xberg 1.2.6

//! Core table reconstruction types and algorithms.
//!
//! This module provides the `HocrWord` type and table reconstruction functions
//! that are shared between the OCR and PDF modules. The algorithms detect
//! column/row structure from word bounding boxes and reconstruct tabular layouts.
//!
//! Originally adapted from the `hocr` module of `html-to-markdown-rs` (removed in v3).

/// Represents a word extracted from hOCR (or any source) with position and confidence information.
#[cfg_attr(alef, alef(skip))]
#[derive(Debug, Clone)]
pub struct HocrWord {
    /// Recognized word text.
    pub text: String,
    /// Left edge of the word bounding box in pixels.
    pub left: u32,
    /// Top edge of the word bounding box in pixels.
    pub top: u32,
    /// Bounding box width in pixels.
    pub width: u32,
    /// Bounding box height in pixels.
    pub height: u32,
    /// OCR confidence score (0.0–100.0).
    pub confidence: f64,
}

impl HocrWord {
    /// Get the right edge position.
    #[cfg(test)]
    #[inline]
    pub(crate) fn right(&self) -> u32 {
        self.left + self.width
    }

    /// Get the bottom edge position.
    #[cfg(test)]
    #[inline]
    pub(crate) fn bottom(&self) -> u32 {
        self.top + self.height
    }

    /// Get the vertical center position.
    #[inline]
    pub(crate) fn y_center(&self) -> f64 {
        self.top as f64 + (self.height as f64 / 2.0)
    }

    /// Get the horizontal center position.
    #[cfg(test)]
    #[inline]
    pub(crate) fn x_center(&self) -> f64 {
        self.left as f64 + (self.width as f64 / 2.0)
    }
}

/// Detect column positions from word x-coordinates.
///
/// Groups words by approximate x-position (within `column_threshold` pixels)
/// and returns the median x-position for each detected column, sorted left to right.
pub(crate) fn detect_columns(words: &[HocrWord], column_threshold: u32) -> Vec<u32> {
    if words.is_empty() {
        return Vec::new();
    }

    let mut position_groups: Vec<Vec<u32>> = Vec::new();

    for word in words {
        let x_pos = word.left;

        let mut found_group = false;
        for group in &mut position_groups {
            if let Some(&first_pos) = group.first()
                && x_pos.abs_diff(first_pos) <= column_threshold
            {
                group.push(x_pos);
                found_group = true;
                break;
            }
        }

        if !found_group {
            position_groups.push(vec![x_pos]);
        }
    }

    let mut columns: Vec<u32> = position_groups
        .iter()
        .filter(|group| !group.is_empty())
        .map(|group| {
            let mut sorted = group.clone();
            sorted.sort_unstable();
            let mid = sorted.len() / 2;
            sorted[mid]
        })
        .collect();

    columns.sort_unstable();
    columns
}

/// Compute the median word height. Returns 0 for an empty slice.
///
/// Extracted so `detect_rows`'s row-grouping threshold,
/// `group_words_into_cell_tokens`'s cell-merge threshold, and (for OCR
/// callers) `merge_disjoint_numeric_columns`'s column-merge threshold are
/// always computed from the same statistic, rather than duplicating the
/// sort-and-index in more than one place.
pub(crate) fn median_word_height(words: &[HocrWord]) -> u32 {
    if words.is_empty() {
        return 0;
    }
    let mut heights: Vec<u32> = words.iter().map(|w| w.height).collect();
    heights.sort_unstable();
    heights[heights.len() / 2]
}

/// Detect row positions from word y-coordinates.
///
/// Groups words by their vertical center position and returns the median
/// y-position for each detected row. The `row_threshold_ratio` is multiplied
/// by the median word height to determine the grouping threshold.
pub(crate) fn detect_rows(words: &[HocrWord], row_threshold_ratio: f64) -> Vec<u32> {
    if words.is_empty() {
        return Vec::new();
    }

    let median_height = median_word_height(words);
    let row_threshold = (median_height as f64 * row_threshold_ratio) as u32;

    let mut position_groups: Vec<Vec<f64>> = Vec::new();

    for word in words {
        let y_center = word.y_center();

        let mut found_group = false;
        for group in &mut position_groups {
            if let Some(&first_pos) = group.first()
                && (y_center - first_pos).abs() <= row_threshold as f64
            {
                group.push(y_center);
                found_group = true;
                break;
            }
        }

        if !found_group {
            position_groups.push(vec![y_center]);
        }
    }

    let mut rows: Vec<u32> = position_groups
        .iter()
        .filter(|group| !group.is_empty())
        .map(|group| {
            let mut sorted = group.clone();
            sorted.sort_by(|a, b| a.total_cmp(b));
            let mid = sorted.len() / 2;
            sorted[mid] as u32
        })
        .collect();

    rows.sort_unstable();
    rows
}

/// Find which row a word belongs to based on its y-center.
fn find_row_index(row_positions: &[u32], word: &HocrWord) -> Option<usize> {
    let y_center = word.y_center() as u32;

    row_positions
        .iter()
        .enumerate()
        .min_by_key(|&(_, row_y)| row_y.abs_diff(y_center))
        .map(|(idx, _)| idx)
}

/// Find which column a word belongs to based on its x-position.
fn find_column_index(col_positions: &[u32], word: &HocrWord) -> Option<usize> {
    let x_pos = word.left;

    col_positions
        .iter()
        .enumerate()
        .min_by_key(|&(_, col_x)| col_x.abs_diff(x_pos))
        .map(|(idx, _)| idx)
}

/// Determine which columns of a table grid contain at least one non-empty
/// cell. Shared by `remove_empty_rows_and_columns` and
/// `reconstruct_table_with_columns`, which both need to drop the same set of
/// all-empty columns — the latter also has to filter a parallel
/// column-position vector by the identical mask so positions keep lining up
/// 1:1 with the grid's surviving columns.
fn non_empty_column_mask(table: &[Vec<String>]) -> Vec<bool> {
    let num_cols = table.first().map_or(0, Vec::len);
    let mut mask = vec![false; num_cols];

    for row in table {
        for (col_idx, cell) in row.iter().enumerate() {
            if !cell.trim().is_empty() {
                mask[col_idx] = true;
            }
        }
    }

    mask
}

/// Remove empty rows and columns from a table grid.
fn remove_empty_rows_and_columns(table: Vec<Vec<String>>) -> Vec<Vec<String>> {
    if table.is_empty() {
        return table;
    }

    let non_empty_cols = non_empty_column_mask(&table);

    table
        .into_iter()
        .filter(|row| row.iter().any(|cell| !cell.trim().is_empty()))
        .map(|row| {
            row.into_iter()
                .enumerate()
                .filter(|(idx, _)| non_empty_cols[*idx])
                .map(|(_, cell)| cell)
                .collect()
        })
        .collect()
}

/// Fraction of the median word height used as the maximum horizontal gap (in
/// pixels) for merging two horizontally-adjacent words within the same
/// detected row into a single cell token before column detection
/// (xberg-io/xberg#688).
///
/// Without this pre-merge, a table cell containing more than one word (e.g.
/// "ABX Air n") puts that cell's second and third words at x-positions that
/// match no genuine column start, so [`detect_columns`] mints a spurious
/// near-empty column per extra word — exactly what the downstream structural
/// validator then rejects, discarding a genuine table entirely.
///
/// 0.6 was chosen because it recovers the exact ground-truth column count
/// (10) on a 215-word-box fixture captured from a real scanned newspaper
/// stock table (ground truth: 16 rows x 10 columns; two side-by-side
/// sub-tables give true column starts at x = {46, 89, 135, 230, 272} and
/// {340, 386, 428, 526, 567}). Measured sensitivity of this factor on that
/// fixture: 0.4 -> 12 columns, 0.6 -> 10 columns (correct), 0.8 -> 9 columns,
/// 1.0 -> 6 columns.
pub(crate) const CELL_MERGE_GAP_HEIGHT_RATIO: f64 = 0.6;

/// Multiplier applied to the normal cell-merge gap when one of the two words being considered
/// is pure punctuation (xberg-io/xberg#1649).
///
/// A lone dash/hyphen glyph rendered as its own OCR word (e.g. the en dash in a date range like
/// "Jan 1 - Jan 31, 2026") is drawn with extra kerning on both sides relative to normal
/// inter-word spacing, so its neighboring gap can land just over `CELL_MERGE_GAP_HEIGHT_RATIO *
/// median height` (measured: a 35px gap against a 34.8px threshold on the reported fixture).
/// Missing that merge leaves the punctuation word to start its own spurious column, which then
/// steals nearby words from the real column during final cell assignment (`find_column_index`
/// picks nearest by raw x-position, not token membership) -- corrupting cells that never
/// contained a multi-word ambiguity in the first place. 2x comfortably covers the measured
/// overshoot without approaching the gaps between genuinely different words (#688's tuned
/// fixture has no punctuation-only tokens, so this multiplier does not affect it). ~keep
const PUNCTUATION_GLUE_GAP_MULTIPLIER: f64 = 2.0;

/// A token/word whose trimmed text is non-empty and entirely ASCII punctuation
/// (a lone "-", ":", "&", etc.), used to widen the cell-merge gap around it (#1649).
fn is_pure_punctuation(text: &str) -> bool {
    let trimmed = text.trim();
    !trimmed.is_empty() && trimmed.chars().all(|ch| ch.is_ascii_punctuation())
}

/// Group horizontally-adjacent words within the same detected row into cell clusters, returning
/// each cluster's merged summary bbox/text (used for column detection, see [`detect_columns`])
/// paired with the original words the cluster contains (xberg-io/xberg#688, xberg-io/xberg#1649).
///
/// Two words in the same row are merged when the horizontal gap between them
/// (the left word's right edge to the right word's left edge) is at most
/// `CELL_MERGE_GAP_HEIGHT_RATIO * median word height`. Rows are the same
/// `row_positions` the caller already computed via `detect_rows` — this does
/// not invent a second row model.
///
/// Column *membership* is decided once per cluster from the cluster's own left edge (its
/// leftmost word), not independently per word: a wide multi-word cell (e.g. a long left-aligned
/// description) has no per-column width bound in [`find_column_index`], only a single
/// representative x-position per column, so a trailing word deep inside such a cell can sit
/// closer by raw x-distance to an unrelated column's anchor than to its own column's anchor.
/// Deciding once per cluster and placing every member word there keeps that trailing word from
/// drifting into a neighboring column at cell-assignment time.
fn group_words_into_cell_tokens<'a>(
    words: &'a [HocrWord],
    row_positions: &[u32],
) -> Vec<(HocrWord, Vec<&'a HocrWord>)> {
    if words.len() <= 1 || row_positions.is_empty() {
        return words.iter().map(|word| (word.clone(), vec![word])).collect();
    }

    let merge_gap = median_word_height(words) as f64 * CELL_MERGE_GAP_HEIGHT_RATIO;

    let mut rows: Vec<Vec<&HocrWord>> = vec![Vec::new(); row_positions.len()];
    for word in words {
        if let Some(row_index) = find_row_index(row_positions, word) {
            rows[row_index].push(word);
        }
    }

    let mut groups: Vec<(HocrWord, Vec<&HocrWord>)> = Vec::with_capacity(words.len());
    for mut row_words in rows {
        row_words.sort_by_key(|w| w.left);
        groups.extend(merge_row_into_cell_tokens(&row_words, merge_gap));
    }

    groups
}

/// Decide the maximum horizontal gap allowed for merging `word` into the token immediately
/// preceding it, and whether `word` itself is glued in as a punctuation *connector* (so a
/// following word may in turn glue to it at the widened gap).
///
/// The widened [`PUNCTUATION_GLUE_GAP_MULTIPLIER`] gap applies only when punctuation genuinely
/// bridges two real words, never to an isolated punctuation cell sitting between two column
/// gaps (xberg-io/xberg#1649 review follow-up):
/// - `word` is pure punctuation: bridging requires `next_word` (the word after it, same row) to
///   itself be within the widened gap of `word` -- i.e. the punctuation has a real neighbour on
///   both sides. A lone `-` with nothing close on the far side keeps the normal gap.
/// - `word` is an ordinary word merging into a token that ends in punctuation: only widened when
///   that trailing punctuation was itself glued in as a connector (`previous_was_connector`),
///   never for an isolated leading punctuation cell. ~keep
fn allowed_merge_gap(
    word: &HocrWord,
    next_word: Option<&HocrWord>,
    base_gap: f64,
    previous_was_connector: bool,
) -> (f64, bool) {
    if is_pure_punctuation(&word.text) {
        let bridges_forward = next_word.is_some_and(|next| {
            let gap_to_next = next.left as f64 - (word.left + word.width) as f64;
            gap_to_next <= base_gap * PUNCTUATION_GLUE_GAP_MULTIPLIER
        });
        return if bridges_forward {
            (base_gap * PUNCTUATION_GLUE_GAP_MULTIPLIER, true)
        } else {
            (base_gap, false)
        };
    }

    if previous_was_connector {
        (base_gap * PUNCTUATION_GLUE_GAP_MULTIPLIER, false)
    } else {
        (base_gap, false)
    }
}

/// Merge one row's words (already sorted left-to-right) into cell-token clusters, applying
/// [`allowed_merge_gap`]'s punctuation-connector rule.
fn merge_row_into_cell_tokens<'a>(row_words: &[&'a HocrWord], merge_gap: f64) -> Vec<(HocrWord, Vec<&'a HocrWord>)> {
    let mut groups: Vec<(HocrWord, Vec<&HocrWord>)> = Vec::new();
    let mut current: Option<(HocrWord, Vec<&HocrWord>)> = None;
    let mut previous_was_connector = false;

    for (index, &word) in row_words.iter().enumerate() {
        let next_word = row_words.get(index + 1).copied();
        current = Some(match current.take() {
            None => (word.clone(), vec![word]),
            Some((mut token, mut members)) => {
                let gap = word.left as f64 - (token.left + token.width) as f64;
                let (allowed_gap, is_connector) = allowed_merge_gap(word, next_word, merge_gap, previous_was_connector);
                if gap <= allowed_gap {
                    let new_right = (word.left + word.width).max(token.left + token.width);
                    let new_bottom = (word.top + word.height).max(token.top + token.height);
                    token.top = token.top.min(word.top);
                    token.width = new_right.saturating_sub(token.left);
                    token.height = new_bottom.saturating_sub(token.top);
                    token.text.push(' ');
                    token.text.push_str(&word.text);
                    members.push(word);
                    previous_was_connector = is_connector;
                    (token, members)
                } else {
                    groups.push((token, members));
                    previous_was_connector = false;
                    (word.clone(), vec![word])
                }
            }
        });
    }
    if let Some(group) = current {
        groups.push(group);
    }

    groups
}

/// Reconstruct a table grid from words with bounding box positions.
///
/// Takes detected words and reconstructs a 2D table by:
/// 1. Detecting row positions (grouping by y-center within `row_threshold_ratio` * median height)
/// 2. Merging horizontally-adjacent words within the same row into cell tokens
///    (see [`CELL_MERGE_GAP_HEIGHT_RATIO`]) so a multi-word cell does not mint
///    spurious columns
/// 3. Detecting column positions from those tokens (grouping by x-coordinate
///    within `column_threshold`)
/// 4. Assigning the *original* words to cells based on closest row/column
/// 5. Combining words within the same cell
///
/// Returns a `Vec<Vec<String>>` where each inner `Vec` is a row of cell texts.
///
/// Thin wrapper over [`reconstruct_table_with_columns`] for callers that only
/// need the grid; use that function instead when the column x-positions used
/// to build the grid must stay correlated with it (e.g. a caller that later
/// indexes `column_positions[column]` against `grid[row][column]`).
#[cfg(any(feature = "pdf", paddle_ocr, test))]
pub(crate) fn reconstruct_table(
    words: &[HocrWord],
    column_threshold: u32,
    row_threshold_ratio: f64,
) -> Vec<Vec<String>> {
    reconstruct_table_with_columns(words, column_threshold, row_threshold_ratio).0
}

/// Like [`reconstruct_table`], but also returns the column x-positions
/// actually used to build the grid, filtered to the same set of columns that
/// survived empty-column removal — so `result.1[i]` corresponds exactly to
/// `result.0[row][i]` for every row.
///
/// Callers that separately call [`detect_columns`] on the raw input words and
/// then index a returned grid by those positions (e.g. to compare adjacent
/// columns) MUST use this function instead: `reconstruct_table` now detects
/// columns from post-merge cell tokens, not from `words` directly, so a
/// `detect_columns(words, ...)` call sitting next to a plain `reconstruct_table`
/// call no longer corresponds to that grid's column count or positions.
pub(crate) fn reconstruct_table_with_columns(
    words: &[HocrWord],
    column_threshold: u32,
    row_threshold_ratio: f64,
) -> (Vec<Vec<String>>, Vec<u32>) {
    if words.is_empty() {
        return (Vec::new(), Vec::new());
    }

    let row_positions = detect_rows(words, row_threshold_ratio);
    let groups = group_words_into_cell_tokens(words, &row_positions);
    let cell_tokens: Vec<HocrWord> = groups.iter().map(|(token, _)| token.clone()).collect();
    let mut col_positions = detect_columns(&cell_tokens, column_threshold);

    if col_positions.is_empty() || row_positions.is_empty() {
        return (Vec::new(), Vec::new());
    }

    let mut result = assign_grouped_words_to_cells(&groups, &row_positions, &col_positions);
    merge_header_fragments_by_geometry(&mut result, &mut col_positions);

    let non_empty_cols = non_empty_column_mask(&result);
    let kept_col_positions: Vec<u32> = col_positions
        .iter()
        .zip(non_empty_cols.iter())
        .filter(|&(_, &keep)| keep)
        .map(|(&pos, _)| pos)
        .collect();

    (remove_empty_rows_and_columns(result), kept_col_positions)
}

const MIN_HEADER_NEIGHBOR_SUPPORT: usize = 2;

/// A header word can start left of the data values it labels and form a header-only x-track.
/// Attach that fragment to the nearest column supported by multiple data rows, while retaining
/// the correlated x-position vector used by downstream table geometry checks. ~keep
fn merge_header_fragments_by_geometry(table: &mut [Vec<String>], column_positions: &mut Vec<u32>) {
    let column_count = table.first().map_or(0, Vec::len);
    if table.len() < 3 || column_count < 3 || column_positions.len() != column_count {
        return;
    }
    if !has_split_invoice_header(&table[0]) {
        return;
    }

    let support: Vec<usize> = (0..column_count)
        .map(|column| {
            table
                .iter()
                .skip(1)
                .filter(|row| row.get(column).is_some_and(|cell| !cell.trim().is_empty()))
                .count()
        })
        .collect();
    let targets = header_fragment_targets(table, column_positions, &support);
    if targets.iter().all(Option::is_none) {
        return;
    }
    rebuild_table_without_header_fragments(table, column_positions, &targets);
}

fn has_split_invoice_header(header: &[String]) -> bool {
    let has = |expected: &str| header.iter().any(|cell| cell.trim().eq_ignore_ascii_case(expected));
    has("QTY") && ((has("UNIT") && has("PRICE")) || (has("LINE") && has("TOTAL")))
}

fn header_fragment_targets(table: &[Vec<String>], column_positions: &[u32], support: &[usize]) -> Vec<Option<usize>> {
    let column_count = support.len();
    let mut targets = vec![None; column_count];
    let mut left = None;
    for column in 0..column_count {
        targets[column] = left;
        if support[column] >= MIN_HEADER_NEIGHBOR_SUPPORT {
            left = Some(column);
        }
    }
    let mut right = None;
    for column in (0..column_count).rev() {
        if support[column] >= MIN_HEADER_NEIGHBOR_SUPPORT {
            targets[column] = None;
            right = Some(column);
            continue;
        }
        let Some(left) = targets[column] else {
            continue;
        };
        let Some(right) = right else {
            continue;
        };
        if support[column] != 0 || !is_split_invoice_fragment(&table[0][column], &table[0][right]) {
            targets[column] = None;
            continue;
        }
        targets[column] = sufficiently_near_matching_column(column_positions, left, column, right);
    }
    targets
}

fn is_split_invoice_fragment(fragment: &str, destination: &str) -> bool {
    let fragment = fragment.trim();
    let destination = destination.trim();
    (fragment.eq_ignore_ascii_case("UNIT") && destination.eq_ignore_ascii_case("PRICE"))
        || (fragment.eq_ignore_ascii_case("LINE") && destination.eq_ignore_ascii_case("TOTAL"))
}

fn sufficiently_near_matching_column(positions: &[u32], left: usize, fragment: usize, right: usize) -> Option<usize> {
    let span = positions[right].abs_diff(positions[left]);
    let right_distance = positions[fragment].abs_diff(positions[right]);
    (right_distance.saturating_mul(3) < span).then_some(right)
}

fn rebuild_table_without_header_fragments(
    table: &mut [Vec<String>],
    column_positions: &mut Vec<u32>,
    targets: &[Option<usize>],
) {
    let column_count = targets.len();
    let (prefixes, suffixes) = header_fragment_affixes(table, targets);
    let kept_columns = targets.iter().filter(|target| target.is_none()).count();
    for (row_index, row) in table.iter_mut().enumerate() {
        let mut rebuilt = Vec::with_capacity(kept_columns);
        for column in 0..column_count {
            if targets[column].is_some() {
                continue;
            }
            if row_index == 0 {
                rebuilt.push(rebuilt_header(&row[column], &prefixes[column], &suffixes[column]));
            } else {
                rebuilt.push(std::mem::take(&mut row[column]));
            }
        }
        *row = rebuilt;
    }
    let mut position_index = 0usize;
    column_positions.retain(|_| {
        let keep_position = targets[position_index].is_none();
        position_index += 1;
        keep_position
    });
}

fn header_fragment_affixes(table: &[Vec<String>], targets: &[Option<usize>]) -> (Vec<String>, Vec<String>) {
    let column_count = targets.len();
    let mut prefixes = vec![String::new(); column_count];
    let mut suffixes = vec![String::new(); column_count];
    for (column, target) in targets.iter().copied().enumerate() {
        let Some(target) = target else {
            continue;
        };
        let fragment = table[0][column].trim();
        let destination = if target < column {
            &mut suffixes[target]
        } else {
            &mut prefixes[target]
        };
        if !destination.is_empty() {
            destination.push(' ');
        }
        destination.push_str(fragment);
    }
    (prefixes, suffixes)
}

fn rebuilt_header(original: &str, prefix: &str, suffix: &str) -> String {
    let original = original.trim();
    let capacity = prefix.len() + original.len() + suffix.len() + 2;
    let mut header = String::with_capacity(capacity);
    for part in [prefix, original, suffix] {
        if part.is_empty() {
            continue;
        }
        if !header.is_empty() {
            header.push(' ');
        }
        header.push_str(part);
    }
    header
}

/// Fraction of a cell's median word height within which two words belong to the same visual line.
///
/// Sized to sit between the two deltas it must tell apart: a sub/superscript is offset from its
/// base's vertical centre by a fraction of its own height (0.6pt against an 11.59pt line on the
/// GH#1628 reproducer), while a wrapped second line is a full line height away. Shares the shape,
/// but not the value, of [`CELL_MERGE_GAP_HEIGHT_RATIO`], which measures a horizontal gap. ~keep
const CELL_LINE_GROUP_HEIGHT_RATIO: f64 = 0.5;

/// Order one cell's words for joining: top-to-bottom by visual line, left-to-right within a line.
///
/// Arrival order is wrong for a cell containing a sub/superscript (xberg-io/xberg#1628). A
/// sub/superscript is drawn as its own content-stream segment sitting a fraction of a point below
/// the line it annotates, so every reading-order sort upstream of here places it after the whole
/// line rather than beside the symbol it belongs to — `eta_S %` joins as `eta % S`.
///
/// Sorting the cell by `left` alone fixes that and introduces something worse: a cell holding two
/// wrapped lines interleaves them (`Hello world` / `again here` becomes `Hello again world here`),
/// which would scramble every wrapped cell in the corpus rather than only cells with scripts.
/// Grouping into visual lines first separates the two cases, because the vertical deltas differ by
/// an order of magnitude — see [`CELL_LINE_GROUP_HEIGHT_RATIO`]. ~keep
fn order_cell_words_in_reading_order(mut cell_words: Vec<&HocrWord>) -> Vec<&HocrWord> {
    if cell_words.len() <= 1 {
        return cell_words;
    }

    let mut heights: Vec<u32> = cell_words.iter().map(|word| word.height).collect();
    heights.sort_unstable();
    let line_gap = heights[heights.len() / 2] as f64 * CELL_LINE_GROUP_HEIGHT_RATIO;

    cell_words.sort_by(|a, b| a.y_center().total_cmp(&b.y_center()).then_with(|| a.left.cmp(&b.left)));

    let mut lines: Vec<Vec<&HocrWord>> = Vec::new();
    for word in cell_words {
        let same_line = lines.last().is_some_and(|line| {
            let centre = line.iter().map(|w| w.y_center()).sum::<f64>() / line.len() as f64;
            (word.y_center() - centre).abs() <= line_gap
        });
        match lines.last_mut() {
            Some(line) if same_line => line.push(word),
            _ => lines.push(vec![word]),
        }
    }

    for line in &mut lines {
        line.sort_by_key(|word| word.left);
    }
    lines.into_iter().flatten().collect()
}

/// Assign each original word independently to its nearest detected row/column and combine
/// same-cell words into space-joined cell text. Retained as a direct-unit-test seam for that
/// per-word assignment and cell-ordering behavior in isolation; the production path
/// (`reconstruct_table_with_columns`) calls [`assign_grouped_words_to_cells`] instead, which
/// assigns a whole merged cell cluster to one column at a time (xberg-io/xberg#1649). ~keep
#[cfg(test)]
fn assign_words_to_cells(words: &[HocrWord], row_positions: &[u32], col_positions: &[u32]) -> Vec<Vec<String>> {
    let num_rows = row_positions.len();
    let num_cols = col_positions.len();
    let mut table: Vec<Vec<Vec<&HocrWord>>> = vec![vec![vec![]; num_cols]; num_rows];

    for word in words {
        if let (Some(r), Some(c)) = (
            find_row_index(row_positions, word),
            find_column_index(col_positions, word),
        ) && r < num_rows
            && c < num_cols
        {
            table[r][c].push(word);
        }
    }

    finish_cell_assignment(table)
}

/// Like [`assign_words_to_cells`], but decides row/column membership once per merged cell
/// cluster (see [`group_words_into_cell_tokens`]) from the cluster's own summary token, then
/// places every original word the cluster contains into that same cell — instead of resolving
/// each original word's column independently, which lets a multi-word cell's trailing word drift
/// into a neighboring column (xberg-io/xberg#1649).
fn assign_grouped_words_to_cells<'a>(
    groups: &[(HocrWord, Vec<&'a HocrWord>)],
    row_positions: &[u32],
    col_positions: &[u32],
) -> Vec<Vec<String>> {
    let num_rows = row_positions.len();
    let num_cols = col_positions.len();
    let mut table: Vec<Vec<Vec<&'a HocrWord>>> = vec![vec![vec![]; num_cols]; num_rows];

    for (token, members) in groups {
        let Some(row) = find_row_index(row_positions, token) else {
            continue;
        };
        if row >= num_rows {
            continue;
        }
        if row == 0 {
            // Row 0 is conventionally the header row throughout this module (see
            // `data_support_count`, `table_to_markdown`). A multi-word header label can still
            // legitimately span more than one detected column even after merging into one
            // cluster for column detection -- its second word's x-position may line up with the
            // data column it labels rather than with its own first word (xberg-io/xberg#2219).
            // Keep independent per-word placement here so `merge_header_fragments_by_geometry`
            // can still reconcile that split; only data rows get whole-cluster placement. ~keep
            for word in members {
                if let Some(col) = find_column_index(col_positions, word)
                    && col < num_cols
                {
                    table[row][col].push(word);
                }
            }
        } else if let Some(col) = find_column_index(col_positions, token)
            && col < num_cols
        {
            table[row][col].extend(members.iter().copied());
        }
    }

    finish_cell_assignment(table)
}

/// Shared tail of [`assign_words_to_cells`] and [`assign_grouped_words_to_cells`]: order each
/// cell's collected words and join them into the final cell text.
fn finish_cell_assignment(table: Vec<Vec<Vec<&HocrWord>>>) -> Vec<Vec<String>> {
    table
        .into_iter()
        .map(|row| {
            row.into_iter()
                .map(|cell_words| {
                    // A "word" whose whole segment held no whitespace keeps that segment's own
                    // trailing space (`split_segment_to_words_lifted` returns it unsplit), which
                    // then stacks on the separator below and renders `Q_HE GJ` as `Q  HE GJ`.
                    // Collapse each word's own whitespace so the join alone owns the spacing. ~keep
                    order_cell_words_in_reading_order(cell_words)
                        .into_iter()
                        .flat_map(|word| word.text.split_whitespace())
                        .collect::<Vec<_>>()
                        .join(" ")
                })
                .collect()
        })
        .collect()
}

/// Convert a table grid to markdown format.
///
/// The first row is treated as the header row, with a separator line added after it.
///
/// Delegates to the crate's single table renderer
/// ([`crate::rendering::common::render_table_markdown`]) so a reconstructed
/// OCR/PDF table serialises identically to one from any other source
/// (xberg-io/xberg#220). That renderer also pads every row to the widest row,
/// which this reconstruction path did not do — ragged rows used to emit fewer
/// pipe columns than the header, misaligning the table for downstream parsers.
pub(crate) fn table_to_markdown(table: &[Vec<String>]) -> String {
    crate::rendering::common::render_table_markdown(table)
}

/// Multiple of the page's average word height used as the vertical-gap
/// threshold for splitting table candidate words into separate regions
/// (#177). Normal row spacing inside one table rarely exceeds ~1.5x the
/// average word height, so a wider multiple avoids splitting a single
/// table's own row gaps while still separating genuinely distinct tables
/// (or a table from surrounding prose) on the same page.
//
// This module also compiles for `feature = "pdf"` (shared with the native PDF table path), but
// the only real callers of this constant -- `ocr::processor::execution` (gated `feature = "ocr"`)
// and `paddle_ocr::backend` (gated `paddle_ocr`) -- need neither `pdf` nor each other's gate. A
// gate as wide as the module's left this compiled with zero callers on a `pdf`-only leg. ~keep
#[cfg(any(feature = "ocr", paddle_ocr))]
pub(crate) const TABLE_REGION_GAP_HEIGHT_MULTIPLIER: u32 = 3;

/// Minimum number of words for a spatial region to be treated as a table
/// candidate. Mirrors the previous whole-page threshold so a single small
/// table on an otherwise text-only page is not over-fabricated.
//
// Same reasoning as `TABLE_REGION_GAP_HEIGHT_MULTIPLIER` above: real callers are gated `ocr` or
// `paddle_ocr`, not `pdf`. ~keep
#[cfg(any(feature = "ocr", paddle_ocr))]
pub(crate) const MIN_TABLE_CANDIDATE_WORDS: usize = 6;

/// Split table-candidate words into vertically separated regions.
///
/// Tesseract's TSV output has no notion of "this is a separate table from
/// that one" — [`reconstruct_table`] previously ran once over every
/// table-confidence word on the page, producing at most one table whose
/// bounding box spanned the union of all such words, even when the page had
/// several independent tables separated by paragraphs of prose (#177).
///
/// This groups words by contiguous vertical extent: a gap between one row's
/// bottom edge and the next word's top edge wider than
/// `TABLE_REGION_GAP_HEIGHT_MULTIPLIER` times the average word height starts
/// a new region. Each region is reconstructed independently, giving each
/// table its own bounding box.
///
// Same reasoning as `TABLE_REGION_GAP_HEIGHT_MULTIPLIER` above: the two real callers --
// `ocr::processor::execution::perform_ocr`'s table-detection branch (gated `feature = "ocr"`)
// and `PaddleOcrBackend::build_ocr_tables_from_words` (gated `paddle_ocr`) -- never need the
// wider `pdf` gate this module carries. ~keep
#[cfg(any(feature = "ocr", paddle_ocr))]
pub(crate) fn cluster_words_into_table_regions(words: &[HocrWord]) -> Vec<Vec<HocrWord>> {
    if words.is_empty() {
        return Vec::new();
    }

    let mut sorted: Vec<&HocrWord> = words.iter().collect();
    sorted.sort_by(|a, b| a.top.cmp(&b.top).then(a.left.cmp(&b.left)));

    let avg_height: u32 = {
        let total: u32 = sorted.iter().map(|w| w.height).sum();
        (total / sorted.len() as u32).max(1)
    };
    let region_gap_threshold = avg_height * TABLE_REGION_GAP_HEIGHT_MULTIPLIER;

    let mut regions: Vec<Vec<HocrWord>> = Vec::new();
    let mut current_region: Vec<HocrWord> = Vec::new();
    let mut current_bottom: u32 = 0;

    for word in sorted {
        let word_bottom = word.top + word.height;
        let is_new_region =
            !current_region.is_empty() && word.top.saturating_sub(current_bottom) > region_gap_threshold;

        if is_new_region {
            regions.push(std::mem::take(&mut current_region));
            current_bottom = 0;
        }

        current_bottom = current_bottom.max(word_bottom);
        current_region.push(word.clone());
    }
    if !current_region.is_empty() {
        regions.push(current_region);
    }

    regions
}

/// Multiple of the region's median word height used as the maximum gap for merging two
/// adjacent OCR table columns that are very likely one logical column split by word-start
/// drift (xberg-io/xberg#1649).
///
/// Right-aligned numeric columns (amounts, balances) have a left edge that shifts with digit
/// count -- `$1,250.00` starts well left of `$87.32` in the same visual column -- so
/// [`detect_columns`]'s absolute `column_threshold` (tuned for left-aligned label columns, and
/// applied in raster-pixel space after DPI normalization can upscale the page 2x+) can end up
/// splitting one logical column into several. 3.0 mirrors [`TABLE_REGION_GAP_HEIGHT_MULTIPLIER`]'s
/// use of median word height as the scale-invariant unit: measured gaps of 95-120px against a
/// 51px median height (roughly 1.9-2.4x) on the reported fixture merge comfortably under this
/// multiplier, while the gap to a genuinely different column (WITHDRAWAL to DEPOSIT, ~454px,
/// ~8.9x) stays well clear of it.
#[cfg(feature = "ocr")]
pub(crate) const DISJOINT_NUMERIC_COLUMN_MERGE_HEIGHT_MULTIPLIER: f64 = 3.0;

/// A non-empty, trimmed cell that looks like a plain number or currency amount: at least one
/// ASCII digit, and no characters outside a small currency/number charset. Deliberately narrow
/// so it never matches ordinary prose (a lone `-` or `()` alone does not count -- see
/// [`merge_disjoint_numeric_columns`], which relies on this to avoid merging text columns).
#[cfg(feature = "ocr")]
fn looks_like_amount(cell: &str) -> bool {
    let trimmed = cell.trim();
    !trimmed.is_empty()
        && trimmed.chars().any(|ch| ch.is_ascii_digit())
        && trimmed
            .chars()
            .all(|ch| ch.is_ascii_digit() || "$,.-()%€£¥+".contains(ch))
}

/// Whether columns `left` and `right` of `table` (row 0 is the header, excluded from this
/// check) are mutually exclusive -- no data row has both populated -- and every populated data
/// cell in either column looks like a number/currency amount. Both conditions must hold for
/// [`merge_disjoint_numeric_columns`] to treat the pair as one logical column split in two.
#[cfg(feature = "ocr")]
fn columns_are_disjoint_and_numeric(table: &[Vec<String>], left: usize, right: usize) -> bool {
    let mut any_data = false;
    for row in table.iter().skip(1) {
        let (Some(left_cell), Some(right_cell)) = (row.get(left), row.get(right)) else {
            return false;
        };
        let left_empty = left_cell.trim().is_empty();
        let right_empty = right_cell.trim().is_empty();
        match (left_empty, right_empty) {
            (true, true) => {}
            (false, false) => return false,
            (false, true) => {
                if !looks_like_amount(left_cell) {
                    return false;
                }
                any_data = true;
            }
            (true, false) => {
                if !looks_like_amount(right_cell) {
                    return false;
                }
                any_data = true;
            }
        }
    }
    any_data
}

/// Whether at most one of columns `left`/`right` carries its own (non-empty) header label in
/// `table`'s row 0.
///
/// A drift-split numeric column (this function's target) has its header text in only one of the
/// split pieces -- the other piece's header cell is empty, because the source document had one
/// label for the whole logical column. Two genuinely distinct, independently-labeled columns
/// (e.g. a ledger's separately headed "Debit" and "Credit") must never be folded together by
/// [`merge_disjoint_numeric_columns`] even when their data happens to be mutually exclusive per
/// row and their x-positions sit close together -- that shape is a normal compact table layout,
/// not a drift artifact, and this guard keeps it untouched regardless of the gap/exclusivity
/// checks. ~keep
#[cfg(feature = "ocr")]
fn at_most_one_column_has_its_own_header(table: &[Vec<String>], left: usize, right: usize) -> bool {
    let Some(header) = table.first() else {
        return true;
    };
    let left_labeled = header.get(left).is_some_and(|cell| !cell.trim().is_empty());
    let right_labeled = header.get(right).is_some_and(|cell| !cell.trim().is_empty());
    !(left_labeled && right_labeled)
}

/// Fold column `right` into column `left` in place: the header row's non-empty fragments join
/// with a space (order preserved), and each data row keeps whichever of the two cells is
/// non-empty (both are never non-empty at once -- callers only reach here after
/// [`columns_are_disjoint_and_numeric`] confirms that). Column `right` is then dropped from
/// every row.
#[cfg(feature = "ocr")]
fn merge_column_into(table: &mut [Vec<String>], left: usize, right: usize) {
    for row in table.iter_mut() {
        let right_cell = row[right].trim().to_string();
        if !right_cell.is_empty() {
            let left_cell = row[left].trim();
            row[left] = if left_cell.is_empty() {
                right_cell
            } else {
                format!("{left_cell} {right_cell}")
            };
        }
        row.remove(right);
    }
}

/// Count of non-empty data cells (row 0, the header, excluded) in `column`.
#[cfg(feature = "ocr")]
fn data_support_count(table: &[Vec<String>], column: usize) -> usize {
    table
        .iter()
        .skip(1)
        .filter(|row| !row[column].trim().is_empty())
        .count()
}

/// Merge adjacent OCR table columns that are very likely one logical (typically right-aligned,
/// numeric) column split by word-start drift (xberg-io/xberg#1649): see
/// [`DISJOINT_NUMERIC_COLUMN_MERGE_HEIGHT_MULTIPLIER`] for the distance bound and
/// [`columns_are_disjoint_and_numeric`] for the content/exclusivity bound. Both must hold, so
/// two genuinely distinct columns (different labels, or data that ever coexists in the same
/// row) are left untouched.
///
/// A right-aligned amount column can split into more than two pieces (e.g. a header-only
/// column with no data support at all, plus two data-bearing pieces at different digit-count
/// widths), so this keeps re-checking the same left index against its new right neighbor after
/// each merge rather than advancing past it. Each merge re-anchors `column_positions[column]`
/// on whichever side actually carries more data, not always the geometrically-leftmost side —
/// otherwise a header-only column's position (with zero data support) would stay the anchor and
/// the next real data column could land just outside `max_gap` of it, even though both data
/// pieces are close together.
#[cfg(feature = "ocr")]
pub(crate) fn merge_disjoint_numeric_columns(
    table: &mut [Vec<String>],
    column_positions: &mut Vec<u32>,
    median_height: u32,
) {
    if table.len() < 2 || column_positions.len() < 2 {
        return;
    }
    let max_gap = median_height as f64 * DISJOINT_NUMERIC_COLUMN_MERGE_HEIGHT_MULTIPLIER;
    let mut column = 0;
    while column + 1 < column_positions.len() {
        let gap = column_positions[column + 1].abs_diff(column_positions[column]) as f64;
        if gap <= max_gap
            && at_most_one_column_has_its_own_header(table, column, column + 1)
            && columns_are_disjoint_and_numeric(table, column, column + 1)
        {
            if data_support_count(table, column + 1) > data_support_count(table, column) {
                column_positions[column] = column_positions[column + 1];
            }
            merge_column_into(table, column, column + 1);
            column_positions.remove(column + 1);
        } else {
            column += 1;
        }
    }
}

/// Drop a leading section-caption row that [`cluster_words_into_table_regions`] joined to the
/// same region as a genuine header row (xberg-io/xberg#1649): a section title (e.g.
/// "TRANSACTIONS") sitting close enough above a table's header to share its region has only one
/// populated cell, spanning what OCR resolved as the leftmost column, while a real header row
/// populates most of the grid's columns.
///
/// Deliberately narrow: requires at least 3 columns and 3 rows (so this never fires on a
/// two-row summary card, which has no separate caption to begin with), the first row's only
/// non-empty cell in column 0, and the second row populating at least 3 columns.
#[cfg(feature = "ocr")]
pub(crate) fn drop_leading_caption_row(table: &mut Vec<Vec<String>>) {
    if table.len() < 3 {
        return;
    }
    let column_count = table[0].len();
    if column_count < 3 {
        return;
    }
    let first_non_empty: Vec<usize> = table[0]
        .iter()
        .enumerate()
        .filter(|(_, cell)| !cell.trim().is_empty())
        .map(|(index, _)| index)
        .collect();
    let second_non_empty_count = table[1].iter().filter(|cell| !cell.trim().is_empty()).count();
    if first_non_empty == [0] && second_non_empty_count >= 3 {
        table.remove(0);
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    /// Build a `HocrWord` with a fixed confidence, for tests that only care
    /// about position/text (matches the `word`/`make_word`/`hocr_word_at`
    /// helper convention used by the other files in this module's blast
    /// radius, e.g. `pdf::native::table` and `paddle_ocr::backend`).
    fn word(text: &str, left: u32, top: u32, width: u32, height: u32) -> HocrWord {
        HocrWord {
            text: text.to_string(),
            left,
            top,
            width,
            height,
            confidence: 95.0,
        }
    }

    #[test]
    fn test_detect_rows_zero_height_words_grouped_into_one_row() {
        let words = vec![
            HocrWord {
                text: "A".to_string(),
                left: 0,
                top: 10,
                width: 5,
                height: 0,
                confidence: 0.0,
            },
            HocrWord {
                text: "B".to_string(),
                left: 0,
                top: 10,
                width: 5,
                height: 0,
                confidence: 0.0,
            },
        ];
        let rows = detect_rows(&words, 0.5);
        assert_eq!(rows.len(), 1);
    }

    #[test]
    fn test_nan_safe_sort_does_not_panic() {
        let mut values: Vec<f64> = vec![1.0, f64::NAN, 2.0];
        values.sort_by(|a, b| a.total_cmp(b));
        assert_eq!(values.len(), 3);
        assert!(!values[0].is_nan());
        assert!(!values[1].is_nan());
        assert!(values[2].is_nan(), "NaN sorts last in ascending total_cmp order");
    }

    #[test]
    fn test_hocr_word_methods() {
        let word = HocrWord {
            text: "Hello".to_string(),
            left: 100,
            top: 50,
            width: 80,
            height: 30,
            confidence: 95.5,
        };

        assert_eq!(word.right(), 180);
        assert_eq!(word.bottom(), 80);
        assert_eq!(word.y_center(), 65.0);
        assert_eq!(word.x_center(), 140.0);
    }

    #[test]
    fn test_detect_columns() {
        let words = vec![
            HocrWord {
                text: "A".to_string(),
                left: 100,
                top: 50,
                width: 20,
                height: 30,
                confidence: 95.0,
            },
            HocrWord {
                text: "B".to_string(),
                left: 300,
                top: 50,
                width: 20,
                height: 30,
                confidence: 95.0,
            },
            HocrWord {
                text: "C".to_string(),
                left: 105,
                top: 100,
                width: 20,
                height: 30,
                confidence: 95.0,
            },
            HocrWord {
                text: "D".to_string(),
                left: 295,
                top: 100,
                width: 20,
                height: 30,
                confidence: 95.0,
            },
        ];

        let cols = detect_columns(&words, 20);
        assert_eq!(cols.len(), 2);
    }

    #[test]
    fn test_detect_rows() {
        let words = vec![
            HocrWord {
                text: "A".to_string(),
                left: 100,
                top: 50,
                width: 20,
                height: 30,
                confidence: 95.0,
            },
            HocrWord {
                text: "B".to_string(),
                left: 200,
                top: 52,
                width: 20,
                height: 30,
                confidence: 95.0,
            },
            HocrWord {
                text: "C".to_string(),
                left: 100,
                top: 100,
                width: 20,
                height: 30,
                confidence: 95.0,
            },
        ];

        let rows = detect_rows(&words, 0.5);
        assert_eq!(rows.len(), 2);
    }

    #[test]
    fn test_reconstruct_table_basic() {
        let words = vec![
            HocrWord {
                text: "Name".to_string(),
                left: 100,
                top: 50,
                width: 40,
                height: 20,
                confidence: 95.0,
            },
            HocrWord {
                text: "Value".to_string(),
                left: 300,
                top: 50,
                width: 40,
                height: 20,
                confidence: 95.0,
            },
            HocrWord {
                text: "Alice".to_string(),
                left: 100,
                top: 100,
                width: 40,
                height: 20,
                confidence: 95.0,
            },
            HocrWord {
                text: "42".to_string(),
                left: 300,
                top: 100,
                width: 20,
                height: 20,
                confidence: 95.0,
            },
        ];

        let table = reconstruct_table(&words, 20, 0.5);
        assert_eq!(table.len(), 2);
        assert_eq!(table[0].len(), 2);
        assert_eq!(table[0][0], "Name");
        assert_eq!(table[0][1], "Value");
        assert_eq!(table[1][0], "Alice");
        assert_eq!(table[1][1], "42");
    }

    #[test]
    fn test_table_to_markdown_basic() {
        let table = vec![
            vec!["Name".to_string(), "Value".to_string()],
            vec!["Alice".to_string(), "42".to_string()],
        ];

        let md = table_to_markdown(&table);
        assert!(md.contains("| Name | Value |"));
        assert!(md.contains("| --- | --- |"));
        assert!(md.contains("| Alice | 42 |"));
    }

    #[test]
    fn test_table_to_markdown_empty() {
        assert_eq!(table_to_markdown(&[]), String::new());
    }

    #[test]
    fn test_table_to_markdown_escapes_pipes() {
        let table = vec![vec!["Header".to_string()], vec!["a|b".to_string()]];

        let md = table_to_markdown(&table);
        assert!(md.contains("a\\|b"));
    }

    /// Regression test for issue where intra-cell word spacing ("Chose 1")
    /// was incorrectly split into separate columns.
    /// The word "1" should stay in the same cell as "Chose" despite having
    /// a different left position, because they're separated by a small gap.
    #[test]
    fn test_reconstruct_table_intra_cell_word_spacing() {
        let words = vec![
            HocrWord {
                text: "Chose".to_string(),
                left: 57,
                top: 496,
                width: 30,
                height: 12,
                confidence: 95.0,
            },
            HocrWord {
                text: "Truc".to_string(),
                left: 306,
                top: 496,
                width: 23,
                height: 12,
                confidence: 95.0,
            },
            HocrWord {
                text: "Chose".to_string(),
                left: 57,
                top: 510,
                width: 28,
                height: 12,
                confidence: 95.0,
            },
            HocrWord {
                text: "1".to_string(),
                left: 90,
                top: 510,
                width: 6,
                height: 12,
                confidence: 95.0,
            },
            HocrWord {
                text: "Truc".to_string(),
                left: 306,
                top: 510,
                width: 21,
                height: 12,
                confidence: 95.0,
            },
            HocrWord {
                text: "1".to_string(),
                left: 332,
                top: 510,
                width: 5,
                height: 12,
                confidence: 95.0,
            },
            HocrWord {
                text: "Chose".to_string(),
                left: 57,
                top: 524,
                width: 28,
                height: 12,
                confidence: 95.0,
            },
            HocrWord {
                text: "2".to_string(),
                left: 90,
                top: 524,
                width: 6,
                height: 12,
                confidence: 95.0,
            },
            HocrWord {
                text: "Truc".to_string(),
                left: 306,
                top: 524,
                width: 21,
                height: 12,
                confidence: 95.0,
            },
            HocrWord {
                text: "2".to_string(),
                left: 332,
                top: 524,
                width: 5,
                height: 12,
                confidence: 95.0,
            },
        ];

        let table = reconstruct_table(&words, 60, 0.5);

        assert_eq!(table.len(), 3, "Expected 3 rows, got {}", table.len());
        assert_eq!(table[0].len(), 2, "Expected 2 columns in row 0, got {}", table[0].len());

        assert_eq!(table[0][0], "Chose", "Header row 1, col 1");
        assert_eq!(table[0][1], "Truc", "Header row 1, col 2");
        assert_eq!(table[1][0], "Chose 1", "Row 2, col 1 should contain merged text");
        assert_eq!(table[1][1], "Truc 1", "Row 2, col 2 should contain merged text");
        assert_eq!(table[2][0], "Chose 2", "Row 3, col 1 should contain merged text");
        assert_eq!(table[2][1], "Truc 2", "Row 3, col 2 should contain merged text");
    }

    /// Regression/verification test for the claim in xberg-io/xberg#183 that
    /// `reconstruct_table` "drops any word for which `find_row_index` or
    /// `find_column_index` returns `None`".
    ///
    /// That claim does not hold for the current implementation: both
    /// `find_row_index` and `find_column_index` resolve to the *nearest*
    /// row/column via `min_by_key`, which is `Some` for every word whenever
    /// `row_positions`/`col_positions` are non-empty — and `reconstruct_table`
    /// already early-returns before this loop if either is empty. There is no
    /// path through this function that silently discards a word; even a word
    /// wildly outside the detected column/row bands is force-assigned to its
    /// nearest band instead of being dropped. This test proves that with an
    /// outlier word far outside the main table extent: every input word is
    /// still present, exactly once, in the reconstructed table.
    #[test]
    fn test_reconstruct_table_never_drops_words_including_far_outliers() {
        let mut words = vec![
            HocrWord {
                text: "A1".to_string(),
                left: 0,
                top: 0,
                width: 10,
                height: 10,
                confidence: 95.0,
            },
            HocrWord {
                text: "B1".to_string(),
                left: 200,
                top: 0,
                width: 10,
                height: 10,
                confidence: 95.0,
            },
            HocrWord {
                text: "A2".to_string(),
                left: 0,
                top: 200,
                width: 10,
                height: 10,
                confidence: 95.0,
            },
            HocrWord {
                text: "B2".to_string(),
                left: 200,
                top: 200,
                width: 10,
                height: 10,
                confidence: 95.0,
            },
        ];
        // Far outside every detected row/column band and every other word's
        // neighborhood — the scenario #183 claims gets silently dropped.
        words.push(HocrWord {
            text: "Outlier".to_string(),
            left: 50_000,
            top: 50_000,
            width: 10,
            height: 10,
            confidence: 95.0,
        });

        let input_word_count = words.len();
        let table = reconstruct_table(&words, 20, 0.5);

        let output_word_count: usize = table
            .iter()
            .flat_map(|row| row.iter())
            .flat_map(|cell| cell.split_whitespace())
            .count();

        assert_eq!(
            output_word_count, input_word_count,
            "every input word (including the far outlier) must appear exactly once in the output; \
             reconstruct_table's nearest-row/nearest-column assignment never returns None here"
        );

        let all_text: Vec<&str> = table
            .iter()
            .flat_map(|row| row.iter())
            .flat_map(|c| c.split_whitespace())
            .collect();
        assert!(
            all_text.contains(&"Outlier"),
            "the outlier word must not be silently dropped"
        );
    }

    /// Regression test for xberg-io/xberg#688: a table cell containing more
    /// than one word ("Alice Smith") must not mint a spurious extra column
    /// from its second word's x-position.
    ///
    /// TEST HONESTY: without the merge-before-column-detection fix, this
    /// fails at the `table[0].len() == 2` assertion — `detect_columns` sees
    /// "Alice"@100 and "Smith"@145 as two distinct groups (gap 45 > the
    /// column_threshold of 20), producing 3 columns instead of 2. `table[0]`
    /// comes out as `["Name", "", "Value"]` (len 3) instead of `["Name",
    /// "Value"]`.
    #[test]
    fn test_reconstruct_table_multiword_cell_no_spurious_column() {
        let words = vec![
            word("Name", 100, 50, 40, 20),
            word("Value", 300, 50, 40, 20),
            word("Alice", 100, 100, 40, 20),
            word("Smith", 145, 100, 30, 20),
            word("42", 300, 100, 20, 20),
        ];

        let table = reconstruct_table(&words, 20, 0.5);

        assert_eq!(table.len(), 2);
        assert_eq!(
            table[0].len(),
            2,
            "expected 2 columns; without the fix 'Smith' (x=145) mints a spurious 3rd column"
        );
        assert_eq!(table[0], vec!["Name".to_string(), "Value".to_string()]);
        assert_eq!(table[1], vec!["Alice Smith".to_string(), "42".to_string()]);
    }

    #[test]
    fn reconstruct_table_attaches_header_fragments_to_nearest_supported_column() {
        let words = vec![
            word("DESCRIPTION", 279, 100, 250, 40),
            word("QTY", 1_699, 100, 90, 40),
            word("UNIT", 2_096, 100, 105, 40),
            word("PRICE", 2_219, 100, 135, 40),
            word("LINE", 2_663, 100, 98, 40),
            word("TOTAL", 2_778, 100, 143, 40),
            word("Espresso Beans", 279, 200, 400, 40),
            word("10", 1_740, 200, 48, 40),
            word("$45.00", 2_206, 200, 150, 40),
            word("$450.00", 2_744, 200, 175, 40),
            word("Cups", 279, 300, 100, 40),
            word("200", 1_709, 300, 75, 40),
            word("$1.20", 2_233, 300, 125, 40),
            word("$240.00", 2_744, 300, 175, 40),
            word("Cleaning Tablets", 279, 400, 360, 40),
            word("$18.50", 2_206, 400, 150, 40),
            word("$92.50", 2_772, 400, 150, 40),
        ];

        let (table, column_positions) = reconstruct_table_with_columns(&words, 50, 0.5);

        assert_eq!(table[0], vec!["DESCRIPTION", "QTY", "UNIT PRICE", "LINE TOTAL"]);
        assert_eq!(table[3], vec!["Cleaning Tablets", "", "$18.50", "$92.50"]);
        // The merged column keeps `detect_columns`' own anchor for the data cluster (2_206, the
        // median of the "$45.00"/"$1.20"/"$18.50" lefts), not the PRICE header word's own left
        // (2_219). `assign_words_to_cells` already used that anchor -- not the header word's
        // position -- to decide every word's column membership, and downstream geometry (the OCR
        // blank-quantity retry's `cell_axis_bounds`) derives an adjacent cell's crop boundary from
        // this same value, so it must track where the data actually sits, not where the header text
        // happened to land. ~keep
        assert_eq!(column_positions, vec![279, 1_709, 2_206, 2_744]);
    }

    #[test]
    fn reconstruct_table_keeps_fragment_far_from_matching_right_column() {
        let words = vec![
            word("DESCRIPTION", 100, 100, 250, 40),
            word("QTY", 1_000, 100, 90, 40),
            word("UNIT", 1_150, 100, 105, 40),
            word("PRICE", 1_500, 100, 135, 40),
            word("Item A", 100, 200, 160, 40),
            word("10", 1_000, 200, 48, 40),
            word("$45.00", 1_500, 200, 150, 40),
            word("Item B", 100, 300, 160, 40),
            word("20", 1_000, 300, 48, 40),
            word("$50.00", 1_500, 300, 150, 40),
        ];

        let (table, column_positions) = reconstruct_table_with_columns(&words, 50, 0.5);

        assert_eq!(table[0], vec!["DESCRIPTION", "QTY", "UNIT", "PRICE"]);
        assert_eq!(column_positions, vec![100, 1_000, 1_150, 1_500]);
    }

    #[test]
    fn reconstruct_table_keeps_far_standalone_header_column() {
        let words = vec![
            word("QTY", 0, 100, 80, 40),
            word("UNIT", 400, 100, 120, 40),
            word("PRICE", 1_000, 100, 100, 40),
            word("1", 0, 200, 30, 40),
            word("$10", 1_000, 200, 60, 40),
            word("2", 0, 300, 30, 40),
            word("$20", 1_000, 300, 60, 40),
        ];

        let (table, column_positions) = reconstruct_table_with_columns(&words, 50, 0.5);

        assert_eq!(table[0], vec!["QTY", "UNIT", "PRICE"]);
        assert_eq!(column_positions, vec![0, 400, 1_000]);
    }

    #[test]
    fn reconstruct_table_keeps_centered_group_header_column() {
        let words = vec![
            word("QTY", 0, 100, 80, 40),
            word("UNIT", 500, 100, 120, 40),
            word("PRICE", 1_000, 100, 100, 40),
            word("1", 0, 200, 30, 40),
            word("$10", 1_000, 200, 60, 40),
            word("2", 0, 300, 30, 40),
            word("$20", 1_000, 300, 60, 40),
        ];

        let (table, column_positions) = reconstruct_table_with_columns(&words, 50, 0.5);

        assert_eq!(table[0], vec!["QTY", "UNIT", "PRICE"]);
        assert_eq!(column_positions, vec![0, 500, 1_000]);
    }

    #[test]
    fn reconstruct_table_keeps_non_header_singleton_column() {
        let words = vec![
            word("QTY", 0, 100, 80, 40),
            word("note", 700, 100, 80, 40),
            word("UNIT", 850, 100, 80, 40),
            word("PRICE", 1_000, 100, 100, 40),
            word("1", 0, 200, 30, 40),
            word("$10", 1_000, 200, 60, 40),
            word("2", 0, 300, 30, 40),
            word("$20", 1_000, 300, 60, 40),
        ];

        let (table, column_positions) = reconstruct_table_with_columns(&words, 50, 0.5);

        assert_eq!(table[0], vec!["QTY", "note", "UNIT PRICE"]);
        assert_eq!(column_positions, vec![0, 700, 1_000]);
    }

    #[test]
    fn reconstruct_table_keeps_unrelated_uppercase_header_column() {
        let words = vec![
            word("QTY", 0, 100, 80, 40),
            word("DISCOUNT", 700, 100, 120, 40),
            word("UNIT", 850, 100, 80, 40),
            word("PRICE", 1_000, 100, 100, 40),
            word("1", 0, 200, 30, 40),
            word("$10", 1_000, 200, 60, 40),
            word("2", 0, 300, 30, 40),
            word("$20", 1_000, 300, 60, 40),
        ];

        let (table, column_positions) = reconstruct_table_with_columns(&words, 50, 0.5);

        assert_eq!(table[0], vec!["QTY", "DISCOUNT", "UNIT PRICE"]);
        assert_eq!(column_positions, vec![0, 700, 1_000]);
    }

    #[test]
    fn reconstruct_table_keeps_every_supported_column() {
        let words = vec![
            word("A", 0, 100, 40, 40),
            word("B", 500, 100, 40, 40),
            word("C", 1_000, 100, 40, 40),
            word("1", 0, 200, 30, 40),
            word("2", 500, 200, 30, 40),
            word("3", 1_000, 200, 30, 40),
            word("4", 0, 300, 30, 40),
            word("5", 500, 300, 30, 40),
            word("6", 1_000, 300, 30, 40),
        ];

        let (table, column_positions) = reconstruct_table_with_columns(&words, 50, 0.5);

        assert_eq!(table[0], vec!["A", "B", "C"]);
        assert_eq!(column_positions, vec![0, 500, 1_000]);
    }

    /// A genuinely separate, closely-spaced column pair ("Wid"/"Zone" at
    /// x=150 and x=180, gap 15px) must survive intact even once a multi-word
    /// cell elsewhere in the same table ("Foo"+"Bar", gap 5px) gets merged.
    ///
    /// TEST HONESTY: without the fix, "Bar" (x=25) mints its own spurious
    /// column, producing 4 columns total ([0, 25, 150, 180] once sorted).
    /// `table[0]` comes out as `["Name", "", "Wid", "Zone"]` (len 4) instead
    /// of `["Name", "Wid", "Zone"]` (len 3), and `table[1]` as `["Foo",
    /// "Bar", "Val1", "Val2"]` instead of `["Foo Bar", "Val1", "Val2"]`.
    #[test]
    fn test_reconstruct_table_close_but_distinct_columns_not_merged() {
        let words = vec![
            word("Name", 0, 0, 20, 20),
            word("Wid", 150, 0, 15, 20),
            word("Zone", 180, 0, 20, 20),
            word("Foo", 0, 40, 20, 20),
            word("Bar", 25, 40, 15, 20),
            word("Val1", 150, 40, 15, 20),
            word("Val2", 180, 40, 20, 20),
        ];

        let table = reconstruct_table(&words, 10, 0.5);

        assert_eq!(table.len(), 2);
        assert_eq!(
            table[0].len(),
            3,
            "Wid/Zone must stay 2 distinct columns, not collapsed by the Foo/Bar merge fix"
        );
        assert_eq!(
            table[0],
            vec!["Name".to_string(), "Wid".to_string(), "Zone".to_string()]
        );
        assert_eq!(
            table[1],
            vec!["Foo Bar".to_string(), "Val1".to_string(), "Val2".to_string()],
            "Foo+Bar merge into one cell; Val1 and Val2 remain distinct cells, not merged together"
        );
    }

    /// Column count must be stable across different `column_threshold`
    /// values once multi-word cells are merged before column detection.
    ///
    /// TEST HONESTY: without the fix, threshold 10 yields 4 columns
    /// ([0, 15, 30, 300], "A"/"B"/"C" each minting a separate column) while
    /// threshold 30 over-merges the spurious columns into the real ones,
    /// yielding only 2 columns ([15, 300]) — i.e. 4 != 2, so the equality
    /// assertion below fails on unfixed code even though 2 happens to match
    /// one of the two unfixed outputs by coincidence.
    #[test]
    fn test_reconstruct_table_column_count_stable_across_thresholds() {
        let words = vec![
            word("H1", 0, 0, 20, 20),
            word("H2", 300, 0, 30, 20),
            word("A", 0, 40, 10, 20),
            word("B", 15, 40, 10, 20),
            word("C", 30, 40, 10, 20),
            word("Val", 300, 40, 20, 20),
        ];

        let table_tight = reconstruct_table(&words, 10, 0.5);
        let table_loose = reconstruct_table(&words, 30, 0.5);

        assert_eq!(table_tight[0].len(), 2);
        assert_eq!(
            table_tight[0].len(),
            table_loose[0].len(),
            "column count must not depend on column_threshold once cells are pre-merged"
        );
        assert_eq!(table_tight, table_loose);
        assert_eq!(table_tight[0], vec!["H1".to_string(), "H2".to_string()]);
        assert_eq!(table_tight[1], vec!["A B C".to_string(), "Val".to_string()]);
    }

    /// Degenerate-input guard: empty input. This early return predates the
    /// fix and behaves identically before and after it — included for
    /// completeness of degenerate-input coverage, not as a fix-detecting
    /// regression test (there is nothing for the fix to change here).
    #[test]
    fn test_reconstruct_table_empty_words_returns_empty_table() {
        let table = reconstruct_table(&[], 20, 0.5);
        assert!(table.is_empty());
    }

    /// Degenerate-input guard: a single word can never trigger the merge
    /// step (nothing to merge with), so this is identical before and after
    /// the fix. It exists to prove `group_words_into_cell_tokens`'s
    /// `words.len() <= 1` guard doesn't panic or drop the word.
    #[test]
    fn test_reconstruct_table_single_word_no_panic() {
        let words = vec![word("Solo", 10, 10, 20, 20)];
        let table = reconstruct_table(&words, 20, 0.5);
        assert_eq!(table, vec![vec!["Solo".to_string()]]);
    }

    /// Degenerate-input guard: zero-height words drive the merge-gap
    /// threshold to exactly 0 (`0.6 * 0`), so only touching/overlapping
    /// words (gap <= 0) merge, and the computation must not panic (no
    /// division by a zero median height, no overflow in the saturating
    /// bbox-union arithmetic).
    ///
    /// TEST HONESTY: without the fix, "X" (x=0..10) and "Y" (x=10..20) —
    /// which touch with a zero-pixel gap — are still 2 separate columns
    /// under raw `detect_columns` (their left edges differ by 10, over the
    /// column_threshold of 5), giving 3 columns (`["X", "Y", "Z"]`) instead
    /// of the correct 2 (`["X Y", "Z"]`).
    #[test]
    fn test_reconstruct_table_zero_height_words_merge_on_touching_gap() {
        let words = vec![word("X", 0, 0, 10, 0), word("Y", 10, 0, 10, 0), word("Z", 50, 0, 10, 0)];

        let table = reconstruct_table(&words, 5, 0.5);

        assert_eq!(table.len(), 1);
        assert_eq!(
            table[0].len(),
            2,
            "X and Y touch with a 0px gap and must merge into one cell"
        );
        assert_eq!(table[0], vec!["X Y".to_string(), "Z".to_string()]);
    }

    /// GH#1628: a cell's words must read left-to-right, not in the order they happened to
    /// arrive in `words`.
    ///
    /// A sub/superscript is drawn as its own content-stream segment sitting a fraction of a point
    /// below the line it annotates, so every reading-order sort upstream of here places it after
    /// the whole line rather than beside the symbol it belongs to. By the time words reach this
    /// function the subscript's `left` still says exactly where it goes --- `eta`(206),
    /// `S`(211), `%`(253) --- but slice order says `eta`, `%`, `S`, and the join followed slice
    /// order. The reported output was `E % S` for `eta_S %` and `Q GJ HE` for `Q_HE GJ`.
    ///
    /// Left ordering is the correct rule here regardless of subscripts: every word in one cell
    /// sits in the same row band by construction (`find_row_index` put it there), so `left` is
    /// reading order for that cell.
    #[test]
    fn issue_1628_cell_words_join_in_left_to_right_order() {
        let words = vec![
            word("eta", 206, 100, 7, 12),
            word("%", 253, 100, 6, 12),
            word("S", 211, 104, 3, 7),
        ];

        let table = assign_words_to_cells(&words, &[100], &[206]);

        assert_eq!(
            table,
            vec![vec!["eta S %".to_string()]],
            "a subscript must join between the symbol it annotates and the next unit, not after it"
        );
    }

    /// GH#1628 negative control: a cell holding two visual lines must NOT interleave them.
    ///
    /// This is the failure mode a naive left-only sort introduces, and it is strictly worse than
    /// the bug being fixed: it scrambles every wrapped cell in the corpus, not just cells with
    /// sub/superscripts. Line grouping is what separates the two cases --- a subscript sits a
    /// fraction of its own height off its base's centre, a second line sits a full line height
    /// away.
    #[test]
    fn issue_1628_cell_with_two_visual_lines_does_not_interleave_them() {
        let words = vec![
            word("Hello", 100, 100, 30, 12),
            word("world", 140, 100, 30, 12),
            word("again", 100, 116, 30, 12),
            word("here", 140, 116, 30, 12),
        ];

        let table = assign_words_to_cells(&words, &[106], &[100]);

        assert_eq!(table, vec![vec!["Hello world again here".to_string()]]);
    }

    /// GH#1628: a segment whose text holds no whitespace is returned unsplit, so its word keeps
    /// the segment's own trailing space. Joining those words then stacks that space on the
    /// separator and renders `Q_HE GJ` as `Q  HE GJ`. Each word's own whitespace must collapse so
    /// the join alone owns the spacing.
    #[test]
    fn issue_1628_cell_join_does_not_stack_a_word_s_own_trailing_space() {
        let words = vec![word("Q ", 206, 100, 8, 12), word("HE", 213, 104, 7, 7)];

        let table = assign_words_to_cells(&words, &[100], &[206]);

        assert_eq!(table, vec![vec!["Q HE".to_string()]]);
    }

    /// Degenerate-input guard: every word on one row is a degenerate case
    /// for the per-row bucketing inside `group_words_into_cell_tokens` (a
    /// single bucket holding every word) — must not panic and must still
    /// merge/split correctly within that one row.
    ///
    /// TEST HONESTY: without the fix, "A" (x=0) and "B" (x=15) are 2
    /// separate columns under raw `detect_columns` (gap 15 > the
    /// column_threshold of 5), giving 3 columns (`["A", "B", "C"]`) instead
    /// of the correct 2 (`["A B", "C"]`).
    #[test]
    fn test_reconstruct_table_all_words_one_row() {
        let words = vec![
            word("A", 0, 0, 10, 10),
            word("B", 15, 0, 10, 10),
            word("C", 100, 0, 10, 10),
        ];

        let table = reconstruct_table(&words, 5, 0.5);

        assert_eq!(table.len(), 1);
        assert_eq!(table[0].len(), 2);
        assert_eq!(table[0], vec!["A B".to_string(), "C".to_string()]);
    }

    /// `reconstruct_table_with_columns` must return column positions that
    /// stay index-correlated with the returned grid: `positions[i]`
    /// corresponds to `grid[row][i]` for every row. Callers like
    /// `pdf::native::table::reconstruct_region_table_with_column_gap` rely on
    /// this to index `column_positions[column]` against `grid[row][column]`
    /// (xberg-io/xberg#688 blast-radius fix: `reconstruct_table` now detects
    /// columns from post-merge cell tokens, so a caller's own separate
    /// `detect_columns(words, ...)` call no longer corresponds to the grid
    /// `reconstruct_table` returns — this sibling function is how such a
    /// caller gets a mutually consistent pair instead).
    ///
    /// Note: row/column assignment always resolves to the nearest position
    /// (xberg-io/xberg#183), and every column detect_columns reports is the
    /// exact position of one of its member tokens, so that token's word
    /// always maps back to its own column at distance 0. In practice this
    /// means a detect_columns-derived column can never end up entirely
    /// empty across all rows, so remove_empty_rows_and_columns cannot
    /// actually strip a column reachable through this path. This test
    /// therefore checks the correspondence invariant directly rather than
    /// depending on triggering that specific removal branch.
    #[test]
    fn test_reconstruct_table_with_columns_positions_correlate_with_grid() {
        let words = vec![
            word("Left", 0, 0, 20, 20),
            word("Right", 300, 0, 20, 20),
            word("L2", 0, 40, 20, 20),
            word("R2", 300, 40, 20, 20),
        ];
        let (grid, positions) = reconstruct_table_with_columns(&words, 20, 0.5);

        assert!(!grid.is_empty());
        assert_eq!(
            positions.len(),
            grid[0].len(),
            "column_positions must have exactly one entry per surviving grid column"
        );
        assert_eq!(positions, vec![0, 300]);
        assert_eq!(grid[0], vec!["Left".to_string(), "Right".to_string()]);
        assert_eq!(grid[1], vec!["L2".to_string(), "R2".to_string()]);
    }

    /// xberg-io/xberg#1649: a wide multi-word cell's trailing word must stay in its own cell's
    /// column even when its raw x-position sits numerically closer to a neighboring column's
    /// anchor. Mirrors the reported fixture's "ACH Deposit - EMPLOYER INC" description cell
    /// sitting immediately left of a narrow "WITHDRAWAL" amount column.
    #[test]
    fn issue_1649_wide_description_word_stays_in_its_own_column_not_the_nearest_anchor() {
        // Coordinates measured on the reported fixture's transaction table (xberg-io/xberg#1649):
        // the header row, and the "ACH Deposit - EMPLOYER INC" data row whose trailing word
        // ("INC") sits closer by raw x-distance to the WITHDRAWAL column anchor (2394) than to
        // the DESCRIPTION column anchor (746).
        let words = vec![
            word("DATE", 126, 1292, 152, 43),
            word("DESCRIPTION", 746, 1292, 412, 44),
            word("WITHDRAWAL", 2394, 1292, 418, 43),
            word("2026-01-05", 125, 1673, 350, 49),
            word("ACH", 741, 1671, 139, 52),
            word("Deposit", 908, 1672, 228, 63),
            word("-", 1159, 1699, 20, 7),
            word("EMPLOYER", 1206, 1671, 360, 52),
            word("INC", 1591, 1671, 108, 52),
        ];

        let table = reconstruct_table(&words, 50, 0.5);

        assert_eq!(table[0], vec!["DATE", "DESCRIPTION", "WITHDRAWAL"]);
        assert_eq!(
            table[1],
            vec!["2026-01-05", "ACH Deposit - EMPLOYER INC", ""],
            "INC must join the description cell it visually belongs to, not the far column its \
             own x-position happens to be nearest to"
        );
    }

    /// xberg-io/xberg#1649: a right-aligned amount column split into two x-position buckets by
    /// digit-width drift (one header-only, one holding every value) must be folded back into one
    /// logical column.
    #[cfg(feature = "ocr")]
    #[test]
    fn issue_1649_merge_disjoint_numeric_columns_folds_a_drift_split_column() {
        let mut table = vec![
            vec!["WITHDRAWAL".to_string(), "".to_string(), "DEPOSIT".to_string()],
            vec!["".to_string(), "$1,250.00".to_string(), "".to_string()],
            vec!["".to_string(), "$87.32".to_string(), "".to_string()],
            vec!["".to_string(), "".to_string(), "$500.00".to_string()],
        ];
        let mut positions = vec![100_u32, 120, 600];

        merge_disjoint_numeric_columns(&mut table, &mut positions, 50);

        assert_eq!(positions.len(), 2, "the drift-split pair must fold into one column");
        assert_eq!(table[0], vec!["WITHDRAWAL".to_string(), "DEPOSIT".to_string()]);
        assert_eq!(table[1], vec!["$1,250.00".to_string(), "".to_string()]);
        assert_eq!(table[2], vec!["$87.32".to_string(), "".to_string()]);
        assert_eq!(table[3], vec!["".to_string(), "$500.00".to_string()]);
    }

    /// Negative control for xberg-io/xberg#1649's fix: two independently-labeled, mutually
    /// exclusive numeric columns (a ledger's "Debit" and "Credit") must NOT be folded together
    /// just because they sit close together and never both hold a value on the same row --
    /// that shape is an ordinary compact ledger layout, not a drift-split artifact, and each
    /// column here carries its own header text.
    #[cfg(feature = "ocr")]
    #[test]
    fn issue_1649_merge_disjoint_numeric_columns_leaves_independently_headed_columns_alone() {
        let mut table = vec![
            vec!["Debit".to_string(), "Credit".to_string()],
            vec!["$1,250.00".to_string(), "".to_string()],
            vec!["".to_string(), "$87.32".to_string()],
        ];
        let mut positions = vec![100_u32, 120];

        merge_disjoint_numeric_columns(&mut table, &mut positions, 50);

        assert_eq!(
            positions.len(),
            2,
            "two independently headed columns must not be merged"
        );
        assert_eq!(table[0], vec!["Debit".to_string(), "Credit".to_string()]);
    }

    /// xberg-io/xberg#1649: a section caption ("TRANSACTIONS") that shares a table region with
    /// the real header row must be dropped, leaving the header as row 0.
    #[cfg(feature = "ocr")]
    #[test]
    fn issue_1649_drop_leading_caption_row_removes_a_lone_leading_caption() {
        let mut table = vec![
            vec!["TRANSACTIONS".to_string(), "".to_string(), "".to_string()],
            vec!["DATE".to_string(), "DESCRIPTION".to_string(), "BALANCE".to_string()],
            vec![
                "2026-01-02".to_string(),
                "Opening Balance".to_string(),
                "$10,500.00".to_string(),
            ],
        ];

        drop_leading_caption_row(&mut table);

        assert_eq!(table.len(), 2, "the caption row must be dropped");
        assert_eq!(table[0], vec!["DATE", "DESCRIPTION", "BALANCE"]);
    }

    /// Negative control: a genuine two-row summary card (header + one data row) must never lose
    /// its header row -- there is no separate caption to drop here, only a legitimate header.
    #[cfg(feature = "ocr")]
    #[test]
    fn issue_1649_drop_leading_caption_row_leaves_a_two_row_summary_card_alone() {
        let mut table = vec![
            vec![
                "ACCOUNT TYPE".to_string(),
                "STATEMENT PERIOD".to_string(),
                "CLOSING BALANCE".to_string(),
            ],
            vec![
                "Checking".to_string(),
                "Jan 1 - Jan 31, 2026".to_string(),
                "$12,847.65".to_string(),
            ],
        ];

        drop_leading_caption_row(&mut table);

        assert_eq!(table.len(), 2, "a two-row table has no caption to drop");
        assert_eq!(table[0][0], "ACCOUNT TYPE");
    }

    /// xberg-io/xberg#1649: a lone punctuation glyph (e.g. the dash in a date range) must merge
    /// into its neighboring cell even when its gap is slightly wider than the normal cell-merge
    /// threshold, so it does not mint a spurious extra column.
    #[test]
    fn issue_1649_punctuation_glyph_merges_across_a_widened_gap() {
        let words = vec![
            word("Jan", 0, 0, 30, 20),
            word("1", 35, 0, 10, 20),
            // Gap from "1"'s right edge (45) to "-"'s left edge (57) is 12px; gap from "-"'s
            // right edge (67) to "Jan"'s left edge (80) is 13px. Median height is 20, so the
            // normal merge_gap (0.6 * 20 = 12) alone would leave the second gap (13) unmerged.
            word("-", 57, 0, 10, 20),
            word("Jan", 80, 0, 30, 20),
            word("31,", 115, 0, 25, 20),
            word("2026", 145, 0, 40, 20),
        ];

        let table = reconstruct_table(&words, 500, 0.5);

        assert_eq!(table.len(), 1);
        assert_eq!(
            table[0],
            vec!["Jan 1 - Jan 31, 2026".to_string()],
            "the dash must glue the date range into a single cell, not split it"
        );
    }

    /// Regression for the punctuation-glue-gap connector rule (xberg-io/xberg#1649 review
    /// follow-up): a lone punctuation cell (e.g. a "-" placeholder for a zero amount) sitting
    /// near a real column boundary must not steal the preceding word into its cell just because
    /// it is punctuation. The widened gap must apply only when the punctuation genuinely
    /// bridges two neighboring words on both sides.
    #[test]
    fn issue_1649_lone_punctuation_cell_keeps_its_own_column() {
        let words = vec![
            word("Debit", 0, 0, 50, 20),
            // Gap from "Debit" (right edge 50) to "-" (left 63) is 13px -- just over the normal
            // merge_gap (0.6 * 20 = 12) but under the punctuation-widened gap (24). Gap from "-"
            // (right edge 73) to "Credit" (left 130) is 57px -- nowhere near even the widened
            // gap, so "-" has no genuine neighbour to bridge and must not glue to "Debit"
            // either. ~keep
            word("-", 63, 0, 10, 20),
            word("Credit", 130, 0, 60, 20),
        ];

        let table = reconstruct_table(&words, 20, 0.5);

        assert_eq!(table.len(), 1);
        assert_eq!(
            table[0],
            vec!["Debit".to_string(), "-".to_string(), "Credit".to_string()],
            "an isolated punctuation cell with no genuine neighbour must not glue to the preceding word"
        );
    }
}