return false;
+    }
+    rows.iter()
+        .skip(1)
+        .any(|row| !row.left.is_empty() && !row.right.is_empty())
+}
+
+fn looks_like_key_value_header_cell(cell: &str) -> bool {
+    let trimmed = cell.trim();
+    if trimmed.len() < 2 || trimmed.len() > 40 {
+        return false;
+    }
+    let words = word_count_simple(trimmed);
+    if !(1..=4).contains(&words) {
+        return false;
+    }
+    let lower = trimmed.to_ascii_lowercase();
+    if matches!(
+        lower.as_str(),
+        "yes" | "no" | "true" | "false" | "none" | "n/a" | "na"
+    ) {
+        return false;
+    }
+    trimmed.chars().any(|c| c.is_alphabetic())
+        && !trimmed.chars().any(|c| c.is_ascii_digit())
+        && !trimmed.ends_with(['.', ',', ';', ':'])
+}
+
+fn looks_like_key_value_label(cell: &str) -> bool {
+    let trimmed = cell.trim();
+    if trimmed.len() < 2 || trimmed.len() > 90 {
+        return false;
+    }
+    let words = word_count_simple(trimmed);
+    if words == 0 || words > 10 {
+        return false;
+    }
+    if trimmed.ends_with(['.', ',', ';']) {
+        return false;
+    }
+    trimmed.chars().any(|c| c.is_alphabetic())
+}
+
+fn key_value_rows_look_like_prose(rows: &[KeyValueRow], header_inferred: bool) -> bool {
+    let mut long_sentence_cells = 0usize;
+    let mut total_cells = 0usize;
+    let mut total_chars = 0usize;
+    let mut paired_rows = 0usize;
+    let mut solo_prose_rows = 0usize;
+
+    for row in rows.iter().skip(usize::from(header_inferred)) {
+        if !row.left.is_empty() && !row.right.is_empty() {
+            paired_rows += 1;
+        } else {
+            let solo = if row.left.is_empty() {
+                row.right.trim()
+            } else {
+                row.left.trim()
+            };
+            if solo.chars().count() > 70
+                || word_count_simple(solo) > 9
+                || (solo.chars().count() > 35 && solo.ends_with(['.', '!', '?']))
+            {
+                solo_prose_rows += 1;
+            }
+        }
+        for cell in [&row.left, &row.right] {
+            let trimmed = cell.trim();
+            if trimmed.is_empty() {
+                continue;
+            }
+            total_cells += 1;
+            total_chars += trimmed.chars().count();
+            if trimmed.chars().count() > 100
+                || (trimmed.chars().count() > 55 && trimmed.ends_with(['.', '!', '?']))
+            {
+                long_sentence_cells += 1;
+            }
+        }
+    }
+
+    if paired_rows < 1 || total_cells == 0 {
+        return true;
+    }
+    if solo_prose_rows >= 3 {
+        return true;
+    }
+
+    let avg_chars = total_chars as f32 / total_cells as f32;
+    avg_chars > 75.0 || long_sentence_cells * 2 >= total_cells
+}
+
+fn marker_matrix_value_rows(rows: &[KeyValueRow]) -> usize {
+    rows.iter()
+        .filter(|row| !row.left.is_empty() && compact_marker_value(&row.right))
+        .count()
+}
+
+fn compact_marker_value(cell: &str) -> bool {
+    let trimmed = cell.trim();
+    if trimmed.is_empty() || trimmed.chars().count() > 80 {
+        return false;
+    }
+    if trimmed.chars().any(|ch| ch.is_alphabetic()) {
+        return false;
+    }
+    trimmed
+        .chars()
+        .any(|ch| ch.is_ascii_digit() || matches!(ch, '•' | '●' | '·'))
+}
+
+fn significant_side_x_clusters(rows: &[VisualRow], split_x: f32, left_side: bool) -> usize {
+    let mut xs = Vec::new();
+    for row in rows {
+        for item in &row.items {
+            let is_left = item.item.x < split_x;
+            if is_left == left_side {
+                xs.push(item.item.x);
+            }
+        }
+    }
+    xs.sort_by(|a, b| a.total_cmp(b));
+
+    let mut counts = Vec::new();
+    let mut center = None::<f32>;
+    let mut count = 0usize;
+    for x in xs {
+        match center {
+            Some(current) if (x - current).abs() <= 8.0 => {
+                center = Some((current * count as f32 + x) / (count as f32 + 1.0));
+                count += 1;
+            }
+            Some(_) => {
+                counts.push(count);
+                center = Some(x);
+                count = 1;
+            }
+            None => {
+                center = Some(x);
+                count = 1;
+            }
+        }
+    }
+    if count > 0 {
+        counts.push(count);
+    }
+
+    counts.into_iter().filter(|&count| count >= 2).count()
+}
+
+fn word_count_simple(cell: &str) -> usize {
+    cell.split_whitespace()
+        .filter(|word| word.chars().any(|c| c.is_alphanumeric()))
+        .count()
+}
+
+fn median_f32(mut values: Vec<f32>) -> Option<f32> {
+    values.retain(|value| value.is_finite());
+    if values.is_empty() {
+        return None;
+    }
+    values.sort_by(|a, b| a.total_cmp(b));
+    Some(values[values.len() / 2])
+}
+
+fn merge_superscript_marker_rows(row_ys: &mut Vec<f32>, cells: &mut Vec<Vec<String>>) {
+    let mut row_idx = 0;
+    while row_idx < cells.len() {
+        let non_empty: Vec<(usize, String)> = cells[row_idx]
+            .iter()
+            .enumerate()
+            .filter_map(|(col_idx, cell)| {
+                let trimmed = cell.trim();
+                (!trimmed.is_empty()).then_some((col_idx, trimmed.to_string()))
+            })
+            .collect();
+
+        if non_empty.len() != 1 || !is_superscript_marker_cell(&non_empty[0].1) {
+            row_idx += 1;
+            continue;
+        }
+
+        let (marker_col, marker) = &non_empty[0];
+        let prev =
+            (row_idx > 0).then(|| (row_idx - 1, (row_ys[row_idx - 1] - row_ys[row_idx]).abs()));
+        let next = (row_idx + 1 < cells.len())
+            .then(|| (row_idx + 1, (row_ys[row_idx] - row_ys[row_idx + 1]).abs()));
+        let target = [prev, next]
+            .into_iter()
+            .flatten()
+            .filter(|(_, gap)| *gap <= 10.0)
+            .min_by(|(_, gap_a), (_, gap_b)| gap_a.total_cmp(gap_b))
+            .map(|(idx, _)| idx);
+
+        let Some(target_idx) = target else {
+            row_idx += 1;
+            continue;
+        };
+
+        let target_cell = &mut cells[target_idx][*marker_col];
+        if target_cell.trim().is_empty() {
+            *target_cell = marker.to_string();
+        } else {
+            target_cell.push_str(marker);
+        }
+        cells.remove(row_idx);
+        row_ys.remove(row_idx);
+    }
+}
+
+fn is_superscript_marker_cell(value: &str) -> bool {
+    let trimmed = value.trim();
+    !trimmed.is_empty()
+        && trimmed.chars().count() <= 2
+        && trimmed
+            .chars()
+            .all(|ch| matches!(ch, '*' | '#' | 'o' | 'O' | '°' | 'º' | '†' | '‡'))
+}
+
 /// What kind of structure a detected `Table` represents. Classification is
 /// computed once at construction so consumers don't have to re-analyze the
 /// cells (and stay consistent across detection backends).
@@ -689,6 +1220,190 @@ mod tests {
         assert!(md.contains("|Cell 1|"));
     }
 
+    #[test]
+    fn test_merge_superscript_marker_rows() {
+        let mut rows = vec![506.0, 500.0, 480.0];
+        let mut cells = vec![
+            vec!["".into(), "".into(), "*".into()],
+            vec!["Name".into(), "Method".into(), "Typical values".into()],
+            vec!["Flow".into(), "ASTM D1238".into(), "3.0".into()],
+        ];
+
+        merge_superscript_marker_rows(&mut rows, &mut cells);
+
+        assert_eq!(rows, vec![500.0, 480.0]);
+        assert_eq!(cells[0][2], "Typical values*");
+    }
+
+    #[test]
+    fn test_column_builder_handles_borderless_specs_table() {
+        let items = vec![
+            make_char("*", 458.1, 544.2, 8.0, 4.4),
+            make_char("Properties", 36.0, 538.6, 12.0, 53.1),
+            make_char("Conditions", 195.8, 538.6, 12.0, 55.0),
+            make_char("Method", 297.2, 538.6, 12.0, 39.4),
+            make_char("Typical values", 384.1, 538.6, 12.0, 74.0),
+            make_char("Units", 510.6, 538.6, 8.0, 17.9),
+            make_char("Rheology", 36.0, 508.3, 10.0, 40.6),
+            make_char("o", 209.8, 492.5, 6.5, 3.5),
+            make_char("Melt Flow Rate", 36.0, 488.0, 10.0, 65.2),
+            make_char("230 ", 190.4, 488.0, 10.0, 19.4),
+            make_char("C/2.16 kg", 213.3, 488.0, 10.0, 42.8),
+            make_char("ASTM D1238", 288.4, 488.0, 10.0, 56.8),
+            make_char("3.0 ", 416.4, 488.0, 10.0, 16.9),
+            make_char("g/10 min", 504.1, 488.0, 10.0, 39.5),
+            make_char("Mechanical", 36.0, 451.5, 10.0, 48.3),
+            make_char("Tensile Stress at Yield", 36.0, 431.3, 10.0, 96.7),
+            make_char("50 mm/min", 197.9, 431.3, 10.0, 50.8),
+            make_char("ASTM D638", 291.2, 431.3, 10.0, 51.3),
+            make_char("31 ", 417.9, 431.3, 10.0, 13.9),
+            make_char("MPa", 514.7, 431.3, 10.0, 18.4),
+            make_char("Elongation at Yield", 36.0, 403.0, 10.0, 82.2),
+            make_char("50 mm/min", 197.9, 403.0, 10.0, 50.8),
+            make_char("ASTM D638", 291.2, 403.0, 10.0, 51.3),
+            make_char("8 ", 420.6, 403.0, 10.0, 8.5),
+            make_char("%", 519.1, 403.0, 10.0, 9.7),
+            make_char("Flexural Modulus", 36.0, 374.6, 10.0, 74.0),
+            make_char("ASTM D790", 291.2, 374.6, 10.0, 51.3),
+            make_char("1400", 412.4, 374.6, 10.0, 21.8),
+            make_char("MPa", 514.7, 374.6, 10.0, 18.4),
+        ];
+
+        let table = try_build_table_from_columns(&items, 1).unwrap();
+        let md = table_to_markdown(&table);
+
+        assert!(
+            md.contains("|Properties|Conditions|Method|Typical values*|Units|"),
+            "{md}"
+        );
+        assert!(md.contains("|Mechanical|||||"), "{md}");
+        assert!(
+            md.contains("|Flexural Modulus||ASTM D790|1400|MPa|"),
+            "{md}"
+        );
+    }
+
+    #[test]
+    fn test_key_value_builder_recovers_sectioned_specs_table() {
+        let items = vec![
+            make_char("Ordering Information", 69.0, 700.0, 9.0, 96.0),
+            make_char("Package Contents", 69.0, 680.0, 9.0, 82.0),
+            make_char(
+                "CCH Adapter Panel with 3 m pigtail; installation guide",
+                200.0,
+                680.0,
+                9.0,
+                245.0,
+            ),
+            make_char("Units per Delivery", 69.0, 660.0, 9.0, 78.0),
+            make_char("1/1", 200.0, 660.0, 9.0, 18.0),
+        ];
+
+        let table = try_build_key_value_table_from_rows(&items, 1).unwrap();
+        let md = table_to_markdown(&table);
+
+        assert!(md.contains("|Field|Value|"), "{md}");
+        assert!(md.contains("|Section|Ordering Information|"), "{md}");
+        assert!(
+            md.contains(
+                "|Package Contents|CCH Adapter Panel with 3 m pigtail; installation guide|"
+            ),
+            "{md}"
+        );
+        assert!(md.contains("|Units per Delivery|1/1|"), "{md}");
+    }
+
+    #[test]
+    fn test_key_value_builder_preserves_two_column_header() {
+        let items = vec![
+            make_char("Media", 86.0, 700.0, 10.0, 36.0),
+            make_char("Options", 311.0, 700.0, 10.0, 44.0),
+            make_char("BACnet/IP (Annex J)", 86.0, 680.0, 10.0, 115.0),
+            make_char("Register as Foreign Device", 311.0, 680.0, 10.0, 138.0),
+        ];
+
+        let table = try_build_key_value_table_from_rows(&items, 1).unwrap();
+        let md = table_to_markdown(&table);
+
+        assert!(md.starts_with("|Media|Options|"), "{md}");
+        assert!(
+            md.contains("|BACnet/IP (Annex J)|Register as Foreign Device|"),
+            "{md}"
+        );
+    }
+
+    #[test]
+    fn test_key_value_builder_keeps_repeated_spec_sections() {
+        let items = vec![
+            make_char("1.33 DUAL VVT-i", 90.0, 700.0, 9.0, 82.0),
+            make_char("Engine Code", 90.0, 682.0, 9.0, 62.0),
+            make_char("1NR-FE", 406.0, 682.0, 9.0, 42.0),
+            make_char("Type", 90.0, 664.0, 9.0, 24.0),
+            make_char("Four cylinders in-line", 376.0, 664.0, 9.0, 104.0),
+            make_char("1.6 VALVEMATIC", 90.0, 636.0, 9.0, 78.0),
+            make_char("Engine Code", 90.0, 618.0, 9.0, 62.0),
+            make_char("1ZR-FAE", 404.0, 618.0, 9.0, 44.0),
+        ];
+
+        let table = try_build_key_value_table_from_rows(&items, 1).unwrap();
+        let md = table_to_markdown(&table);
+
+        assert!(md.contains("|Section|1.33 DUAL VVT-i|"), "{md}");
+        assert!(md.contains("|Engine Code|1NR-FE|"), "{md}");
+        assert!(md.contains("|Section|1.6 VALVEMATIC|"), "{md}");
+        assert!(md.contains("|Engine Code|1ZR-FAE|"), "{md}");
+    }
+
+    #[test]
+    fn test_key_value_builder_rejects_split_prose() {
+        let items = vec![
+            make_char(
+                "This paragraph describes an operational process and continues without a field label.",
+                70.0,
+                700.0,
+                10.0,
+                350.0,
+            ),
+            make_char(
+                "It was split only because the text wrapped across a wide line.",
+                455.0,
+                700.0,
+                10.0,
+                300.0,
+            ),
+            make_char(
+                "Another sentence explains background context rather than a measurable property.",
+                70.0,
+                680.0,
+                10.0,
+                350.0,
+            ),
+            make_char(
+                "The neighboring phrase is not a value and should not form a table.",
+                455.0,
+                680.0,
+                10.0,
+                300.0,
+            ),
+            make_char(
+                "Finally, this narrative line keeps flowing with normal prose content.",
+                70.0,
+                660.0,
+                10.0,
+                350.0,
+            ),
+            make_char(
+                "It has punctuation and complete sentences on both sides of the gap.",
+                455.0,
+                660.0,
+                10.0,
+                300.0,
+            ),
+        ];
+
+        assert!(try_build_key_value_table_from_rows(&items, 1).is_none());
+    }
+
     #[test]
     fn test_body_font_table_detected() {
         let items = vec![