pdf_inspector/

lib.rs

// Rust 1.95 introduced collapsible_match for `if` inside match arms.
// The content-stream parsers use this pattern extensively (match on operator
// name, then check `in_text_block && !op.operands.is_empty()`). Collapsing
// these into match guards would hurt readability. Allow crate-wide.
#![allow(clippy::collapsible_match)]

//! Smart PDF detection and text extraction using lopdf
//!
//! # Quick start
//!
//! ```no_run
//! // Full processing (detect + extract + markdown) with defaults
//! let result = pdf_inspector::process_pdf("document.pdf").unwrap();
//! println!("type: {:?}, pages: {}", result.pdf_type, result.page_count);
//! if let Some(md) = &result.markdown {
//!     println!("{md}");
//! }
//!
//! // Fast metadata-only detection (no text extraction)
//! let info = pdf_inspector::detect_pdf("document.pdf").unwrap();
//! println!("type: {:?}, pages: {}", info.pdf_type, info.page_count);
//!
//! // Custom options via builder
//! use pdf_inspector::{PdfOptions, ProcessMode};
//! let result = pdf_inspector::process_pdf_with_options(
//!     "document.pdf",
//!     PdfOptions::new().mode(ProcessMode::Analyze),
//! ).unwrap();
//! ```

#[cfg(feature = "python")]
pub mod python;

pub mod adobe_korea1;
mod bidi;
mod bidi_mirroring;
pub mod detector;
pub mod extractor;
mod form_bbox_repair;
pub mod glyph_names;
mod mac_glyph_order;
pub mod markdown;
mod overlong_numerals;
pub mod process_mode;
pub mod structure_tree;
pub mod tables;
mod text_quality;
pub mod text_utils;
pub mod tounicode;
pub mod types;
pub mod vision;
mod xref_repair;

pub use detector::{
    detect_pdf_type, detect_pdf_type_mem, detect_pdf_type_mem_with_config,
    detect_pdf_type_with_config, DetectionConfig, PdfType, PdfTypeResult, ScanStrategy,
};
pub use extractor::geometry::PageRotation;
pub use extractor::{
    extract_text, extract_text_with_positions, extract_text_with_positions_and_rotations_mem,
    extract_text_with_positions_and_rotations_mem_in_frame,
    extract_text_with_positions_and_rotations_mem_with_options, extract_text_with_positions_mem,
    extract_text_with_positions_mem_in_frame, extract_text_with_positions_mem_with_options,
    extract_text_with_positions_pages, extract_text_with_positions_pages_with_password,
    PositionFrame, PositionOptions,
};
pub use markdown::{
    to_markdown, to_markdown_from_items, to_markdown_from_items_with_rects,
    to_markdown_from_items_with_rects_and_page_count, MarkdownOptions, MarkdownProfile,
};
pub use process_mode::ProcessMode;
pub use types::{BoldSource, LayoutComplexity, PdfLine, PdfRect, TextItem};

use lopdf::Document;
use std::collections::{BTreeMap, HashMap, HashSet};
use std::path::Path;
use text_quality::{
    analyze_text_quality, detect_encoding_issues, is_cid_garbage, is_garbage_text,
    region_items_have_decoding_issue,
};
use tounicode::FontCMaps;

#[cfg(not(target_arch = "wasm32"))]
struct ProcessingTimer(std::time::Instant);

#[cfg(target_arch = "wasm32")]
struct ProcessingTimer;

impl ProcessingTimer {
    fn start() -> Self {
        #[cfg(not(target_arch = "wasm32"))]
        {
            Self(std::time::Instant::now())
        }

        #[cfg(target_arch = "wasm32")]
        {
            Self
        }
    }

    fn elapsed_ms(&self) -> u64 {
        #[cfg(not(target_arch = "wasm32"))]
        {
            self.0.elapsed().as_millis() as u64
        }

        #[cfg(target_arch = "wasm32")]
        {
            // The wasm32-unknown-unknown standard library has no clock.
            // Browser bindings measure with JavaScript's host clock.
            0
        }
    }
}

/// OCR reason emitted when the extracted text layer appears garbled due to
/// broken font decoding or mojibake.
pub const OCR_REASON_SUSPECTED_GARBLED_TEXT: &str = "suspected_garbled_text";

/// OCR reason: the page is a scanned image (a full-page raster / image-only
/// page) with no usable text layer.
pub const OCR_REASON_SCANNED: &str = "scanned";

/// OCR reason: the page has no extractable text and no image to OCR — blank,
/// or content the parser cannot reach.
pub const OCR_REASON_NO_TEXT: &str = "no_text";

/// OCR reason: the page's text is drawn as vector outlines (path operators)
/// rather than real text operators, so it cannot be extracted as characters.
pub const OCR_REASON_VECTOR_TEXT: &str = "vector_text";

/// OCR reason: every text-showing operator on the page leaves nothing to
/// see — text render mode 3 (invisible), or mode 7 (clip only) with
/// nothing painted through the clip — while an image covers at least half
/// of the page: a scan carrying a text layer nobody sees. What that layer
/// says is not what the page shows, so the page is read from its raster.
pub const OCR_REASON_INVISIBLE_TEXT_LAYER: &str = "invisible_text_layer";

// =========================================================================
// Result type
// =========================================================================

/// OCR reasons for a single 1-indexed page.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct PageOcrReasons {
    /// 1-indexed page number.
    pub page: u32,
    /// Machine-readable OCR reason identifiers.
    pub reasons: Vec<String>,
}

/// A font whose ToUnicode CMap (or, for a font without one, the embedded
/// program's own cmap table) had no entry for some of the codes the document
/// shows through it, and what became of those codes.
///
/// A CMap written for some of a font's glyphs but not all of them loses the
/// others' letters from the text. A code without an entry is read from the
/// mapped codes around it when they spell it out — a CMap mapping code 36
/// to `A` and code 38 to `C` says code 37 is `B`, for a run of digits or of
/// letters of one case whose glyph order follows the alphabet — and is a
/// U+FFFD in the text otherwise, so the loss stays visible. The counts let
/// a caller weigh text read from such a font: `codes - interpolated -
/// unmapped` of its codes had an entry.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct FontCMapGaps {
    /// The font's `/BaseFont` name, or its resource name when it has none
    /// or an empty one.
    pub font: String,
    /// Codes shown through the font's CMap, repeats included: two-byte
    /// codes, or the bytes of a single-byte CMap.
    pub codes: u32,
    /// Codes without an entry that were read from the mapped codes around
    /// them.
    pub interpolated: u32,
    /// Codes without an entry that could not be read; each is a U+FFFD in
    /// the text.
    pub unmapped: u32,
}

/// High-level PDF processing result.
#[derive(Debug)]
pub struct PdfProcessResult {
    /// The detected PDF type.
    pub pdf_type: PdfType,
    /// Markdown output (populated in [`ProcessMode::Full`], `None` otherwise).
    pub markdown: Option<String>,
    /// Page count.
    pub page_count: u32,
    /// Processing time in milliseconds.
    pub processing_time_ms: u64,
    /// 1-indexed page numbers that need OCR.
    pub pages_needing_ocr: Vec<u32>,
    /// Machine-readable OCR reasons by 1-indexed page.
    pub ocr_reasons_by_page: Vec<PageOcrReasons>,
    /// The `/Title` of the document information dictionary, decoded as a
    /// PDF text string (UTF-16 or UTF-8 after a byte order mark,
    /// PDFDocEncoding otherwise; see [`PdfTypeResult::title`]). `None` when
    /// the entry is missing or not a string. The entries below follow the
    /// same decoding and missing-value rule.
    pub title: Option<String>,
    /// The document information dictionary's `/Author`.
    pub author: Option<String>,
    /// The document information dictionary's `/Subject`.
    pub subject: Option<String>,
    /// The document information dictionary's `/Keywords`.
    pub keywords: Option<String>,
    /// The document information dictionary's `/Creator`: the application
    /// the document was authored in.
    pub creator: Option<String>,
    /// The document information dictionary's `/Producer`: the application
    /// that wrote the PDF.
    pub producer: Option<String>,
    /// The document information dictionary's `/CreationDate` as written, a
    /// PDF date string such as `D:20240115103000+01'00'`, neither validated
    /// nor converted.
    pub creation_date: Option<String>,
    /// The document information dictionary's `/ModDate` as written, like
    /// `creation_date`.
    pub mod_date: Option<String>,
    /// Detection confidence score (0.0–1.0).
    pub confidence: f32,
    /// Layout complexity analysis (tables, multi-column detection).
    pub layout: LayoutComplexity,
    /// `true` when broken font encodings are detected (garbled text,
    /// replacement characters). Clients should fall back to OCR.
    pub has_encoding_issues: bool,
    /// The fonts whose ToUnicode CMap — or, for a font without one, the
    /// embedded program's cmap table — lacked an entry for a code the
    /// document shows through it, with the counts of codes shown, read from
    /// their neighbours and left as U+FFFD (see [`FontCMapGaps`]). Always
    /// empty in [`ProcessMode::DetectOnly`], which decodes no text;
    /// otherwise empty when every such code had an entry.
    pub cmap_gaps: Vec<FontCMapGaps>,
}

// =========================================================================
// Options builder
// =========================================================================

/// Configuration for [`process_pdf_with_options`] and friends.
///
/// Use the builder methods to customise behaviour:
///
/// ```
/// use pdf_inspector::{PdfOptions, ProcessMode};
///
/// let opts = PdfOptions::new()
///     .mode(ProcessMode::Analyze)
///     .pages([1, 3, 5]);
/// ```
#[derive(Clone)]
pub struct PdfOptions {
    /// How far the pipeline should run (default: [`ProcessMode::Full`]).
    pub mode: ProcessMode,
    /// Detection configuration.
    pub detection: DetectionConfig,
    /// Markdown formatting options (only used in [`ProcessMode::Full`]).
    pub markdown: MarkdownOptions,
    /// Optional set of 1-indexed pages to process.  `None` = all pages.
    pub page_filter: Option<HashSet<u32>>,
    /// Password for decrypting an encrypted PDF. `None` falls back to the
    /// empty password (owner-only encryption).
    pub password: Option<String>,
}

// Manual `Debug` so the password is never leaked through debug logging or a
// panic that formats the options; it renders as `Some("[REDACTED]")`.
impl std::fmt::Debug for PdfOptions {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("PdfOptions")
            .field("mode", &self.mode)
            .field("detection", &self.detection)
            .field("markdown", &self.markdown)
            .field("page_filter", &self.page_filter)
            .field("password", &self.password.as_ref().map(|_| "[REDACTED]"))
            .finish()
    }
}

impl Default for PdfOptions {
    fn default() -> Self {
        Self {
            mode: ProcessMode::Full,
            detection: DetectionConfig::default(),
            markdown: MarkdownOptions::default(),
            page_filter: None,
            password: None,
        }
    }
}

impl PdfOptions {
    /// Create options with all defaults ([`ProcessMode::Full`]).
    pub fn new() -> Self {
        Self::default()
    }

    /// Shorthand for detect-only options.
    pub fn detect_only() -> Self {
        Self {
            mode: ProcessMode::DetectOnly,
            ..Self::default()
        }
    }

    /// Set the processing mode.
    pub fn mode(mut self, mode: ProcessMode) -> Self {
        self.mode = mode;
        self
    }

    /// Set detection configuration.
    pub fn detection(mut self, config: DetectionConfig) -> Self {
        self.detection = config;
        self
    }

    /// Set markdown formatting options.
    pub fn markdown(mut self, options: MarkdownOptions) -> Self {
        self.markdown = options;
        self
    }

    /// Limit processing to specific 1-indexed pages.
    pub fn pages(mut self, pages: impl IntoIterator<Item = u32>) -> Self {
        self.page_filter = Some(pages.into_iter().collect());
        self
    }

    /// Set the password used to decrypt an encrypted PDF.
    pub fn password(mut self, password: impl Into<String>) -> Self {
        self.password = Some(password.into());
        self
    }
}

// =========================================================================
// Public convenience functions
// =========================================================================

/// Process a PDF file with full extraction (detect → extract → markdown).
///
/// This is the most common entry point.  Equivalent to
/// `process_pdf_with_options(path, PdfOptions::new())`.
pub fn process_pdf<P: AsRef<Path>>(path: P) -> Result<PdfProcessResult, PdfError> {
    process_pdf_with_options(path, PdfOptions::new())
}

/// Fast metadata-only detection — no text extraction or markdown generation.
///
/// Equivalent to `process_pdf_with_options(path, PdfOptions::detect_only())`.
pub fn detect_pdf<P: AsRef<Path>>(path: P) -> Result<PdfProcessResult, PdfError> {
    process_pdf_with_options(path, PdfOptions::detect_only())
}

/// Process a PDF file with custom options.
///
/// The document is loaded **once** and shared between detection and extraction.
pub fn process_pdf_with_options<P: AsRef<Path>>(
    path: P,
    options: PdfOptions,
) -> Result<PdfProcessResult, PdfError> {
    let start = ProcessingTimer::start();
    validate_pdf_file(&path)?;

    // Load the document once — shared by detection AND extraction.
    let (doc, page_count) =
        load_document_from_path_with_password(&path, options.password.as_deref())?;

    process_document(doc, page_count, options, start)
}

/// Process a PDF from a memory buffer with full extraction.
pub fn process_pdf_mem(buffer: &[u8]) -> Result<PdfProcessResult, PdfError> {
    process_pdf_mem_with_options(buffer, PdfOptions::new())
}

/// Fast metadata-only detection from a memory buffer.
pub fn detect_pdf_mem(buffer: &[u8]) -> Result<PdfProcessResult, PdfError> {
    process_pdf_mem_with_options(buffer, PdfOptions::detect_only())
}

/// Process a PDF from a memory buffer with custom options.
///
/// The buffer is parsed **once** and shared between detection and extraction.
pub fn process_pdf_mem_with_options(
    buffer: &[u8],
    options: PdfOptions,
) -> Result<PdfProcessResult, PdfError> {
    let start = ProcessingTimer::start();
    validate_pdf_bytes(buffer)?;

    let (doc, page_count) =
        load_document_from_mem_with_password(buffer, options.password.as_deref())?;

    process_document(doc, page_count, options, start)
}

// =========================================================================
// Deprecated compat shims
// =========================================================================

/// Process a PDF file with custom detection and markdown configuration.
#[deprecated(since = "0.2.0", note = "Use process_pdf_with_options instead")]
pub fn process_pdf_with_config<P: AsRef<Path>>(
    path: P,
    config: DetectionConfig,
    markdown_options: MarkdownOptions,
) -> Result<PdfProcessResult, PdfError> {
    process_pdf_with_options(
        path,
        PdfOptions::new()
            .detection(config)
            .markdown(markdown_options),
    )
}

/// Process a PDF file with custom configuration and optional page filter.
#[deprecated(since = "0.2.0", note = "Use process_pdf_with_options instead")]
pub fn process_pdf_with_config_pages<P: AsRef<Path>>(
    path: P,
    config: DetectionConfig,
    markdown_options: MarkdownOptions,
    page_filter: Option<&HashSet<u32>>,
) -> Result<PdfProcessResult, PdfError> {
    let mut opts = PdfOptions::new()
        .detection(config)
        .markdown(markdown_options);
    opts.page_filter = page_filter.cloned();
    process_pdf_with_options(path, opts)
}

/// Process PDF from memory buffer with custom detection and markdown configuration.
#[deprecated(since = "0.2.0", note = "Use process_pdf_mem_with_options instead")]
pub fn process_pdf_mem_with_config(
    buffer: &[u8],
    config: DetectionConfig,
    markdown_options: MarkdownOptions,
) -> Result<PdfProcessResult, PdfError> {
    process_pdf_mem_with_options(
        buffer,
        PdfOptions::new()
            .detection(config)
            .markdown(markdown_options),
    )
}

// =========================================================================
// Region-based text extraction (for hybrid OCR pipelines)
// =========================================================================

/// Lightweight classification result for routing decisions.
#[derive(Debug)]
pub struct PdfClassification {
    /// The detected PDF type.
    pub pdf_type: PdfType,
    /// Total page count.
    pub page_count: u32,
    /// 0-indexed page numbers that need OCR (scanned/image pages).
    pub pages_needing_ocr: Vec<u32>,
    /// Detection confidence score (0.0–1.0).
    pub confidence: f32,
}

/// Classify a PDF from a memory buffer without extracting text.
/// Returns the PDF type and which pages need OCR (~10-50ms).
pub fn classify_pdf_mem(buffer: &[u8]) -> Result<PdfClassification, PdfError> {
    validate_pdf_bytes(buffer)?;
    let (doc, page_count) = load_document_from_mem(buffer)?;
    let detection = detector::detect_from_document(&doc, page_count, &DetectionConfig::default())?;
    Ok(PdfClassification {
        pdf_type: detection.pdf_type,
        page_count,
        // Convert from 1-indexed to 0-indexed for caller convenience
        pages_needing_ocr: detection.pages_needing_ocr.iter().map(|&p| p - 1).collect(),
        confidence: detection.confidence,
    })
}

/// The PDF written back out with the `/BBox` of its Form XObjects repaired
/// — a zero-area box widened, numerals too large for any parser saturated
/// — for callers that render the document with their own renderer. Rust
/// only: the Python, Node.js and WebAssembly bindings do not expose it.
///
/// Some producers write `/BBox [0 0 0 0]` on a form XObject that holds a
/// page's content; taken as the clip it declares, the box hides the form
/// entirely, and a page drawn through it renders blank. Others write the
/// box as ±(DBL_MAX / 2) in full, 308-digit numerals no integer parser
/// holds: the form drops out of the document for one reader and clips to
/// nothing for another. pdf-inspector repairs both whenever it loads a
/// document, so its own extraction and the renderer of its OCR pipeline
/// see the content; a renderer given the original bytes does not, and can
/// be given these instead.
///
/// Returns `Ok(None)` when no form needs the repair, and for an encrypted
/// document — whether or not it opens without a password — since a plain
/// serialization would drop its protection; the OCR pipeline renders a
/// decrypted copy of such a document in memory instead. Otherwise
/// `Ok(Some(bytes))` holds a plain serialization of the loaded document:
/// object streams and incremental updates are flattened, and the file's
/// own repairs — a recovered cross-reference table, a missing end-of-file
/// marker — are folded in.
pub fn widen_degenerate_form_bboxes_mem(buffer: &[u8]) -> Result<Option<Vec<u8>>, PdfError> {
    validate_pdf_bytes(buffer)?;
    let (mut doc, _page_count, repairs) = match load_document_from_mem_with_repairs(buffer, None) {
        Ok(loaded) => loaded,
        Err(PdfError::Encrypted) => return Ok(None),
        Err(error) => return Err(error),
    };
    if !repairs.repaired_forms() || doc.is_encrypted() || doc.encryption_state.is_some() {
        return Ok(None);
    }
    Ok(form_bbox_repair::serialize_for_rendering(&mut doc))
}

// =========================================================================
// Per-page markdown extraction
// =========================================================================

/// Per-page markdown extraction result.
#[derive(Debug)]
pub struct PageMarkdown {
    /// 0-indexed page number.
    pub page: u32,
    /// Formatted markdown for this page.
    pub markdown: String,
    /// `true` when text on this page is unreliable (GID-encoded fonts,
    /// encoding issues, garbage text, or empty extraction).
    pub needs_ocr: bool,
    /// Machine-readable OCR reason when the cause is known.
    pub ocr_reason: Option<String>,
}

/// Combined per-page markdown extraction and layout classification result.
#[derive(Debug)]
pub struct PagesExtractionResult {
    /// Per-page markdown results.
    pub pages: Vec<PageMarkdown>,
    /// 1-indexed pages where tables were detected.
    pub pages_with_tables: Vec<u32>,
    /// 1-indexed pages where multi-column layout was detected.
    pub pages_with_columns: Vec<u32>,
    /// 1-indexed pages that need OCR (scanned/image-based).
    pub pages_needing_ocr: Vec<u32>,
    /// Machine-readable OCR reasons by 1-indexed page.
    pub ocr_reasons_by_page: Vec<PageOcrReasons>,
    /// True if any page has tables or columns.
    pub is_complex: bool,
}

pub(crate) struct InternalPagesExtraction {
    pub(crate) result: PagesExtractionResult,
    #[cfg(all(feature = "ocr", not(target_arch = "wasm32")))]
    pub(crate) page_count: u32,
    #[cfg(all(feature = "ocr", not(target_arch = "wasm32")))]
    pub(crate) supplemental_ocr_regions: BTreeMap<u32, Vec<PdfRect>>,
    /// The document written back out for the renderer when form XObjects
    /// were repaired at load (see `form_bbox_repair`); `None` when the
    /// original bytes render as loaded.
    #[cfg(all(feature = "ocr", not(target_arch = "wasm32")))]
    pub(crate) render_bytes: Option<Vec<u8>>,
}

/// Extract formatted markdown for pages of a PDF, with layout
/// classification metadata.
///
/// Unlike [`process_pdf_mem`] which returns one concatenated markdown string,
/// this returns per-page markdown so callers can mix direct extraction
/// (for simple text pages) with GPU OCR (for complex/scanned pages).
///
/// When `pages` is `None`, every page (0-indexed, in document order) is
/// returned. When `Some(&[...])`, only the listed 0-indexed pages are
/// returned, in the caller's order.
///
/// Font statistics are computed from the full document so header
/// detection thresholds are consistent regardless of which pages are
/// requested. Per-page `needs_ocr` is set when the page has GID-encoded
/// fonts, encoding issues, or garbage text.
///
/// Layout complexity (tables, columns) is computed from the full document
/// at near-zero cost since the items/rects/lines are already in memory.
pub fn extract_pages_markdown_mem(
    buffer: &[u8],
    pages: Option<&[u32]>,
) -> Result<PagesExtractionResult, PdfError> {
    extract_pages_markdown_mem_impl(
        buffer,
        pages,
        None,
        &MarkdownOptions::default(),
        false,
        false,
        false,
    )
    .map(|extraction| extraction.result)
}

#[cfg(all(feature = "ocr", not(target_arch = "wasm32")))]
/// `render_repairs` asks for the repaired document to be written back out
/// for the renderer when the loader changed it (see `form_bbox_repair`);
/// a caller that will not render leaves it off.
pub(crate) fn extract_pages_markdown_mem_for_ocr(
    buffer: &[u8],
    pages: Option<&[u32]>,
    password: Option<&str>,
    markdown_options: &MarkdownOptions,
    render_repairs: bool,
) -> Result<InternalPagesExtraction, PdfError> {
    extract_pages_markdown_mem_impl(
        buffer,
        pages,
        password,
        markdown_options,
        markdown_options.strip_headers_footers,
        true,
        render_repairs,
    )
}

fn extract_pages_markdown_mem_impl(
    buffer: &[u8],
    pages: Option<&[u32]>,
    password: Option<&str>,
    markdown_options: &MarkdownOptions,
    strip_repeated_headers_footers: bool,
    preserve_ocr_candidates: bool,
    render_repairs: bool,
) -> Result<InternalPagesExtraction, PdfError> {
    validate_pdf_bytes(buffer)?;
    let (doc, page_count, repairs) = load_document_from_mem_with_repairs(buffer, password)?;
    #[cfg(all(feature = "ocr", not(target_arch = "wasm32")))]
    let mut doc = doc;
    #[cfg(not(all(feature = "ocr", not(target_arch = "wasm32"))))]
    let _ = (repairs, render_repairs);
    let font_cmaps = FontCMaps::from_doc(&doc);

    // Extract ALL pages to get accurate, document-wide font stats. A malformed
    // unselected page cannot make a valid requested page fail, but errors on a
    // requested page retain the normal extraction semantics.
    let required_pages: Option<HashSet<u32>> = pages.map(|pages| {
        pages
            .iter()
            .filter_map(|page| page.checked_add(1))
            .collect()
    });
    let ((all_items, all_rects, all_lines), page_thresholds, gid_pages, _page_rotations, _) =
        if let Some(required_pages) = required_pages.as_ref() {
            extractor::extract_positioned_text_for_document_analysis(
                &doc,
                &font_cmaps,
                required_pages,
            )?
        } else {
            extractor::extract_positioned_text_from_doc(&doc, &font_cmaps, None)?
        };
    let text_quality = analyze_text_quality(&all_items);

    // Resolve page numbers with full-document context before partitioning.
    // Per-page Markdown receives the original items plus these decisions so
    // table detection can retain legitimate numeric cells.
    let (filtered_items, removed_page_number_pages, page_number_removal_mask) =
        extractor::filter_markdown_page_numbers_with_removed_pages(all_items.clone(), page_count);

    // Tables need the original numeric cells; columns use folio-cleaned
    // evidence so removed page numbers cannot create false layout metadata.
    let chart_regions = markdown::chart_regions_by_page(&all_items, &all_rects, &all_lines);
    let complexity = compute_layout_complexity_with_chart_regions(
        &all_items,
        &filtered_items,
        &all_rects,
        &all_lines,
        &chart_regions,
    );

    // Compute font stats from full document (cross-page consistency).
    let font_stats = markdown::analysis::calculate_font_stats_from_items(&filtered_items);
    let repeated_header_footer_items = if strip_repeated_headers_footers {
        repeated_header_footer_item_keys(&all_items, &page_thresholds, &chart_regions, page_count)
    } else {
        HashSet::new()
    };

    // When caller doesn't specify pages, return every page in document order.
    let all_pages: Vec<u32>;
    let pages_slice: &[u32] = match pages {
        Some(p) => p,
        None => {
            all_pages = (0..page_count).collect();
            &all_pages
        }
    };

    let mut results = Vec::with_capacity(pages_slice.len());
    let mut pages_needing_ocr = Vec::new();
    let mut ocr_reasons_by_page = BTreeMap::new();
    #[cfg(all(feature = "ocr", not(target_arch = "wasm32")))]
    let mut supplemental_ocr_regions = BTreeMap::new();
    let lopdf_pages = doc.get_pages();

    for &page_0idx in pages_slice {
        // Out-of-range pages → empty + needs_ocr
        if page_0idx >= page_count {
            pages_needing_ocr.push(page_0idx + 1);
            results.push(PageMarkdown {
                page: page_0idx,
                markdown: String::new(),
                needs_ocr: true,
                ocr_reason: None,
            });
            continue;
        }

        let page_1idx = page_0idx + 1;

        // Partition items, removal decisions, and rects for this page only.
        let (page_items, page_number_removal_mask): (Vec<TextItem>, Vec<bool>) = all_items
            .iter()
            .zip(&page_number_removal_mask)
            .filter(|(item, _)| {
                item.page == page_1idx
                    && !repeated_header_footer_items.contains(&HeaderFooterItemKey::from(*item))
            })
            .map(|(item, remove)| (item.clone(), *remove))
            .unzip();

        let page_rects: Vec<PdfRect> = all_rects
            .iter()
            .filter(|r| r.page == page_1idx)
            .cloned()
            .collect();

        let page_lines: Vec<types::PdfLine> = all_lines
            .iter()
            .filter(|l| l.page == page_1idx)
            .cloned()
            .collect();

        #[cfg(all(feature = "ocr", not(target_arch = "wasm32")))]
        {
            let image_regions: Vec<PdfRect> = page_items
                .iter()
                .filter_map(supplemental_ocr_image_region)
                .collect();
            if !image_regions.is_empty() {
                supplemental_ocr_regions.insert(page_1idx, image_regions);
            }
        }

        let has_gid = gid_pages.contains(&page_1idx);
        let has_text_quality_issue = text_quality.pages_needing_ocr.contains(&page_1idx);

        // A page can extract cleanly (no decoding issues, non-empty text)
        // while still being fundamentally a scan: a full-page raster with
        // a little genuine native text drawn over it (a header, a stamp, a
        // cover-sheet annotation). Text-quality signals alone can't see
        // that — consult the same "large background image" signal
        // classify_pdf/detect_pdf_type already uses, so the two APIs can't
        // silently disagree on whether a page needs OCR. See #227.
        // Also covers vector-outlined text (glyphs drawn as paths, not
        // shown via a text-showing operator): a hybrid page with real
        // embedded-font body text elsewhere would otherwise still extract
        // non-empty, non-garbled markdown and miss OCR routing entirely.
        // detect_from_document's Mixed-type per-page routing always sends
        // these pages to OCR; mirror that here too. And a page whose every
        // text-showing operator is invisible under a covering image: the
        // text layer it extracts describes the raster rather than being
        // the page's content. All three signals share one
        // analyze_page_content pass — see page_ocr_signals's doc comment.
        let signals = lopdf_pages
            .get(&page_1idx)
            .map(|&page_id| detector::page_ocr_signals(&doc, page_id))
            .unwrap_or_default();
        let has_template_image = signals.template_image_needs_ocr;
        let has_vector_text = signals.has_vector_text;
        let has_invisible_text_layer = signals.has_invisible_text_layer;

        // Build markdown with document-wide font stats
        let options = MarkdownOptions {
            base_font_size: Some(font_stats.most_common_size),
            include_page_numbers: false,
            strip_headers_footers: false,
            ..markdown_options.clone()
        };

        let md = if has_text_quality_issue {
            String::new()
        } else {
            markdown::to_markdown_from_items_with_rects_and_lines(
                page_items,
                options,
                &page_rects,
                &page_lines,
                markdown::MarkdownDocumentContext {
                    page_thresholds: &page_thresholds,
                    struct_roles: None,
                    struct_tables: &[],
                    page_count,
                    prefiltered_page_number_pages: Some(&removed_page_number_pages),
                    prefiltered_page_number_mask: Some(&page_number_removal_mask),
                    precomputed_chart_regions: Some(&chart_regions),
                },
            )
        };

        let has_decoding_issue = has_text_quality_issue
            || (!md.is_empty() && (is_cid_garbage(&md) || detect_encoding_issues(&md)));
        // First among a page's reasons, as classification's
        // `page_ocr_reasons` lists it too, so a page whose whole text layer
        // is hidden under a scan — a scan whatever its fonts are — gets the
        // same first reason from both surfaces; the reasons after it keep
        // this surface's own order.
        if has_invisible_text_layer {
            add_ocr_reason(
                &mut ocr_reasons_by_page,
                page_1idx,
                OCR_REASON_INVISIBLE_TEXT_LAYER,
            );
        }
        if has_decoding_issue {
            add_ocr_reason(
                &mut ocr_reasons_by_page,
                page_1idx,
                OCR_REASON_SUSPECTED_GARBLED_TEXT,
            );
        }
        if has_template_image {
            add_ocr_reason(&mut ocr_reasons_by_page, page_1idx, OCR_REASON_SCANNED);
        }
        if has_vector_text {
            add_ocr_reason(&mut ocr_reasons_by_page, page_1idx, OCR_REASON_VECTOR_TEXT);
        }
        let ocr_reason = page_ocr_reason(&ocr_reasons_by_page, page_1idx);

        let needs_ocr = ocr_reason.is_some()
            || md.trim().is_empty()
            || has_gid
            || is_garbage_text(&md)
            || has_template_image
            || has_vector_text
            || has_invisible_text_layer;

        if needs_ocr {
            pages_needing_ocr.push(page_1idx);
        }

        results.push(PageMarkdown {
            page: page_0idx,
            // The public native extractor continues to suppress unreliable
            // text. The OCR orchestrator retains clean partial text
            // internally so it can compare/fuse it with OCR before deciding
            // what is safe to return.
            markdown: if needs_ocr && !preserve_ocr_candidates {
                String::new()
            } else {
                md
            },
            needs_ocr,
            ocr_reason,
        });
    }

    Ok(InternalPagesExtraction {
        result: PagesExtractionResult {
            pages: results,
            pages_with_tables: complexity.pages_with_tables,
            pages_with_columns: complexity.pages_with_columns,
            pages_needing_ocr,
            ocr_reasons_by_page: page_ocr_reasons_vec(ocr_reasons_by_page),
            is_complex: complexity.is_complex,
        },
        #[cfg(all(feature = "ocr", not(target_arch = "wasm32")))]
        page_count,
        #[cfg(all(feature = "ocr", not(target_arch = "wasm32")))]
        supplemental_ocr_regions,
        // A renderer reading the original bytes would clip a repaired form
        // to nothing, so the OCR pipeline renders the repaired document.
        #[cfg(all(feature = "ocr", not(target_arch = "wasm32")))]
        render_bytes: if render_repairs && repairs.repaired_forms() {
            // A renderer given the original bytes would clip the repaired
            // forms to nothing again, so a copy that cannot be written is
            // an error, not a fallback.
            Some(
                form_bbox_repair::serialize_for_rendering(&mut doc).ok_or_else(|| {
                    PdfError::Parse(
                        "the repaired document could not be written for rendering".to_string(),
                    )
                })?,
            )
        } else {
            None
        },
    })
}

/// Image regions large enough to plausibly contain rasterized document
/// structure such as a table. Logos, icons, and decorative rules remain below
/// these physical-size gates. OCR still has to produce a valid table inside
/// the region before any text is fused into a clean native page.
#[cfg(any(test, all(feature = "ocr", not(target_arch = "wasm32"))))]
fn supplemental_ocr_image_region(item: &TextItem) -> Option<PdfRect> {
    const MIN_WIDTH_PT: f32 = 108.0;
    const MIN_HEIGHT_PT: f32 = 72.0;
    const MIN_AREA_PT2: f32 = 20_000.0;

    if !matches!(item.item_type, types::ItemType::Image)
        || !item.x.is_finite()
        || !item.y.is_finite()
        || !item.width.is_finite()
        || !item.height.is_finite()
    {
        return None;
    }
    let x = item.x.min(item.x + item.width);
    let y = item.y.min(item.y + item.height);
    let width = item.width.abs();
    let height = item.height.abs();
    (width >= MIN_WIDTH_PT && height >= MIN_HEIGHT_PT && width * height >= MIN_AREA_PT2).then_some(
        PdfRect {
            x,
            y,
            width,
            height,
            page: item.page,
        },
    )
}

#[derive(Debug, Clone, PartialEq, Eq, Hash)]
struct HeaderFooterItemKey {
    page: u32,
    x: u32,
    y: u32,
    text: String,
}

impl From<&TextItem> for HeaderFooterItemKey {
    fn from(item: &TextItem) -> Self {
        Self {
            page: item.page,
            x: item.x.to_bits(),
            y: item.y.to_bits(),
            text: item.text.clone(),
        }
    }
}

fn repeated_header_footer_item_keys(
    items: &[TextItem],
    page_thresholds: &HashMap<u32, f32>,
    chart_regions: &HashMap<u32, Vec<(f32, f32, f32, f32)>>,
    page_count: u32,
) -> HashSet<HeaderFooterItemKey> {
    let candidates = items
        .iter()
        .filter(|item| {
            matches!(
                item.item_type,
                types::ItemType::Text | types::ItemType::FormField
            )
        })
        .cloned()
        .collect();
    let lines = extractor::group_prefiltered_items_into_lines_with_thresholds_and_charts(
        candidates,
        page_thresholds,
        &HashSet::new(),
        chart_regions,
    );
    let all_items: HashSet<_> = lines
        .iter()
        .flat_map(|line| line.items.iter().map(HeaderFooterItemKey::from))
        .collect();
    let kept = markdown::strip_repeated_header_footer_lines(lines, page_count);
    let kept_items: HashSet<_> = kept
        .iter()
        .flat_map(|line| line.items.iter().map(HeaderFooterItemKey::from))
        .collect();
    all_items.difference(&kept_items).cloned().collect()
}

#[cfg(all(test, feature = "ocr", not(target_arch = "wasm32")))]
mod ocr_header_footer_tests {
    use super::*;

    fn item(page: u32, text: &str, y: f32) -> TextItem {
        TextItem {
            text: text.to_string(),
            x: 10.0,
            y,
            width: 120.0,
            height: 10.0,
            font: "Test".to_string(),
            font_tag: String::new(),
            legacy_symbol_rewrite: false,
            font_size: 10.0,
            page,
            is_bold: false,
            is_italic: false,
            font_weight: None,
            bold_source: None,
            fixed_pitch: None,
            fill_color: None,
            stroke_color: None,
            render_mode: None,
            is_underline: false,
            is_strikeout: false,
            rotation: 0.0,
            advance_known: true,
            item_type: types::ItemType::Text,
            mcid: None,
            baseline_shift: 0.0,
        }
    }

    #[test]
    fn local_pipeline_prefilters_document_wide_repeated_headers() {
        let mut items = Vec::new();
        let mut thresholds = HashMap::new();
        for page in 1..=3 {
            items.push(item(page, "Repeated report header", 800.0));
            for line in 0..12 {
                items.push(item(
                    page,
                    &format!("Page {page} paragraph {line} unique content"),
                    700.0 - line as f32 * 40.0,
                ));
            }
            thresholds.insert(page, 0.1);
        }

        let removed = repeated_header_footer_item_keys(&items, &thresholds, &HashMap::new(), 3);
        assert_eq!(removed.len(), 2);
        for page in 1..=3 {
            assert_eq!(
                removed.contains(&HeaderFooterItemKey::from(&item(
                    page,
                    "Repeated report header",
                    800.0,
                ))),
                page > 1,
            );
            assert!(!removed.contains(&HeaderFooterItemKey::from(&item(
                page,
                &format!("Page {page} paragraph 5 unique content"),
                500.0,
            ))));
        }
    }
}

/// Path-based wrapper for [`extract_pages_markdown_mem`].
///
/// Reads the PDF from disk and extracts per-page markdown. Pass `None` for
/// `pages` to return every page in document order, or `Some(&[...])` to
/// restrict to specific 0-indexed pages (in caller-supplied order).
pub fn extract_pages_markdown<P: AsRef<Path>>(
    path: P,
    pages: Option<&[u32]>,
) -> Result<PagesExtractionResult, PdfError> {
    validate_pdf_file(&path)?;
    let buffer = read_file(path.as_ref())?;
    extract_pages_markdown_mem(&buffer, pages)
}

// =========================================================================
// Structure-tree element extraction (tagged PDFs)
// =========================================================================

/// One structure-tree element reference from a tagged PDF, resolved to a
/// page and Marked Content ID.
///
/// Join `(page, mcid)` against [`TextItem::page`] / [`TextItem::mcid`] from
/// [`extract_text_with_positions`] to attach semantic roles (heading levels,
/// paragraphs, table cells, …) to extracted text.
#[derive(Debug, Clone)]
pub struct StructureElement {
    /// 1-indexed page number (matches [`TextItem::page`]).
    pub page: u32,
    /// Marked Content ID from the page's content stream (matches
    /// [`TextItem::mcid`]).
    pub mcid: i64,
    /// Standard structure type name ("H1".."H6", "P", "Table", "TD", …).
    /// Custom tags are resolved through the document's `/RoleMap`; tags
    /// with no standard mapping are returned verbatim.
    pub role: String,
}

/// Extract structure-tree element references from a tagged PDF in memory.
///
/// Parses `/StructTreeRoot` (when present) and returns one entry per
/// marked-content reference, resolved to its 1-indexed page, MCID, and
/// structure type name. Returns an empty list when the PDF is not tagged.
///
/// Pass `Some(&[...])` with 1-indexed page numbers (matching
/// [`TextItem::page`]) to restrict output to those pages; pass `None` for
/// the whole document. Entries are sorted by `(page, mcid)`.
pub fn extract_structure_elements_mem(
    buffer: &[u8],
    pages: Option<&[u32]>,
) -> Result<Vec<StructureElement>, PdfError> {
    validate_pdf_bytes(buffer)?;
    let (doc, _page_count) = load_document_from_mem(buffer)?;
    let Some(tree) = structure_tree::StructTree::from_doc(&doc) else {
        return Ok(Vec::new());
    };
    let page_ids = doc.get_pages();
    let roles = tree.mcid_to_roles(&page_ids);

    let page_filter: Option<HashSet<u32>> = pages.map(|p| p.iter().copied().collect());
    let mut elements: Vec<StructureElement> = roles
        .into_iter()
        .filter(|(page, _)| page_filter.as_ref().is_none_or(|f| f.contains(page)))
        .flat_map(|(page, mcids)| {
            mcids.into_iter().map(move |(mcid, role)| StructureElement {
                page,
                mcid,
                role: role.name().to_string(),
            })
        })
        .collect();
    elements.sort_unstable_by_key(|e| (e.page, e.mcid));
    Ok(elements)
}

/// Path-based wrapper for [`extract_structure_elements_mem`].
///
/// Reads the PDF from disk and extracts structure-tree element references.
/// Pass `None` for `pages` to return the whole document, or `Some(&[...])`
/// to restrict to specific 1-indexed pages.
pub fn extract_structure_elements<P: AsRef<Path>>(
    path: P,
    pages: Option<&[u32]>,
) -> Result<Vec<StructureElement>, PdfError> {
    validate_pdf_file(&path)?;
    let buffer = read_file(path.as_ref())?;
    extract_structure_elements_mem(&buffer, pages)
}

// =========================================================================
// Region-based text extraction (for hybrid OCR pipelines)
// =========================================================================

/// Result for a single region's text extraction.
#[derive(Debug)]
pub struct RegionText {
    /// Extracted text (may be empty if region has no text items).
    pub text: String,
    /// `true` when the text should not be trusted and OCR should be used instead.
    /// Set when: the region is empty, the page uses GID-encoded fonts, or the
    /// extracted text fails garbage/encoding checks.
    pub needs_ocr: bool,
    /// Machine-readable OCR reason when the cause is known.
    pub ocr_reason: Option<String>,
}

/// Result for a page's region extractions.
#[derive(Debug)]
pub struct PageRegionResult {
    /// 0-indexed page number.
    pub page: u32,
    /// Per-region results, parallel to the input regions.
    pub regions: Vec<RegionText>,
}

/// Minimum alphanumeric mass an invisible (Tr 3) text layer must carry for
/// the OCR-layer fallback in [`extract_text_in_regions_mem`] to adopt it. A
/// real OCR layer carries far more; a stray watermark or artifact does not.
const OCR_LAYER_MIN_ALNUM: usize = 40;

/// Alphanumeric mass of extracted items, ignoring raster placeholders.
/// `[Image: ...]` items (ItemType::Image) are synthesized for image
/// XObjects — they mark that pixels exist, not that text was read, so they
/// must not count as coverage.
fn non_placeholder_alnum(items: &[TextItem]) -> usize {
    items
        .iter()
        .filter(|it| !matches!(it.item_type, types::ItemType::Image))
        .map(|it| it.text.chars().filter(|c| c.is_alphanumeric()).count())
        .sum()
}

/// Extract text within bounding-box regions from a PDF in memory.
///
/// This is designed for hybrid OCR pipelines: a layout model detects regions
/// in a rendered page image, and this function extracts the PDF text that
/// falls within each region — avoiding GPU OCR for text-based pages.
///
/// Each region result includes a `needs_ocr` flag that is set when extraction
/// quality is suspect (empty text, GID-encoded fonts, garbage/encoding issues).
///
/// # Arguments
///
/// * `buffer` — PDF file bytes
/// * `page_regions` — list of `(page_number_0indexed, Vec<[x1, y1, x2, y2]>)`.
///   Coordinates are in **PDF points** with **top-left origin**, relative to
///   the page's **visible page box** (`CropBox ∩ MediaBox`, else the
///   MediaBox) as laid out in the content stream — the frame
///   [`extract_text_with_positions_mem`] reports items in (with `y` flipped
///   by the box height). `/Rotate` is not applied and a page whose text is
///   predominantly rotated is turned (see [`PageRotation`]), so this matches
///   a rendered page image only for pages with `/Rotate 0` whose text is not
///   predominantly rotated; rects taken from a rendered page go through
///   [`extract_text_in_regions_mem_in_frame`] with [`PositionFrame::Display`].
///
/// # Returns
///
/// A `Vec<PageRegionResult>` parallel to `page_regions`.
pub fn extract_text_in_regions_mem(
    buffer: &[u8],
    page_regions: &[(u32, Vec<[f32; 4]>)],
) -> Result<Vec<PageRegionResult>, PdfError> {
    extract_text_in_regions_mem_in_frame(buffer, page_regions, PositionFrame::Sheet)
}

/// [`extract_text_in_regions_mem`] with the frame of the region rects given
/// explicitly. [`PositionFrame::Sheet`] reads them as that function does;
/// [`PositionFrame::Display`] reads them on the rendered page — the visible
/// page box turned clockwise by the page's inheritable `/Rotate`, top-left
/// origin, `y` down — the frame a layout model working on a page image
/// reports boxes in. Pages whose text is predominantly rotated are handled
/// the same way in both frames.
pub fn extract_text_in_regions_mem_in_frame(
    buffer: &[u8],
    page_regions: &[(u32, Vec<[f32; 4]>)],
    frame: PositionFrame,
) -> Result<Vec<PageRegionResult>, PdfError> {
    extract_text_in_regions_mem_with_options(
        buffer,
        page_regions,
        PositionOptions::new().frame(frame),
    )
}

/// [`extract_text_in_regions_mem_in_frame`] with every option given as a
/// [`PositionOptions`]: the frame the region rects are read in, and whether
/// bold is also read from the font's weight class (`bold_from_weight`: a
/// weight class of `bold_weight_threshold` or more, 600 by default, is bold,
/// and a run the weight makes bold is then its own item while the region's
/// lines are assembled). The default options are
/// [`extract_text_in_regions_mem`].
pub fn extract_text_in_regions_mem_with_options(
    buffer: &[u8],
    page_regions: &[(u32, Vec<[f32; 4]>)],
    options: PositionOptions,
) -> Result<Vec<PageRegionResult>, PdfError> {
    let frame = options.frame;
    validate_pdf_bytes(buffer)?;
    let (doc, _page_count) = load_document_from_mem(buffer)?;
    let pages = doc.get_pages();

    // Build a set of pages we need to extract (1-indexed for lopdf)
    let needed_pages: HashSet<u32> = page_regions.iter().map(|(p, _)| p + 1).collect();

    // Fast mode: skip expensive TrueType font fallback parsing.
    // Fonts that can't be decoded from ToUnicode alone will produce empty/garbage
    // text, triggering needs_ocr=true → GPU OCR fallback in the pipeline.
    let font_cmaps = FontCMaps::from_doc_pages_fast(&doc, Some(&needed_pages));

    // Extract text items for needed pages only
    let mut items_by_page: HashMap<u32, Vec<TextItem>> = HashMap::new();
    let mut page_heights: HashMap<u32, f32> = HashMap::new();
    let mut gid_pages: HashSet<u32> = HashSet::new();
    let mut page_thresholds: HashMap<u32, f32> = HashMap::new();
    let mut rotated_pages: HashMap<u32, RegionCoordSpace> = HashMap::new();
    let mut display_pages: HashMap<u32, extractor::DisplayPage> = HashMap::new();
    let mut style_cache = extractor::FontStyleCache::new();

    for (page_num, &page_id) in pages.iter() {
        if !needed_pages.contains(page_num) {
            continue;
        }

        // Extract text items for this page in the visible-page-box frame;
        // region bounds flip y with that box's height. The Form XObject
        // budget is shared with the invisible-layer retry below so one page
        // cannot consume two full expansion budgets.
        let mut form_budget = extractor::FormWalkBudget::new();
        let extractor::PageBoxExtraction {
            mut items,
            has_gid_fonts: mut has_gid,
            mut coords_rotated,
            skipped_invisible,
            page_box,
            ..
        } = extractor::extract_page_text_items_in_page_box_with_options(
            &doc,
            page_id,
            *page_num,
            &font_cmaps,
            options.text_extraction(false),
            &mut style_cache,
            &mut form_budget,
        )?;
        page_heights.insert(*page_num, page_box.height());
        if frame == PositionFrame::Display {
            display_pages.insert(
                *page_num,
                extractor::DisplayPage::new(&doc, page_id, page_box),
            );
        }
        // OCR-layer fallback: scanned pages often carry their text as an
        // invisible (Tr 3) layer behind the page raster. The visible-only
        // pass sees nothing there but `[Image: ...]` placeholders, so every
        // region on the page reports needs_ocr even though the exact text is
        // embedded in the PDF — and this extractor then disagrees with the
        // markdown path, which already retries Mixed PDFs with the invisible
        // layer included. Retry page-scoped, and only when (a) the first
        // pass actually SKIPPED invisible text — blank pages and image-only
        // scans without an OCR layer must not pay a second content-stream
        // parse (review catch) — and (b) the page has NO visible text item
        // at all (punctuation counts, whitespace-only artifacts don't): an
        // invisible OCR layer transcribes the raster, so any visible glyph
        // has an invisible twin there and adoption would duplicate it
        // (review catches — strict gate, no fuzzy dedupe). Adopt the retry
        // only when it contributes real, non-garbage text.
        let has_visible_text = items.iter().any(|it| {
            !matches!(it.item_type, types::ItemType::Image) && !it.text.trim().is_empty()
        });
        if skipped_invisible && !has_visible_text {
            if let Ok(invisible) = extractor::extract_page_text_items_in_page_box_with_options(
                &doc,
                page_id,
                *page_num,
                &font_cmaps,
                options.text_extraction(true),
                &mut style_cache,
                &mut form_budget,
            ) {
                let inv_items = invisible.items;
                let inv_alnum = non_placeholder_alnum(&inv_items);
                // Judge the WHOLE recovered layer, not a prefix — a broken
                // OCR layer can hide its garbage past any fixed sample size
                // (review catch).
                let sample: String = inv_items
                    .iter()
                    .filter(|it| !matches!(it.item_type, types::ItemType::Image))
                    .map(|it| it.text.as_str())
                    .collect();
                if inv_alnum >= OCR_LAYER_MIN_ALNUM && !is_garbage_text(&sample) {
                    items = inv_items;
                    has_gid = invisible.has_gid_fonts;
                    coords_rotated = invisible.coords_rotated;
                }
            }
        }
        let threshold = text_utils::fix_letterspaced_items(&mut items);
        if threshold > 0.10 {
            page_thresholds.insert(*page_num, threshold);
        }
        if has_gid {
            gid_pages.insert(*page_num);
        }
        if coords_rotated != extractor::geometry::PageRotation::Upright {
            rotated_pages.insert(*page_num, coords_rotated.into());
        }
        items_by_page.insert(*page_num, items);
    }

    // For each page's regions, filter and assemble text
    let mut results = Vec::with_capacity(page_regions.len());

    for (page_0idx, regions) in page_regions {
        let page_1idx = page_0idx + 1;
        let items = items_by_page.get(&page_1idx);
        let page_h = page_heights.get(&page_1idx).copied().unwrap_or(792.0);
        let _page_has_gid = gid_pages.contains(&page_1idx);
        let adaptive_threshold = page_thresholds.get(&page_1idx).copied().unwrap_or(0.10);
        let coords = rotated_pages
            .get(&page_1idx)
            .copied()
            .unwrap_or(RegionCoordSpace::Standard);
        let display = display_pages.get(&page_1idx).copied();

        let mut page_results = Vec::with_capacity(regions.len());

        // Exclusive item->region assignment: overlapping layout regions used
        // to extract shared items into EVERY region they touched (the
        // 1.5pt inclusion margin makes borders generous), duplicating whole
        // lines in the final markdown on 21% of bench docs — and downstream
        // duplicate-handling sometimes dropped the variant holding a
        // sentence tail, turning duplication into content LOSS. Each item
        // now belongs to the single region with the largest overlap area;
        // items are partitioned, never suppressed, so no content can vanish.
        let all_bounds: Vec<RegionBounds> = regions
            .iter()
            .map(|rect| {
                let [rx1, ry1, rx2, ry2] = sheet_region_rect(*rect, display);
                region_bounds(rx1, ry1, rx2, ry2, page_h, coords)
            })
            .collect();
        // Single pass over items: assign each to the best-overlap region and
        // bucket the clone directly (review: avoid a second O(items x
        // regions) traversal). `had_candidates` marks regions that touched
        // at least one item even if every one was assigned elsewhere.
        let mut region_items: Vec<Vec<TextItem>> = vec![Vec::new(); regions.len()];
        let mut had_candidates: Vec<bool> = vec![false; regions.len()];
        if let Some(items) = items {
            for item in items {
                let mut best: Option<usize> = None;
                let mut best_area = 0.0_f32;
                for (ri, b) in all_bounds.iter().enumerate() {
                    if !region_overlaps_item(item, *b) {
                        continue;
                    }
                    had_candidates[ri] = true;
                    let area = region_item_overlap_area(item, *b);
                    if area > best_area {
                        best_area = area;
                        best = Some(ri);
                    }
                }
                if let Some(ri) = best {
                    region_items[ri].push(item.clone());
                }
            }
        }

        for (region_idx, _rect) in regions.iter().enumerate() {
            let matched: Vec<TextItem> = std::mem::take(&mut region_items[region_idx]);
            let assigned_count = matched.len();
            let has_text_quality_issue = region_items_have_decoding_issue(&matched);
            let text = collect_text_from_matched_items(matched, adaptive_threshold);
            let has_cid_issue = is_cid_garbage(&text);
            let has_encoding_issue = detect_encoding_issues(&text);
            let ocr_reason = if has_text_quality_issue || has_cid_issue || has_encoding_issue {
                Some(suspected_garbled_reason())
            } else {
                None
            };

            // Check per-region text quality instead of blanket page-level
            // GID rejection. A GID font in a logo elsewhere on the page
            // shouldn't force GPU OCR for clean text regions.
            // A region whose ONLY overlapping items were assigned to a
            // better-overlapping neighbor must not fall back to OCR: the
            // pixels it would re-read belong to that neighbor, and OCR
            // would reintroduce the duplication exclusivity removed.
            // Before exclusive assignment these regions were non-empty
            // native (no OCR), so this preserves the old OCR load too.
            // Requires ZERO items assigned HERE: a region whose own
            // assigned items materialize to empty text (whitespace-only,
            // collector-filtered) keeps its OCR fallback.
            let lost_to_neighbor = text.trim().is_empty()
                && ocr_reason.is_none()
                && assigned_count == 0
                && had_candidates[region_idx];
            let needs_ocr = !lost_to_neighbor
                && (ocr_reason.is_some() || text.trim().is_empty() || is_garbage_text(&text));

            page_results.push(RegionText {
                text,
                needs_ocr,
                ocr_reason,
            });
        }

        results.push(PageRegionResult {
            page: *page_0idx,
            regions: page_results,
        });
    }

    Ok(results)
}

/// Extract tables within bounding-box regions from a PDF in memory.
///
/// Similar to [`extract_text_in_regions_mem`] but runs table detection on items
/// within each region and returns markdown pipe-tables instead of flat text.
/// Regions use the same visible-page-box coordinate frame.
///
/// When table structure is detected, `text` contains a markdown pipe-table and
/// `needs_ocr` is `false`. When no table is found (too few items, poor alignment,
/// GID fonts, etc.), `text` is empty and `needs_ocr` is `true` so the caller can
/// fall back to GPU OCR.
pub fn extract_tables_in_regions_mem(
    buffer: &[u8],
    page_regions: &[(u32, Vec<[f32; 4]>)],
) -> Result<Vec<PageRegionResult>, PdfError> {
    extract_tables_in_regions_mem_in_frame(buffer, page_regions, PositionFrame::Sheet)
}

/// [`extract_tables_in_regions_mem`] with the frame of the region rects given
/// explicitly — see [`extract_text_in_regions_mem_in_frame`].
pub fn extract_tables_in_regions_mem_in_frame(
    buffer: &[u8],
    page_regions: &[(u32, Vec<[f32; 4]>)],
    frame: PositionFrame,
) -> Result<Vec<PageRegionResult>, PdfError> {
    extract_tables_in_regions_mem_with_options(
        buffer,
        page_regions,
        PositionOptions::new().frame(frame),
    )
}

/// [`extract_tables_in_regions_mem_in_frame`] with every option given as a
/// [`PositionOptions`] — see [`extract_text_in_regions_mem_with_options`].
pub fn extract_tables_in_regions_mem_with_options(
    buffer: &[u8],
    page_regions: &[(u32, Vec<[f32; 4]>)],
    options: PositionOptions,
) -> Result<Vec<PageRegionResult>, PdfError> {
    let frame = options.frame;
    validate_pdf_bytes(buffer)?;
    let (doc, _page_count) = load_document_from_mem(buffer)?;
    let pages = doc.get_pages();

    let needed_pages: HashSet<u32> = page_regions.iter().map(|(p, _)| p + 1).collect();
    let font_cmaps = FontCMaps::from_doc_pages_fast(&doc, Some(&needed_pages));

    let mut items_by_page: HashMap<u32, Vec<TextItem>> = HashMap::new();
    let mut rects_by_page: HashMap<u32, Vec<PdfRect>> = HashMap::new();
    let mut lines_by_page: HashMap<u32, Vec<PdfLine>> = HashMap::new();
    let mut page_heights: HashMap<u32, f32> = HashMap::new();
    let mut gid_pages: HashSet<u32> = HashSet::new();
    let mut page_thresholds: HashMap<u32, f32> = HashMap::new();
    let mut rotated_pages: HashMap<u32, RegionCoordSpace> = HashMap::new();
    let mut display_pages: HashMap<u32, extractor::DisplayPage> = HashMap::new();
    let mut style_cache = extractor::FontStyleCache::new();

    for (page_num, &page_id) in pages.iter() {
        if !needed_pages.contains(page_num) {
            continue;
        }
        let extractor::PageBoxExtraction {
            mut items,
            rects,
            lines,
            has_gid_fonts: has_gid,
            coords_rotated,
            page_box,
            ..
        } = extractor::extract_page_text_items_in_page_box_with_options(
            &doc,
            page_id,
            *page_num,
            &font_cmaps,
            options.text_extraction(false),
            &mut style_cache,
            &mut extractor::FormWalkBudget::new(),
        )?;
        page_heights.insert(*page_num, page_box.height());
        if frame == PositionFrame::Display {
            display_pages.insert(
                *page_num,
                extractor::DisplayPage::new(&doc, page_id, page_box),
            );
        }
        let threshold = text_utils::fix_letterspaced_items(&mut items);
        if threshold > 0.10 {
            page_thresholds.insert(*page_num, threshold);
        }
        if has_gid {
            gid_pages.insert(*page_num);
        }
        if coords_rotated != extractor::geometry::PageRotation::Upright {
            rotated_pages.insert(*page_num, coords_rotated.into());
        }
        items_by_page.insert(*page_num, items);
        rects_by_page.insert(*page_num, rects);
        lines_by_page.insert(*page_num, lines);
    }

    let mut results = Vec::with_capacity(page_regions.len());

    for (page_0idx, regions) in page_regions {
        let page_1idx = page_0idx + 1;
        let items = items_by_page.get(&page_1idx);
        let page_h = page_heights.get(&page_1idx).copied().unwrap_or(792.0);
        let _page_has_gid = gid_pages.contains(&page_1idx);
        let coords = rotated_pages
            .get(&page_1idx)
            .copied()
            .unwrap_or(RegionCoordSpace::Standard);
        let display = display_pages.get(&page_1idx).copied();

        let mut page_results = Vec::with_capacity(regions.len());

        for rect in regions {
            let [rx1, ry1, rx2, ry2] = sheet_region_rect(*rect, display);

            // Note: we intentionally DO NOT bail on page_has_gid here.
            // The GID flag means some font on the page uses unresolvable
            // glyph IDs, but that font may only appear in a logo or
            // header — not in the table region. Instead we let the
            // per-region text quality checks (is_garbage_text, is_cid_garbage,
            // detect_encoding_issues) reject based on the actual extracted
            // content. This avoids rejecting clean tables just because an
            // unrelated decorative font on the same page is GID-encoded.

            let bounds = region_bounds(rx1, ry1, rx2, ry2, page_h, coords);
            let matched: Vec<TextItem> = match items {
                Some(items) => items
                    .iter()
                    .filter(|item| region_overlaps_item(item, bounds))
                    .cloned()
                    .collect(),
                None => Vec::new(),
            };

            if matched.is_empty() {
                page_results.push(RegionText {
                    text: String::new(),
                    needs_ocr: true,
                    ocr_reason: None,
                });
                continue;
            }

            if region_items_have_decoding_issue(&matched) {
                page_results.push(RegionText {
                    text: String::new(),
                    needs_ocr: true,
                    ocr_reason: Some(suspected_garbled_reason()),
                });
                continue;
            }

            // Compute base_font_size as most common font size in the region
            let base_font_size = {
                let mut freq: HashMap<i32, usize> = HashMap::new();
                for item in &matched {
                    *freq.entry((item.font_size * 10.0) as i32).or_default() += 1;
                }
                freq.into_iter()
                    .max_by_key(|(_, count)| *count)
                    .map(|(size, _)| size as f32 / 10.0)
                    .unwrap_or(12.0)
            };

            // Try rect-backed and line-backed vector-grid detectors first,
            // then fall back to the heuristic text-only detector. Each
            // candidate's markdown is quality-gated by the same
            // needs_ocr checks the heuristic-only path used: if a vector
            // detector produces a partial/garbled table, we ignore it and
            // try the next path rather than degrade the output.
            // needs_ocr fires on any of:
            //   - garbage text (non-alphanumeric heavy)
            //   - CID/Latin-1 mojibake
            //   - encoding issues (U+FFFD, dollar-as-space)
            //   - structural giveaways that the table is partial /
            //     mis-detected (numeric "header", empty header cells,
            //     duplicate header cells).
            // skip_body_font = false / layout_assisted = true because the
            // layout model already identified this region as a table.
            let region_rects: Vec<PdfRect> = rects_by_page
                .get(&page_1idx)
                .map(|rs| {
                    rs.iter()
                        .filter(|r| region_overlaps_rect(r, bounds))
                        .cloned()
                        .collect()
                })
                .unwrap_or_default();
            let region_lines: Vec<PdfLine> = lines_by_page
                .get(&page_1idx)
                .map(|ls| {
                    ls.iter()
                        .filter(|l| region_overlaps_line(l, bounds))
                        .cloned()
                        .collect()
                })
                .unwrap_or_default();

            // Total length of text the page extractor saw inside this
            // region, used by the captured-fragment guard below.
            let region_text_chars: usize = matched.iter().map(|i| i.text.chars().count()).sum();
            let region_area = (rx2 - rx1).max(0.0) * (ry2 - ry1).max(0.0);
            let line_region_has_vertical_rules = has_vertical_rules(&region_lines);

            let evaluate =
                |source: TableCandidateSource, t: &tables::Table| -> Option<TableCandidate> {
                    let md = tables::table_to_markdown(t);
                    let trimmed = md.trim();
                    if trimmed.is_empty() {
                        return None;
                    }
                    if is_garbage_text(&md)
                        || is_cid_garbage(&md)
                        || detect_encoding_issues(&md)
                        || looks_like_partial_table_ex(&md, true)
                    {
                        return None;
                    }
                    // Reject extractions that only captured a small fraction
                    // of the text actually in the region. Two recurring
                    // failure shapes this catches:
                    //   - "header-only": detector found the column-header band
                    //     cleanly but missed every data row below (financial
                    //     statements with multi-line column headers + many
                    //     data rows are the dominant case).
                    //   - "sparse": detector returned a couple of fragmentary
                    //     cells even though the region has many lines of text.
                    // The region floor (200 chars) keeps short legitimate
                    // tables (timestamps, units, axis labels) from being
                    // rejected as partial.
                    if captured_only_a_fragment(&md, region_text_chars) {
                        return None;
                    }
                    // Text-density floor: when a region has lots of pixel
                    // real estate but very few text items, the page
                    // extractor likely hit a font-CMap failure (Identity-H
                    // fonts with missing or broken ToUnicode entries,
                    // Type-3 fonts without unicode metadata, etc.). The
                    // page extractor returns punctuation-only fragments or
                    // single-glyph strings; the rendered image still has
                    // the visible text, so the region should fall back to
                    // OCR. `captured_only_a_fragment` compares captured
                    // chars against text the extractor saw, which is
                    // symmetrically low under font-decode failure — this
                    // guard breaks that symmetry by comparing against
                    // bbox area, which is independent of extraction.
                    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);
                    let issue = if source == TableCandidateSource::Line
                        && line_region_has_vertical_rules
                        && line_table_collapses_text_rows(t, &matched, shape)
                    {
                        Some(TableCandidateIssue::LineRowUndercount)
                    } else if wide_table_sparse_prefix_undercount(&md) {
                        Some(TableCandidateIssue::SparseWideUndercount)
                    } 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)
                    } else {
                        None
                    };
                    Some(TableCandidate {
                        markdown: md,
                        source,
                        shape,
                        issue,
                    })
                };

            let mut candidates: Vec<TableCandidate> = Vec::new();
            if !region_rects.is_empty() {
                let (rect_tables, _) =
                    tables::detect_tables_from_rects(&matched, &region_rects, page_1idx);
                if let Some(candidate) = rect_tables
                    .iter()
                    .find_map(|t| evaluate(TableCandidateSource::Rect, t))
                {
                    candidates.push(candidate);
                }
            }
            if !region_lines.is_empty() {
                let line_tables =
                    tables::detect_tables_from_lines(&matched, &region_lines, page_1idx);
                if let Some(candidate) = line_tables
                    .iter()
                    .find_map(|t| evaluate(TableCandidateSource::Line, t))
                {
                    candidates.push(candidate);
                }
            }
            let detected = tables::detect_tables_with_page_width(
                &matched,
                base_font_size,
                false,
                items.map_or(1.0, |items| tables::content_width(items)),
            );
            if let Some(candidate) = detected
                .iter()
                .find_map(|t| evaluate(TableCandidateSource::Heuristic, t))
            {
                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 {
                    text: candidate.markdown.clone(),
                    needs_ocr: false,
                    ocr_reason: None,
                }),
                None => page_results.push(RegionText {
                    text: String::new(),
                    needs_ocr: true,
                    ocr_reason: None,
                }),
            }
        }

        results.push(PageRegionResult {
            page: *page_0idx,
            regions: page_results,
        });
    }

    Ok(results)
}

/// Region-scoped vector grid detection result for TSR-compatible callers.
#[derive(Debug, Clone)]
pub struct VectorGridDetection {
    /// HTML-like structure tokens consumed by the TSR path.
    pub structure_tokens: Vec<String>,
    /// One crop-pixel bbox per `<td>` token, in document order.
    pub cell_bboxes: Vec<Vec<f32>>,
}

#[derive(Clone, Copy)]
enum VectorGridSource {
    Rects,
    Lines,
}

/// Detect a vector ruled-line / rectangle grid inside one page region.
///
/// The returned shape intentionally matches [`TsrTableInput`]'s structure
/// fields so callers can hand it to `extract_tables_with_structure_*` and let
/// the existing PDF-text cell fill path populate contents.
///
/// `region_pdf_pt_bbox` is `[x1, y1, x2, y2]` in PDF points, top-left origin,
/// relative to the visible page box (see [`extract_text_in_regions_mem`]).
/// The returned cell bboxes are crop-image pixels.
pub fn detect_vector_grid_in_region_mem(
    buffer: &[u8],
    page_idx: u32,
    region_pdf_pt_bbox: [f32; 4],
    render_dpi: f32,
) -> Result<Option<VectorGridDetection>, PdfError> {
    validate_pdf_bytes(buffer)?;
    let (doc, _page_count) = load_document_from_mem(buffer)?;
    let pages = doc.get_pages();

    let page_1idx = page_idx + 1;
    let Some(&page_id) = pages.get(&page_1idx) else {
        return Ok(None);
    };

    let needed_pages = HashSet::from([page_1idx]);
    let font_cmaps = FontCMaps::from_doc_pages_fast(&doc, Some(&needed_pages));
    let extractor::PageBoxExtraction {
        mut items,
        rects,
        lines,
        coords_rotated,
        page_box,
        ..
    } = extractor::extract_page_text_items_in_page_box(
        &doc,
        page_id,
        page_1idx,
        &font_cmaps,
        false,
        &mut extractor::FontStyleCache::new(),
        &mut extractor::FormWalkBudget::new(),
    )?;
    let page_h = page_box.height();
    text_utils::fix_letterspaced_items(&mut items);

    let coords = RegionCoordSpace::from(coords_rotated);
    if coords != RegionCoordSpace::Standard {
        // TODO: add a rotated-page vector-grid fixture before enabling this.
        // The TSR crop contract is top-left page coordinates, while rotated
        // extraction normalizes vector geometry into a synthetic coordinate
        // space. Returning None is safer than emitting misleading bboxes.
        return Ok(None);
    }
    let [rx1, ry1, rx2, ry2] = region_pdf_pt_bbox;
    let bounds = region_bounds(rx1, ry1, rx2, ry2, page_h, coords);

    let items_in_region: Vec<TextItem> = items
        .iter()
        .filter(|item| region_overlaps_item(item, bounds))
        .cloned()
        .collect();
    if items_in_region.is_empty() {
        return Ok(None);
    }

    let rects_in_region: Vec<PdfRect> = rects
        .iter()
        .filter(|rect| region_overlaps_rect(rect, bounds))
        .cloned()
        .collect();
    let lines_in_region: Vec<PdfLine> = lines
        .iter()
        .filter(|line| region_overlaps_line(line, bounds))
        .cloned()
        .collect();

    // Match the existing geometry pipeline priority: rect-backed grids first,
    // then line-backed grids. The detector output is only the validity gate;
    // bboxes below are rebuilt from the filtered vector geometry because Table
    // stores centers for some rect paths and row starts for line paths.
    let (rect_tables, _) =
        tables::detect_tables_from_rects(&items_in_region, &rects_in_region, page_1idx);
    for table in rect_tables {
        if let Some(result) = vector_grid_result_from_table(
            &table,
            VectorGridSource::Rects,
            &rects_in_region,
            &lines_in_region,
            region_pdf_pt_bbox,
            render_dpi,
            page_h,
            coords,
        ) {
            return Ok(Some(result));
        }
    }

    let line_tables =
        tables::detect_vector_grid_tables_from_lines(&items_in_region, &lines_in_region, page_1idx);
    for table in line_tables {
        if let Some(result) = vector_grid_result_from_table(
            &table,
            VectorGridSource::Lines,
            &rects_in_region,
            &lines_in_region,
            region_pdf_pt_bbox,
            render_dpi,
            page_h,
            coords,
        ) {
            return Ok(Some(result));
        }
    }

    Ok(None)
}

#[allow(clippy::too_many_arguments)]
fn vector_grid_result_from_table(
    table: &tables::Table,
    source: VectorGridSource,
    rects: &[PdfRect],
    lines: &[PdfLine],
    crop_pdf_pt_bbox: [f32; 4],
    render_dpi: f32,
    page_height: f32,
    coord_space: RegionCoordSpace,
) -> Option<VectorGridDetection> {
    let num_rows = table.cells.len();
    let num_cols = table.cells.first().map_or(0, Vec::len);
    if num_rows == 0 || num_cols == 0 || table.cells.iter().any(|row| row.len() != num_cols) {
        return None;
    }

    let (x_edges, y_edges) = match source {
        VectorGridSource::Rects => rect_grid_edges(rects, num_cols, num_rows)
            .or_else(|| inferred_grid_edges(table, rects, lines, num_cols, num_rows))?,
        VectorGridSource::Lines => line_grid_edges(table, lines, num_cols, num_rows)
            .or_else(|| inferred_grid_edges(table, rects, lines, num_cols, num_rows))?,
    };

    if x_edges.len() != num_cols + 1 || y_edges.len() != num_rows + 1 {
        return None;
    }

    let mut structure_tokens = Vec::with_capacity(num_rows * (num_cols + 2) + 2);
    let mut cell_bboxes = Vec::with_capacity(num_rows * num_cols);
    structure_tokens.push("<table>".to_string());

    // TODO: refactor vector detectors to return normalized `(x_edges, y_edges)`.
    // Today `Table.columns` / `Table.rows` have detector-specific semantics,
    // so this export reconstructs edges from the validated table plus geometry.
    // Keep v1 structural output uniform: the downstream TSR text-fill path
    // does not require header semantics, and reliable header detection can be
    // layered later without changing the geometry contract.
    for r in 0..num_rows {
        structure_tokens.push("<tr>".to_string());
        for c in 0..num_cols {
            structure_tokens.push("<td></td>".to_string());
            let bbox_px = extracted_cell_to_crop_px(
                [x_edges[c], y_edges[r + 1], x_edges[c + 1], y_edges[r]],
                crop_pdf_pt_bbox,
                render_dpi,
                page_height,
                coord_space,
            )?;
            if !crop_px_bbox_is_plausible(bbox_px, crop_pdf_pt_bbox, render_dpi) {
                return None;
            }
            cell_bboxes.push(bbox_px.to_vec());
        }
        structure_tokens.push("</tr>".to_string());
    }

    structure_tokens.push("</table>".to_string());
    Some(VectorGridDetection {
        structure_tokens,
        cell_bboxes,
    })
}

fn crop_px_bbox_is_plausible(
    bbox_px: [f32; 4],
    crop_pdf_pt_bbox: [f32; 4],
    render_dpi: f32,
) -> bool {
    let ppi = if render_dpi > 0.0 {
        render_dpi / 72.0
    } else {
        1.0
    };
    let crop_w = (crop_pdf_pt_bbox[2] - crop_pdf_pt_bbox[0]).abs() * ppi;
    let crop_h = (crop_pdf_pt_bbox[3] - crop_pdf_pt_bbox[1]).abs() * ppi;
    let slack = 1.0;
    bbox_px[0] >= -slack
        && bbox_px[1] >= -slack
        && bbox_px[2] <= crop_w + slack
        && bbox_px[3] <= crop_h + slack
}

#[cfg(test)]
mod vector_grid_tests {
    use super::crop_px_bbox_is_plausible;

    /// Regression for `forecast_table_chart.pdf` (doc 128 from the
    /// opendataloader-bench corpus). The table has six visual columns, but
    /// text X-clustering in the cell-rect fallback previously split wide
    /// columns into ten spurious columns.
    #[test]
    fn forecast_table_chart_six_cols() {
        use crate::extractor::content_stream::extract_page_text_items;
        use crate::tables::detect_tables_from_rects;
        use crate::tounicode::FontCMaps;
        use lopdf::Document;
        use std::collections::HashSet;
        use std::fs;

        let path = "tests/fixtures/forecast_table_chart.pdf";
        let buf = fs::read(path).unwrap();
        let doc = Document::load_mem(&buf).unwrap();
        let pages = doc.get_pages();
        let &page_id = pages.get(&1).unwrap();
        let needed: HashSet<u32> = HashSet::from([1]);
        let cmaps = FontCMaps::from_doc_pages_fast(&doc, Some(&needed));
        let ((items, rects, _lines), _has_gid, _rotated, _skipped_invisible) =
            extract_page_text_items(
                &doc,
                page_id,
                1,
                &cmaps,
                false,
                &mut crate::extractor::FontStyleCache::new(),
                &mut crate::extractor::FormWalkBudget::new(),
            )
            .unwrap();

        let (rect_tables, _) = detect_tables_from_rects(&items, &rects, 1);
        assert_eq!(rect_tables.len(), 1, "expected one rect-detected table");
        let t = &rect_tables[0];
        assert_eq!(
            t.columns.len(),
            6,
            "doc 128 has a 6-column table; got {} edges: {:?}",
            t.columns.len(),
            t.columns
        );
        assert!(
            t.rows.len() >= 14 && t.rows.len() <= 17,
            "row count drift: {}",
            t.rows.len()
        );
    }

    /// Helper: load a fixture PDF and run the rect-based table detector.
    fn detect_rect_tables_in_fixture_page(path: &str, page_num: u32) -> Vec<crate::tables::Table> {
        use crate::extractor::content_stream::extract_page_text_items;
        use crate::tables::detect_tables_from_rects;
        use crate::tounicode::FontCMaps;
        use lopdf::Document;
        use std::collections::HashSet;
        use std::fs;

        let buf = fs::read(path).unwrap();
        let doc = Document::load_mem(&buf).unwrap();
        let pages = doc.get_pages();
        let &page_id = pages.get(&page_num).unwrap();
        let needed: HashSet<u32> = HashSet::from([page_num]);
        let cmaps = FontCMaps::from_doc_pages_fast(&doc, Some(&needed));
        let ((items, rects, _lines), _has_gid, _rotated, _skipped_invisible) =
            extract_page_text_items(
                &doc,
                page_id,
                page_num,
                &cmaps,
                false,
                &mut crate::extractor::FontStyleCache::new(),
                &mut crate::extractor::FormWalkBudget::new(),
            )
            .unwrap();

        let (rect_tables, _) = detect_tables_from_rects(&items, &rects, page_num);
        rect_tables
    }

    fn detect_rect_tables_in_fixture(path: &str) -> Vec<crate::tables::Table> {
        detect_rect_tables_in_fixture_page(path, 1)
    }

    /// Regression for the prose-in-a-frame failure mode introduced by the
    /// shaded-header detection lift (PR #76). The accessory_building permit
    /// form has a paragraph of legal text laid out in a 2-column justified
    /// block; the new fill-priority + dedup changes start producing rects
    /// for it, and the rect detector then admits a 10×2 fake table where
    /// every cell holds a sentence fragment ("I agree to comply...", "I",
    /// "It is the property owner's responsibility..."). This test asserts
    /// the detector REJECTS that fake table — only the real 5×3 form data
    /// table (TYPE / SIZE / SETBACKS) should survive. See pdf-evals PR #30
    /// for the original score regression that surfaced this.
    #[test]
    fn accessory_building_rejects_prose_in_frame() {
        let tables = detect_rect_tables_in_fixture(
            "tests/fixtures/accessory_building_permit_prose_frame.pdf",
        );
        // Real form data table (TYPE / SIZE / SETBACKS) must still be detected.
        let data_table = tables.iter().find(|t| t.columns.len() == 3);
        assert!(
            data_table.is_some(),
            "expected to keep the 5×3 TYPE/SIZE/SETBACKS data table; got {:?}",
            tables
                .iter()
                .map(|t| (t.rows.len(), t.columns.len()))
                .collect::<Vec<_>>()
        );
        // Prose paragraph laid out in 2 cols must NOT be detected as a table.
        // If the rejection regresses, the 10×2 fake table reappears and
        // produces fragmented markdown like "I agree to comply..." | "I"
        // that fragments mid-sentence.
        let prose_table = tables.iter().find(|t| t.columns.len() == 2);
        assert!(
            prose_table.is_none(),
            "expected the 10×2 prose-in-frame block to be rejected; got rows×cols = {:?}",
            prose_table.map(|t| (t.rows.len(), t.columns.len()))
        );
    }

    #[test]
    fn td9264_insurance_prose_not_rect_table() {
        let tables = detect_rect_tables_in_fixture_page("tests/fixtures/td9264.pdf", 4);
        assert!(
            tables.is_empty(),
            "expected no rect-detected tables for the insurance-company prose; got {:?}",
            tables
                .iter()
                .map(|t| (t.rows.len(), t.columns.len(), t.cells.clone()))
                .collect::<Vec<_>>()
        );
    }

    /// Wireless table regression: decorative/text-region rects may provide row
    /// bands, but without a real rect-derived column scaffold they must not be
    /// accepted as a vector grid.
    #[test]
    fn wireless_two_col_rejects_rect_grid() {
        let tables = detect_rect_tables_in_fixture("tests/fixtures/wireless_two_col_no_rects.pdf");
        assert!(
            tables.is_empty(),
            "expected no rect-detected tables for wireless content; got {:?}",
            tables
                .iter()
                .map(|t| (t.rows.len(), t.columns.len()))
                .collect::<Vec<_>>()
        );
    }

    #[test]
    fn wireless_two_col_region_rejects_vector_grid() {
        let buf = std::fs::read("tests/fixtures/wireless_two_col_no_rects.pdf").unwrap();
        let crops = [
            [49.32_f32, 52.92, 558.72, 214.2],
            [49.32_f32, 288.72, 556.56, 378.0],
            [51.48_f32, 478.44, 558.36, 567.36],
        ];
        for crop in crops {
            let detected = crate::detect_vector_grid_in_region_mem(&buf, 0, crop, 200.0).unwrap();
            assert!(
                detected.is_none(),
                "expected no vector grid for wireless crop {crop:?}; got {} cells",
                detected.map(|grid| grid.cell_bboxes.len()).unwrap_or(0)
            );
        }
    }

    /// Wireless dense table regression: text-position columns alone are not
    /// enough evidence for a rect-derived grid.
    #[test]
    fn wireless_dense_rejects_rect_grid() {
        let tables = detect_rect_tables_in_fixture("tests/fixtures/wireless_dense_no_rects.pdf");
        assert!(
            tables.is_empty(),
            "expected no rect-detected tables for wireless content; got {:?}",
            tables
                .iter()
                .map(|t| (t.rows.len(), t.columns.len()))
                .collect::<Vec<_>>()
        );
    }

    #[test]
    fn wireless_dense_region_rejects_vector_grid() {
        let buf = std::fs::read("tests/fixtures/wireless_dense_no_rects.pdf").unwrap();
        let crops = [
            [72.36_f32, 177.48, 243.72, 333.36],
            [72.0_f32, 390.24, 286.92, 417.6],
        ];
        for crop in crops {
            let detected = crate::detect_vector_grid_in_region_mem(&buf, 0, crop, 200.0).unwrap();
            assert!(
                detected.is_none(),
                "expected no vector grid for wireless crop {crop:?}; got {} cells",
                detected.map(|grid| grid.cell_bboxes.len()).unwrap_or(0)
            );
        }
    }

    #[test]
    fn multiline_indent_cell_rect_grid_fixture_detects_table() {
        let tables = detect_rect_tables_in_fixture_page(
            "tests/fixtures/multiline_indent_cell_rect_grid.pdf",
            30,
        );
        let table = tables
            .iter()
            .max_by_key(|t| t.rows.len() * t.columns.len())
            .expect("expected a rect-detected table");
        assert_eq!(
            table.columns.len(),
            5,
            "expected the Controls Version / Control / IG table shape; got {:?}",
            tables
                .iter()
                .map(|t| (t.rows.len(), t.columns.len()))
                .collect::<Vec<_>>()
        );
        assert!(
            table.rows.len() >= 3,
            "expected at least header plus data rows; got {}",
            table.rows.len()
        );
    }

    #[test]
    fn multiline_indent_cell_rect_grid_region_detects_vector_grid() {
        let buf = std::fs::read("tests/fixtures/multiline_indent_cell_rect_grid.pdf").unwrap();
        let detected =
            crate::detect_vector_grid_in_region_mem(&buf, 29, [0.0, 0.0, 612.0, 792.0], 200.0)
                .unwrap()
                .expect("expected vector grid for multiline indented description table");
        let rows = detected
            .structure_tokens
            .iter()
            .filter(|token| token.as_str() == "<tr>")
            .count();
        assert_eq!(detected.cell_bboxes.len() % rows, 0);
        assert_eq!(detected.cell_bboxes.len() / rows, 5);
        assert!(rows >= 3);
        assert!(!detected.cell_bboxes.is_empty());
    }

    /// Regression for `greencomp_competence.pdf` — a 2-column "Area / Competence"
    /// glossary with a green-shaded header row and plain (line-drawn) body cells.
    /// Mirrors the production failure cohort #1 (Contractions glossary) and #6
    /// (BIO 350 course header): a few colored header rects sit in a horizontal
    /// strip while body rows are drawn with `m`/`l` operators, so the rect
    /// cluster has only 2 Y-edges and `try_build_grid` rejects.
    ///
    /// The page repeats a full-page background many times. Those fills must be
    /// removed from clustering before chart/table evidence is evaluated, or
    /// they swamp the real cell rectangles and make this table look chart-like.
    #[test]
    fn greencomp_competence_two_cols() {
        let tables = detect_rect_tables_in_fixture("tests/fixtures/greencomp_competence.pdf");
        assert!(
            !tables.is_empty(),
            "expected at least one rect-detected table for shaded-header + plain-body shape"
        );
        let t = tables
            .iter()
            .max_by_key(|t| t.rows.len() * t.columns.len())
            .unwrap();
        assert_eq!(
            t.columns.len(),
            2,
            "GreenComp competence is a 2-column table; got {}: {:?}",
            t.columns.len(),
            t.columns
        );
        assert!(
            t.rows.len() >= 6,
            "expected at least 6 rows of competences; got {}",
            t.rows.len()
        );
    }

    /// Regression for `upstage_key_functions.pdf` — a 4-column "Service Stage /
    /// Function Name / Explanation / Expected Benefit" table with a blue-shaded
    /// header band plus alternating row backgrounds. Mirrors production crops
    /// #2 (Parameter / Value with alternating blue rows) and #7 (Spanish XML
    /// schema with shaded header). Currently `pdf2md` returns zero markdown
    /// table rows.
    #[test]
    fn upstage_key_functions_four_cols() {
        let tables = detect_rect_tables_in_fixture("tests/fixtures/upstage_key_functions.pdf");
        assert!(
            !tables.is_empty(),
            "expected at least one rect-detected table for shaded-header + alt-row shape"
        );
        let t = tables
            .iter()
            .max_by_key(|t| t.rows.len() * t.columns.len())
            .unwrap();
        assert_eq!(
            t.columns.len(),
            4,
            "Service Flow is a 4-column table; got {}: {:?}",
            t.columns.len(),
            t.columns
        );
        assert!(
            t.rows.len() >= 8,
            "expected at least 8 visible body rows; got {}",
            t.rows.len()
        );
    }

    /// Regression for `wired_header_data_misalign.pdf` — a single page from a
    /// parts catalog with a 4-column wire-bordered table (`Item | EAN | Nombre
    /// | Cant`). Column headers are centered/right-aligned inside their cells
    /// while data is left-aligned, so cluster_x_positions merges or drops
    /// columns and the cell-rect fallback used to assign text to the wrong
    /// columns (lost a column, fragmented neighbor cells). The fix prefers
    /// rect-border-derived column edges when they're well-distributed across
    /// the actual text items. This test asserts the detector keeps all 4
    /// columns and every column ends up populated.
    #[test]
    fn wired_header_data_misalign_keeps_all_columns() {
        let tables = detect_rect_tables_in_fixture("tests/fixtures/wired_header_data_misalign.pdf");
        let table = tables
            .iter()
            .find(|t| t.columns.len() == 4 && t.rows.len() >= 5)
            .unwrap_or_else(|| {
                panic!(
                    "expected a 4-column ≥5-row table; got {:?}",
                    tables
                        .iter()
                        .map(|t| (t.rows.len(), t.columns.len()))
                        .collect::<Vec<_>>()
                )
            });
        for c in 0..4 {
            let populated_rows = table
                .cells
                .iter()
                .filter(|row| !row[c].trim().is_empty())
                .count();
            assert!(
                populated_rows >= 2,
                "column {} only populated in {} rows; cells: {:?}",
                c,
                populated_rows,
                table.cells
            );
        }
    }

    #[test]
    fn test_crop_px_bbox_is_plausible_bounds() {
        let crop = [10.0, 20.0, 110.0, 220.0];

        assert!(crop_px_bbox_is_plausible(
            [0.0, 0.0, 100.0, 200.0],
            crop,
            72.0
        ));
        assert!(crop_px_bbox_is_plausible(
            [0.0, 0.0, 200.0, 400.0],
            crop,
            144.0
        ));
        assert!(!crop_px_bbox_is_plausible(
            [-2.0, 0.0, 50.0, 100.0],
            crop,
            72.0
        ));
        assert!(!crop_px_bbox_is_plausible(
            [0.0, -2.0, 50.0, 100.0],
            crop,
            72.0
        ));
        assert!(!crop_px_bbox_is_plausible(
            [0.0, 0.0, 102.0, 200.0],
            crop,
            72.0
        ));
        assert!(!crop_px_bbox_is_plausible(
            [0.0, 0.0, 100.0, 202.0],
            crop,
            72.0
        ));

        // Boundary values at the existing 1px slack should remain valid.
        assert!(crop_px_bbox_is_plausible(
            [-1.0, -1.0, 101.0, 201.0],
            crop,
            72.0
        ));

        // Non-positive DPI falls back to 1.0 ppi, so crop points equal pixels.
        assert!(crop_px_bbox_is_plausible(
            [0.0, 0.0, 100.0, 200.0],
            crop,
            0.0
        ));
        assert!(crop_px_bbox_is_plausible(
            [0.0, 0.0, 100.0, 200.0],
            crop,
            -144.0
        ));
        assert!(!crop_px_bbox_is_plausible(
            [0.0, 0.0, 102.0, 200.0],
            crop,
            -144.0
        ));
    }
}

fn line_grid_edges(
    table: &tables::Table,
    lines: &[PdfLine],
    num_cols: usize,
    num_rows: usize,
) -> Option<(Vec<f32>, Vec<f32>)> {
    if lines.is_empty() || table.columns.len() != num_cols + 1 || table.rows.len() != num_rows {
        return None;
    }

    let angle_tolerance = 2.0_f32.to_radians().tan();
    let mut ys = Vec::new();

    for line in lines {
        let dx = (line.x2 - line.x1).abs();
        let dy = (line.y2 - line.y1).abs();
        let length = (dx * dx + dy * dy).sqrt();
        if length < 20.0 {
            continue;
        }
        if dx > 0.01 && dy / dx <= angle_tolerance {
            ys.push((line.y1 + line.y2) * 0.5);
        }
    }

    let mut x_edges = table.columns.clone();
    x_edges.sort_by(|a, b| a.total_cmp(b));

    let snapped_y = snap_vector_edges(ys, true);
    let mut y_edges = Vec::with_capacity(num_rows + 1);
    for &row_top in &table.rows {
        let matched = snapped_y
            .iter()
            .copied()
            .find(|y| (*y - row_top).abs() <= 3.0)
            .unwrap_or(row_top);
        y_edges.push(matched);
    }

    let last_top = *y_edges.last()?;
    let bottom = snapped_y
        .iter()
        .copied()
        .filter(|y| *y < last_top - 3.0)
        .max_by(|a, b| a.total_cmp(b))?;
    y_edges.push(bottom);

    if x_edges.len() == num_cols + 1 && y_edges.len() == num_rows + 1 {
        Some((x_edges, y_edges))
    } else {
        None
    }
}

fn rect_grid_edges(
    rects: &[PdfRect],
    num_cols: usize,
    num_rows: usize,
) -> Option<(Vec<f32>, Vec<f32>)> {
    if rects.is_empty() {
        return None;
    }

    let mut xs = Vec::new();
    let mut ys = Vec::new();
    for rect in rects {
        let (x1, y1, x2, y2) = normalized_rect_edges(rect);
        if (x2 - x1) < 5.0 || (y2 - y1) < 5.0 {
            continue;
        }
        xs.push(x1);
        xs.push(x2);
        ys.push(y1);
        ys.push(y2);
    }

    let x_edges = snap_vector_edges(xs, false);
    let y_edges = snap_vector_edges(ys, true);
    if x_edges.len() == num_cols + 1 && y_edges.len() == num_rows + 1 {
        Some((x_edges, y_edges))
    } else {
        None
    }
}

fn inferred_grid_edges(
    table: &tables::Table,
    rects: &[PdfRect],
    lines: &[PdfLine],
    num_cols: usize,
    num_rows: usize,
) -> Option<(Vec<f32>, Vec<f32>)> {
    let bounds = vector_geometry_bounds(rects, lines);
    let x_edges = if table.columns.len() == num_cols + 1 {
        let mut edges = table.columns.clone();
        edges.sort_by(|a, b| a.total_cmp(b));
        Some(edges)
    } else {
        infer_ascending_edges(&table.columns, num_cols, bounds.map(|b| (b.x_min, b.x_max)))
    }?;

    let y_edges = if table.rows.len() == num_rows + 1 {
        let mut edges = table.rows.clone();
        edges.sort_by(|a, b| b.total_cmp(a));
        Some(edges)
    } else {
        infer_descending_edges(&table.rows, num_rows, bounds.map(|b| (b.y_min, b.y_max)))
    }?;

    Some((x_edges, y_edges))
}

fn infer_ascending_edges(
    positions: &[f32],
    expected_centers: usize,
    bounds: Option<(f32, f32)>,
) -> Option<Vec<f32>> {
    if positions.len() != expected_centers || positions.is_empty() {
        return None;
    }
    let mut centers = positions.to_vec();
    centers.sort_by(|a, b| a.total_cmp(b));
    if centers.len() == 1 {
        return None;
    }

    let mut edges = Vec::with_capacity(centers.len() + 1);
    let first_gap = centers[1] - centers[0];
    let last_gap = centers[centers.len() - 1] - centers[centers.len() - 2];
    let left = bounds
        .map(|(min, _)| min)
        .filter(|min| min.is_finite() && *min < centers[0])
        .unwrap_or(centers[0] - first_gap * 0.5);
    let right = bounds
        .map(|(_, max)| max)
        .filter(|max| max.is_finite() && *max > *centers.last().unwrap())
        .unwrap_or(*centers.last().unwrap() + last_gap * 0.5);

    edges.push(left);
    for pair in centers.windows(2) {
        edges.push((pair[0] + pair[1]) * 0.5);
    }
    edges.push(right);

    strictly_ordered(&edges, false).then_some(edges)
}

fn infer_descending_edges(
    positions: &[f32],
    expected_centers: usize,
    bounds: Option<(f32, f32)>,
) -> Option<Vec<f32>> {
    if positions.len() != expected_centers || positions.is_empty() {
        return None;
    }
    let mut centers = positions.to_vec();
    centers.sort_by(|a, b| b.total_cmp(a));
    if centers.len() == 1 {
        return None;
    }

    let mut edges = Vec::with_capacity(centers.len() + 1);
    let first_gap = centers[0] - centers[1];
    let last_gap = centers[centers.len() - 2] - centers[centers.len() - 1];
    let top = bounds
        .map(|(_, max)| max)
        .filter(|max| max.is_finite() && *max > centers[0])
        .unwrap_or(centers[0] + first_gap * 0.5);
    let bottom = bounds
        .map(|(min, _)| min)
        .filter(|min| min.is_finite() && *min < *centers.last().unwrap())
        .unwrap_or(*centers.last().unwrap() - last_gap * 0.5);

    edges.push(top);
    for pair in centers.windows(2) {
        edges.push((pair[0] + pair[1]) * 0.5);
    }
    edges.push(bottom);

    strictly_ordered(&edges, true).then_some(edges)
}

fn snap_vector_edges(mut values: Vec<f32>, descending: bool) -> Vec<f32> {
    values.retain(|v| v.is_finite());
    values.sort_by(|a, b| a.total_cmp(b));

    let mut snapped: Vec<f32> = Vec::new();
    let mut cluster: Vec<f32> = Vec::new();
    for value in values {
        if cluster
            .last()
            .is_some_and(|last| (value - *last).abs() <= 3.0)
        {
            cluster.push(value);
        } else {
            if !cluster.is_empty() {
                snapped.push(cluster.iter().sum::<f32>() / cluster.len() as f32);
            }
            cluster = vec![value];
        }
    }
    if !cluster.is_empty() {
        snapped.push(cluster.iter().sum::<f32>() / cluster.len() as f32);
    }
    if descending {
        snapped.sort_by(|a, b| b.total_cmp(a));
    }
    snapped
}

fn strictly_ordered(values: &[f32], descending: bool) -> bool {
    values.windows(2).all(|pair| {
        pair[0].is_finite()
            && pair[1].is_finite()
            && if descending {
                pair[0] > pair[1]
            } else {
                pair[0] < pair[1]
            }
    })
}

fn normalized_rect_edges(rect: &PdfRect) -> (f32, f32, f32, f32) {
    let x2 = rect.x + rect.width;
    let y2 = rect.y + rect.height;
    (
        rect.x.min(x2),
        rect.y.min(y2),
        rect.x.max(x2),
        rect.y.max(y2),
    )
}

fn vector_geometry_bounds(rects: &[PdfRect], lines: &[PdfLine]) -> Option<RegionBounds> {
    let mut bounds: Option<RegionBounds> = None;
    let mut include = |x1: f32, y1: f32, x2: f32, y2: f32| {
        let next = RegionBounds {
            x_min: x1.min(x2),
            y_min: y1.min(y2),
            x_max: x1.max(x2),
            y_max: y1.max(y2),
        };
        bounds = Some(if let Some(prev) = bounds {
            RegionBounds {
                x_min: prev.x_min.min(next.x_min),
                y_min: prev.y_min.min(next.y_min),
                x_max: prev.x_max.max(next.x_max),
                y_max: prev.y_max.max(next.y_max),
            }
        } else {
            next
        });
    };

    for rect in rects {
        let (x1, y1, x2, y2) = normalized_rect_edges(rect);
        include(x1, y1, x2, y2);
    }
    for line in lines {
        include(line.x1, line.y1, line.x2, line.y2);
    }

    bounds
}

fn extracted_cell_to_crop_px(
    bbox: [f32; 4],
    crop_pdf_pt_bbox: [f32; 4],
    render_dpi: f32,
    page_height: f32,
    coord_space: RegionCoordSpace,
) -> Option<[f32; 4]> {
    let [x1, y1, x2, y2] = extracted_bbox_to_page_top_left(bbox, page_height, coord_space);
    if !(x1.is_finite() && y1.is_finite() && x2.is_finite() && y2.is_finite()) {
        return None;
    }
    if x1 >= x2 || y1 >= y2 {
        return None;
    }

    let ppi = if render_dpi > 0.0 {
        render_dpi / 72.0
    } else {
        1.0
    };
    let [crop_x1, crop_y1, _, _] = crop_pdf_pt_bbox;
    Some([
        (x1 - crop_x1) * ppi,
        (y1 - crop_y1) * ppi,
        (x2 - crop_x1) * ppi,
        (y2 - crop_y1) * ppi,
    ])
}

fn extracted_bbox_to_page_top_left(
    bbox: [f32; 4],
    page_height: f32,
    coord_space: RegionCoordSpace,
) -> [f32; 4] {
    let [x1, y1, x2, y2] = bbox;
    let x_min = x1.min(x2);
    let x_max = x1.max(x2);
    let y_min = y1.min(y2);
    let y_max = y1.max(y2);

    match coord_space {
        RegionCoordSpace::Standard => [x_min, page_height - y_max, x_max, page_height - y_min],
        RegionCoordSpace::Rotated90Ccw => {
            [-y_max, page_height - x_max, -y_min, page_height - x_min]
        }
        RegionCoordSpace::Rotated90Cw => [y_min, page_height + x_min, y_max, page_height + x_max],
    }
}

// =========================================================================
// Region-based table extraction with external structure recovery (TSR)
// =========================================================================

/// Input for [`extract_tables_with_structure_mem`]: one cropped table region
/// plus the raw structure-recovery output for it.
///
/// The structure tokens and bboxes are typically produced by an external
/// table-structure recognition model (e.g. SLANet on PaddleOCR) running on
/// a rendered crop of the page. pdf-inspector uses the structure to lay out
/// the cells and pulls the cell text from the native PDF — no OCR involved.
#[derive(Debug, Clone)]
pub struct TsrTableInput {
    /// 0-indexed page number where the crop was taken from.
    pub page: u32,
    /// Crop bbox on the page, `[x1, y1, x2, y2]` in PDF points with
    /// **top-left origin**, relative to the visible page box — the frame of
    /// a rendered page image (see [`extract_text_in_regions_mem`]).
    pub crop_pdf_pt_bbox: [f32; 4],
    /// DPI the crop image was rendered at (e.g. `200.0`). Used to convert
    /// cell bboxes from image-pixels back to PDF points.
    pub render_dpi: f32,
    /// Raw structure tokens emitted by the TSR model, in document order.
    /// See [`tables::structured::parse_structure`] for the accepted grammar.
    pub structure_tokens: Vec<String>,
    /// One bbox per cell (in document order, parallel to the cell open-tags
    /// in `structure_tokens`). May be 4-element `[x1,y1,x2,y2]` or
    /// 8-element 4-corner polygon, in **crop image-pixel space**.
    pub cell_bboxes: Vec<Vec<f32>>,
}

/// Extract structured cells using externally-supplied structure recovery.
///
/// For each input, this:
/// 1. Pairs each cell open-tag in `structure_tokens` with the next bbox in
///    `cell_bboxes` (document order), tracking row/col with rowspan/colspan
///    awareness.
/// 2. Converts each cell bbox from crop image-pixels into page PDF-points.
/// 3. Pulls the cell's text by overlap-testing PDF text items inside that
///    bbox — same primitives used by [`extract_text_in_regions_mem`].
///
/// Returns one `Vec<StructuredCell>` per input, in input order. Each cell
/// carries its (row, col, rowspan, colspan, is_header) metadata, the
/// extracted text, and its page-PDF-pt bbox so callers can do their own
/// rendering, debug overlays, or per-cell post-processing.
///
/// Inputs whose page is out of range or whose tokens parse to zero cells
/// produce an empty `Vec`.
///
/// See [`extract_tables_with_structure_mem`] if you just want the rendered
/// markdown.
pub fn extract_tables_with_structure_cells_mem(
    buffer: &[u8],
    inputs: &[TsrTableInput],
) -> Result<Vec<Vec<tables::StructuredCell>>, PdfError> {
    use tables::structured::{
        cell_px_to_page_pt, normalize_cell_bands, parse_structure, polygon_to_aabb, StructuredCell,
    };

    validate_pdf_bytes(buffer)?;
    let (doc, _page_count) = load_document_from_mem(buffer)?;
    let pages = doc.get_pages();

    let needed_pages: HashSet<u32> = inputs.iter().map(|t| t.page + 1).collect();
    let font_cmaps = FontCMaps::from_doc_pages_fast(&doc, Some(&needed_pages));

    let mut items_by_page: HashMap<u32, Vec<TextItem>> = HashMap::new();
    let mut page_heights: HashMap<u32, f32> = HashMap::new();
    let mut page_thresholds: HashMap<u32, f32> = HashMap::new();
    let mut rotated_pages: HashMap<u32, RegionCoordSpace> = HashMap::new();
    let mut style_cache = extractor::FontStyleCache::new();

    for (page_num, &page_id) in pages.iter() {
        if !needed_pages.contains(page_num) {
            continue;
        }
        let extractor::PageBoxExtraction {
            mut items,
            coords_rotated,
            page_box,
            ..
        } = extractor::extract_page_text_items_in_page_box(
            &doc,
            page_id,
            *page_num,
            &font_cmaps,
            false,
            &mut style_cache,
            &mut extractor::FormWalkBudget::new(),
        )?;
        page_heights.insert(*page_num, page_box.height());
        let threshold = text_utils::fix_letterspaced_items(&mut items);
        if threshold > 0.10 {
            page_thresholds.insert(*page_num, threshold);
        }
        if coords_rotated != extractor::geometry::PageRotation::Upright {
            rotated_pages.insert(*page_num, coords_rotated.into());
        }
        items_by_page.insert(*page_num, items);
    }

    let mut results: Vec<Vec<StructuredCell>> = Vec::with_capacity(inputs.len());

    for input in inputs {
        let page_1idx = input.page + 1;
        let Some(items) = items_by_page.get(&page_1idx) else {
            // Out-of-range page or page with no extractable text — emit empty.
            results.push(Vec::new());
            continue;
        };
        let page_h = page_heights.get(&page_1idx).copied().unwrap_or(792.0);
        let adaptive_threshold = page_thresholds.get(&page_1idx).copied().unwrap_or(0.10);
        let coords = rotated_pages
            .get(&page_1idx)
            .copied()
            .unwrap_or(RegionCoordSpace::Standard);

        let crop_origin = [input.crop_pdf_pt_bbox[0], input.crop_pdf_pt_bbox[1]];

        let slots = parse_structure(&input.structure_tokens);
        if slots.is_empty() {
            results.push(Vec::new());
            continue;
        }

        let mut cells: Vec<StructuredCell> = Vec::with_capacity(slots.len());
        for slot in &slots {
            let page_pt_bbox;

            if let Some(coords_arr) = input.cell_bboxes.get(slot.bbox_idx) {
                if let Some(aabb_px) = polygon_to_aabb(coords_arr) {
                    page_pt_bbox = cell_px_to_page_pt(aabb_px, input.render_dpi, crop_origin);
                } else {
                    page_pt_bbox = [0.0, 0.0, 0.0, 0.0];
                }
            } else {
                page_pt_bbox = [0.0, 0.0, 0.0, 0.0];
            }

            cells.push(StructuredCell {
                row: slot.row,
                col: slot.col,
                rowspan: slot.rowspan,
                colspan: slot.colspan,
                is_header: slot.is_header,
                text: String::new(),
                page_pt_bbox,
            });
        }

        normalize_cell_bands(&mut cells);

        // Stage 1: exclusive per-token assignment. Each PDF text item is
        // first split into whitespace-separated tokens with estimated x
        // positions (see `split_item_into_token_subitems`). For each token
        // we find the cell(s) whose (band-clamped) bbox satisfies the
        // strict membership rule (`tsr_region_contains_item`: center
        // inside OR >=60% overlap on both axes). If multiple cells
        // qualify, the closest-center wins. Tokens that don't land in any
        // cell are eligible for stage-2 orphan recovery.
        //
        // Per-token (rather than per-item) routing is what prevents the
        // dense-grid collapse bug: a row rendered as one wide Tj
        // ("Marshall Islands 0.9 0.9 0.9") produces one TextItem whose
        // CENTER lies in only one cell, so per-item routing parks the
        // whole row in that single cell and leaves the rest of the row
        // empty. Per-token routing distributes the words to whichever
        // cells their (estimated) positions fall into. Single-token items
        // collapse to a one-element token list and behave exactly as
        // before.
        //
        // The token-level exclusivity (one token → one cell) still
        // prevents the cell-overlap bug where SLANet emits cells whose
        // y-extents overlap between rows. Closest-center disambiguation
        // routes each token to the correct row.

        // Pre-compute each cell's bounds + center (in PDF-pt-flipped space)
        // so we don't redo the work per token.
        let cell_meta: Vec<Option<(RegionBounds, f32, f32)>> = cells
            .iter()
            .map(|cell| {
                let [x1, y1, x2, y2] = cell.page_pt_bbox;
                if x1 >= x2 || y1 >= y2 {
                    return None;
                }
                let bounds = region_bounds(x1, y1, x2, y2, page_h, coords);
                let cx = (bounds.x_min + bounds.x_max) * 0.5;
                let cy = (bounds.y_min + bounds.y_max) * 0.5;
                Some((bounds, cx, cy))
            })
            .collect();

        let mut per_cell_items: Vec<Vec<TextItem>> = vec![Vec::new(); cells.len()];
        // Tokens of an item that did NOT land in any cell during stage 1.
        // These are the orphan candidates handed to `tsr_assign_orphan_items`.
        // Using token-grain orphan candidates (rather than the original wide
        // item) lets stage 2 recover individual words that fell just outside
        // their cell's clamped band, without re-attributing already-claimed
        // tokens.
        let mut orphan_token_subitems: Vec<TextItem> = Vec::new();

        for item in items.iter() {
            let token_subitems = split_item_into_token_subitems(item);
            for token_item in token_subitems {
                let token_w = text_utils::effective_width(&token_item);
                let token_cx = token_item.x + token_w * 0.5;
                let token_cy = token_item.y + text_utils::effective_height(&token_item) * 0.5;
                let mut best: Option<(usize, f32)> = None;
                for (cell_idx, meta) in cell_meta.iter().enumerate() {
                    let Some((bounds, ccx, ccy)) = meta else {
                        continue;
                    };
                    if !tsr_region_contains_item(&token_item, *bounds) {
                        continue;
                    }
                    let dx = token_cx - ccx;
                    let dy = token_cy - ccy;
                    let dist_sq = dx * dx + dy * dy;
                    if best.is_none_or(|(_, d)| dist_sq < d) {
                        best = Some((cell_idx, dist_sq));
                    }
                }
                if let Some((ci, _)) = best {
                    per_cell_items[ci].push(token_item);
                } else {
                    orphan_token_subitems.push(token_item);
                }
            }
        }

        // Build per-cell text from the assigned tokens. Markdown cells must
        // be one line — collapse line breaks from the line-grouping pass.
        for (cell_idx, matched) in per_cell_items.into_iter().enumerate() {
            cells[cell_idx].text = collect_text_from_matched_items(matched, adaptive_threshold)
                .replace(['\n', '\r'], " ");
        }

        // Stage 2: orphan assignment — tokens that didn't land in any cell
        // during stage 1 get assigned to their nearest *empty* cell,
        // clamped by a plausibility cap derived from cell geometry.
        //
        // This recovers two failure modes left by `normalize_cell_bands`:
        //   (a) header text positioned to the LEFT of a column whose band
        //       was derived from data cells centered farther right, so the
        //       header text falls outside the clamped band; and
        //   (b) local SLANet row drift where a cell's bbox sits slightly
        //       above/below its target text item, so the strict rules miss.
        // Empty-cell-only is the safety net: a cell already filled by stage 1
        // is never overwritten or augmented, so the cell-bleed case PR #62
        // closed cannot regress.
        tsr_assign_orphan_items(
            &orphan_token_subitems,
            &mut cells,
            &std::collections::HashSet::new(),
            page_h,
            coords,
        );

        results.push(cells);
    }

    Ok(results)
}

/// Split a `TextItem` into one virtual sub-item per whitespace-separated
/// token, with each token's `x` / `width` estimated from the original item's
/// effective width and the token's character offset.
///
/// PDFs often render an entire row's content as a single Tj — e.g.
/// "Marshall Islands 0.9 0.9 0.9" — producing one wide TextItem whose
/// center sits in only one of the model-emitted cells. Per-item routing
/// then parks the whole row in that one cell. Splitting on whitespace
/// gives each word its own approximate position so per-cell routing can
/// distribute the words to whichever cells their estimated centers fall
/// into.
///
/// The character-width estimate is `item.width / char_count`. A run with a
/// known zero advance keeps all its tokens at its origin — that is where
/// they were drawn; only a run whose advance is unknown (and somehow has no
/// box) is spread over half an em per character for the interpolation. The
/// estimate is uniform across the item —
/// fine for routing, since we only need to know which cell each token's
/// center lands in, not its exact position. Single-token items collapse to
/// a one-element vector equivalent to the input item, making this a no-op
/// for the common case.
fn split_item_into_token_subitems(item: &TextItem) -> Vec<TextItem> {
    let total_chars = item.text.chars().count();
    if total_chars == 0 {
        return Vec::new();
    }
    // Token positions are interpolated along +x, the reading direction of
    // an upright run only. A rotated run (vertical table header) or an
    // upside-down one, whose tokens advance towards -x, stays one item: its
    // box already routes it to the right cell.
    if !item.is_upright() {
        return vec![item.clone()];
    }
    let item_w = if item.width > 0.0 || item.advance_known {
        item.width
    } else {
        total_chars as f32 * item.font_size * 0.5
    };
    let char_w = item_w / total_chars as f32;

    let mut tokens: Vec<TextItem> = Vec::new();
    let mut current_token = String::new();
    let mut current_start_idx: Option<usize> = None;

    let push_token =
        |tokens: &mut Vec<TextItem>, text: String, start_idx: usize, end_idx: usize| {
            if text.is_empty() {
                return;
            }
            let mut sub = item.clone();
            sub.text = text;
            sub.x = item.x + start_idx as f32 * char_w;
            sub.width = (end_idx - start_idx) as f32 * char_w;
            tokens.push(sub);
        };

    for (idx, ch) in item.text.chars().enumerate() {
        if ch.is_whitespace() {
            if let Some(start_idx) = current_start_idx.take() {
                let text = std::mem::take(&mut current_token);
                push_token(&mut tokens, text, start_idx, idx);
            }
        } else {
            if current_start_idx.is_none() {
                current_start_idx = Some(idx);
            }
            current_token.push(ch);
        }
    }
    if let Some(start_idx) = current_start_idx {
        let text = std::mem::take(&mut current_token);
        push_token(&mut tokens, text, start_idx, total_chars);
    }

    tokens
}

/// Compute plausibility caps for the orphan-assignment pass. Returns
/// `(cap_x, cap_y)` — the maximum x/y distance from a text item's center
/// to a candidate empty cell's bbox before the candidate is rejected.
///
/// Caps are derived from cell geometry so they scale with the table:
/// dense small-row tables get a tight cap, looser tables get more slack.
/// Floor values guard against degenerate single-cell tables collapsing
/// the cap to zero.
fn tsr_assignment_caps(cells: &[tables::StructuredCell]) -> (f32, f32) {
    let mut widths: Vec<f32> = Vec::with_capacity(cells.len());
    let mut heights: Vec<f32> = Vec::with_capacity(cells.len());
    for cell in cells {
        let [x1, y1, x2, y2] = cell.page_pt_bbox;
        let w = (x2 - x1).abs();
        let h = (y2 - y1).abs();
        if w > 0.0 && h > 0.0 {
            widths.push(w);
            heights.push(h);
        }
    }
    if widths.is_empty() {
        return (0.0, 0.0);
    }
    widths.sort_by(|a, b| a.total_cmp(b));
    heights.sort_by(|a, b| a.total_cmp(b));
    let median_w = widths[widths.len() / 2];
    let median_h = heights[heights.len() / 2];
    // Floor values: even on a dense table, a 5pt floor handles small
    // pixel-level bbox jitter without being so loose that we'd cross
    // into a neighboring row/column. Symmetric in both axes.
    let cap_x = median_w.max(5.0);
    let cap_y = median_h.max(5.0);
    (cap_x, cap_y)
}

/// For each text item that wasn't claimed by any cell during stage 1,
/// find the nearest *empty* cell within `(cap_x, cap_y)` of the item's
/// center and append the item's text to that cell. Cells that already
/// have content are skipped — stage 2 only fills, never augments.
///
/// Distance is point-to-rect: 0 if the item center is inside the cell's
/// bbox, else the axis-aligned gap to the nearest edge. Both x-gap and
/// y-gap must be within their respective caps for a candidate to qualify;
/// among qualifying candidates, the smallest combined euclidean distance
/// wins.
fn tsr_assign_orphan_items(
    items: &[TextItem],
    cells: &mut [tables::StructuredCell],
    claimed: &std::collections::HashSet<usize>,
    page_height: f32,
    coord_space: RegionCoordSpace,
) {
    if cells.is_empty() {
        return;
    }
    let (cap_x, cap_y) = tsr_assignment_caps(cells);
    if cap_x <= 0.0 || cap_y <= 0.0 {
        return;
    }
    // The caps come from the cells' page-coordinate boxes; in a turned frame
    // the page's horizontal extent runs along the frame's y axis and vice
    // versa, so swap them (and the same-line tolerance derived from cap_y).
    let (cap_x, cap_y) = match coord_space {
        RegionCoordSpace::Standard => (cap_x, cap_y),
        RegionCoordSpace::Rotated90Ccw | RegionCoordSpace::Rotated90Cw => (cap_y, cap_x),
    };
    // Y-tolerance for "same line as a previous orphan" — multi-token branch
    // names like "Blue Valley Parkway" are 3 separate text items and should
    // all stack into the same cell. But two orphans on different rows of
    // the PDF (different y values) targeting the same empty cell should
    // NOT merge — that produces the "Mitchell Woonsocket" / "Shawnee Blue
    // Valley Parkway" run-on cells. Half a row of slack is conservative.
    let y_tolerance = (cap_y * 0.5).max(3.0);

    // Pre-compute each empty cell's region bounds so we don't re-flip
    // page coordinates per orphan-candidate pair.
    let cell_bounds: Vec<Option<RegionBounds>> = cells
        .iter()
        .map(|cell| {
            if !cell.text.is_empty() {
                return None;
            }
            let [x1, y1, x2, y2] = cell.page_pt_bbox;
            if x1 >= x2 || y1 >= y2 {
                return None;
            }
            Some(region_bounds(x1, y1, x2, y2, page_height, coord_space))
        })
        .collect();

    // Track the y-center of the FIRST orphan that landed in each cell so
    // subsequent orphans only stack if they're on the same line.
    let mut stage2_first_y: std::collections::HashMap<usize, f32> =
        std::collections::HashMap::new();

    for (i, item) in items.iter().enumerate() {
        if claimed.contains(&i) {
            continue;
        }
        let item_w = text_utils::effective_width(item);
        if item.text.trim().is_empty() {
            continue;
        }
        let cx = item.x + item_w * 0.5;
        let cy = item.y + text_utils::effective_height(item) * 0.5;

        let mut best: Option<(usize, f32)> = None;
        for (ci, bounds_opt) in cell_bounds.iter().enumerate() {
            let Some(bounds) = bounds_opt else {
                continue;
            };
            // If a previous orphan already landed in this cell, only let a
            // new orphan join if it's on the same line. Cross-line orphans
            // need to look elsewhere (next-nearest empty cell).
            if let Some(&first_y) = stage2_first_y.get(&ci) {
                if (first_y - cy).abs() > y_tolerance {
                    continue;
                }
            }
            let dx = (bounds.x_min - cx).max(0.0).max(cx - bounds.x_max);
            let dy = (bounds.y_min - cy).max(0.0).max(cy - bounds.y_max);
            if dx > cap_x || dy > cap_y {
                continue;
            }
            let dist_sq = dx * dx + dy * dy;
            if best.is_none_or(|(_, d)| dist_sq < d) {
                best = Some((ci, dist_sq));
            }
        }

        if let Some((ci, _)) = best {
            // Append, preserving stage 1's content. Same-line orphans
            // stack to support multi-token text (e.g. "Blue Valley
            // Parkway"); cross-line orphans are filtered out above.
            let trimmed = item.text.trim();
            if cells[ci].text.is_empty() {
                cells[ci].text = trimmed.to_string();
            } else {
                cells[ci].text.push(' ');
                cells[ci].text.push_str(trimmed);
            }
            stage2_first_y.entry(ci).or_insert(cy);
        }
    }
}

/// Extract markdown tables using externally-supplied structure recovery.
///
/// Convenience wrapper around [`extract_tables_with_structure_cells_mem`]
/// that renders each cell list to markdown via
/// [`tables::cells_to_markdown`]. Returns one markdown string per input,
/// in input order. Inputs whose page is out of range or whose tokens parse
/// to zero cells produce an empty string.
pub fn extract_tables_with_structure_mem(
    buffer: &[u8],
    inputs: &[TsrTableInput],
) -> Result<Vec<String>, PdfError> {
    let cells_lists = extract_tables_with_structure_cells_mem(buffer, inputs)?;
    Ok(cells_lists
        .into_iter()
        .map(|cells| {
            if cells.is_empty() {
                String::new()
            } else {
                tables::cells_to_markdown(&cells)
            }
        })
        .collect())
}

/// Markdown for one extracted table plus a diagnostic flag describing
/// which path produced it.
///
/// `fallback_reason` is `None` when the TSR-hybrid path produced the
/// markdown directly; `Some(<short identifier>)` when stage 1's quality
/// check fired and either in-place expansion or the heuristic fallback
/// produced the output. The reason string is stable enough to use as a
/// metric label (e.g. `multi_row_in_cell_expanded`, `phantom_empty_row`).
#[derive(Debug, Clone)]
pub struct TableExtractionResult {
    pub markdown: String,
    pub fallback_reason: Option<String>,
}

#[derive(Debug, Clone)]
enum TsrQualityIssue {
    PhantomEmptyRow,
    MultiRowInCell {
        expanded_cells: Option<Vec<tables::StructuredCell>>,
    },
}

impl TsrQualityIssue {
    fn reason(&self) -> &'static str {
        match self {
            Self::PhantomEmptyRow => "phantom_empty_row",
            Self::MultiRowInCell { .. } => "multi_row_in_cell",
        }
    }
}

#[derive(Clone)]
struct TsrCellTextLine {
    center_y: f32,
    half_height: f32,
    items: Vec<TextItem>,
}

impl TsrCellTextLine {
    fn new(item: TextItem) -> Self {
        let center_y = item.y + text_utils::effective_height(&item) * 0.5;
        let half_height = (text_utils::effective_height(&item) * 0.5).max(2.5);
        Self {
            center_y,
            half_height,
            items: vec![item],
        }
    }

    fn add(&mut self, item: TextItem) {
        let center_y = item.y + text_utils::effective_height(&item) * 0.5;
        let existing = self.items.len() as f32;
        self.center_y = (self.center_y * existing + center_y) / (existing + 1.0);
        self.half_height = self
            .half_height
            .max((text_utils::effective_height(&item) * 0.5).max(2.5));
        self.items.push(item);
    }

    fn bottom_y(&self) -> f32 {
        self.items
            .iter()
            .map(|item| item.y)
            .fold(f32::INFINITY, f32::min)
    }
}

#[derive(Clone)]
struct TsrRowExpansion {
    bands: Vec<f32>,
    tolerance: f32,
}

fn collect_items_in_tsr_cell(
    items: &[TextItem],
    cell: &tables::StructuredCell,
    page_height: f32,
    coord_space: RegionCoordSpace,
) -> Vec<TextItem> {
    let [x1, y1, x2, y2] = cell.page_pt_bbox;
    if x1 >= x2 || y1 >= y2 {
        return Vec::new();
    }
    let bounds = region_bounds(x1, y1, x2, y2, page_height, coord_space);
    items
        .iter()
        .filter(|item| !item.text.trim().is_empty() && tsr_region_contains_item(item, bounds))
        .cloned()
        .collect()
}

fn cluster_tsr_cell_text_lines(mut items: Vec<TextItem>) -> Vec<TsrCellTextLine> {
    if items.is_empty() {
        return Vec::new();
    }

    items.sort_by(|a, b| {
        let ay = a.y + a.height * 0.5;
        let by = b.y + b.height * 0.5;
        by.total_cmp(&ay).then(a.x.total_cmp(&b.x))
    });

    let mut lines: Vec<TsrCellTextLine> = Vec::new();
    for item in items {
        let item_top = item.y + text_utils::effective_height(&item);
        let item_half_height = (text_utils::effective_height(&item) * 0.5).max(2.5);
        if let Some(last) = lines.last_mut() {
            let gap = last.bottom_y() - item_top;
            if gap <= last.half_height.max(item_half_height) {
                last.add(item);
                continue;
            }
        }
        lines.push(TsrCellTextLine::new(item));
    }

    lines
}

fn build_tsr_row_expansion(
    row_cells: &[usize],
    cells: &[tables::StructuredCell],
    cell_lines: &[Vec<TsrCellTextLine>],
) -> Option<TsrRowExpansion> {
    if row_cells.is_empty() {
        return None;
    }
    let has_spanning_cell = row_cells
        .iter()
        .any(|&idx| cells[idx].rowspan > 1 || cells[idx].colspan > 1);
    if has_spanning_cell {
        return None;
    }

    let multiline_cells = row_cells
        .iter()
        .filter(|&&idx| cell_lines[idx].len() >= 2)
        .count();
    if row_cells.len() >= 2 && multiline_cells < 2 {
        return None;
    }
    if row_cells.len() == 1 && multiline_cells == 0 {
        return None;
    }

    let mut centers: Vec<(f32, f32)> = row_cells
        .iter()
        .flat_map(|&idx| {
            cell_lines[idx]
                .iter()
                .map(|line| (line.center_y, line.half_height))
        })
        .collect();
    if centers.len() < 2 {
        return None;
    }
    centers.sort_by(|a, b| b.0.total_cmp(&a.0));

    let mut half_heights: Vec<f32> = centers.iter().map(|(_, h)| *h).collect();
    half_heights.sort_by(|a, b| a.total_cmp(b));
    let tolerance = (half_heights[half_heights.len() / 2] * 0.8).max(3.0);

    let mut bands: Vec<(f32, usize)> = Vec::new();
    for (center, _) in centers {
        if let Some((band_center, count)) = bands
            .iter_mut()
            .find(|(band_center, _)| (*band_center - center).abs() <= tolerance)
        {
            *band_center = (*band_center * *count as f32 + center) / (*count as f32 + 1.0);
            *count += 1;
        } else {
            bands.push((center, 1));
        }
    }

    // V1 targets the common 1-2 lost-row cases; larger compressions stay on
    // the existing heuristic fallback path until we have evidence to broaden it.
    if !(2..=4).contains(&bands.len()) {
        return None;
    }

    let min_support = if row_cells.len() >= 2 { 2 } else { 1 };
    let supported = bands.iter().all(|(band, _)| {
        row_cells
            .iter()
            .filter(|&&idx| {
                cell_lines[idx]
                    .iter()
                    .any(|line| (line.center_y - *band).abs() <= tolerance)
            })
            .count()
            >= min_support
    });
    if !supported {
        return None;
    }

    bands.sort_by(|a, b| b.0.total_cmp(&a.0));
    Some(TsrRowExpansion {
        bands: bands.into_iter().map(|(center, _)| center).collect(),
        tolerance,
    })
}

fn text_for_tsr_band(
    lines: &[TsrCellTextLine],
    band: f32,
    tolerance: f32,
    adaptive_threshold: f32,
) -> String {
    let mut matched: Vec<TextItem> = lines
        .iter()
        .filter(|line| (line.center_y - band).abs() <= tolerance)
        .flat_map(|line| line.items.iter().cloned())
        .collect();
    if matched.is_empty() {
        return String::new();
    }
    matched.sort_by(|a, b| b.y.total_cmp(&a.y).then(a.x.total_cmp(&b.x)));
    collect_text_from_matched_items(matched, adaptive_threshold).replace(['\n', '\r'], " ")
}

fn slice_cell_bbox_for_expanded_row(
    cell: &tables::StructuredCell,
    row_idx: usize,
    row_count: usize,
) -> [f32; 4] {
    let mut bbox = cell.page_pt_bbox;
    let top = bbox[1].min(bbox[3]);
    let bottom = bbox[1].max(bbox[3]);
    let height = bottom - top;
    if height <= 0.0 || row_count == 0 {
        return bbox;
    }
    let step = height / row_count as f32;
    bbox[1] = top + step * row_idx as f32;
    bbox[3] = if row_idx + 1 == row_count {
        bottom
    } else {
        top + step * (row_idx + 1) as f32
    };
    bbox
}

fn try_expand_multi_row_cells(
    cells: &[tables::StructuredCell],
    items: &[TextItem],
    page_height: f32,
    coord_space: RegionCoordSpace,
    adaptive_threshold: f32,
) -> Option<Vec<tables::StructuredCell>> {
    if cells.is_empty() {
        return None;
    }

    let cell_lines: Vec<Vec<TsrCellTextLine>> = cells
        .iter()
        .map(|cell| {
            if cell.rowspan > 1 {
                Vec::new()
            } else {
                cluster_tsr_cell_text_lines(collect_items_in_tsr_cell(
                    items,
                    cell,
                    page_height,
                    coord_space,
                ))
            }
        })
        .collect();

    let mut cells_by_row: BTreeMap<usize, Vec<usize>> = BTreeMap::new();
    for (idx, cell) in cells.iter().enumerate() {
        cells_by_row.entry(cell.row).or_default().push(idx);
    }

    let mut expansions: HashMap<usize, TsrRowExpansion> = HashMap::new();
    for (&row, row_cells) in &cells_by_row {
        let covered_by_rowspan = cells
            .iter()
            .any(|cell| cell.rowspan > 1 && cell.row <= row && row < cell.row + cell.rowspan);
        if covered_by_rowspan {
            continue;
        }
        if let Some(expansion) = build_tsr_row_expansion(row_cells, cells, &cell_lines) {
            expansions.insert(row, expansion);
        }
    }

    if expansions.is_empty() {
        return None;
    }

    let mut expanded = Vec::with_capacity(cells.len() + expansions.len());
    let mut row_shift = 0usize;
    for (&row, row_cells) in &cells_by_row {
        if let Some(expansion) = expansions.get(&row) {
            for (band_idx, band) in expansion.bands.iter().enumerate() {
                for &cell_idx in row_cells {
                    let mut cell = cells[cell_idx].clone();
                    cell.row = row + row_shift + band_idx;
                    cell.rowspan = 1;
                    cell.text = text_for_tsr_band(
                        &cell_lines[cell_idx],
                        *band,
                        expansion.tolerance,
                        adaptive_threshold,
                    );
                    cell.page_pt_bbox =
                        slice_cell_bbox_for_expanded_row(&cell, band_idx, expansion.bands.len());
                    expanded.push(cell);
                }
            }
            row_shift += expansion.bands.len() - 1;
        } else {
            for &cell_idx in row_cells {
                let mut cell = cells[cell_idx].clone();
                cell.row += row_shift;
                expanded.push(cell);
            }
        }
    }

    let original_rows = cells
        .iter()
        .map(|cell| cell.row + cell.rowspan.max(1))
        .max()
        .unwrap_or(0);
    let expanded_rows = expanded
        .iter()
        .map(|cell| cell.row + cell.rowspan.max(1))
        .max()
        .unwrap_or(0);
    (expanded_rows > original_rows).then_some(expanded)
}

/// Detect quality issues in the TSR-hybrid output for a single input.
///
/// Returns `Some(issue)` if the cells look like they reflect a known
/// SLANet detection pathology. Reasons (also used as metric labels):
///
/// * `phantom_empty_row` — a row whose every cell is empty, surrounded
///   above and below by rows with content. SLANet sometimes emits an
///   extra row that doesn't correspond to any visible PDF row.
/// * `multi_row_in_cell` — at least one non-label `rowspan==1` cell
///   encloses PDF text items that cluster into two distinct visual lines
///   separated by a whitespace gap larger than the line height. Cells
///   declared as `rowspan>1` are excluded since they are *expected*
///   to span multiple lines. First-row/first-column wraps are ignored
///   unless the in-place row expansion has enough support to repair them,
///   because those are often legitimate wrapped headers or row labels.
///   SLANet's row under-detection on tightly-packed tables produces the
///   rowspan==1-but-multi-line pattern (the FNBO failure mode).
fn detect_tsr_quality_issue(
    buffer: &[u8],
    input: &TsrTableInput,
    cells: &[tables::StructuredCell],
) -> Result<Option<TsrQualityIssue>, PdfError> {
    if cells.is_empty() {
        return Ok(None);
    }

    // Phantom row: cheap, computed from cell metadata alone.
    let max_row = cells.iter().map(|c| c.row).max().unwrap_or(0);
    if max_row >= 2 {
        let mut row_has_content = vec![false; max_row + 1];
        for cell in cells {
            if !cell.text.trim().is_empty() {
                row_has_content[cell.row] = true;
            }
        }
        for r in 1..max_row {
            if !row_has_content[r] && row_has_content[r - 1] && row_has_content[r + 1] {
                return Ok(Some(TsrQualityIssue::PhantomEmptyRow));
            }
        }
    }

    // Multi-row-in-cell: re-extract PDF text items in the page and look
    // for `rowspan==1` cells that contain items grouped into ≥2 visual
    // lines separated by a real whitespace gap. This is the FNBO mode:
    // a tall TSR cell catches text from two adjacent PDF rows that
    // SLANet failed to separate. Cells declared `rowspan>1` are
    // expected to be multi-line and are excluded.
    let (doc, _page_count) = load_document_from_mem(buffer)?;
    let pages = doc.get_pages();
    let page_1idx = input.page + 1;
    let Some(&page_id) = pages.get(&page_1idx) else {
        return Ok(None);
    };
    let mut needed: HashSet<u32> = HashSet::new();
    needed.insert(page_1idx);
    let font_cmaps = FontCMaps::from_doc_pages_fast(&doc, Some(&needed));
    let extractor::PageBoxExtraction {
        mut items,
        coords_rotated,
        page_box,
        ..
    } = extractor::extract_page_text_items_in_page_box(
        &doc,
        page_id,
        page_1idx,
        &font_cmaps,
        false,
        &mut extractor::FontStyleCache::new(),
        &mut extractor::FormWalkBudget::new(),
    )?;
    let page_h = page_box.height();
    let adaptive_threshold = text_utils::fix_letterspaced_items(&mut items);
    let coords = RegionCoordSpace::from(coords_rotated);
    let expanded_cells =
        try_expand_multi_row_cells(cells, &items, page_h, coords, adaptive_threshold);
    let first_row = cells.iter().map(|cell| cell.row).min().unwrap_or(0);
    let first_col = cells.iter().map(|cell| cell.col).min().unwrap_or(0);

    for cell in cells {
        // rowspan>1 cells are intentionally multi-line — skip them.
        if cell.rowspan > 1 {
            continue;
        }
        if cell.text.trim().is_empty() {
            continue;
        }
        let cell_items = collect_items_in_tsr_cell(&items, cell, page_h, coords);
        if cell_items.len() < 2 {
            continue;
        }
        if cluster_tsr_cell_text_lines(cell_items).len() < 2 {
            continue;
        }
        if expanded_cells.is_some() {
            return Ok(Some(TsrQualityIssue::MultiRowInCell { expanded_cells }));
        }
        if !is_wrapped_tsr_label_cell(cell, first_row, first_col) {
            return Ok(Some(TsrQualityIssue::MultiRowInCell {
                expanded_cells: None,
            }));
        }
    }

    Ok(None)
}

fn is_wrapped_tsr_label_cell(
    cell: &tables::StructuredCell,
    first_row: usize,
    first_col: usize,
) -> bool {
    cell.is_header || cell.row == first_row || cell.col == first_col
}

/// Auto-fallback variant of [`extract_tables_with_structure_mem`]:
/// runs the TSR-hybrid path, checks the resulting cells for known
/// SLANet detection pathologies (phantom rows, multi-row-in-cell text),
/// and falls back to the heuristic [`extract_tables_in_regions_mem`]
/// for any input where the TSR path looks compromised.
///
/// On clean inputs this is identical to the markdown variant. On
/// `multi_row_in_cell`, the wrapper first tries to expand over-stuffed
/// rows in place; if that cannot produce a usable table, the heuristic
/// markdown replaces the TSR markdown and `fallback_reason` is set to
/// the diagnostic label.
///
/// Two failure modes are guarded against per-input:
///
/// * **Empty heuristic**: if the heuristic returns empty/whitespace
///   markdown for a flagged region, the original TSR markdown is
///   preserved and `fallback_reason` is suffixed with
///   `_heuristic_empty` (e.g. `multi_row_in_cell_heuristic_empty`).
///   This avoids replacing a usable wrong-but-non-empty TSR output
///   with literally nothing.
/// * **Expanded multi-row cells**: when in-place recovery succeeds, the
///   result is labeled `multi_row_in_cell_expanded` and the heuristic is
///   not consulted.
/// * **Per-input errors**: any failure in detection or heuristic
///   extraction for a single input is contained — that input
///   returns the raw TSR markdown with `fallback_reason` set to
///   an `_error` label so callers can metric on it. Other inputs
///   in the same batch are unaffected.
///
/// Use this from production callers that want self-healing output.
/// Use [`extract_tables_with_structure_mem`] when you want raw TSR
/// output regardless of quality (e.g. eval harnesses comparing the
/// two paths).
pub fn extract_tables_with_structure_auto_mem(
    buffer: &[u8],
    inputs: &[TsrTableInput],
) -> Result<Vec<TableExtractionResult>, PdfError> {
    let tsr_cells = extract_tables_with_structure_cells_mem(buffer, inputs)?;
    let mut results = Vec::with_capacity(inputs.len());

    for (i, input) in inputs.iter().enumerate() {
        let cells = &tsr_cells[i];
        let tsr_md = if cells.is_empty() {
            String::new()
        } else {
            tables::cells_to_markdown(cells)
        };

        let issue = match detect_tsr_quality_issue(buffer, input, cells) {
            Ok(opt) => opt,
            Err(_) => {
                // Detection failed for this input — fall through with
                // the raw TSR markdown so the rest of the batch is
                // unaffected. Tag the reason for caller metrics.
                results.push(TableExtractionResult {
                    markdown: tsr_md,
                    fallback_reason: Some("detection_error".to_string()),
                });
                continue;
            }
        };

        let result = match issue {
            None => TableExtractionResult {
                markdown: tsr_md,
                fallback_reason: None,
            },
            Some(issue) => {
                let reason = issue.reason().to_string();
                if let TsrQualityIssue::MultiRowInCell {
                    expanded_cells: Some(expanded_cells),
                } = issue
                {
                    let expanded_md = tables::cells_to_markdown(&expanded_cells);
                    if !expanded_md.trim().is_empty() {
                        results.push(TableExtractionResult {
                            markdown: expanded_md,
                            fallback_reason: Some("multi_row_in_cell_expanded".to_string()),
                        });
                        continue;
                    }
                }
                // Fall back to heuristic on the input's table region.
                // The crop's PDF-pt bbox IS the table region.
                let heuristic_md = match extract_tables_in_regions_mem(
                    buffer,
                    &[(input.page, vec![input.crop_pdf_pt_bbox])],
                ) {
                    Ok(pages) => pages
                        .into_iter()
                        .next()
                        .and_then(|p| p.regions.into_iter().next().map(|r| r.text))
                        .unwrap_or_default(),
                    Err(_) => {
                        // Heuristic threw — keep raw TSR markdown.
                        results.push(TableExtractionResult {
                            markdown: tsr_md,
                            fallback_reason: Some(format!("{reason}_heuristic_error")),
                        });
                        continue;
                    }
                };
                if heuristic_md.trim().is_empty() {
                    // Heuristic produced nothing useful — keep TSR
                    // markdown rather than ship empty. The reason
                    // suffix lets callers count this case.
                    TableExtractionResult {
                        markdown: tsr_md,
                        fallback_reason: Some(format!("{reason}_heuristic_empty")),
                    }
                } else {
                    TableExtractionResult {
                        markdown: heuristic_md,
                        fallback_reason: Some(reason),
                    }
                }
            }
        };
        results.push(result);
    }

    Ok(results)
}

/// Coordinate frame the extracted items of a page live in: plain page
/// coordinates, or the frame turned by the rotated-page correction (see
/// `extractor::geometry::PageRotation`). Region boxes arrive in page
/// coordinates and are turned the same way before matching items.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
enum RegionCoordSpace {
    Standard,
    /// Frame turned for a counter-clockwise page: `(x, y) → (y, -x)`.
    Rotated90Ccw,
    /// Frame turned for a clockwise page: `(x, y) → (-y, x)`.
    Rotated90Cw,
}

impl From<extractor::geometry::PageRotation> for RegionCoordSpace {
    fn from(rotation: extractor::geometry::PageRotation) -> Self {
        use extractor::geometry::PageRotation;
        match rotation {
            PageRotation::Upright => RegionCoordSpace::Standard,
            PageRotation::Ccw => RegionCoordSpace::Rotated90Ccw,
            PageRotation::Cw => RegionCoordSpace::Rotated90Cw,
        }
    }
}

/// A region rect in the sheet frame's top-left space: unchanged when the
/// caller gave it there, turned back from the rendered page when it came in
/// the display frame (`display` is the page's mapping in that case).
fn sheet_region_rect(rect: [f32; 4], display: Option<extractor::DisplayPage>) -> [f32; 4] {
    match display {
        Some(page) => page.region_to_sheet(rect),
        None => rect,
    }
}

#[derive(Clone, Copy)]
struct RegionBounds {
    x_min: f32,
    y_min: f32,
    x_max: f32,
    y_max: f32,
}

/// Collect text items that fall within a region bbox (top-left origin, PDF points)
/// and return them as a single string in reading order.
///
/// `page_height` is the height of the frame the items live in — the visible
/// page box height for items from [`extract_text_with_positions_mem`].
///
/// Items from a page whose text was predominantly rotated live in a turned
/// coordinate frame (see [`PageRotation`]). This legacy entry point infers
/// only the counter-clockwise turn, from most items sitting at negative y;
/// a clockwise turn is never inferred, because content drawn at negative
/// page x looks the same. Pass the page's frame explicitly with
/// [`collect_text_in_region_in_frame`] (frames come from
/// [`extract_text_with_positions_and_rotations_mem`]), or use
/// [`extract_text_in_regions_mem`], which handles both turns itself.
pub fn collect_text_in_region(
    items: &[TextItem],
    rx1: f32,
    ry1: f32,
    rx2: f32,
    ry2: f32,
    page_height: f32,
) -> String {
    collect_text_in_region_with_options(
        items,
        rx1,
        ry1,
        rx2,
        ry2,
        page_height,
        infer_region_coord_space(items),
        0.10,
    )
}

/// [`collect_text_in_region`] with the page's coordinate frame given
/// explicitly instead of inferred: `rotation` is the turn
/// [`extract_text_with_positions_and_rotations_mem`] reported for the items'
/// page (`PageRotation::Upright` for pages absent from that map). The region
/// bbox stays in top-left page coordinates and is turned to match.
pub fn collect_text_in_region_in_frame(
    items: &[TextItem],
    rx1: f32,
    ry1: f32,
    rx2: f32,
    ry2: f32,
    page_height: f32,
    rotation: PageRotation,
) -> String {
    collect_text_in_region_with_options(
        items,
        rx1,
        ry1,
        rx2,
        ry2,
        page_height,
        RegionCoordSpace::from(rotation),
        0.10,
    )
}

#[allow(clippy::too_many_arguments)]
fn collect_text_in_region_with_options(
    items: &[TextItem],
    rx1: f32,
    ry1: f32,
    rx2: f32,
    ry2: f32,
    page_height: f32,
    coord_space: RegionCoordSpace,
    adaptive_threshold: f32,
) -> String {
    let bounds = region_bounds(rx1, ry1, rx2, ry2, page_height, coord_space);
    let matched: Vec<TextItem> = items
        .iter()
        .filter(|item| region_overlaps_item(item, bounds))
        .cloned()
        .collect();
    collect_text_from_matched_items(matched, adaptive_threshold)
}

#[cfg(test)]
#[allow(clippy::too_many_arguments)]
fn collect_text_in_tsr_cell(
    items: &[TextItem],
    rx1: f32,
    ry1: f32,
    rx2: f32,
    ry2: f32,
    page_height: f32,
    coord_space: RegionCoordSpace,
    adaptive_threshold: f32,
) -> String {
    let bounds = region_bounds(rx1, ry1, rx2, ry2, page_height, coord_space);
    let matched: Vec<TextItem> = items
        .iter()
        .filter(|item| tsr_region_contains_item(item, bounds))
        .cloned()
        .collect();
    collect_text_from_matched_items(matched, adaptive_threshold)
}

fn collect_text_from_matched_items(matched: Vec<TextItem>, adaptive_threshold: f32) -> String {
    if matched.is_empty() {
        return String::new();
    }

    // Simple extraction: the caller (fire-pdf) already handles reading order
    // and column splitting via the layout model. We just need to sort items
    // top-to-bottom, left-to-right and group into lines. Baselines go
    // through `line_y`, which snaps super/subscript glyph runs onto the body
    // baseline they are attached to: sorting raw `y` put every raised
    // affiliation marker of an author line ahead of the names, and the 3pt
    // window then emitted them as an orphan ",2,3,2,4,*" line.
    let mut sorted = matched;
    sorted.sort_by(|a, b| b.line_y().total_cmp(&a.line_y()).then(a.x.total_cmp(&b.x)));
    let region_rtl = text_utils::is_rtl_text(sorted.iter().map(|i| &i.text));

    let y_tolerance = 3.0;
    let mut lines: Vec<extractor::TextLine> = Vec::new();

    for item in sorted {
        let should_merge = lines.last().is_some_and(|last_line: &extractor::TextLine| {
            last_line.page == item.page && (last_line.y - item.line_y()).abs() < y_tolerance
        });
        if should_merge {
            lines.last_mut().unwrap().items.push(item);
        } else {
            let y = item.line_y();
            let page = item.page;
            lines.push(extractor::TextLine {
                items: vec![item],
                y,
                page,
                adaptive_threshold,
            });
        }
    }

    // Sort items within each line by X position
    for line in &mut lines {
        text_utils::sort_line_items(&mut line.items, region_rtl);
    }

    lines
        .into_iter()
        .map(|line| line.text())
        .collect::<Vec<_>>()
        .join("\n")
}

fn infer_region_coord_space(items: &[TextItem]) -> RegionCoordSpace {
    // Rotated-page correction negates one axis: a counter-clockwise turn
    // maps y = -(old right edge), a clockwise turn maps x = -(old top edge),
    // so most items land at negative y or negative x respectively. Use this
    // to keep `collect_text_in_region` behavior compatible for direct
    // callers that do not have extractor metadata.
    // Legacy heuristic for callers without extractor metadata: a
    // counter-clockwise turn maps y = -(old right edge), so most items land
    // at negative y (the pre-existing rule, kept as is). A clockwise turn is
    // never inferred — its signature (negative x) cannot be told apart from
    // content drawn at negative page coordinates. Callers pass the frame
    // explicitly through `collect_text_in_region_in_frame` or use
    // `extract_text_in_regions_mem`, which carries the page rotation itself.
    let negative_y = items.iter().filter(|item| item.y < 0.0).count();
    if !items.is_empty() && negative_y * 2 >= items.len() {
        RegionCoordSpace::Rotated90Ccw
    } else {
        RegionCoordSpace::Standard
    }
}

/// Convert a top-left-origin region into bounds in the item frame.
/// `page_height` is the visible page box height: items were shifted into
/// that box, so flipping `y` by its height lands regions on them.
fn region_bounds(
    rx1: f32,
    ry1: f32,
    rx2: f32,
    ry2: f32,
    page_height: f32,
    coord_space: RegionCoordSpace,
) -> RegionBounds {
    let tx_min = rx1.min(rx2);
    let tx_max = rx1.max(rx2);
    let ty_min = ry1.min(ry2);
    let ty_max = ry1.max(ry2);
    let by_min = page_height - ty_max;
    let by_max = page_height - ty_min;
    match coord_space {
        RegionCoordSpace::Standard => RegionBounds {
            x_min: tx_min,
            y_min: by_min,
            x_max: tx_max,
            y_max: by_max,
        },
        RegionCoordSpace::Rotated90Ccw => RegionBounds {
            x_min: by_min,
            x_max: by_max,
            y_min: -tx_max,
            y_max: -tx_min,
        },
        RegionCoordSpace::Rotated90Cw => RegionBounds {
            x_min: -by_max,
            x_max: -by_min,
            y_min: tx_min,
            y_max: tx_max,
        },
    }
}

/// Inclusion margin shared by the region/item overlap predicates and the
/// exclusive-assignment area score — these MUST stay in sync: an item that
/// passes the boolean guard must always have positive overlap area.
const REGION_MARGIN: f32 = 1.5;

/// Overlap area between an item and region bounds (same margin as the
/// boolean test) — the exclusive-assignment score.
fn region_item_overlap_area(item: &TextItem, bounds: RegionBounds) -> f32 {
    let item_x_max = item.x + text_utils::effective_width(item);
    let item_y_max = item.y + text_utils::effective_height(item);
    let x_overlap = (item_x_max.min(bounds.x_max + REGION_MARGIN)
        - item.x.max(bounds.x_min - REGION_MARGIN))
    .max(0.0);
    let y_overlap = (item_y_max.min(bounds.y_max + REGION_MARGIN)
        - item.y.max(bounds.y_min - REGION_MARGIN))
    .max(0.0);
    x_overlap * y_overlap
}

fn region_overlaps_item(item: &TextItem, bounds: RegionBounds) -> bool {
    let item_x_min = item.x;
    let item_x_max = item.x + text_utils::effective_width(item);
    let item_y_min = item.y;
    let item_y_max = item.y + text_utils::effective_height(item);

    let x_overlap = (item_x_max.min(bounds.x_max + REGION_MARGIN)
        - item_x_min.max(bounds.x_min - REGION_MARGIN))
    .max(0.0);
    let y_overlap = (item_y_max.min(bounds.y_max + REGION_MARGIN)
        - item_y_min.max(bounds.y_min - REGION_MARGIN))
    .max(0.0);
    x_overlap > 0.0 && y_overlap > 0.0
}

fn region_overlaps_rect(rect: &PdfRect, bounds: RegionBounds) -> bool {
    let (x_min, y_min, x_max, y_max) = normalized_rect_edges(rect);
    ranges_overlap(
        x_min,
        x_max,
        bounds.x_min - REGION_MARGIN,
        bounds.x_max + REGION_MARGIN,
    ) && ranges_overlap(
        y_min,
        y_max,
        bounds.y_min - REGION_MARGIN,
        bounds.y_max + REGION_MARGIN,
    )
}

fn region_overlaps_line(line: &PdfLine, bounds: RegionBounds) -> bool {
    let x_min = line.x1.min(line.x2);
    let x_max = line.x1.max(line.x2);
    let y_min = line.y1.min(line.y2);
    let y_max = line.y1.max(line.y2);
    ranges_overlap(
        x_min,
        x_max,
        bounds.x_min - REGION_MARGIN,
        bounds.x_max + REGION_MARGIN,
    ) && ranges_overlap(
        y_min,
        y_max,
        bounds.y_min - REGION_MARGIN,
        bounds.y_max + REGION_MARGIN,
    )
}

fn ranges_overlap(a_min: f32, a_max: f32, b_min: f32, b_max: f32) -> bool {
    a_max >= b_min && b_max >= a_min
}

fn tsr_region_contains_item(item: &TextItem, bounds: RegionBounds) -> bool {
    let item_x_min = item.x;
    let item_x_max = item.x + text_utils::effective_width(item);
    let item_y_min = item.y;
    let item_y_max = item.y + text_utils::effective_height(item);

    let center_x = (item_x_min + item_x_max) * 0.5;
    let center_y = (item_y_min + item_y_max) * 0.5;
    if center_x >= bounds.x_min
        && center_x <= bounds.x_max
        && center_y >= bounds.y_min
        && center_y <= bounds.y_max
    {
        return true;
    }

    let x_overlap = (item_x_max.min(bounds.x_max) - item_x_min.max(bounds.x_min)).max(0.0);
    let y_overlap = (item_y_max.min(bounds.y_max) - item_y_min.max(bounds.y_min)).max(0.0);
    let item_width = (item_x_max - item_x_min).max(0.1);
    let item_height = (item_y_max - item_y_min).max(0.1);

    x_overlap / item_width >= 0.6 && y_overlap / item_height >= 0.6
}

// =========================================================================
// Internal: single-load document pipeline
// =========================================================================

/// Load a PDF from disk, returning the parsed document and page count.
///
/// `Document::load_metadata` for page count + `Document::load` for content
/// are combined here, but lopdf loads the full doc in `load()` so we extract
/// page count from it directly to avoid the metadata-only round-trip.
pub(crate) fn load_document_from_path<P: AsRef<Path>>(
    path: P,
) -> Result<(Document, u32), PdfError> {
    load_document_from_path_with_password(path, None)
}

/// Load a PDF file, decrypting with `password` if the file is encrypted.
pub(crate) fn load_document_from_path_with_password<P: AsRef<Path>>(
    path: P,
    password: Option<&str>,
) -> Result<(Document, u32), PdfError> {
    let buffer = read_file(path.as_ref())?;
    load_document_from_mem_with_password(&buffer, password)
}

/// Load a PDF from a memory buffer.
pub(crate) fn load_document_from_mem(buffer: &[u8]) -> Result<(Document, u32), PdfError> {
    load_document_from_mem_with_password(buffer, None)
}

/// Load a PDF from a memory buffer, decrypting with `password` if encrypted.
pub(crate) fn load_document_from_mem_with_password(
    buffer: &[u8],
    password: Option<&str>,
) -> Result<(Document, u32), PdfError> {
    load_document_from_mem_with_repairs(buffer, password)
        .map(|(doc, page_count, _)| (doc, page_count))
}

/// Repairs applied to a document's objects once it is loaded, beyond the
/// container repairs the loader tries when a file does not parse.
#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
pub(crate) struct LoadRepairs {
    /// Form XObjects whose zero-area `/BBox` was widened
    /// (see `form_bbox_repair`).
    pub(crate) widened_form_bboxes: usize,
    /// `/BBox` numerals too large for any parser, saturated in the file's
    /// bytes before it was read (see `overlong_numerals`).
    pub(crate) saturated_bbox_numerals: usize,
}

impl LoadRepairs {
    /// Whether a Form XObject was repaired: a renderer given the original
    /// bytes would lose it again.
    pub(crate) fn repaired_forms(&self) -> bool {
        self.widened_form_bboxes > 0 || self.saturated_bbox_numerals > 0
    }
}

/// [`load_document_from_mem_with_password`], also reporting the repairs
/// applied to the loaded objects.
pub(crate) fn load_document_from_mem_with_repairs(
    buffer: &[u8],
    password: Option<&str>,
) -> Result<(Document, u32, LoadRepairs), PdfError> {
    // Drop anything before the `%PDF-` header. Cross-reference offsets are
    // header-relative in every major reader (mupdf, pdfium, poppler, pdf.js;
    // lopdf slices at `%PDF-` internally as well), so this keeps them exact
    // for lopdf and for the repair candidates below. A file whose offsets are
    // off by a constant anyway — they count the leading bytes, or a
    // version-like mention in the leading bytes was taken for the header — is
    // recovered by lopdf's cross-reference reconstruction, the same path every
    // reader takes for it.
    let buffer = strip_leading_bytes_before_header(buffer);

    // Fix malformed struct element names before parsing. Some PDF generators
    // write bare names (/S Code) instead of proper PDF names (/S /Code), which
    // causes lopdf to silently drop the entire object.
    let fixed = structure_tree::fix_bare_struct_names(buffer);
    let buf = fixed.as_ref();

    match load_document_bytes(buf, password) {
        Ok(doc) => {
            let (doc, saturated) = reload_after_saturating_bbox_numerals(doc, buf, password);
            finish_loaded_document(doc, saturated)
        }
        Err(first_err) => {
            for repaired in repair_pdf_container_candidates(buf) {
                match load_document_bytes(&repaired, password) {
                    Ok(doc) => {
                        log::debug!("loaded PDF after repairing malformed container bytes");
                        let (doc, saturated) =
                            reload_after_saturating_bbox_numerals(doc, &repaired, password);
                        return finish_loaded_document(doc, saturated);
                    }
                    Err(e) => {
                        if is_encrypted_lopdf_error(&e) {
                            return Err(e.into());
                        }
                    }
                }
            }
            Err(first_err.into())
        }
    }
}

/// The document loaded again from `bytes` with the `/BBox` numerals of its
/// unloaded objects saturated (see `overlong_numerals`), when that brings
/// objects in — otherwise `doc` as it came — and then with the `/BBox`
/// arrays its object streams could not yield recovered into it; with the
/// count of numerals saturated either way.
fn reload_after_saturating_bbox_numerals(
    doc: Document,
    bytes: &[u8],
    password: Option<&str>,
) -> (Document, usize) {
    let (mut doc, count) = match overlong_numerals::saturate_overlong_bbox_numerals(bytes, &doc) {
        Some((rewritten, count)) => match load_document_bytes(&rewritten, password) {
            Ok(reloaded) if reloaded.objects.len() > doc.objects.len() => {
                log::debug!("loaded PDF after saturating {count} /BBox numeral(s) no parser holds");
                (reloaded, count)
            }
            _ => (doc, 0),
        },
        None => (doc, 0),
    };
    let in_object_streams =
        overlong_numerals::recover_referenced_bboxes_in_object_streams(&mut doc);
    if in_object_streams > 0 {
        log::debug!(
            "recovered /BBox array(s) from object streams after saturating {in_object_streams} \
             numeral(s) no parser holds"
        );
    }
    (doc, count + in_object_streams)
}

/// A loaded document with zero pages is unusable by every caller, and with
/// the decompression bound in place it can also mean the page tree lived in
/// an object stream lopdf skipped for exceeding the bound. Fail the load
/// either way rather than letting a pageless document masquerade as a
/// successful parse.
fn finish_loaded_document(
    mut doc: Document,
    saturated_bbox_numerals: usize,
) -> Result<(Document, u32, LoadRepairs), PdfError> {
    let page_count = doc.get_pages().len() as u32;
    if page_count == 0 {
        return Err(PdfError::Parse(
            "document has no readable pages".to_string(),
        ));
    }
    // lopdf drops the contents of object streams it could not expand (over
    // the decompression bound, or unparseable) without any signal on the
    // returned document. The only trace either loader path leaves is a
    // deficit between the xref size (`max_id`) and the objects actually
    // loaded — which free xref entries also contribute to, so this is a
    // heuristic: surface large deficits for diagnosability, worded so a
    // legitimately sparse xref isn't reported as data loss. The document is
    // still usable either way; unloaded objects resolve as not-found.
    let expected = doc.max_id as usize;
    let loaded = doc.objects.len();
    let missing = expected.saturating_sub(loaded);
    if missing > 1000 && missing.saturating_mul(64) > expected {
        log::warn!(
            "loaded {loaded} of {expected} xref object slots; the rest are free \
             xref entries or object streams skipped over the \
             {MAX_STREAM_DECOMPRESSED_BYTES}-byte decompression bound, and will \
             resolve as not-found"
        );
    }
    let repairs = LoadRepairs {
        widened_form_bboxes: form_bbox_repair::widen_degenerate_form_bboxes(&mut doc),
        saturated_bbox_numerals,
    };
    Ok((doc, page_count, repairs))
}

/// Per-stream decompression budget applied while loading (object streams and
/// xref streams), and the bound within which the detector reads a font's
/// ToUnicode CMap (`detector::font_decoder`), object streams are read again
/// for overlong numerals (`overlong_numerals`), and a CID font's embedded
/// program is read to tell whether its subset was renumbered
/// (`tounicode::try_remap_subset_cmap`): none of those streams costs more
/// than a stream the loader materializes, and changing the bound changes
/// them all. Some tagged PDFs pack their structure tree into object
/// streams that inflate to hundreds of MB each from a ~20MB file; lopdf
/// materializes every object stream eagerly at load, so without a bound one
/// such document exhausts memory before any of our code runs. lopdf skips an
/// object stream that would exceed the bound (its objects resolve as
/// not-found), which the zero-page check in `load_document_from_mem_with_password`
/// turns into a load error instead of a silently wrong answer.
pub(crate) const MAX_STREAM_DECOMPRESSED_BYTES: usize = 8 * 1024 * 1024;

fn bounded_load_options() -> lopdf::LoadOptions {
    lopdf::LoadOptions {
        max_decompressed_size: Some(MAX_STREAM_DECOMPRESSED_BYTES),
        ..Default::default()
    }
}

fn load_document_bytes(buf: &[u8], password: Option<&str>) -> Result<Document, lopdf::Error> {
    match Document::load_mem_with_options(buf, bounded_load_options()) {
        // Some encrypted PDFs load structurally but leave their streams
        // encrypted (`is_encrypted()` stays true); reading them yields garbage
        // until we re-load with a password. Others fail load_mem outright with
        // an encryption error. Handle both by re-loading with the password.
        Ok(doc) if doc.is_encrypted() => decrypt_document_bytes(buf, password),
        Ok(doc) => Ok(doc),
        Err(ref e) if is_encrypted_lopdf_error(e) => decrypt_document_bytes(buf, password),
        Err(e) => Err(e),
    }
}

/// Re-load an encrypted PDF, decrypting with `password`. Falls back to the
/// empty password (owner-only encryption, the common "protected" case) when a
/// non-empty password was supplied but rejected.
fn decrypt_document_bytes(buf: &[u8], password: Option<&str>) -> Result<Document, lopdf::Error> {
    let pw = password.unwrap_or("");
    let with_password = |pw: &str| lopdf::LoadOptions {
        password: Some(pw.to_string()),
        ..bounded_load_options()
    };
    match Document::load_mem_with_options(buf, with_password(pw)) {
        Ok(doc) => Ok(doc),
        Err(inner) if !pw.is_empty() => {
            Document::load_mem_with_options(buf, with_password("")).map_err(|_| inner)
        }
        Err(inner) => Err(inner),
    }
}

fn repair_pdf_container_candidates(buf: &[u8]) -> Vec<Vec<u8>> {
    let mut candidates = Vec::new();

    add_repair_candidate(&mut candidates, append_missing_eof_marker(buf), buf);
    add_repair_candidate(&mut candidates, recover_startxref_pointer(buf), buf);
    add_repair_candidate(
        &mut candidates,
        xref_repair::rebuild_short_xref_entries(buf),
        buf,
    );

    candidates
}

/// Some PDF writers emit a `startxref` pointer that doesn't actually point
/// at the cross-reference table — a single corrupted byte in the offset is
/// enough. lopdf trusts that pointer outright and fails to load rather than
/// searching for the real table, unlike pypdf/pdfium which both recover by
/// locating it directly. This finds the real (classic, non-stream) `xref`
/// table by scanning for the keyword — validating that a plausible
/// subsection header follows, not just any standalone "xref" token, since
/// this crate processes untrusted input and a coincidental match inside
/// unrelated stream/string content must not get "repaired" against a bogus
/// offset (lopdf would then load successfully against garbage instead of
/// returning a clean error) — and appends a corrected trailing
/// `startxref`/`%%EOF` block. lopdf's own `get_xref_start` always uses the
/// *last* `%%EOF` in the final 512 bytes of the buffer, so ours
/// transparently supersedes the broken one without needing to touch
/// anything already in the file.
///
/// Doesn't cover cross-reference *streams* (`N 0 obj << /Type /XRef ...`,
/// used by some PDF 1.5+ writers instead of a classic table) — recovering
/// those needs the containing object's number, not just a byte offset.
fn recover_startxref_pointer(buf: &[u8]) -> Option<Vec<u8>> {
    let xref_pos = find_last_valid_xref_table_start(buf)?;

    let mut repaired = Vec::with_capacity(buf.len() + 32);
    repaired.extend_from_slice(buf);
    if !repaired.ends_with(b"\n") {
        repaired.push(b'\n');
    }
    repaired.extend_from_slice(format!("startxref\n{xref_pos}\n%%EOF\n").as_bytes());
    Some(repaired)
}

/// Finds the last standalone `xref` token in `buf` that is immediately
/// followed by a plausible classic cross-reference subsection header
/// (`<start-id> <count>`, e.g. "0 6") — the shape every real classic xref
/// table starts with. A single reverse byte scan: O(n) even on a
/// pathological buffer with many non-matching or non-standalone "xref"
/// occurrences, unlike repeatedly re-searching a shrinking prefix.
fn find_last_valid_xref_table_start(buf: &[u8]) -> Option<usize> {
    const KEYWORD: &[u8] = b"xref";
    if buf.len() < KEYWORD.len() {
        return None;
    }
    let mut pos = buf.len() - KEYWORD.len();
    loop {
        if &buf[pos..pos + KEYWORD.len()] == KEYWORD {
            let before_ok = pos == 0 || buf[pos - 1].is_ascii_whitespace();
            let after_ok = buf
                .get(pos + KEYWORD.len())
                .is_none_or(|c| c.is_ascii_whitespace());
            if before_ok && after_ok && looks_like_xref_subsection_header(buf, pos + KEYWORD.len())
            {
                return Some(pos);
            }
        }
        if pos == 0 {
            return None;
        }
        pos -= 1;
    }
}

/// Checks that `buf[pos..]` starts (after whitespace) with two
/// whitespace-separated runs of ASCII digits — `<start-id> <count>`, the
/// first subsection header of a classic PDF cross-reference table.
fn looks_like_xref_subsection_header(buf: &[u8], pos: usize) -> bool {
    fn skip_ws(buf: &[u8], mut pos: usize) -> usize {
        while buf.get(pos).is_some_and(u8::is_ascii_whitespace) {
            pos += 1;
        }
        pos
    }
    fn skip_digits(buf: &[u8], mut pos: usize) -> usize {
        while buf.get(pos).is_some_and(u8::is_ascii_digit) {
            pos += 1;
        }
        pos
    }

    let pos = skip_ws(buf, pos);
    let after_first_digits = skip_digits(buf, pos);
    if after_first_digits == pos {
        return false; // no start-id
    }
    let sep = skip_ws(buf, after_first_digits);
    if sep == after_first_digits {
        return false; // start-id and count must be whitespace-separated
    }
    let after_count = skip_digits(buf, sep);
    if after_count == sep {
        return false; // no count
    }
    // The count run must end at whitespace/buffer-end, not run into trailing
    // garbage (e.g. a coincidental "xref\n0 6garbage" in stream content).
    buf.get(after_count).is_none_or(u8::is_ascii_whitespace)
}

fn add_repair_candidate(
    candidates: &mut Vec<Vec<u8>>,
    candidate: Option<Vec<u8>>,
    original: &[u8],
) {
    let Some(candidate) = candidate else {
        return;
    };
    if candidate.as_slice() == original {
        return;
    }
    if candidates.iter().any(|existing| existing == &candidate) {
        return;
    }
    candidates.push(candidate);
}

fn append_missing_eof_marker(buf: &[u8]) -> Option<Vec<u8>> {
    if contains_recent_eof_marker(buf) {
        return None;
    }

    let mut end = buf.len();
    while end > 0 && buf[end - 1].is_ascii_whitespace() {
        end -= 1;
    }

    if !buf[..end].ends_with(b"%%EO") {
        return None;
    }

    let mut repaired = Vec::with_capacity(end + 2);
    repaired.extend_from_slice(&buf[..end]);
    repaired.extend_from_slice(b"F\n");
    Some(repaired)
}

fn contains_recent_eof_marker(buf: &[u8]) -> bool {
    let start = buf.len().saturating_sub(1024);
    buf[start..].windows(b"%%EOF".len()).any(|w| w == b"%%EOF")
}

/// Core processing pipeline operating on a pre-loaded document.
fn process_document(
    doc: Document,
    page_count: u32,
    options: PdfOptions,
    start: ProcessingTimer,
) -> Result<PdfProcessResult, PdfError> {
    // Step 1 — Detection (cheap: scans content streams for text operators)
    let detection = detector::detect_from_document(&doc, page_count, &options.detection)?;
    let pdf_type = detection.pdf_type;
    let pages_needing_ocr = detection.pages_needing_ocr;
    let title = detection.title;
    let author = detection.author;
    let subject = detection.subject;
    let keywords = detection.keywords;
    let creator = detection.creator;
    let producer = detection.producer;
    let creation_date = detection.creation_date;
    let mod_date = detection.mod_date;
    let confidence = detection.confidence;
    let detection_ocr_reasons = detection.ocr_reasons_by_page;

    // DetectOnly → return immediately
    if options.mode == ProcessMode::DetectOnly {
        return Ok(PdfProcessResult {
            pdf_type,
            markdown: None,
            page_count,
            processing_time_ms: start.elapsed_ms(),
            pages_needing_ocr,
            ocr_reasons_by_page: page_ocr_reasons_vec(detection_ocr_reasons),
            title,
            author,
            subject,
            keywords,
            creator,
            producer,
            creation_date,
            mod_date,
            confidence,
            layout: LayoutComplexity::default(),
            has_encoding_issues: false,
            cmap_gaps: Vec::new(),
        });
    }

    // Scanned / ImageBased → nothing to extract
    if matches!(pdf_type, PdfType::Scanned | PdfType::ImageBased) {
        return Ok(PdfProcessResult {
            pdf_type,
            markdown: None,
            page_count,
            processing_time_ms: start.elapsed_ms(),
            pages_needing_ocr,
            ocr_reasons_by_page: page_ocr_reasons_vec(detection_ocr_reasons),
            title,
            author,
            subject,
            keywords,
            creator,
            producer,
            creation_date,
            mod_date,
            confidence,
            layout: LayoutComplexity::default(),
            has_encoding_issues: false,
            cmap_gaps: Vec::new(),
        });
    }

    // Step 2 — Extraction (reuses the already-loaded document)
    let extracted = {
        let font_cmaps = FontCMaps::from_doc(&doc);
        // Most page-filtered requests extract only the selected pages. Gather
        // other pages only when a selected contextual folio needs cross-page
        // evidence; failures on those context-only pages are non-fatal.
        let result = extractor::extract_positioned_text_with_folio_context(
            &doc,
            &font_cmaps,
            options.page_filter.as_ref(),
        );

        // For Mixed/template PDFs: if normal extraction produces garbage text
        // (mostly non-alphanumeric), retry with invisible (Tr=3) text included.
        // This unlocks OCR text layers behind scanned images.
        if pdf_type == PdfType::Mixed {
            if let Ok((ref items, _, _)) = result.as_ref().map(|(e, _, _, _, _)| e) {
                let sample: String = items
                    .iter()
                    .filter(|item| {
                        options
                            .page_filter
                            .as_ref()
                            .is_none_or(|filter| filter.contains(&item.page))
                    })
                    .take(200)
                    .map(|item| item.text.as_str())
                    .collect();
                if is_garbage_text(&sample) || sample.trim().is_empty() {
                    extractor::extract_positioned_text_include_invisible_with_folio_context(
                        &doc,
                        &font_cmaps,
                        options.page_filter.as_ref(),
                    )
                } else {
                    result
                }
            } else {
                // Normal extraction failed — try invisible as fallback
                extractor::extract_positioned_text_include_invisible_with_folio_context(
                    &doc,
                    &font_cmaps,
                    options.page_filter.as_ref(),
                )
            }
        } else {
            result
        }
    };

    // For Mixed PDFs, extraction failure is non-fatal
    let extracted = if pdf_type == PdfType::Mixed {
        extracted.ok()
    } else {
        Some(extracted?)
    };

    // Parse structure tree for tagged PDFs (reuses the loaded document)
    let (struct_roles, struct_tables) = structure_tree::StructTree::from_doc(&doc)
        .map(|tree| {
            let page_ids = doc.get_pages();
            let roles = tree.mcid_to_roles(&page_ids);
            let tables = tree.extract_tables(&page_ids);
            if !roles.is_empty() {
                log::debug!(
                    "structure tree: {} pages with MCID roles, {} total MCIDs, {} tagged tables",
                    roles.len(),
                    tree.mcid_count(),
                    tables.len()
                );
            }
            let roles = if roles.is_empty() { None } else { Some(roles) };
            (roles, tables)
        })
        .unwrap_or((None, Vec::new()));

    let (
        markdown,
        layout,
        has_encoding_issues,
        gid_pages,
        text_quality_pages,
        text_quality_reasons_by_page,
        cmap_gaps,
    ) = match extracted {
        Some((
            (items, rects, lines),
            page_thresholds,
            gid_encoded_pages,
            _page_rotations,
            cmap_coverage,
        )) => {
            let mut ocr_reasons_by_page = BTreeMap::new();

            // For TextBased PDFs with pages flagged for OCR (Identity-H or
            // Type3 fonts without ToUnicode), check whether the CID-as-Unicode
            // passthrough actually produced readable text.  If a page's text
            // is garbage, strip its items so we don't emit mojibake.
            // Only applies to TextBased — for Mixed PDFs, OCR flags come from
            // template images rather than font encoding issues.
            let (items, rects, lines) =
                if pages_needing_ocr.is_empty() || pdf_type != PdfType::TextBased {
                    (items, rects, lines)
                } else {
                    let ocr_set: std::collections::HashSet<u32> =
                        pages_needing_ocr.iter().copied().collect();
                    // Collect text per OCR-flagged page and check quality
                    let mut garbage_pages: std::collections::HashSet<u32> =
                        std::collections::HashSet::new();
                    for &pg in &ocr_set {
                        if options
                            .page_filter
                            .as_ref()
                            .is_some_and(|filter| !filter.contains(&pg))
                        {
                            continue;
                        }
                        let page_text: String = items
                            .iter()
                            .filter(|i| i.page == pg)
                            .map(|i| i.text.as_str())
                            .collect();
                        if is_cid_garbage(&page_text) {
                            garbage_pages.insert(pg);
                        }
                    }
                    if garbage_pages.is_empty() {
                        (items, rects, lines)
                    } else {
                        log::debug!(
                            "suppressing garbage text from OCR-flagged pages: {:?}",
                            garbage_pages
                        );
                        for page in &garbage_pages {
                            add_ocr_reason(
                                &mut ocr_reasons_by_page,
                                *page,
                                OCR_REASON_SUSPECTED_GARBLED_TEXT,
                            );
                        }
                        let items: Vec<_> = items
                            .into_iter()
                            .filter(|i| !garbage_pages.contains(&i.page))
                            .collect();
                        let rects: Vec<_> = rects
                            .into_iter()
                            .filter(|r| !garbage_pages.contains(&r.page))
                            .collect();
                        let lines: Vec<_> = lines
                            .into_iter()
                            .filter(|l| !garbage_pages.contains(&l.page))
                            .collect();
                        (items, rects, lines)
                    }
                };

            let selected_page = |page: u32| {
                options
                    .page_filter
                    .as_ref()
                    .is_none_or(|filter| filter.contains(&page))
            };
            let rects: Vec<_> = rects
                .into_iter()
                .filter(|rect| selected_page(rect.page))
                .collect();
            let lines: Vec<_> = lines
                .into_iter()
                .filter(|line| selected_page(line.page))
                .collect();
            let gid_encoded_pages: HashSet<_> = gid_encoded_pages
                .into_iter()
                .filter(|page| selected_page(*page))
                .collect();
            let FolioFilteredItems {
                items,
                layout_items,
                removal_mask,
                removed_pages,
            } = select_items_with_document_folio_context(
                items,
                page_count,
                options.page_filter.as_ref(),
            );

            let text_quality = analyze_text_quality(&items);
            merge_ocr_reasons(&mut ocr_reasons_by_page, text_quality.reasons_by_page);
            let chart_regions = markdown::chart_regions_by_page(&items, &rects, &lines);
            let layout = compute_layout_complexity_with_chart_regions(
                &items,
                &layout_items,
                &rects,
                &lines,
                &chart_regions,
            );

            let md = if options.mode == ProcessMode::Analyze {
                None
            } else {
                Some(markdown::to_markdown_from_items_with_rects_and_lines(
                    items,
                    options.markdown,
                    &rects,
                    &lines,
                    markdown::MarkdownDocumentContext {
                        page_thresholds: &page_thresholds,
                        struct_roles: struct_roles.as_ref(),
                        struct_tables: &struct_tables,
                        page_count,
                        prefiltered_page_number_pages: Some(&removed_pages),
                        prefiltered_page_number_mask: Some(removal_mask.as_slice()),
                        precomputed_chart_regions: Some(&chart_regions),
                    },
                ))
            };

            // A code no CMap could read is an encoding issue whether or not
            // the Markdown that would show its U+FFFD is generated in this
            // mode; a gap read from its neighbours is not one.
            let cmap_unmapped = cmap_coverage.values().any(|stats| stats.unmapped > 0);
            let enc = !ocr_reasons_by_page.is_empty()
                || text_quality.has_encoding_issues
                || cmap_unmapped
                || md.as_ref().is_some_and(|m| detect_encoding_issues(m));
            (
                md,
                layout,
                enc,
                gid_encoded_pages,
                text_quality.pages_needing_ocr,
                ocr_reasons_by_page,
                font_cmap_gaps(cmap_coverage),
            )
        }
        None => (
            None,
            LayoutComplexity::default(),
            false,
            std::collections::HashSet::new(),
            Vec::new(),
            BTreeMap::new(),
            Vec::new(),
        ),
    };

    // If the extracted text is predominantly garbage (non-alphanumeric) and
    // the PDF is image-backed (Mixed/template), upgrade to Scanned — the text
    // layer comes from a bad OCR pass, and callers should use proper OCR.
    let (pdf_type, markdown, confidence) =
        if pdf_type == PdfType::Mixed && markdown.as_ref().is_some_and(|m| is_garbage_text(m)) {
            (PdfType::Scanned, None, 0.95)
        } else {
            (pdf_type, markdown, confidence)
        };

    // If a TextBased PDF produces garbage text, the fonts are undecodable
    // (e.g. Identity-H without ToUnicode for non-Latin scripts like Cyrillic).
    // Drop the useless markdown and flag all pages for OCR.
    let (markdown, has_encoding_issues, force_ocr_all) = if pdf_type == PdfType::TextBased
        && markdown.as_ref().is_some_and(|m| is_garbage_text(m))
    {
        log::debug!("TextBased PDF has garbage text — flagging all pages for OCR");
        (None, true, true)
    } else {
        (markdown, has_encoding_issues, false)
    };

    // Add pages with gid-encoded fonts (unresolvable encoding) to OCR list.
    // When ALL pages have gid-encoded fonts, suppress unreliable markdown.
    let all_gid = !gid_pages.is_empty() && gid_pages.len() as u32 >= page_count;
    let mut pages_needing_ocr = pages_needing_ocr;
    if force_ocr_all {
        pages_needing_ocr = (1..=page_count).collect();
    }
    if !gid_pages.is_empty() {
        log::debug!("pages with gid-encoded fonts (need OCR): {:?}", gid_pages);
        for page in gid_pages {
            if !pages_needing_ocr.contains(&page) {
                pages_needing_ocr.push(page);
            }
        }
        pages_needing_ocr.sort_unstable();
    }
    if !text_quality_pages.is_empty() {
        log::debug!(
            "pages with OCR reason {} (need OCR): {:?}",
            OCR_REASON_SUSPECTED_GARBLED_TEXT,
            text_quality_pages
        );
        for page in text_quality_pages {
            if !pages_needing_ocr.contains(&page) {
                pages_needing_ocr.push(page);
            }
        }
        pages_needing_ocr.sort_unstable();
    }

    // Detect sparse extraction: when a TEXT-BASED PDF produces very few
    // characters per page, the text is likely embedded in images/forms
    // that need OCR.  Flag all pages for OCR in this case.
    // Only check when markdown was actually generated (not in Analyze mode).
    if pdf_type == PdfType::TextBased
        && page_count > 0
        && pages_needing_ocr.is_empty()
        && markdown.is_some()
    {
        let md_len = markdown.as_ref().map_or(0, |m| m.len());
        let chars_per_page = md_len as f32 / page_count as f32;
        if chars_per_page < 50.0 && md_len < 500 {
            log::debug!(
                "sparse extraction: {:.0} chars/page — recommending OCR for all {} pages",
                chars_per_page,
                page_count
            );
            pages_needing_ocr = (1..=page_count).collect();
        }
    }

    let markdown = if all_gid {
        log::debug!(
            "all {} pages have gid-encoded fonts — suppressing markdown output",
            page_count
        );
        None
    } else {
        markdown
    };

    Ok(PdfProcessResult {
        pdf_type,
        markdown,
        page_count,
        processing_time_ms: start.elapsed_ms(),
        pages_needing_ocr,
        ocr_reasons_by_page: {
            // Detector reasons (scanned / no_text / vector_text /
            // invisible_text_layer / garbled) merged with the
            // markdown-stage garbled detection, deduped per page.
            let mut merged = detection_ocr_reasons;
            merge_ocr_reasons(&mut merged, text_quality_reasons_by_page);
            page_ocr_reasons_vec(merged)
        },
        title,
        author,
        subject,
        keywords,
        creator,
        producer,
        creation_date,
        mod_date,
        confidence,
        layout,
        has_encoding_issues,
        cmap_gaps,
    })
}

/// The fonts whose CMap lacked an entry for a code shown through it, with
/// their counts, from the per-font coverage of an extraction.
fn font_cmap_gaps(coverage: types::CMapCoverageByFont) -> Vec<FontCMapGaps> {
    coverage
        .into_iter()
        .filter(|(_, stats)| stats.has_gaps())
        .map(|(font, stats)| FontCMapGaps {
            font,
            codes: stats.codes,
            interpolated: stats.interpolated,
            unmapped: stats.unmapped,
        })
        .collect()
}

// =========================================================================
// Internal helpers
// =========================================================================

fn suspected_garbled_reason() -> String {
    OCR_REASON_SUSPECTED_GARBLED_TEXT.to_string()
}

pub(crate) fn add_ocr_reason(
    reasons_by_page: &mut BTreeMap<u32, Vec<String>>,
    page: u32,
    reason: &str,
) {
    let reasons = reasons_by_page.entry(page).or_default();
    if !reasons.iter().any(|existing| existing == reason) {
        reasons.push(reason.to_string());
    }
}

fn merge_ocr_reasons(
    reasons_by_page: &mut BTreeMap<u32, Vec<String>>,
    extra_reasons_by_page: BTreeMap<u32, Vec<String>>,
) {
    for (page, reasons) in extra_reasons_by_page {
        for reason in reasons {
            add_ocr_reason(reasons_by_page, page, &reason);
        }
    }
}

fn page_ocr_reason(reasons_by_page: &BTreeMap<u32, Vec<String>>, page: u32) -> Option<String> {
    reasons_by_page
        .get(&page)
        .and_then(|reasons| reasons.first())
        .cloned()
}

fn page_ocr_reasons_vec(reasons_by_page: BTreeMap<u32, Vec<String>>) -> Vec<PageOcrReasons> {
    reasons_by_page
        .into_iter()
        .map(|(page, reasons)| PageOcrReasons { page, reasons })
        .collect()
}

/// Detect markdown tables with suspicious structure that suggest the heuristic
/// missed/mangled rows or columns. Returns true when the caller should treat
/// the result as `needs_ocr` and fall back to GPU OCR.
///
/// Catches three failure modes observed in production:
///
/// 1. **Header row looks like a data row** — first row starts with a numeric
///    value (e.g. `|2|...`), suggesting we missed the actual header above it.
///    Real headers almost never start with a bare number.
///
/// 2. **Header has empty cells in a multi-column table** — e.g.
///    `|Position||Administration|Administration|` (3+ cols, ≥1 empty cell).
///    Indicates poor column boundary detection.
///
/// 3. **Header has duplicate non-empty cells** in a multi-column table —
///    e.g. `Administration|Administration` appearing as adjacent cells means
///    we collapsed multi-line headers wrong.
///
/// Conservative by design: a few false positives (perfectly fine tables flagged)
/// just mean we run GPU OCR which is the existing safe path.
/// When `layout_assisted` is true (the layout model identified this region
/// as a table), we relax boundary-detection heuristics (numeric header,
/// empty header cells, sparse first data row) because the layout model
/// already gave us the table bbox — we're not guessing "is this a table?"
/// anymore, only "can we extract it correctly?". Paragraph and duplicate-
/// header checks stay, since those indicate genuine extraction quality
/// issues regardless of how the region was identified.
/// Return true when the captured table markdown represents only a small
/// fraction of the text the page extractor actually saw inside the
/// region — typically a header-only band or a sparse fragment where
/// the detector found valid grid structure but missed most of the
/// data rows below.
///
/// Tuned at a 25% floor: tables that captured at least a quarter of
/// the region's text are treated as complete-enough. Below 25%, the
/// caller falls back to `needs_ocr = true` so GLM-OCR can take over.
/// The 200-char region floor keeps short legitimate tables (units,
/// axis labels, single-row stat blocks) from being mis-flagged.
fn captured_only_a_fragment(markdown: &str, region_text_chars: usize) -> bool {
    if region_text_chars <= 200 {
        return false;
    }
    let captured_text_chars: usize = markdown
        .chars()
        .filter(|c| !matches!(c, '|' | '-' | '\n'))
        .count();
    captured_text_chars * 4 < region_text_chars
}

/// Return true when the text the page extractor saw inside this region
/// is far too little for the bbox area — a strong signal that the page
/// has a font-CMap failure: Identity-H fonts with missing or broken
/// ToUnicode entries, Type-3 fonts without unicode metadata, etc. The
/// rendered image still carries the visible text (so GLM-OCR will
/// succeed), but the text extractor returns punctuation-only fragments
/// or single-glyph repeats.
///
/// `captured_only_a_fragment` can't catch this case on its own because
/// `region_text_chars` is itself symmetrically low under font-decode
/// failure — the captured-vs-region ratio still looks fine when both
/// numerator and denominator collapse. The area-based floor breaks the
/// symmetry: bbox area is independent of extraction success.
///
/// Threshold of 0.003 chars/sq pt sits between observed clean
/// extractions (≥0.005 chars/sq pt on full-page A4 tables, key/value
/// layouts, archival catalogs) and observed font-decode failures
/// (≤0.0014 chars/sq pt on the prod-traffic samples that motivated
/// this guard).
///
/// Two char-count + area bounds keep the guard from misfiring:
///   - text_chars < 20: too few chars to distinguish a font-decode
///     failure from a synthetic / fragmentary fixture. Observed
///     prod failures decode ≥30 chars before bottoming out.
///   - area < 30,000 sq pt: tiny stat blocks where density is
///     naturally low even for legitimate extractions.
///   - area > 400,000 sq pt: near-whole-A4 bboxes include large
///     white-space margins, so density is unreliable. Real font-
///     decode failures present at typical table sizes
///     (50k–400k sq pt).
fn region_text_density_too_low(region_text_chars: usize, region_area: f32) -> bool {
    if region_text_chars < 20 {
        return false;
    }
    if !(30_000.0..=400_000.0).contains(&region_area) {
        return false;
    }
    (region_text_chars as f32) / region_area < 0.003
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum TableCandidateSource {
    Rect,
    Line,
    Heuristic,
    Column,
    KeyValue,
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum TableCandidateIssue {
    LineRowUndercount,
    SparseWideUndercount,
    TextColumnUndercount,
    ProseGridFragment,
}

#[derive(Debug, Clone)]
struct TableCandidate {
    markdown: String,
    source: TableCandidateSource,
    shape: MarkdownTableShape,
    issue: Option<TableCandidateIssue>,
}

#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
struct MarkdownTableShape {
    rows: usize,
    cols: usize,
    raw_cols: usize,
}

fn select_table_candidate(candidates: &[TableCandidate]) -> Option<&TableCandidate> {
    let first = candidates.first()?;

    // A line grid that visibly collapses several captured text baselines
    // into too few rows is structurally undercounted. Prefer a clearly
    // wider clean heuristic if one exists; otherwise force OCR rather than
    // serving a tidy-looking fragment.
    if first.issue == Some(TableCandidateIssue::LineRowUndercount) {
        return candidates.iter().find(|candidate| {
            matches!(
                candidate.source,
                TableCandidateSource::Heuristic
                    | TableCandidateSource::Column
                    | TableCandidateSource::KeyValue
            ) && candidate.issue.is_none()
                && candidate.shape.cols * 10 >= first.shape.cols * 13
        });
    }

    let mut accepted = candidates
        .iter()
        .find(|candidate| candidate.issue.is_none())?;

    // Keep the vector-first behavior from PR #85 unless the text heuristic
    // is also clean and has substantially more structure. This catches
    // vector grids that pass text quality checks while missing implicit rows
    // or sparse columns, without swapping on small shape noise.
    if matches!(
        accepted.source,
        TableCandidateSource::Rect | TableCandidateSource::Line
    ) {
        if let Some(heuristic) = candidates.iter().find(|candidate| {
            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)
}

fn heuristic_substantially_better(
    heuristic: MarkdownTableShape,
    accepted: MarkdownTableShape,
) -> bool {
    (accepted.rows > 0 && heuristic.rows * 2 >= accepted.rows * 3)
        || (accepted.cols > 0 && heuristic.cols * 10 >= accepted.cols * 13)
}

fn markdown_table_shape(markdown: &str) -> MarkdownTableShape {
    let mut shape = MarkdownTableShape::default();
    for cells in markdown_pipe_rows(markdown) {
        shape.rows += 1;
        shape.raw_cols = shape.raw_cols.max(cells.len());
        shape.cols = shape
            .cols
            .max(cells.iter().filter(|cell| !cell.trim().is_empty()).count());
    }
    shape
}

fn markdown_pipe_rows(markdown: &str) -> Vec<Vec<&str>> {
    markdown
        .lines()
        .filter_map(|line| {
            let trimmed = line.trim();
            if !trimmed.starts_with('|') || !trimmed.ends_with('|') {
                return None;
            }
            if is_markdown_separator_row(trimmed) {
                return None;
            }
            let parts: Vec<&str> = trimmed.split('|').collect();
            if parts.len() < 3 {
                return None;
            }
            Some(parts[1..parts.len() - 1].to_vec())
        })
        .collect()
}

fn is_markdown_separator_row(line: &str) -> bool {
    let mut saw_dash = false;
    for ch in line.chars() {
        match ch {
            '-' => saw_dash = true,
            '|' | ':' | ' ' => {}
            _ => return false,
        }
    }
    saw_dash
}

fn line_table_collapses_text_rows(
    table: &tables::Table,
    items: &[TextItem],
    shape: MarkdownTableShape,
) -> bool {
    let table_rows = shape.rows.max(
        table
            .cells
            .iter()
            .filter(|row| row.iter().any(|cell| !cell.trim().is_empty()))
            .count(),
    );
    let table_cols = shape.raw_cols.max(shape.cols);
    if !(2..=4).contains(&table_rows) || table_cols < 3 {
        return false;
    }

    let captured_items: Vec<&TextItem> = table
        .item_indices
        .iter()
        .filter_map(|&idx| items.get(idx))
        .filter(|item| !item.text.trim().is_empty())
        .collect();
    let captured_chars: usize = captured_items
        .iter()
        .map(|item| item.text.chars().count())
        .sum();
    if captured_chars <= 200 {
        return false;
    }

    let implicit_rows = y_cluster_count(&captured_items);
    implicit_rows >= 4 && table_rows * 2 <= implicit_rows
}

fn y_cluster_count(items: &[&TextItem]) -> usize {
    if items.is_empty() {
        return 0;
    }
    let mut ys: Vec<f32> = items.iter().map(|item| item.y).collect();
    ys.sort_by(|a, b| a.total_cmp(b));

    let mut clusters = 1;
    let mut center = ys[0];
    let mut count = 1usize;
    for &y in &ys[1..] {
        if (y - center).abs() > 3.0 {
            clusters += 1;
            center = y;
            count = 1;
        } else {
            center = (center * count as f32 + y) / (count as f32 + 1.0);
            count += 1;
        }
    }
    clusters
}

fn has_vertical_rules(lines: &[PdfLine]) -> bool {
    let angle_tolerance = 2.0_f32.to_radians().tan();
    lines
        .iter()
        .filter(|line| {
            let dx = (line.x2 - line.x1).abs();
            let dy = (line.y2 - line.y1).abs();
            let length = (dx * dx + dy * dy).sqrt();
            length >= 20.0 && dy > 0.01 && dx / dy <= angle_tolerance
        })
        .count()
        >= 2
}

fn wide_table_sparse_prefix_undercount(markdown: &str) -> bool {
    let rows = markdown_pipe_rows(markdown);
    if rows.len() < 4 {
        return false;
    }
    let header = &rows[0];
    let raw_cols = header.len();
    if raw_cols < 8 {
        return false;
    }

    let empty_headers: Vec<usize> = header
        .iter()
        .enumerate()
        .filter_map(|(idx, cell)| cell.trim().is_empty().then_some(idx))
        .collect();
    if empty_headers.len() != 1 {
        return false;
    }
    let empty_header_idx = empty_headers[0];
    if empty_header_idx == 0 || empty_header_idx >= raw_cols / 2 {
        return false;
    }

    let prefix_end = (raw_cols / 2).max(empty_header_idx + 1).min(raw_cols);
    if prefix_end <= 2 {
        return false;
    }

    let mut data_rows = 0usize;
    let mut sparse_prefix_rows = 0usize;
    for row in rows.iter().skip(1) {
        if row.iter().all(|cell| cell.trim().is_empty()) {
            continue;
        }
        data_rows += 1;
        let empty_prefix_cells = row
            .iter()
            .skip(1)
            .take(prefix_end.saturating_sub(1))
            .filter(|cell| cell.trim().is_empty())
            .count();
        if empty_prefix_cells >= 2 {
            sparse_prefix_rows += 1;
        }
    }

    data_rows >= 3 && sparse_prefix_rows * 2 >= data_rows
}

fn text_cluster_column_undercount(items: &[TextItem], shape: MarkdownTableShape) -> bool {
    let table_cols = shape.raw_cols.max(shape.cols);
    if table_cols < 2 || items.len() < table_cols * 2 {
        return false;
    }

    // Count "significant" x-clusters — clusters whose item count is at
    // least 1/4 of the dominant cluster. Filters out within-cell text
    // variation (wrapped continuations, bullet starts, indents) that
    // produces many small x-clusters not corresponding to real columns.
    let cluster_counts = x_cluster_item_counts(items);
    let Some(&dominant) = cluster_counts.iter().max() else {
        return false;
    };
    let min_cluster_size = (dominant / 4).max(2);
    let significant_clusters: usize = cluster_counts
        .iter()
        .filter(|&&n| n >= min_cluster_size)
        .count();

    // Two regimes:
    //   - Wide table undercount (e.g. 11-col dropped to 9): fire when the
    //     clustering surfaces ≥2 extra columns AND ≥1.2× the markdown
    //     column count. Inherits the old wide-table heuristic.
    //   - Narrow table undercount (e.g. 4-col dropped to 2): fire when
    //     the clustering surfaces ≥2× the markdown column count and at
    //     least 3 real columns. Catches narrow numeric columns
    //     (dates, amounts, IDs) collapsed into adjacent ones.
    let wide_undercount = table_cols >= 6
        && significant_clusters >= table_cols + 2
        && significant_clusters * 10 >= table_cols * 12;
    let narrow_undercount = significant_clusters >= 3 && significant_clusters >= table_cols * 2;
    wide_undercount || narrow_undercount
}

/// Cluster x-positions of non-empty text items with 8pt tolerance and
/// return the item count of each cluster. Same clustering policy as
/// `x_cluster_count`; the array form lets the caller weight clusters
/// by population to filter out single-item outliers.
fn x_cluster_item_counts(items: &[TextItem]) -> Vec<usize> {
    let mut xs: Vec<f32> = items
        .iter()
        .filter(|item| !item.text.trim().is_empty())
        .map(|item| item.x)
        .collect();
    if xs.is_empty() {
        return Vec::new();
    }
    xs.sort_by(|a, b| a.total_cmp(b));

    let mut counts: Vec<usize> = Vec::new();
    let mut center = xs[0];
    let mut count = 1usize;
    for &x in &xs[1..] {
        if (x - center).abs() > 8.0 {
            counts.push(count);
            center = x;
            count = 1;
        } else {
            center = (center * count as f32 + x) / (count as f32 + 1.0);
            count += 1;
        }
    }
    counts.push(count);
    counts
}

fn prose_grid_fragment_needs_ocr(markdown: &str) -> bool {
    let rows = markdown_pipe_rows(markdown);
    if rows.len() < 2 {
        return false;
    }
    let raw_cols = rows[0].len();
    if !(2..=4).contains(&raw_cols) {
        return false;
    }

    let mut seen_by_col = vec![0usize; raw_cols];
    let mut compact_by_col = vec![0usize; raw_cols];
    let mut long_prose = 0usize;
    let mut total = 0usize;

    for row in rows.iter().skip(1) {
        for (col, cell) in row.iter().take(raw_cols).enumerate() {
            let trimmed = cell.trim();
            if trimmed.is_empty() {
                continue;
            }
            total += 1;
            seen_by_col[col] += 1;
            if compact_identifier_cell(trimmed) {
                compact_by_col[col] += 1;
            }
            if long_prose_cell(trimmed) {
                long_prose += 1;
            }
        }
    }

    let min_total = if raw_cols == 2 { 2 } else { raw_cols * 2 };
    if total < min_total {
        return false;
    }
    if long_prose * 3 < total * 2 {
        return false;
    }

    !seen_by_col
        .iter()
        .zip(compact_by_col.iter())
        .any(|(&seen, &compact)| seen >= 3 && compact * 2 >= seen)
}

fn compact_identifier_cell(cell: &str) -> bool {
    let trimmed = cell.trim();
    if trimmed.is_empty() {
        return false;
    }
    let words = word_count(trimmed);
    if words <= 3 && trimmed.chars().count() <= 40 {
        return true;
    }

    let chars = trimmed.chars().filter(|c| !c.is_whitespace()).count();
    if chars == 0 || chars > 48 {
        return false;
    }
    let compact_marks = trimmed
        .chars()
        .filter(|c| {
            c.is_ascii_digit()
                || matches!(c, '.' | ',' | ':' | ';' | '/' | '-' | '(' | ')' | '[' | ']')
        })
        .count();
    compact_marks * 2 >= chars
}

fn long_prose_cell(cell: &str) -> bool {
    let trimmed = cell.trim();
    if compact_identifier_cell(trimmed) {
        return false;
    }
    let words = word_count(trimmed);
    let alpha = trimmed.chars().filter(|c| c.is_alphabetic()).count();
    words >= 4 && alpha >= 12
}

fn word_count(text: &str) -> usize {
    text.split_whitespace()
        .filter(|word| word.chars().any(|c| c.is_alphanumeric()))
        .count()
}

fn starts_with_numbered_table_label(cell: &str) -> bool {
    let trimmed = cell.trim_start();
    let digit_count = trimmed.chars().take_while(|c| c.is_ascii_digit()).count();

    digit_count > 0
        && digit_count <= 3
        && trimmed
            .chars()
            .nth(digit_count)
            .is_some_and(|c| matches!(c, '.' | ')' | '-' | ':'))
}

fn starts_with_uppercase_alpha(cell: &str) -> bool {
    cell.chars()
        .find(|c| c.is_alphabetic())
        .is_some_and(|c| c.is_uppercase())
}

fn compact_title_like_cell(cell: &str) -> bool {
    let trimmed = cell.trim();
    if trimmed.len() < 3 || trimmed.len() > 80 || !starts_with_uppercase_alpha(trimmed) {
        return false;
    }
    let words = trimmed
        .split_whitespace()
        .filter(|word| word.chars().any(|c| c.is_alphabetic()))
        .count();
    (1..=6).contains(&words)
}

fn numbered_rowspan_hierarchy_needs_ocr(markdown: &str) -> bool {
    let rows = markdown_pipe_rows(markdown);
    if rows.len() < 6 {
        return false;
    }

    let n_cols = rows[0].len();
    if !(3..=6).contains(&n_cols) {
        return false;
    }

    let data_rows = &rows[1..];
    let numbered_group_rows = data_rows
        .iter()
        .filter(|row| {
            row.first()
                .is_some_and(|cell| starts_with_numbered_table_label(cell))
        })
        .count();
    let blank_first_subrows = data_rows
        .iter()
        .filter(|row| {
            row.first().is_some_and(|cell| cell.trim().is_empty())
                && row.get(1).is_some_and(|cell| compact_title_like_cell(cell))
        })
        .count();

    // Native Markdown cannot express rowspans.  In numbered hierarchical
    // tables, repeated blank first-column sub-rows are a strong signal that
    // wrapped content from sibling rows can be assigned to the wrong row even
    // after continuation merging is conservative.  Let OCR handle this shape
    // instead of serving a tidy-looking but lossy native table.
    numbered_group_rows >= 2 && blank_first_subrows >= 2
}

fn looks_like_partial_table_ex(markdown: &str, layout_assisted: bool) -> bool {
    let lines: Vec<&str> = markdown.lines().filter(|l| l.starts_with('|')).collect();
    if lines.len() < 2 {
        return false;
    }
    // Header is the first pipe-line; separator is the second
    let header_line = lines[0];
    let separator_line = lines.get(1).copied().unwrap_or("");
    let is_separator = |l: &str| l.chars().all(|c| matches!(c, '|' | '-' | ' '));
    if !is_separator(separator_line) {
        // No separator after the first line — not a well-formed pipe-table.
        // table_to_markdown always emits one when it returns content, so this
        // shouldn't happen in practice. If it does, fall through to OCR.
        return true;
    }

    // Parse header cells: split on '|', drop the leading/trailing empty pieces
    let cells: Vec<&str> = header_line.split('|').map(|s| s.trim()).collect::<Vec<_>>();
    // The first and last items are always empty (string starts and ends with '|')
    if cells.len() < 3 {
        return false;
    }
    let header_cells: Vec<&str> = cells[1..cells.len() - 1].to_vec();
    let n_cols = header_cells.len();
    if n_cols < 2 {
        // Single-column tables are usually lists/keys, not tables. Keep them
        // (caller can decide), but multi-column header checks below don't
        // apply.
        return false;
    }

    if layout_assisted && numbered_rowspan_hierarchy_needs_ocr(markdown) {
        return true;
    }

    // Failure mode 1: header starts with a bare number (likely we missed
    // the real header row above). Skip when layout-assisted — the layout
    // model's bbox includes the real header; a numeric first cell (e.g.,
    // a year "2024") is legitimate.
    if !layout_assisted {
        if let Some(first) = header_cells.first() {
            let trimmed = first.trim();
            if !trimmed.is_empty() && trimmed.chars().all(|c| c.is_ascii_digit()) {
                return true;
            }
        }
    }

    // Failure mode 2: header has empty cells in a multi-column table.
    // 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 {
        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 {
        return true;
    }

    // Failure mode 3: header has duplicate non-empty cells
    let mut seen: std::collections::HashSet<&str> = std::collections::HashSet::new();
    for cell in &header_cells {
        if cell.is_empty() {
            continue;
        }
        if !seen.insert(cell) {
            return true;
        }
    }

    // Failure mode 4: first data row has many empty cells in a multi-column
    // table. Real tables rarely have a leading row with most cells blank;
    // when this happens it usually means the heuristic split a multi-row
    // header (e.g. "Position\nAdministration (1986-1992) | Administration
    // (1992-1998)") into a single-row header + a sparse data row.
    if let Some(first_data_line) = lines.get(2) {
        let data_cells: Vec<&str> = first_data_line
            .split('|')
            .map(|s| s.trim())
            .collect::<Vec<_>>();
        if data_cells.len() >= 3 {
            let data_inner = &data_cells[1..data_cells.len() - 1];
            let empty_data = data_inner.iter().filter(|c| c.is_empty()).count();
            // ≥3 cols, and significant portion of cells in the first data
            // row are empty → likely we mis-split a multi-row header.
            // When layout-assisted, relax from 33% to 50% — the bbox is
            // more reliable, and real tables with one sparse first row
            // (totals, subtotals) are common.
            let threshold = if layout_assisted { 2 } else { 3 };
            if n_cols >= 3 && empty_data * threshold >= n_cols {
                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;
                }
            }
        }
    }

    // Failure mode 5: cells flow as continuation paragraph (text wrapping
    // mistaken for column structure). When a paragraph of prose gets mis-
    // detected as a multi-column table, cells in the same column tend to
    // start with lowercase letters or punctuation (continuation), not
    // capital letters / digits (new entries). Real tables almost never
    // have most data cells starting lowercase.
    //
    // Signal: ≥2 cols, ≥4 data rows, and ≥60% of non-empty data cells
    // start with a lowercase letter or continuation punctuation.
    let data_rows: Vec<Vec<&str>> = lines
        .iter()
        .skip(2) // header + separator
        .map(|l| {
            let parts: Vec<&str> = l.split('|').map(|s| s.trim()).collect();
            if parts.len() >= 3 {
                parts[1..parts.len() - 1].to_vec()
            } else {
                Vec::new()
            }
        })
        .filter(|cells| !cells.is_empty())
        .collect();

    if n_cols >= 2 && data_rows.len() >= 4 {
        let mut continuation = 0;
        let mut total = 0;
        for row in &data_rows {
            for cell in row {
                let trimmed = cell.trim();
                if trimmed.is_empty() {
                    continue;
                }
                total += 1;
                let first = trimmed.chars().next().unwrap();
                // Continuation indicators: lowercase letter, common
                // mid-sentence punctuation, closing quote
                if first.is_lowercase()
                    || matches!(first, ',' | '.' | ';' | ')' | '"' | '\'' | '”' | '’')
                {
                    continuation += 1;
                }
            }
        }
        if total > 0 && continuation * 5 >= total * 3 {
            // ≥60% of cells look like sentence continuations → paragraph
            // misread as table.
            return true;
        }
    }

    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)]
fn looks_like_partial_table(markdown: &str) -> bool {
    looks_like_partial_table_ex(markdown, false)
}

#[cfg(test)]
mod text_cluster_column_undercount_tests {
    use super::{text_cluster_column_undercount, MarkdownTableShape, TextItem};
    use crate::types::ItemType;

    fn item(x: f32, y: f32, text: &str) -> TextItem {
        TextItem {
            text: text.to_string(),
            x,
            y,
            width: text.len() as f32 * 5.0,
            height: 10.0,
            font: "F".into(),
            font_tag: String::new(),
            legacy_symbol_rewrite: false,
            font_size: 10.0,
            page: 1,
            is_bold: false,
            is_italic: false,
            font_weight: None,
            bold_source: None,
            fixed_pitch: None,
            fill_color: None,
            stroke_color: None,
            render_mode: None,
            is_underline: false,
            is_strikeout: false,
            rotation: 0.0,
            advance_known: true,
            item_type: ItemType::Text,
            mcid: None,
            baseline_shift: 0.0,
        }
    }

    /// 22 rows × 4 columns of items, neutral synthetic content. Simulates
    /// the "narrow numeric columns dropped" shape: heuristic detector
    /// produced 2 markdown columns but the page geometry shows 4.
    fn make_four_column_grid() -> Vec<TextItem> {
        let xs = [60.0, 145.0, 330.0, 430.0];
        let mut items = Vec::new();
        for row in 0..22 {
            let y = 700.0 - row as f32 * 20.0;
            for &x in &xs {
                items.push(item(x, y, "x"));
            }
        }
        items
    }

    fn shape(cols: usize) -> MarkdownTableShape {
        MarkdownTableShape {
            rows: 22,
            cols,
            raw_cols: cols,
        }
    }

    #[test]
    fn narrow_undercount_2col_markdown_4col_geometry_fires() {
        // The shape this PR targets: heuristic detector collapsed a
        // 4-column page into 2 markdown columns because narrow numeric
        // columns (e.g. dates, IDs) didn't survive x-position
        // clustering. Significant clusters (4) >= markdown_cols (2) * 2.
        let items = make_four_column_grid();
        assert!(text_cluster_column_undercount(&items, shape(2)));
    }

    #[test]
    fn matching_geometry_does_not_fire() {
        // 4-col page with 4-col markdown — no undercount.
        let items = make_four_column_grid();
        assert!(!text_cluster_column_undercount(&items, shape(4)));
    }

    #[test]
    fn single_column_skipped() {
        // 1-col markdown is not a table — never flag.
        let items = make_four_column_grid();
        assert!(!text_cluster_column_undercount(&items, shape(1)));
    }

    #[test]
    fn insufficient_items_skipped() {
        // < 2*table_cols items — not enough signal to claim undercount.
        let items = vec![
            item(60.0, 700.0, "x"),
            item(145.0, 700.0, "x"),
            item(330.0, 700.0, "x"),
        ];
        assert!(!text_cluster_column_undercount(&items, shape(2)));
    }

    #[test]
    fn outlier_singletons_filtered_out() {
        // 2-col grid (22 items each col) plus a few stray outliers at
        // other x positions. Significant clusters = 2; markdown=2;
        // doesn't fire (significant >= cols*2 = 4 is false).
        let mut items = Vec::new();
        for row in 0..22 {
            let y = 700.0 - row as f32 * 20.0;
            items.push(item(60.0, y, "x"));
            items.push(item(330.0, y, "x"));
        }
        // Add a handful of strays (continuation fragments, footnotes).
        items.push(item(145.0, 100.0, "footnote"));
        items.push(item(430.0, 50.0, "page#"));
        assert!(!text_cluster_column_undercount(&items, shape(2)));
    }

    #[test]
    fn wide_table_path_still_fires() {
        // The original wide-table case: 6+ markdown cols with ≥2 extra
        // significant clusters. Builds a 12-col geometry with 8-col
        // markdown — the legacy wide-undercount path catches this.
        let mut items = Vec::new();
        let xs: Vec<f32> = (0..12).map(|i| 50.0 + i as f32 * 40.0).collect();
        for row in 0..22 {
            let y = 700.0 - row as f32 * 20.0;
            for &x in &xs {
                items.push(item(x, y, "x"));
            }
        }
        assert!(text_cluster_column_undercount(&items, shape(8)));
    }
}

#[cfg(test)]
mod region_text_density_tests {
    use super::region_text_density_too_low;

    #[test]
    fn small_region_skips_check() {
        // Tiny stat blocks below the area floor are never flagged —
        // can't distinguish "font failure" from "small legitimate table".
        // 100×100 = 10,000 sq pt, below the 30,000 floor.
        assert!(!region_text_density_too_low(5, 10_000.0));
    }

    #[test]
    fn dense_full_table_passes() {
        // Observed clean extractions sit at 0.005–0.015 chars/sq pt.
        // Full A4 ledger: 438,000 sq pt with 6,500 chars → density 0.015.
        assert!(!region_text_density_too_low(6_500, 438_000.0));
    }

    #[test]
    fn moderate_density_table_passes() {
        // Key/value tables with multi-line cells sit at ~0.005 chars/sq pt.
        // 291,000 sq pt with 1,600 chars → density 0.0055.
        assert!(!region_text_density_too_low(1_600, 291_000.0));
    }

    #[test]
    fn font_decode_failure_caught() {
        // Big region, almost no extractable text — page extractor hit a
        // CMap failure. 102,000 sq pt with 46 chars → density 0.0005.
        assert!(region_text_density_too_low(46, 102_000.0));
    }

    #[test]
    fn sparse_glyph_repeat_caught() {
        // Full-page Cyrillic where every glyph decoded to "Т". The text
        // extractor returned a few hundred chars, but they're all the
        // same letter. Density 0.00025 — well under the floor.
        assert!(region_text_density_too_low(89, 353_000.0));
    }

    #[test]
    fn sparse_layout_band_caught() {
        // Specimen-materials horizontal band: real text decoded fine for
        // the slice that's in-bbox, but the table extends below the
        // bbox. 73,000 sq pt with 96 chars → density 0.0013.
        assert!(region_text_density_too_low(96, 73_000.0));
    }

    #[test]
    fn boundary_at_density_floor() {
        // Right at the 0.003 floor: 90 chars / 30,000 sq pt = 0.003
        // exactly. The check rejects when density is strictly less than
        // the floor, so the boundary is treated as acceptable.
        assert!(!region_text_density_too_low(90, 30_000.0));
        // Just under: 89 chars / 30,000 = 0.00297 — flagged.
        assert!(region_text_density_too_low(89, 30_000.0));
    }

    #[test]
    fn whole_page_bbox_skips_check() {
        // Near-whole-A4 bboxes (>400,000 sq pt) include large white-
        // space margins, so text density is unreliable. Layout
        // typically produces tight per-table bboxes; whole-page
        // bboxes appear in fixtures and edge cases where this signal
        // would misfire. 612×792 = 484,704 sq pt with 423 chars in
        // a real table inside it.
        assert!(!region_text_density_too_low(423, 484_704.0));
    }

    #[test]
    fn tiny_text_chars_skips_check() {
        // Synthetic / fragmentary fixtures with <20 chars in a
        // generously-sized bbox aren't font-decode failures — they're
        // unit-test artifacts. Real prod failures decode ≥30 chars.
        // 8 chars in a 127,800 sq pt bbox would otherwise flag at
        // density 0.00006.
        assert!(!region_text_density_too_low(8, 127_800.0));
    }
}

#[cfg(test)]
mod captured_only_a_fragment_tests {
    use super::captured_only_a_fragment;

    #[test]
    fn small_region_skips_check() {
        // Short legitimate tables (axis labels, unit blocks) shouldn't be
        // flagged even when the captured markdown is tiny.
        let md = "|Year|Value|\n|---|---|\n|2024|10|";
        assert!(!captured_only_a_fragment(md, 50));
    }

    #[test]
    fn full_table_passes() {
        // Captured markdown matches the region text — full extraction.
        let md =
            "|Name|Year|Country|\n|---|---|---|\n|Alice|2020|US|\n|Bob|2021|UK|\n|Carol|2019|FR|";
        // Region had ~50 chars of text (rough estimate of just the data words).
        assert!(!captured_only_a_fragment(md, 50));
        // Even a much larger region matched by the markdown content passes.
        assert!(!captured_only_a_fragment(md, md.len()));
    }

    #[test]
    fn header_only_extraction_rejected() {
        // Captured the column-header band (~30 chars) while the region
        // actually has many rows of data (~1500 chars).
        let md = "|Description|Year|Amount|\n|---|---|---|";
        assert!(captured_only_a_fragment(md, 1500));
    }

    #[test]
    fn sparse_fragment_rejected() {
        // A couple of fragment cells captured from a content-rich region.
        let md = "|percent|for|\n|---|---|\n|sites|15|";
        assert!(captured_only_a_fragment(md, 2000));
    }

    #[test]
    fn boundary_at_25_percent_floor() {
        // Right at the 25% line: 250 captured chars of 1000 region chars.
        // The check rejects when captured*4 < region, so 250*4=1000 is NOT
        // less than 1000 — boundary is treated as acceptable.
        let md = "x".repeat(250);
        assert!(!captured_only_a_fragment(&md, 1000));
        // Just under 25%: 249*4=996 < 1000 — flagged.
        let md_under = "x".repeat(249);
        assert!(captured_only_a_fragment(&md_under, 1000));
    }
}

#[cfg(test)]
mod table_candidate_selection_tests {
    use super::{
        line_table_collapses_text_rows, markdown_table_shape, prose_grid_fragment_needs_ocr,
        select_table_candidate, text_cluster_column_undercount,
        wide_table_sparse_prefix_undercount, MarkdownTableShape, TableCandidate,
        TableCandidateIssue, TableCandidateSource,
    };
    use crate::tables::Table;
    use crate::types::{ItemType, TextItem};

    fn candidate(
        source: TableCandidateSource,
        rows: usize,
        cols: usize,
        issue: Option<TableCandidateIssue>,
    ) -> TableCandidate {
        TableCandidate {
            markdown: format!("{source:?}-{rows}x{cols}"),
            source,
            shape: MarkdownTableShape {
                rows,
                cols,
                raw_cols: cols,
            },
            issue,
        }
    }

    fn item(text: &str, y: f32) -> TextItem {
        item_at(text, 10.0, y)
    }

    fn item_at(text: &str, x: f32, y: f32) -> TextItem {
        TextItem {
            text: text.to_string(),
            x,
            y,
            width: 50.0,
            height: 10.0,
            font: "F1".to_string(),
            font_tag: String::new(),
            legacy_symbol_rewrite: false,
            font_size: 10.0,
            page: 1,
            is_bold: false,
            is_italic: false,
            font_weight: None,
            bold_source: None,
            fixed_pitch: None,
            fill_color: None,
            stroke_color: None,
            render_mode: None,
            is_underline: false,
            is_strikeout: false,
            rotation: 0.0,
            advance_known: true,
            item_type: ItemType::Text,
            mcid: None,
            baseline_shift: 0.0,
        }
    }

    #[test]
    fn markdown_shape_counts_rows_and_non_empty_columns() {
        let md = "|A|B||D|\n|---|---|---|---|\n|one|two|three||";
        let shape = markdown_table_shape(md);
        assert_eq!(shape.rows, 2);
        assert_eq!(shape.cols, 3);
        assert_eq!(shape.raw_cols, 4);
    }

    #[test]
    fn prefers_clean_heuristic_when_substantially_wider_than_vector() {
        let candidates = vec![
            candidate(TableCandidateSource::Rect, 5, 3, None),
            candidate(TableCandidateSource::Heuristic, 5, 4, None),
        ];
        let selected = select_table_candidate(&candidates).unwrap();
        assert_eq!(selected.source, TableCandidateSource::Heuristic);
    }

    #[test]
    fn line_row_undercount_routes_to_ocr_without_wider_heuristic() {
        let candidates = vec![
            candidate(
                TableCandidateSource::Line,
                2,
                3,
                Some(TableCandidateIssue::LineRowUndercount),
            ),
            candidate(TableCandidateSource::Heuristic, 4, 3, None),
        ];
        assert!(select_table_candidate(&candidates).is_none());
    }

    #[test]
    fn line_row_undercount_can_use_clearly_wider_heuristic() {
        let candidates = vec![
            candidate(
                TableCandidateSource::Line,
                2,
                3,
                Some(TableCandidateIssue::LineRowUndercount),
            ),
            candidate(TableCandidateSource::Heuristic, 2, 4, None),
        ];
        let selected = select_table_candidate(&candidates).unwrap();
        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";
        let items = vec![
            item(long, 100.0),
            item(long, 112.0),
            item(long, 124.0),
            item(long, 136.0),
        ];
        let table = Table::new(
            vec![0.0, 100.0, 200.0, 300.0],
            vec![150.0, 120.0],
            vec![
                vec!["A".to_string(), "B".to_string(), "C".to_string()],
                vec!["D".to_string(), "E".to_string(), "F".to_string()],
            ],
            vec![0, 1, 2, 3],
        );
        let shape = MarkdownTableShape {
            rows: 2,
            cols: 3,
            raw_cols: 3,
        };
        assert!(line_table_collapses_text_rows(&table, &items, shape));
    }

    #[test]
    fn wide_sparse_prefix_with_blank_header_is_suspicious() {
        let md = "|Name|Flag A|Flag B||Metric A|Metric B|Metric C|Metric D|\n\
                  |---|---|---|---|---|---|---|---|\n\
                  |Row 1|Y|||1|2|3|4|\n\
                  |Row 2|||N|5|6|7|8|\n\
                  |Row 3||||9|10|11|12|";
        assert!(wide_table_sparse_prefix_undercount(md));
    }

    #[test]
    fn wide_blank_header_with_dense_body_is_allowed() {
        let md = "|Name|Flag A|Flag B||Metric A|Metric B|Metric C|Metric D|\n\
                  |---|---|---|---|---|---|---|---|\n\
                  |Row 1|Y|N|Y|1|2|3|4|\n\
                  |Row 2|N|Y|N|5|6|7|8|\n\
                  |Row 3|Y|Y|N|9|10|11|12|";
        assert!(!wide_table_sparse_prefix_undercount(md));
    }

    #[test]
    fn x_clusters_can_signal_wide_column_undercount() {
        let mut items = Vec::new();
        for y in [100.0, 112.0, 124.0] {
            for x in [
                10.0, 45.0, 80.0, 115.0, 150.0, 185.0, 220.0, 255.0, 290.0, 325.0,
            ] {
                items.push(item_at("value", x, y));
            }
        }
        let shape = MarkdownTableShape {
            rows: 4,
            cols: 8,
            raw_cols: 8,
        };
        assert!(text_cluster_column_undercount(&items, shape));
    }

    #[test]
    fn wrapped_prose_grid_fragment_is_suspicious() {
        let md = "|A useful capability with several words|Another descriptive capability column|A final descriptive capability column|\n\
                  |---|---|---|\n\
                  |The group includes experienced specialists|received a strong neutral recommendation|system in a neutral evaluation setting|\n\
                  |presented several papers in public venues|shown stronger performance than alternatives|ranked highly in a neutral benchmark|\n\
                  |recognized by external reviewers|delivered useful results for operators|used in production style workflows|";
        assert!(prose_grid_fragment_needs_ocr(md));
    }

    #[test]
    fn compact_identifier_columns_are_not_prose_fragments() {
        let md = "|Box 1, F-7|Description|Date|\n\
                  |---|---|---|\n\
                  |Box 1, F-8|Long neutral description with several words for one record|2020 Jan 1|\n\
                  |Box 1, F-9|Another neutral description with several words for another record|2021 Feb 2|\n\
                  |Box 1, F-10|Additional neutral description with several words for a record|n.d.|";
        assert!(!prose_grid_fragment_needs_ocr(md));
    }

    #[test]
    fn two_column_all_prose_fragment_is_suspicious() {
        let md = "|A long descriptive header fragment|Another long descriptive header fragment|\n\
                  |---|---|\n\
                  |Long wrapped prose content from one visual cell|More wrapped prose content from a neighboring visual cell|";
        assert!(prose_grid_fragment_needs_ocr(md));
    }

    #[test]
    fn two_column_key_value_table_is_allowed() {
        let md = "|Field|Detail|\n\
                  |---|---|\n\
                  |Status|Long neutral explanation with several words for this value|\n\
                  |Owner|Another neutral explanation with several words for this value|";
        assert!(!prose_grid_fragment_needs_ocr(md));
    }
}

#[cfg(test)]
mod looks_like_partial_table_tests {
    use super::{
        looks_like_partial_table, looks_like_partial_table_ex, markdown_table_body_is_dense,
    };

    #[test]
    fn good_table_passes() {
        let md = "|Name|Year|Country|\n|---|---|---|\n|Alice|2020|US|\n|Bob|2021|UK|";
        assert!(
            !looks_like_partial_table(md),
            "should not flag well-formed table"
        );
    }

    #[test]
    fn header_starting_with_number_is_partial() {
        // Heuristic missed the actual header row above
        let md = "|2|Cambodian Women for Peace|9,835|\n|---|---|---|\n|3|Association|711|";
        assert!(looks_like_partial_table(md));
    }

    #[test]
    fn header_with_empty_cells_in_3col_is_partial() {
        // Empty cell in 3+ column header → bad column detection
        let md =
            "|Position||Administration|Administration|\n|---|---|---|---|\n|Senate|24|8.3|16.7|";
        assert!(looks_like_partial_table(md));
    }

    #[test]
    fn header_with_duplicate_cells_is_partial() {
        // Duplicate "Administration" → collapsed multi-line header wrong
        let md =
            "|Position|Administration|Administration|Notes|\n|---|---|---|---|\n|Senate|24|16|x|";
        assert!(looks_like_partial_table(md));
    }

    #[test]
    fn two_column_with_one_empty_cell_passes() {
        // Many real two-column tables have key-only rows; don't penalise.
        let md = "|Key||\n|---|---|\n|Alice|123|\n|Bob|456|";
        // Header "Key|" has one empty cell but only 2 cols total — keep it.
        assert!(!looks_like_partial_table(md));
    }

    #[test]
    fn single_column_table_is_kept() {
        // Single-column "tables" are common (lists). Caller can decide; we
        // don't second-guess based on column count alone.
        let md = "|Item|\n|---|\n|First|\n|Second|";
        assert!(!looks_like_partial_table(md));
    }

    #[test]
    fn no_table_at_all_returns_true() {
        // table_to_markdown should never produce this, but defensive — if
        // there's no separator, treat as not-a-table.
        let md = "Just some text\nWith multiple lines";
        // No lines start with '|' so we return false (no header to inspect).
        assert!(!looks_like_partial_table(md));
    }

    #[test]
    fn first_data_row_with_many_empty_cells_is_partial() {
        // Multi-row header collapsed to single-row → first "data row" has
        // most cells empty (the actual sub-header values).
        let md = "|Government|No. of Seats|Aquino|Ramos|\n|---|---|---|---|\n|Position|||(1986-1992)|\n|Senate|24|8.3|16.7|";
        assert!(looks_like_partial_table(md));
    }

    #[test]
    fn first_data_row_with_one_empty_cell_in_4col_passes() {
        // Real data rows can have one empty cell (e.g. missing value);
        // only flag when ≥1/3 of cells are empty.
        let md = "|A|B|C|D|\n|---|---|---|---|\n|x|y||z|\n|p|q|r|s|";
        assert!(!looks_like_partial_table(md));
    }

    #[test]
    fn paragraph_misread_as_two_column_table_is_partial() {
        // Real production failure: text-wrapped paragraph mis-detected as
        // 2-col table. Each cell continues the previous one as prose.
        let md = "|Approval is needed from the|Acquisitions of|\n\
                  |---|---|\n\
                  |Treasurer if the acquisition|residential and|\n\
                  |constitutes a \"significant|agricultural|\n\
                  |action,\" including acquiring an|land by foreign|\n\
                  |interest in different types of|persons must be|\n\
                  |land where the monetary|reported to the|";
        assert!(looks_like_partial_table(md));
    }

    #[test]
    fn real_multi_word_table_is_kept() {
        // Real table with multi-word entries — cells start with capital
        // letters / proper nouns, NOT lowercase continuations.
        let md = "|Country|Capital|Notes|\n\
                  |---|---|---|\n\
                  |United States|Washington DC|Federal capital|\n\
                  |United Kingdom|London|City of London is a separate|\n\
                  |France|Paris|Île-de-France region|\n\
                  |Germany|Berlin|Reunified 1990|\n\
                  |Spain|Madrid|Largest city in Spain|";
        assert!(!looks_like_partial_table(md));
    }

    // --- layout_assisted relaxation tests ---

    #[test]
    fn numeric_header_accepted_when_layout_assisted() {
        // Year as first header cell is valid when layout model gave us the bbox.
        let md = "|2024|Revenue|Growth|\n|---|---|---|\n|Q1|1.2M|5%|\n|Q2|1.4M|8%|";
        assert!(
            looks_like_partial_table(md),
            "strict mode rejects numeric header"
        );
        assert!(
            !looks_like_partial_table_ex(md, true),
            "layout-assisted should accept"
        );
    }

    #[test]
    fn one_empty_header_accepted_when_layout_assisted() {
        // Common in merged-header tables: one spanning cell leaves a gap.
        let md = "|Position||Senate|House|\n|---|---|---|---|\n|Chair|1|2|3|\n|Vice|4|5|6|";
        assert!(
            looks_like_partial_table(md),
            "strict rejects 1 empty header"
        );
        assert!(
            !looks_like_partial_table_ex(md, true),
            "layout-assisted allows 1 empty"
        );
    }

    #[test]
    fn two_empty_headers_still_rejected_when_layout_assisted() {
        // 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 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"
        );
    }

    #[test]
    fn sparse_first_row_relaxed_when_layout_assisted() {
        // 1/4 empty = 25%, below strict 33% threshold but accepted by layout-assisted 50%.
        let md = "|A|B|C|D|\n|---|---|---|---|\n|x||y|z|\n|p|q|r|s|";
        assert!(!looks_like_partial_table(md), "strict: 25% empty is OK");
        // 2/4 = 50%, strict would flag (2*3>=4), relaxed threshold (2*2>=4) would also flag.
        let md2 = "|A|B|C|D|\n|---|---|---|---|\n|||y|z|\n|p|q|r|s|";
        assert!(looks_like_partial_table(md2), "strict: 50% empty flagged");
        assert!(
            looks_like_partial_table_ex(md2, true),
            "layout-assisted: 50% also flagged"
        );
        // 2/6 = 33%, strict flags (2*3>=6), relaxed does not (2*2<6)
        let md3 = "|A|B|C|D|E|F|\n|---|---|---|---|---|---|\n|x|||y|z|w|\n|a|b|c|d|e|f|";
        assert!(looks_like_partial_table(md3), "strict: 33% flagged");
        assert!(
            !looks_like_partial_table_ex(md3, true),
            "layout-assisted: 33% accepted"
        );
    }

    #[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.
        let md = "|Approval is needed from the|Acquisitions of|\n\
                  |---|---|\n\
                  |Treasurer if the acquisition|residential and|\n\
                  |constitutes a \"significant|agricultural|\n\
                  |action,\" including acquiring an|land by foreign|\n\
                  |interest in different types of|persons must be|\n\
                  |land where the monetary|reported to the|";
        assert!(
            looks_like_partial_table_ex(md, true),
            "paragraph rejection stays strict"
        );
    }

    #[test]
    fn numbered_rowspan_hierarchy_rejected_when_layout_assisted() {
        let md = "|Group|Task|Detail|Benefit|\n\
                  |---|---|---|---|\n\
                  |1. Group alpha|Task setup|Begin setup|Faster start|\n\
                  |2. Group beta|Storage setup|Provides tools||\n\
                  ||Label workspace|Creates review sets|Lets teams review|\n\
                  ||Model training|Builds model|Supports rollout|\n\
                  |3. Group gamma|Pipeline setup|Configures flow|Improves control|";
        assert!(
            looks_like_partial_table_ex(md, true),
            "layout-assisted native extraction should fall back for numbered row-spanned hierarchies"
        );
    }

    #[test]
    fn plain_numbered_table_kept_when_layout_assisted() {
        let md = "|Step|Task|Detail|Benefit|\n\
                  |---|---|---|---|\n\
                  |1. Group alpha|Task setup|Begin setup|Faster start|\n\
                  |2. Group beta|Storage setup|Provides tools|Easier review|\n\
                  |3. Group gamma|Pipeline setup|Configures flow|Improves control|";
        assert!(
            !looks_like_partial_table_ex(md, true),
            "numbered rows alone are fine; the guard requires blank-first-column sub-rows"
        );
    }

    #[test]
    fn duplicate_headers_still_rejected_when_layout_assisted() {
        let md =
            "|Position|Administration|Administration|Notes|\n|---|---|---|---|\n|Senate|24|16|x|";
        assert!(
            looks_like_partial_table_ex(md, true),
            "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));
    }
}

struct FolioFilteredItems {
    items: Vec<types::TextItem>,
    layout_items: Vec<types::TextItem>,
    removal_mask: Vec<bool>,
    removed_pages: HashSet<u32>,
}

/// Resolve folios with complete document context, then select the caller's
/// requested pages without losing those decisions.
fn select_items_with_document_folio_context(
    all_items: Vec<types::TextItem>,
    page_count: u32,
    page_filter: Option<&HashSet<u32>>,
) -> FolioFilteredItems {
    let (all_layout_items, all_removed_pages, all_removal_mask) =
        extractor::filter_markdown_page_numbers_with_removed_pages(all_items.clone(), page_count);
    let selected_page = |page: u32| page_filter.is_none_or(|filter| filter.contains(&page));

    let (items, removal_mask) = all_items
        .into_iter()
        .zip(all_removal_mask)
        .filter(|(item, _)| selected_page(item.page))
        .unzip();
    let layout_items = all_layout_items
        .into_iter()
        .filter(|item| selected_page(item.page))
        .collect();
    let removed_pages = all_removed_pages
        .into_iter()
        .filter(|page| selected_page(*page))
        .collect();

    FolioFilteredItems {
        items,
        layout_items,
        removal_mask,
        removed_pages,
    }
}

/// Analyse extracted items and rects for layout complexity.
#[cfg(test)]
fn compute_layout_complexity(
    items: &[types::TextItem],
    column_items: &[types::TextItem],
    rects: &[types::PdfRect],
    lines: &[types::PdfLine],
) -> LayoutComplexity {
    let page_chart_regions = markdown::chart_regions_by_page(items, rects, lines);
    compute_layout_complexity_with_chart_regions(
        items,
        column_items,
        rects,
        lines,
        &page_chart_regions,
    )
}

fn compute_layout_complexity_with_chart_regions(
    items: &[types::TextItem],
    column_items: &[types::TextItem],
    rects: &[types::PdfRect],
    lines: &[types::PdfLine],
    page_chart_regions: &markdown::PageChartRegions,
) -> LayoutComplexity {
    use markdown::analysis::calculate_font_stats_from_items;

    // --- Collect unique pages ---
    let mut seen_pages: Vec<u32> = items.iter().map(|i| i.page).collect();
    seen_pages.sort();
    seen_pages.dedup();

    let font_stats = calculate_font_stats_from_items(items);
    let base_size = font_stats.most_common_size;

    // --- Tables: use rect-based → line-based → heuristic detectors per page,
    //     with side-by-side band splitting ---
    let mut pages_with_tables: Vec<u32> = Vec::new();
    for &page in &seen_pages {
        let page_items: Vec<&types::TextItem> = items.iter().filter(|i| i.page == page).collect();

        // Check for side-by-side layout
        let owned_items: Vec<types::TextItem> = page_items.iter().map(|i| (*i).clone()).collect();
        let page_content_width = tables::content_width(&owned_items);
        let bands = markdown::split_side_by_side(&owned_items);
        let chart_regions = page_chart_regions
            .get(&page)
            .map(Vec::as_slice)
            .unwrap_or_default();

        let band_ranges: Vec<(f32, f32)> = if bands.is_empty() {
            // Single region — use sentinel range that includes everything
            vec![(f32::MIN, f32::MAX)]
        } else {
            bands
        };

        let mut found_table = false;
        for &(x_lo, x_hi) in &band_ranges {
            let margin = 2.0;
            let band_items: Vec<types::TextItem> = owned_items
                .iter()
                .filter(|item| {
                    (x_lo == f32::MIN || (item.x >= x_lo - margin && item.x < x_hi + margin))
                        && !markdown::item_is_in_chart_region(item, chart_regions)
                })
                .cloned()
                .collect();

            let band_rects: Vec<types::PdfRect> = if x_lo == f32::MIN {
                rects.iter().filter(|r| r.page == page).cloned().collect()
            } else {
                markdown::filter_rects_to_band(rects, page, x_lo, x_hi)
            };

            let band_lines: Vec<types::PdfLine> = if x_lo == f32::MIN {
                lines.iter().filter(|l| l.page == page).cloned().collect()
            } else {
                markdown::filter_lines_to_band(lines, page, x_lo, x_hi)
            };

            // TOC pages route through the table detector but render as flat
            // lists. They aren't tables in any user-facing sense, so don't
            // count them toward LayoutComplexity (would also trip the
            // table-page guard in column detection below).
            let has_data_table =
                |tables: &[tables::Table]| tables.iter().any(|t| t.kind == tables::TableKind::Data);

            let (rect_tables, _) = tables::detect_tables_from_rects(&band_items, &band_rects, page);
            if has_data_table(&rect_tables) {
                found_table = true;
                break;
            }
            let line_tables = tables::detect_tables_from_lines(&band_items, &band_lines, page);
            if has_data_table(&line_tables) {
                found_table = true;
                break;
            }
            // Heuristic fallback for borderless tables
            let heuristic_tables = tables::detect_tables_with_page_width(
                &band_items,
                base_size,
                false,
                page_content_width,
            );
            if has_data_table(&heuristic_tables) {
                found_table = true;
                break;
            }
        }
        if found_table {
            pages_with_tables.push(page);
        }
    }

    let mut pages_with_columns: Vec<u32> = Vec::new();
    for &page in &seen_pages {
        let chart_regions = page_chart_regions
            .get(&page)
            .map(Vec::as_slice)
            .unwrap_or_default();
        let page_column_items: Vec<types::TextItem> = column_items
            .iter()
            .filter(|item| {
                item.page == page && !markdown::item_is_in_chart_region(item, chart_regions)
            })
            .cloned()
            .collect();
        let cols =
            extractor::detect_columns(&page_column_items, page, pages_with_tables.contains(&page));
        if cols.len() >= 2 {
            pages_with_columns.push(page);
        }
    }

    let is_complex = !pages_with_tables.is_empty() || !pages_with_columns.is_empty();

    LayoutComplexity {
        is_complex,
        pages_with_tables,
        pages_with_columns,
    }
}

#[derive(Debug, thiserror::Error)]
pub enum PdfError {
    #[error("IO error: {0}")]
    Io(#[from] std::io::Error),
    #[error("PDF parsing error: {0}")]
    Parse(String),
    #[error("PDF is encrypted")]
    Encrypted,
    #[error("Invalid PDF structure")]
    InvalidStructure,
    #[error("Not a PDF: {0}")]
    NotAPdf(String),
}

impl From<lopdf::Error> for PdfError {
    fn from(e: lopdf::Error) -> Self {
        match e {
            lopdf::Error::IO(io_err) => PdfError::Io(io_err),
            lopdf::Error::Decryption(_)
            | lopdf::Error::InvalidPassword
            | lopdf::Error::AlreadyEncrypted
            | lopdf::Error::UnsupportedSecurityHandler(_) => PdfError::Encrypted,
            lopdf::Error::Unimplemented(msg) if msg.contains("encrypted") => PdfError::Encrypted,
            lopdf::Error::Parse(ref pe) if pe.to_string().contains("invalid file header") => {
                PdfError::NotAPdf("invalid PDF file header".to_string())
            }
            lopdf::Error::MissingXrefEntry
            | lopdf::Error::Xref(_)
            | lopdf::Error::IndirectObject { .. }
            | lopdf::Error::ObjectIdMismatch
            | lopdf::Error::InvalidObjectStream(_)
            | lopdf::Error::InvalidOffset(_) => PdfError::InvalidStructure,
            other => PdfError::Parse(other.to_string()),
        }
    }
}

/// Check whether a `lopdf::Error` represents an encryption-related failure.
pub(crate) fn is_encrypted_lopdf_error(e: &lopdf::Error) -> bool {
    matches!(
        e,
        lopdf::Error::Decryption(_)
            | lopdf::Error::InvalidPassword
            | lopdf::Error::AlreadyEncrypted
            | lopdf::Error::UnsupportedSecurityHandler(_)
    ) || matches!(e, lopdf::Error::Unimplemented(msg) if msg.contains("encrypted"))
}

// ---------------------------------------------------------------------------
// PDF validation helpers
// ---------------------------------------------------------------------------

/// Strip UTF-8 BOM and leading ASCII whitespace from a byte slice.
fn strip_bom_and_whitespace(bytes: &[u8]) -> &[u8] {
    let b = if bytes.starts_with(&[0xEF, 0xBB, 0xBF]) {
        &bytes[3..]
    } else {
        bytes
    };
    let start = b
        .iter()
        .position(|&c| !c.is_ascii_whitespace())
        .unwrap_or(b.len());
    &b[start..]
}

/// Case-insensitive prefix check on byte slices.
fn starts_with_ci(haystack: &[u8], needle: &[u8]) -> bool {
    if haystack.len() < needle.len() {
        return false;
    }
    haystack[..needle.len()]
        .iter()
        .zip(needle)
        .all(|(a, b)| a.eq_ignore_ascii_case(b))
}

/// Try to identify what kind of file the bytes represent.
fn detect_file_type_hint(bytes: &[u8]) -> String {
    if bytes.is_empty() {
        return "file is empty".to_string();
    }

    let trimmed = strip_bom_and_whitespace(bytes);

    // HTML
    if starts_with_ci(trimmed, b"<!doctype html")
        || starts_with_ci(trimmed, b"<html")
        || starts_with_ci(trimmed, b"<head")
        || starts_with_ci(trimmed, b"<body")
    {
        return "file appears to be HTML".to_string();
    }

    // XML (but not HTML)
    if trimmed.starts_with(b"<?xml") || trimmed.starts_with(b"<") {
        if starts_with_ci(trimmed, b"<?xml") {
            return "file appears to be XML".to_string();
        }
        if trimmed.starts_with(b"<") && !trimmed.starts_with(b"<%") {
            return "file appears to be XML".to_string();
        }
    }

    // JSON
    if trimmed.starts_with(b"{") || trimmed.starts_with(b"[") {
        return "file appears to be JSON".to_string();
    }

    // PNG
    if bytes.starts_with(&[0x89, 0x50, 0x4E, 0x47]) {
        return "file appears to be a PNG image".to_string();
    }

    // JPEG
    if bytes.starts_with(&[0xFF, 0xD8, 0xFF]) {
        return "file appears to be a JPEG image".to_string();
    }

    // ZIP / Office documents
    if bytes.starts_with(&[0x50, 0x4B, 0x03, 0x04]) {
        return "file appears to be a ZIP archive (possibly an Office document)".to_string();
    }

    // If it looks like mostly printable ASCII/UTF-8, call it plain text
    let sample = &bytes[..bytes.len().min(512)];
    let printable = sample
        .iter()
        .filter(|&&b| b.is_ascii_graphic() || b.is_ascii_whitespace())
        .count();
    if printable > sample.len() * 3 / 4 {
        return "file appears to be plain text".to_string();
    }

    "file is not a PDF".to_string()
}

/// The `%PDF` header must *start* within this many bytes of the buffer.
///
/// mupdf, pdfium and poppler all accept a header that is not at offset 0
/// (leading bytes before the header, e.g. an echoed multipart envelope or a
/// UTF-8 BOM); this mirrors that tolerance with a bounded search.
const PDF_HEADER_SEARCH_WINDOW: usize = 1024;

/// Longest header the locator inspects: `%PDF-M.N` plus its line ending.
const PDF_HEADER_PROBE_LEN: usize = 9;

/// Byte offset of the `%PDF-` header within the first
/// [`PDF_HEADER_SEARCH_WINDOW`] bytes of `buffer`.
///
/// Candidates are ranked by how much they look like a real header line, and
/// by buffer order within a rank: a canonical `%PDF-M.N` line (starting a
/// line and ending at a line break) first, then `%PDF-` followed by a version
/// digit, then any other `%PDF-`. A version-like mention inside leading text
/// (`X-Note: %PDF-1.4`, say) therefore does not outrank the actual header.
/// lopdf's own header parser requires the literal `%PDF-`, so a bare `%PDF`
/// is not a candidate.
fn pdf_header_offset(buffer: &[u8]) -> Option<usize> {
    // Only the start of the marker is bounded by the window; let the marker
    // itself run past it.
    let probe_end = (PDF_HEADER_SEARCH_WINDOW + PDF_HEADER_PROBE_LEN).min(buffer.len());
    let probe = &buffer[..probe_end];

    let mut best: Option<(u8, usize)> = None;
    for (offset, window) in probe
        .windows(b"%PDF-".len())
        .enumerate()
        .take(PDF_HEADER_SEARCH_WINDOW)
    {
        if window != b"%PDF-" {
            continue;
        }
        let version = &probe[offset + b"%PDF-".len()..];
        let at_line_start = offset == 0 || matches!(probe[offset - 1], b'\r' | b'\n');
        let rank = if at_line_start && is_canonical_pdf_version_line(version) {
            return Some(offset);
        } else if version.first().is_some_and(u8::is_ascii_digit) {
            1
        } else {
            2
        };
        if best.is_none_or(|(best_rank, _)| rank < best_rank) {
            best = Some((rank, offset));
        }
    }
    best.map(|(_, offset)| offset)
}

/// `M.N` immediately followed by a line ending (or the end of the probe),
/// i.e. the bytes after `%PDF-` on a canonical header line.
fn is_canonical_pdf_version_line(version: &[u8]) -> bool {
    matches!(version, [major, b'.', minor, rest @ ..]
        if major.is_ascii_digit()
            && minor.is_ascii_digit()
            && rest.first().is_none_or(|b| matches!(b, b'\r' | b'\n')))
}

/// Return `buffer` starting at its `%PDF-` header, dropping any leading bytes.
///
/// Returns `buffer` unchanged when no header is found so the loader reports
/// its own error.
fn strip_leading_bytes_before_header(buffer: &[u8]) -> &[u8] {
    match pdf_header_offset(buffer) {
        Some(offset) if offset > 0 => {
            log::debug!("dropping {offset} leading bytes before the %PDF- header");
            &buffer[offset..]
        }
        _ => buffer,
    }
}

/// Validate that a byte buffer looks like a PDF (has `%PDF-` magic).
///
/// The header must start within the first [`PDF_HEADER_SEARCH_WINDOW`] bytes;
/// see [`pdf_header_offset`]. Leading bytes before it are tolerated here and
/// dropped by the loader.
pub(crate) fn validate_pdf_bytes(buffer: &[u8]) -> Result<(), PdfError> {
    if pdf_header_offset(buffer).is_some() {
        Ok(())
    } else {
        Err(PdfError::NotAPdf(detect_file_type_hint(buffer)))
    }
}

/// Attach `path` to an IO error so a missing input is distinguishable from
/// a missing built-in resource. Relative paths also name the working
/// directory they were resolved against.
pub(crate) fn io_error_at(path: &Path, err: std::io::Error) -> PdfError {
    let location = if path.is_absolute() {
        path.display().to_string()
    } else if let Ok(cwd) = std::env::current_dir() {
        format!("{} (working directory {})", path.display(), cwd.display())
    } else {
        path.display().to_string()
    };
    PdfError::Io(std::io::Error::new(
        err.kind(),
        format!("{location}: {err}"),
    ))
}

/// Read a whole file, reporting `path` in the error.
pub(crate) fn read_file(path: &Path) -> Result<Vec<u8>, PdfError> {
    std::fs::read(path).map_err(|err| io_error_at(path, err))
}

/// Validate that a file on disk looks like a PDF.
///
/// Reads only the header search window and delegates to [`validate_pdf_bytes`].
pub(crate) fn validate_pdf_file<P: AsRef<Path>>(path: P) -> Result<(), PdfError> {
    use std::io::Read;
    let path = path.as_ref();
    let mut file = std::fs::File::open(path).map_err(|err| io_error_at(path, err))?;
    let mut buf = [0u8; PDF_HEADER_SEARCH_WINDOW + PDF_HEADER_PROBE_LEN];
    let n = file.read(&mut buf).map_err(|err| io_error_at(path, err))?;
    validate_pdf_bytes(&buf[..n])
}

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

    #[test]
    fn missing_file_error_names_the_path() {
        let err = validate_pdf_file("this-file-does-not-exist.pdf").unwrap_err();
        let msg = err.to_string();
        assert!(
            msg.contains("this-file-does-not-exist.pdf"),
            "missing-file error should name the path, got {msg}"
        );
        assert!(
            msg.contains("working directory"),
            "relative path should name the working directory, got {msg}"
        );
    }

    #[test]
    fn pdf_header_offset_finds_header_at_start() {
        assert_eq!(pdf_header_offset(b"%PDF-1.4\n%%EOF"), Some(0));
    }

    #[test]
    fn pdf_header_offset_skips_bom_and_whitespace() {
        let mut buf = vec![0xEF, 0xBB, 0xBF, b'\n', b'\t'];
        buf.extend_from_slice(b"%PDF-1.7\n");
        assert_eq!(pdf_header_offset(&buf), Some(5));
    }

    #[test]
    fn pdf_header_offset_finds_header_after_leading_bytes() {
        // Leading bytes before the header, e.g. an echoed multipart envelope.
        let prefix = b"--boundary\r\nContent-Type: application/pdf\r\n\r\n";
        let mut buf = prefix.to_vec();
        buf.extend_from_slice(b"%PDF-1.5\n%%EOF");
        assert_eq!(pdf_header_offset(&buf), Some(prefix.len()));
    }

    #[test]
    fn pdf_header_offset_ignores_bare_marker_in_leading_text() {
        let buf = b"filename=\"%PDF.pdf\"\r\n\r\n%PDF-1.6\n";
        let expected = buf.len() - b"%PDF-1.6\n".len();
        assert_eq!(pdf_header_offset(buf), Some(expected));
        assert_eq!(pdf_header_offset(b"xx%PDF\n1 0 obj"), None);
    }

    #[test]
    fn pdf_header_offset_accepts_dash_without_version_digit() {
        assert_eq!(pdf_header_offset(b"%PDF-\n1 0 obj"), Some(0));
    }

    #[test]
    fn pdf_header_offset_accepts_header_at_window_edge() {
        let mut buf = vec![b' '; PDF_HEADER_SEARCH_WINDOW - 1];
        buf.extend_from_slice(b"%PDF-1.4\n");
        assert_eq!(pdf_header_offset(&buf), Some(PDF_HEADER_SEARCH_WINDOW - 1));
    }

    #[test]
    fn pdf_header_offset_rejects_header_beyond_window() {
        let mut buf = vec![b' '; PDF_HEADER_SEARCH_WINDOW];
        buf.extend_from_slice(b"%PDF-1.4\n");
        assert_eq!(pdf_header_offset(&buf), None);
        assert_eq!(pdf_header_offset(b""), None);
        assert_eq!(pdf_header_offset(b"just some text"), None);
    }

    #[test]
    fn pdf_header_offset_ranks_canonical_header_lines_first() {
        // dashed-only at 0, version digit at 7, canonical lines at 19 and 28.
        let buf = b"%PDF-x\n%PDF-1 note\n%PDF-1.4\n%PDF-2.0\r\n";
        assert_eq!(pdf_header_offset(buf), Some(19));
        // A canonical-looking version mid-line is only a version-digit match.
        let buf = b"X-Note: %PDF-1.4\n%PDF-1.7\n";
        assert_eq!(pdf_header_offset(buf), Some(17));
        assert_eq!(pdf_header_offset(b"%PDF-x\n%PDF-1 note\n"), Some(7));
        assert_eq!(pdf_header_offset(b"%PDF-x\n"), Some(0));

        // Many version-like mentions never outrank the real header line.
        let mut buf = b"x: %PDF-1\n".repeat(50);
        buf.extend_from_slice(b"%PDF-1.4\n");
        assert_eq!(
            pdf_header_offset(&buf),
            Some(buf.len() - b"%PDF-1.4\n".len())
        );
    }

    #[test]
    fn strip_leading_bytes_before_header_slices_to_header() {
        let buf = b"junk\r\n%PDF-1.4\n%%EOF";
        assert_eq!(strip_leading_bytes_before_header(buf), b"%PDF-1.4\n%%EOF");
        let clean = b"%PDF-1.4\n%%EOF";
        assert_eq!(strip_leading_bytes_before_header(clean), clean);
        let none = b"not a pdf";
        assert_eq!(strip_leading_bytes_before_header(none), none);
    }

    #[test]
    fn validate_pdf_bytes_accepts_leading_bytes_and_rejects_text() {
        assert!(validate_pdf_bytes(b"--b\r\n\r\n%PDF-1.4\n").is_ok());
        match validate_pdf_bytes(b"This is plain text mentioning %PDF, not a PDF.") {
            Err(PdfError::NotAPdf(hint)) => assert!(hint.contains("plain text"), "{hint}"),
            other => panic!("expected NotAPdf, got {other:?}"),
        }
    }

    fn test_item(text: &str, x: f32, y: f32, width: f32, height: f32) -> TextItem {
        TextItem {
            text: text.to_string(),
            x,
            y,
            width,
            height,
            font: "Helvetica".to_string(),
            font_tag: String::new(),
            legacy_symbol_rewrite: false,
            font_size: height,
            page: 1,
            is_bold: false,
            is_italic: false,
            font_weight: None,
            bold_source: None,
            fixed_pitch: None,
            fill_color: None,
            stroke_color: None,
            render_mode: None,
            is_underline: false,
            is_strikeout: false,
            rotation: 0.0,
            advance_known: true,
            item_type: ItemType::Text,
            mcid: None,
            baseline_shift: 0.0,
        }
    }

    fn test_text_item_on_page(page: u32, text: &str) -> TextItem {
        TextItem {
            page,
            ..test_item(text, 10.0, 10.0, text.len() as f32 * 5.0, 12.0)
        }
    }

    fn test_image_item(width: f32, height: f32) -> TextItem {
        TextItem {
            item_type: ItemType::Image,
            ..test_item("[Image]", 20.0, 30.0, width, height)
        }
    }

    #[test]
    fn supplemental_ocr_regions_require_substantial_physical_images() {
        let substantial = supplemental_ocr_image_region(&test_image_item(200.0, 120.0)).unwrap();
        assert_eq!(substantial.width, 200.0);
        assert_eq!(substantial.height, 120.0);

        assert!(supplemental_ocr_image_region(&test_image_item(100.0, 200.0)).is_none());
        assert!(supplemental_ocr_image_region(&test_image_item(200.0, 60.0)).is_none());
        assert!(supplemental_ocr_image_region(&test_image_item(120.0, 100.0)).is_none());
        assert!(
            supplemental_ocr_image_region(&test_item("text", 0.0, 0.0, 300.0, 300.0)).is_none()
        );
    }

    #[test]
    fn supplemental_ocr_regions_normalize_negative_image_dimensions() {
        let region = supplemental_ocr_image_region(&test_image_item(-200.0, -120.0)).unwrap();
        assert_eq!(region.x, -180.0);
        assert_eq!(region.y, -90.0);
        assert_eq!(region.width, 200.0);
        assert_eq!(region.height, 120.0);
    }

    #[test]
    fn removed_sparse_folios_leave_no_layout_evidence() {
        let items = vec![
            test_item("1", 25.0, 20.0, 12.0, 10.0),
            test_item("2", 520.0, 60.0, 12.0, 10.0),
        ];
        let (filtered, _, _) =
            extractor::filter_markdown_page_numbers_with_removed_pages(items.clone(), 1);
        assert!(filtered.is_empty());

        let filtered = compute_layout_complexity(&items, &filtered, &[], &[]);

        assert!(!filtered.is_complex);
        assert!(filtered.pages_with_tables.is_empty());
        assert!(filtered.pages_with_columns.is_empty());
    }

    #[test]
    fn dense_chart_panel_is_not_reported_as_a_table() {
        let mut items: Vec<TextItem> = (0..8)
            .flat_map(|row| {
                (0..6).map(move |column| {
                    test_item(
                        &format!("{}", row * 10 + column),
                        105.0 + column as f32 * 35.0,
                        525.0 - row as f32 * 15.0,
                        24.0,
                        10.0,
                    )
                })
            })
            .collect();
        for row in 0..6 {
            items.push(test_item(
                "Left column prose continues here",
                80.0,
                320.0 - row as f32 * 15.0,
                160.0,
                10.0,
            ));
            items.push(test_item(
                "Right column prose continues here",
                300.0,
                320.0 - row as f32 * 15.0,
                160.0,
                10.0,
            ));
        }
        let mut lines: Vec<PdfLine> = (0..30)
            .map(|column| PdfLine {
                x1: 100.0 + column as f32 * 8.0,
                y1: 400.0,
                x2: 100.0 + column as f32 * 8.0,
                y2: 550.0,
                page: 1,
            })
            .collect();
        lines.extend((0..6).map(|row| PdfLine {
            x1: 100.0,
            y1: 400.0 + row as f32 * 30.0,
            x2: 332.0,
            y2: 400.0 + row as f32 * 30.0,
            page: 1,
        }));
        let rects = vec![PdfRect {
            x: 80.0,
            y: 350.0,
            width: 280.0,
            height: 240.0,
            page: 1,
        }];

        let complexity = compute_layout_complexity(&items, &items, &rects, &lines);

        assert!(complexity.pages_with_tables.is_empty());
    }

    #[test]
    fn page_selection_keeps_document_wide_folio_layout_decisions() {
        let mut items = Vec::new();
        for page in 1..=4 {
            for row in 0..8 {
                let y = 20.0 + row as f32 * 8.0;
                let mut folio = test_item(&(row * 10 + page).to_string(), 25.0, y, 12.0, 10.0);
                folio.page = page;
                let mut footer =
                    test_item(&format!("Footer row {row} summary"), 43.0, y, 470.0, 10.0);
                footer.page = page;
                let mut body = test_item(&format!("Body{row}"), 530.0, y, 55.0, 10.0);
                body.page = page;
                items.extend([folio, footer, body]);
            }
        }

        let page_one_items: Vec<_> = items
            .iter()
            .filter(|item| item.page == 1)
            .cloned()
            .collect();
        let (page_local_layout, _, _) =
            extractor::filter_markdown_page_numbers_with_removed_pages(page_one_items.clone(), 4);
        let page_local = compute_layout_complexity(&page_one_items, &page_local_layout, &[], &[]);
        assert!(
            page_local.pages_with_columns.contains(&1),
            "fixture must reproduce page-local folio column evidence"
        );

        let selected =
            select_items_with_document_folio_context(items, 4, Some(&HashSet::from([1])));
        assert_eq!(
            selected
                .removal_mask
                .iter()
                .filter(|remove| **remove)
                .count(),
            8
        );
        let document_wide =
            compute_layout_complexity(&selected.items, &selected.layout_items, &[], &[]);
        assert!(!document_wide.pages_with_columns.contains(&1));
    }

    #[test]
    fn test_detect_encoding_issues_fffd() {
        assert!(detect_encoding_issues(
            "Some text with \u{FFFD} replacement"
        ));
    }

    #[test]
    fn test_detect_encoding_issues_dollar_as_space() {
        // Simulates broken CMap: "$Workshop$on$Chest$Wall$Deformities$and$..."
        let garbled = "Last$advanced$Book$Programm$3th$Workshop$on$Chest$Wall$Deformities$and$More";
        assert!(detect_encoding_issues(garbled));
    }

    #[test]
    fn test_detect_encoding_issues_financial_text() {
        // Legitimate dollar signs in financial text should NOT trigger
        let financial = "Revenue was $100M in Q1, up from $90M. Costs: $50M, $30M, $20M, $15M, $12M, $8M, $5M, $3M, $2M, $1M, $500K.";
        assert!(!detect_encoding_issues(financial));
    }

    #[test]
    fn test_detect_encoding_issues_clean_text() {
        assert!(!detect_encoding_issues(
            "Normal markdown text with no issues."
        ));
    }

    #[test]
    fn test_detect_encoding_issues_few_dollars() {
        // Under threshold of 10 total dollars — should not trigger
        let text = "a$b c$d e$f";
        assert!(!detect_encoding_issues(text));
    }

    /// Real garbled output from ParseBench `text_simple__att10k.pdf`: a broken
    /// ToUnicode CMap shifts every character by a per-range constant, so
    /// "Certificate of Designations with respect to Series C" extracts as
    /// pure-ASCII ciphertext (issue #118).
    const SHIFTED_CIPHER_TEXT: &str =
        "8VceZWZTReVW9VdZXReZdhZeYcVdaVTeeHVcZVd8EcVWVccVUHeT:iYZSZe-(,e2'WZ]V \
        ;VScfRcj2&,*,*$ .'R CZdecfVehYZTYUVWZVdeYVcZXYedWY]UVcdW]X'eVcUVSeYVcVXZdecReRUR]]ed \
        TCBGC@ZUReVUdfSdZUZRcZVdZdWZ]VUYVcVhZeYafcdfReeVXf]ReZV0*S$.$ZZZ$6$&ViTVae \
        WceYVZdecfVedcVWVccVUe.'S&.'T&.'U&.'V&.'WSV]h(EfcdfReeYZdcVXf]ReZYV \
        cVXZdecReYVcVSjRXcVVdeWfcZdYRTajWRjdfTYZdecfVeWZ]VUYVcVhZeYeYVH:8 \
        cVbfVde( .'S fRcRejWTVceRZZXReZTZWZT7V]]IV]VaYV8(RUHfeYhVdeVc7V]]IV]VaYV8 \
        :iYZSZe.'TeYVaVcZUVUZX9VTVSVc-&,*$";

    #[test]
    fn test_detect_encoding_issues_shifted_cipher_text() {
        assert!(detect_encoding_issues(SHIFTED_CIPHER_TEXT));
    }

    #[test]
    fn test_shifted_cipher_below_sample_threshold_not_flagged() {
        // Fewer than 200 ASCII letters: not enough evidence to condemn.
        assert!(!detect_encoding_issues("8VceZWZTReVW9VdZXReZdhZeY"));
    }

    #[test]
    fn test_clean_prose_not_flagged_as_cipher() {
        let prose = "Certificate of Designations with respect to Series C Preferred Stock, \
            filed February 18, 2020. No instrument which defines the rights of holders of \
            long-term debt of the registrant and all of its consolidated subsidiaries is \
            filed herewith pursuant to Regulation S-K, Item 601. Pursuant to this regulation, \
            the registrant hereby agrees to furnish a copy of any such instrument to the SEC \
            upon request.";
        assert!(!detect_encoding_issues(prose));
    }

    #[test]
    fn test_camel_case_code_not_flagged_as_cipher() {
        let code = "The getElementById and querySelectorAll methods return DOM nodes. Use \
            addEventListener with removeEventListener, requestAnimationFrame with \
            cancelAnimationFrame, and setTimeout with clearTimeout. The XMLHttpRequest \
            object exposes onreadystatechange, responseText and getAllResponseHeaders. \
            Prefer createElement, appendChild, insertBefore and replaceChild for DOM \
            manipulation, and getBoundingClientRect for layout measurement.";
        assert!(!detect_encoding_issues(code));
    }

    #[test]
    fn test_accented_european_text_not_flagged_as_cipher() {
        let swedish = "Regeringen föreslår att riksdagen antar förslaget till lag om ändring \
            i skatteförfarandelagen. Bestämmelserna föreslås träda i kraft den första januari. \
            Förslaget innebär att företag med säte i utlandet måste lämna särskilda uppgifter \
            till Skatteverket varje kvartal, och att avgiften höjs för överträdelser av de nya \
            bestämmelserna om rapporteringsskyldighet för gränsöverskridande arrangemang.";
        assert!(!detect_encoding_issues(swedish));
    }

    #[test]
    fn test_all_caps_text_not_flagged_as_cipher() {
        let caps = "EXHIBIT INDEX PURSUANT TO ITEM 601 OF REGULATION SK CERTIFICATE OF \
            DESIGNATIONS WITH RESPECT TO SERIES C PREFERRED STOCK FILED FEBRUARY EIGHTEEN \
            TWENTY TWENTY AND INCORPORATED HEREIN BY REFERENCE TO THE ANNUAL REPORT ON FORM \
            TENK FOR THE PERIOD ENDED DECEMBER THIRTYFIRST TWENTY NINETEEN AS AMENDED";
        assert!(!detect_encoding_issues(caps));
    }

    // A Caesar shift of prose that stays within a single case block does not
    // trigger the case-shift signal, but it permutes the letter histogram: the
    // frequency SHAPE stays English-like while letter POSITIONS scramble, so
    // the permutation signal catches it regardless of case.
    const CAESAR_PROSE: &str =
        "The registrant hereby agrees to furnish a copy of any such instrument to the \
        Commission upon request. This certificate of designations was filed February with \
        respect to Series Preferred Stock and incorporated herein by reference to the annual \
        report on form for the period ended December as amended and restated thereafter.";

    fn caesar_shift(text: &str, k: u8) -> String {
        text.chars()
            .map(|c| match c {
                'a'..='z' => (((c as u8 - b'a' + k) % 26) + b'a') as char,
                'A'..='Z' => (((c as u8 - b'A' + k) % 26) + b'A') as char,
                _ => c,
            })
            .collect()
    }

    #[test]
    fn test_mixed_case_caesar_shift_flagged() {
        assert!(detect_encoding_issues(&caesar_shift(CAESAR_PROSE, 3)));
    }

    #[test]
    fn test_all_lowercase_caesar_shift_flagged() {
        // Uniform all-lowercase garbled prose: the earlier mixed-case guard
        // would have exempted this, so it must be caught by the case-agnostic
        // permutation signal instead.
        assert!(detect_encoding_issues(&caesar_shift(
            &CAESAR_PROSE.to_lowercase(),
            5
        )));
    }

    #[test]
    fn test_all_uppercase_caesar_shift_flagged() {
        assert!(detect_encoding_issues(&caesar_shift(
            &CAESAR_PROSE.to_uppercase(),
            7
        )));
    }

    #[test]
    fn test_dna_sequence_not_flagged_as_cipher() {
        // Non-linguistic ASCII: unlike English (low vowel ratio, low cosine)
        // but not garbled. Its 4-letter alphabet makes the frequency profile
        // too steep, so the shape cosine falls below the permutation threshold.
        let dna = "ACGT".repeat(120);
        assert!(!detect_encoding_issues(&dna));
    }

    #[test]
    fn test_protein_sequence_not_flagged_as_cipher() {
        let protein =
            "MKTAYIAKQRQISFVKSHFSRQLEERLGLIEVQAPILSRVGDGTQDNLSGAEKAVQVKVKALPDAQFEVVHSLAKWKR"
                .repeat(6);
        assert!(!detect_encoding_issues(&protein));
    }

    #[test]
    fn test_ticker_list_not_flagged_as_cipher() {
        let tickers = "AAPL MSFT GOOG TSLA NVDA AMZN META NFLX AMD INTC CSCO ORCL CRM ADBE QCOM \
            TXN AVGO MU LRCX KLAC ASML SNPS CDNS FTNT PANW "
            .repeat(4);
        assert!(!detect_encoding_issues(&tickers));
    }

    #[test]
    fn test_text_quality_flags_shifted_cipher_page() {
        let items = vec![
            test_text_item_on_page(1, SHIFTED_CIPHER_TEXT),
            test_text_item_on_page(2, "A clean second page should not be routed to OCR."),
        ];

        let quality = analyze_text_quality(&items);

        assert!(quality.has_encoding_issues);
        assert_eq!(quality.pages_needing_ocr, vec![1]);
        assert_eq!(
            quality.reasons_by_page.get(&1).cloned(),
            Some(vec![OCR_REASON_SUSPECTED_GARBLED_TEXT.to_string()])
        );
    }

    #[test]
    fn test_text_quality_cipher_stats_accumulate_across_items() {
        // The garbled page arrives as many short spans; no single span has
        // enough letters to flag on its own.
        let items: Vec<TextItem> = SHIFTED_CIPHER_TEXT
            .split_whitespace()
            .map(|chunk| test_text_item_on_page(1, chunk))
            .collect();

        let quality = analyze_text_quality(&items);

        assert_eq!(quality.pages_needing_ocr, vec![1]);
    }

    #[test]
    fn test_text_quality_flags_localized_cid_mojibake_span() {
        let items = vec![
            test_text_item_on_page(
                1,
                "Waiting Period 等待期 Maternity and newborn infant care benefit",
            ),
            test_text_item_on_page(
                1,
                "Inpatient and Day-care Benefits DÂB\u{009B}A4gÉ9¶0ÅDÂB\u{009B}Ê(D>öBÑ9¯",
            ),
            test_text_item_on_page(1, "Covered up to annual maximum. 赔付至年度最高保额。"),
            test_text_item_on_page(2, "A clean second page should not be routed to OCR."),
        ];

        let quality = analyze_text_quality(&items);

        assert!(quality.has_encoding_issues);
        assert_eq!(quality.pages_needing_ocr, vec![1]);
        assert_eq!(
            quality.reasons_by_page.get(&1).cloned(),
            Some(vec![OCR_REASON_SUSPECTED_GARBLED_TEXT.to_string()])
        );
    }

    #[test]
    fn test_text_quality_flags_replacement_and_private_use_runs() {
        let items = vec![
            test_text_item_on_page(1, "broken \u{FFFD}\u{FFFD} text"),
            test_text_item_on_page(3, "\u{E000}\u{E001}\u{E002}"),
        ];

        let quality = analyze_text_quality(&items);

        assert_eq!(quality.pages_needing_ocr, vec![1, 3]);
        assert_eq!(
            quality.reasons_by_page.get(&1).cloned(),
            Some(vec![OCR_REASON_SUSPECTED_GARBLED_TEXT.to_string()])
        );
        assert_eq!(
            quality.reasons_by_page.get(&3).cloned(),
            Some(vec![OCR_REASON_SUSPECTED_GARBLED_TEXT.to_string()])
        );
    }

    #[test]
    fn test_text_quality_allows_clean_multilingual_and_latin1_text() {
        let items = vec![
            test_text_item_on_page(1, "你好世界，这是一段正常的中文文本。"),
            test_text_item_on_page(1, "Résumé déjà vu: façade, São Paulo, año 2026."),
            test_text_item_on_page(1, "A single icon \u{E000} should not force OCR."),
        ];

        let quality = analyze_text_quality(&items);

        assert!(!quality.has_encoding_issues);
        assert!(quality.pages_needing_ocr.is_empty());
        assert!(quality.reasons_by_page.is_empty());
    }

    #[test]
    fn test_text_quality_allows_toc_leaders_form_rules_and_short_math() {
        let items = vec![
            test_text_item_on_page(
                1,
                "Feature Overview ........................................................................................................ 1-5",
            ),
            test_text_item_on_page(1, "Signature __________________________________________"),
            test_text_item_on_page(1, "__________________________________________________"),
            test_text_item_on_page(1, "2×()×"),
        ];

        let quality = analyze_text_quality(&items);

        assert!(!quality.has_encoding_issues);
        assert!(quality.pages_needing_ocr.is_empty());
    }

    #[test]
    fn test_text_quality_allows_isolated_replacement_character() {
        let items = vec![
            test_text_item_on_page(1, "\u{FFFD}2026 FINRA"),
            test_text_item_on_page(1, "A mostly clean page should not be sent to OCR."),
        ];

        let quality = analyze_text_quality(&items);

        assert!(!quality.has_encoding_issues);
        assert!(quality.pages_needing_ocr.is_empty());
    }

    #[test]
    fn test_text_quality_allows_formula_replacement_on_clean_page() {
        let items = vec![
            test_text_item_on_page(
                1,
                "The LCOE of a power plant can be decomposed into three parts and described in prose.",
            ),
            test_text_item_on_page(
                1,
                "This page has enough normal text that a damaged equation should not force OCR.",
            ),
            test_text_item_on_page(1, "\u{FFFD}\u{FFFD}\u{FFFD}\u{FFFD} = x + y"),
            test_text_item_on_page(1, "More normal explanatory text follows after the formula."),
        ];

        let quality = analyze_text_quality(&items);

        assert!(!quality.has_encoding_issues);
        assert!(quality.pages_needing_ocr.is_empty());
    }

    #[test]
    fn test_text_quality_flags_dense_replacement_text_page() {
        let items = vec![
            test_text_item_on_page(1, "\u{FFFD}\u{FFFD}\u{FFFD}\u{FFFD} broken layer"),
            test_text_item_on_page(1, "more \u{FFFD}\u{FFFD}\u{FFFD}\u{FFFD} broken text"),
            test_text_item_on_page(1, "\u{FFFD}\u{FFFD}\u{FFFD}\u{FFFD}"),
        ];

        let quality = analyze_text_quality(&items);

        assert_eq!(quality.pages_needing_ocr, vec![1]);
        assert!(quality.has_encoding_issues);
    }

    #[test]
    fn test_text_quality_allows_tex_ligature_c1_controls_in_words() {
        let items = vec![test_text_item_on_page(
            1,
            "Especialmente de\u{85}ciente es nuestro conocimiento del control. The amniotic \u{87}uid is important.",
        )];

        let quality = analyze_text_quality(&items);

        assert!(!quality.has_encoding_issues);
        assert!(quality.pages_needing_ocr.is_empty());
    }

    #[test]
    fn test_region_text_quality_is_scoped_to_matched_items() {
        let clean_region = vec![
            test_text_item_on_page(1, "Clean native text"),
            test_text_item_on_page(1, "Résumé déjà vu"),
        ];
        let garbled_region = vec![
            test_text_item_on_page(1, "Clean prefix"),
            test_text_item_on_page(1, "DÂB\u{009B}A4gÉ9¶0ÅDÂB\u{009B}Ê(D>öBÑ9¯"),
        ];

        assert!(!region_items_have_decoding_issue(&clean_region));
        assert!(region_items_have_decoding_issue(&garbled_region));
    }

    #[test]
    fn test_garbage_text_detection() {
        // Simulates garbage output from Identity-H fonts without ToUnicode.
        // Needs >= 50 non-whitespace chars and < 50% alphanumeric.
        let garbage = ",&<X ~%5&8-!A ~*(!,-!U (/#!U X ~#/=U 9/%*(!U !(  X \
                       (%U-(-/ V %&((8-#&&< *,(6--< %5&8-!( (,(/! #/<5U X \
                       º&( >/5 /5&(#(8-!5 *,(6--( *,%@/-A W";
        assert!(is_garbage_text(garbage));

        // Normal text should not be garbage
        let normal = "This is a normal paragraph with words and sentences that contains enough characters to pass the threshold.";
        assert!(!is_garbage_text(normal));

        // Cyrillic text should not be garbage
        let cyrillic =
            "Роботизированные технологии комплексы для производства металлургических предприятий";
        assert!(!is_garbage_text(cyrillic));
    }

    #[test]
    fn test_cid_garbage_detection() {
        // Simulates CID garbage from Identity-H fonts: Latin Extended chars
        // mixed with C1 control characters (U+0080–U+009F).
        let cid_garbage = "Ë>íÓ\tý\r\u{0088}æ&Ït\u{0094}äí;\ný;wAL¢©èåD\rü£\
                           qq\u{0096}¶Í Æ\réá; Ô 7G\u{008B}ý;èÕç¢ £ ý;C";
        assert!(
            is_cid_garbage(cid_garbage),
            "CID garbage with C1 controls should be detected"
        );

        // Valid Korean text (CID-as-Unicode passthrough) should NOT be garbage
        let korean = "본 가격표는 국내 거주 중인 외국인을 위한 한국어 가격표의 비공식 번역본입니다";
        assert!(
            !is_cid_garbage(korean),
            "Valid Korean text should not be flagged as garbage"
        );

        // Valid Japanese text should NOT be garbage
        let japanese = "羽田空港新飛行経路に係る航空機騒音の測定結果";
        assert!(
            !is_cid_garbage(japanese),
            "Valid Japanese text should not be flagged as garbage"
        );

        let japanese_toc = "第１章 市政経営方針の位置づけ ···································· 1";
        assert!(
            !is_cid_garbage(japanese_toc),
            "Japanese TOC dot leaders should not be flagged as garbage"
        );

        let tex_ligature = "amniotic \u{87}uid volume regulation";
        assert!(
            !is_cid_garbage(tex_ligature),
            "A single TeX ligature byte inside a word should not be CID garbage"
        );
    }

    #[test]
    fn tsr_text_fill_does_not_pull_neighboring_overlapping_rows() {
        use crate::tables::structured::normalize_cell_bands;
        use crate::tables::StructuredCell;

        let items = vec![
            test_item("Branch Name", 12.0, 88.0, 55.0, 8.0),
            test_item("Deposits", 112.0, 88.0, 36.0, 8.0),
            test_item("Oak Street", 12.0, 72.0, 48.0, 8.0),
            test_item("100", 112.0, 72.0, 18.0, 8.0),
            test_item("Boardwalk", 12.0, 55.2, 46.0, 8.0),
            test_item("200", 112.0, 55.2, 18.0, 8.0),
        ];
        let mut cells = vec![
            StructuredCell {
                row: 0,
                col: 0,
                rowspan: 1,
                colspan: 1,
                is_header: true,
                text: String::new(),
                page_pt_bbox: [10.0, 100.0, 100.0, 125.0],
            },
            StructuredCell {
                row: 0,
                col: 1,
                rowspan: 1,
                colspan: 1,
                is_header: true,
                text: String::new(),
                page_pt_bbox: [100.0, 100.0, 170.0, 125.0],
            },
            StructuredCell {
                row: 1,
                col: 0,
                rowspan: 1,
                colspan: 1,
                is_header: false,
                text: String::new(),
                page_pt_bbox: [10.0, 116.0, 100.0, 141.0],
            },
            StructuredCell {
                row: 1,
                col: 1,
                rowspan: 1,
                colspan: 1,
                is_header: false,
                text: String::new(),
                page_pt_bbox: [100.0, 116.0, 170.0, 141.0],
            },
            StructuredCell {
                row: 2,
                col: 0,
                rowspan: 1,
                colspan: 1,
                is_header: false,
                text: String::new(),
                page_pt_bbox: [10.0, 132.8, 100.0, 157.8],
            },
            StructuredCell {
                row: 2,
                col: 1,
                rowspan: 1,
                colspan: 1,
                is_header: false,
                text: String::new(),
                page_pt_bbox: [100.0, 132.8, 170.0, 157.8],
            },
        ];

        normalize_cell_bands(&mut cells);
        for cell in &mut cells {
            let [x1, y1, x2, y2] = cell.page_pt_bbox;
            cell.text = collect_text_in_tsr_cell(
                &items,
                x1,
                y1,
                x2,
                y2,
                200.0,
                RegionCoordSpace::Standard,
                0.10,
            );
        }

        assert_eq!(cells[0].text, "Branch Name");
        assert_eq!(cells[2].text, "Oak Street");
        assert_eq!(cells[4].text, "Boardwalk");
        assert!(!cells[0].text.contains("Oak Street"));
        assert!(!cells[2].text.contains("Branch Name"));
        assert!(!cells[2].text.contains("Boardwalk"));
    }

    #[test]
    fn multi_row_expansion_splits_overstuffed_tsr_row() {
        use crate::tables::{cells_to_markdown, StructuredCell};

        let items = vec![
            test_item("Branch Name", 20.0, 166.0, 55.0, 8.0),
            test_item("Deposits", 120.0, 166.0, 36.0, 8.0),
            test_item("Oak Street", 20.0, 136.0, 48.0, 8.0),
            test_item("100", 120.0, 136.0, 18.0, 8.0),
            test_item("Boardwalk", 20.0, 116.0, 46.0, 8.0),
            test_item("200", 120.0, 116.0, 18.0, 8.0),
        ];
        let cells = vec![
            StructuredCell {
                row: 0,
                col: 0,
                rowspan: 1,
                colspan: 1,
                is_header: true,
                text: "Branch Name".into(),
                page_pt_bbox: [10.0, 20.0, 100.0, 40.0],
            },
            StructuredCell {
                row: 0,
                col: 1,
                rowspan: 1,
                colspan: 1,
                is_header: true,
                text: "Deposits".into(),
                page_pt_bbox: [110.0, 20.0, 190.0, 40.0],
            },
            StructuredCell {
                row: 1,
                col: 0,
                rowspan: 1,
                colspan: 1,
                is_header: false,
                text: "Oak Street Boardwalk".into(),
                page_pt_bbox: [10.0, 40.0, 100.0, 100.0],
            },
            StructuredCell {
                row: 1,
                col: 1,
                rowspan: 1,
                colspan: 1,
                is_header: false,
                text: "100 200".into(),
                page_pt_bbox: [110.0, 40.0, 190.0, 100.0],
            },
        ];

        let expanded =
            try_expand_multi_row_cells(&cells, &items, 200.0, RegionCoordSpace::Standard, 0.10)
                .expect("overstuffed data row should expand");
        let md = cells_to_markdown(&expanded);

        assert_eq!(expanded.iter().map(|c| c.row).max().unwrap() + 1, 3);
        assert!(
            md.contains("|Oak Street|100|"),
            "missing first data row: {md}"
        );
        assert!(
            md.contains("|Boardwalk|200|"),
            "missing second data row: {md}"
        );
        assert!(
            !md.contains("Oak Street Boardwalk"),
            "compressed cell text should be replaced: {md}"
        );
    }

    #[test]
    fn tsr_assignment_caps_uses_median_geometry() {
        use crate::tables::StructuredCell;
        let cells = vec![
            StructuredCell {
                row: 0,
                col: 0,
                rowspan: 1,
                colspan: 1,
                is_header: false,
                text: String::new(),
                page_pt_bbox: [0.0, 0.0, 100.0, 20.0], // 100x20
            },
            StructuredCell {
                row: 0,
                col: 1,
                rowspan: 1,
                colspan: 1,
                is_header: false,
                text: String::new(),
                page_pt_bbox: [100.0, 0.0, 200.0, 20.0],
            },
            StructuredCell {
                row: 1,
                col: 0,
                rowspan: 1,
                colspan: 1,
                is_header: false,
                text: String::new(),
                page_pt_bbox: [0.0, 20.0, 100.0, 40.0],
            },
        ];
        let (cap_x, cap_y) = tsr_assignment_caps(&cells);
        assert_eq!(cap_x, 100.0);
        assert_eq!(cap_y, 20.0);
    }

    #[test]
    fn tsr_assignment_caps_floor_protects_degenerate_input() {
        use crate::tables::StructuredCell;
        let cells = vec![StructuredCell {
            row: 0,
            col: 0,
            rowspan: 1,
            colspan: 1,
            is_header: false,
            text: String::new(),
            page_pt_bbox: [0.0, 0.0, 1.0, 1.0],
        }];
        let (cap_x, cap_y) = tsr_assignment_caps(&cells);
        assert_eq!(cap_x, 5.0);
        assert_eq!(cap_y, 5.0);
    }

    #[test]
    fn stage2_recovers_left_aligned_header_text_outside_data_band() {
        // Symptom A reproduction: the column band derived from data-cell
        // centers ends up too far right, so header text positioned at the
        // left of the column falls outside the band and stage 1's strict
        // membership rejects it. Stage 2 should re-attach by proximity.
        //
        // Item coords are bottom-left native; cell page_pt_bbox is top-left.
        // page_height=200 so a top-left bbox y=[88, 100] flips to native y
        // bounds [100, 112]; an item at native y=104 (center 108) lands in.
        use crate::tables::StructuredCell;
        let items = vec![
            // Header text — centered in row 0 (native y=104, center 108) but
            // at the LEFT of the column (x=175, far left of the [410, 700]
            // data-derived band).
            test_item("Address", 175.0, 104.0, 50.0, 8.0),
            // Data row 1 — fits its cell.
            test_item("205 W Oak St", 420.0, 84.0, 100.0, 8.0),
            // Data row 2 — fits its cell.
            test_item("155 E Boardwalk Dr", 420.0, 64.0, 100.0, 8.0),
        ];
        // Cells AFTER normalize_cell_bands would have run — col 0 band
        // shifted right by data-cell centers, header cell now excludes
        // the "Address" text at center x=200.
        let mut cells = vec![
            StructuredCell {
                row: 0,
                col: 0,
                rowspan: 1,
                colspan: 1,
                is_header: true,
                text: String::new(),
                page_pt_bbox: [410.0, 88.0, 700.0, 100.0],
            },
            StructuredCell {
                row: 1,
                col: 0,
                rowspan: 1,
                colspan: 1,
                is_header: false,
                text: String::new(),
                page_pt_bbox: [410.0, 108.0, 700.0, 116.0],
            },
            StructuredCell {
                row: 2,
                col: 0,
                rowspan: 1,
                colspan: 1,
                is_header: false,
                text: String::new(),
                page_pt_bbox: [410.0, 128.0, 700.0, 136.0],
            },
        ];
        let page_h = 200.0;

        // Stage 1 mimic — fill cells via the strict rule, track claimed.
        let mut claimed: std::collections::HashSet<usize> = std::collections::HashSet::new();
        for cell in &mut cells {
            let [x1, y1, x2, y2] = cell.page_pt_bbox;
            let bounds = region_bounds(x1, y1, x2, y2, page_h, RegionCoordSpace::Standard);
            let mut matched: Vec<TextItem> = Vec::new();
            for (i, item) in items.iter().enumerate() {
                if tsr_region_contains_item(item, bounds) {
                    claimed.insert(i);
                    matched.push(item.clone());
                }
            }
            cell.text = collect_text_from_matched_items(matched, 0.10).replace(['\n', '\r'], " ");
        }
        // Header is empty after stage 1 (Address fell outside col 0 band).
        assert_eq!(cells[0].text, "", "header should be empty after stage 1");
        // Data rows already populated.
        assert!(
            cells[1].text.contains("Oak"),
            "data row 1 should contain Oak: got {:?}",
            cells[1].text
        );
        assert!(
            cells[2].text.contains("Boardwalk"),
            "data row 2 should contain Boardwalk: got {:?}",
            cells[2].text
        );

        // Stage 2 should fill the orphan "Address" into the empty header.
        tsr_assign_orphan_items(
            &items,
            &mut cells,
            &claimed,
            page_h,
            RegionCoordSpace::Standard,
        );
        assert_eq!(cells[0].text, "Address");
        // Data rows must NOT have been augmented (already filled by stage 1).
        assert!(!cells[1].text.contains("Address"));
        assert!(!cells[2].text.contains("Address"));
    }

    #[test]
    fn stage2_recovers_y_shifted_col0_in_consecutive_rows() {
        // Symptom B reproduction: a stretch of rows where col 0 cell bboxes
        // sit just above the actual branch-name text. After stage 1 those
        // cells are empty; stage 2 should pull the orphan items in by
        // y-proximity.
        //
        // page_height=800. Cells are 14pt tall in top-left; flipped native
        // bounds are [240,254], [220,234], [200,214]. Items sit ~1pt below
        // each cell's native y range (still within ~1pt of the edge), so
        // both center-containment and 60% overlap fail in stage 1.
        use crate::tables::StructuredCell;
        let items = vec![
            // Bellevue: native y=235, center 239 — just below row 0's
            // cell native bottom (240). Closer to row 0 than row 1.
            test_item("Bellevue", 30.0, 235.0, 45.0, 8.0),
            // Glenwood: native y=215, center 219 — just below row 1.
            test_item("Glenwood", 30.0, 215.0, 45.0, 8.0),
            // Metro Crossing: native y=195, center 199 — just below row 2.
            test_item("Metro Crossing", 30.0, 195.0, 70.0, 8.0),
        ];
        let mut cells = vec![
            StructuredCell {
                row: 0,
                col: 0,
                rowspan: 1,
                colspan: 1,
                is_header: false,
                text: String::new(),
                page_pt_bbox: [10.0, 546.0, 200.0, 560.0],
            },
            StructuredCell {
                row: 1,
                col: 0,
                rowspan: 1,
                colspan: 1,
                is_header: false,
                text: String::new(),
                page_pt_bbox: [10.0, 566.0, 200.0, 580.0],
            },
            StructuredCell {
                row: 2,
                col: 0,
                rowspan: 1,
                colspan: 1,
                is_header: false,
                text: String::new(),
                page_pt_bbox: [10.0, 586.0, 200.0, 600.0],
            },
        ];
        let page_h = 800.0;

        let mut claimed: std::collections::HashSet<usize> = std::collections::HashSet::new();
        for cell in &mut cells {
            let [x1, y1, x2, y2] = cell.page_pt_bbox;
            let bounds = region_bounds(x1, y1, x2, y2, page_h, RegionCoordSpace::Standard);
            let mut matched: Vec<TextItem> = Vec::new();
            for (i, item) in items.iter().enumerate() {
                if tsr_region_contains_item(item, bounds) {
                    claimed.insert(i);
                    matched.push(item.clone());
                }
            }
            cell.text = collect_text_from_matched_items(matched, 0.10).replace(['\n', '\r'], " ");
        }
        // All three cells empty after stage 1 (text falls just below each).
        for c in &cells {
            assert!(
                c.text.is_empty(),
                "stage 1 should leave all cells empty: {:?}",
                c
            );
        }

        tsr_assign_orphan_items(
            &items,
            &mut cells,
            &claimed,
            page_h,
            RegionCoordSpace::Standard,
        );
        assert_eq!(cells[0].text, "Bellevue");
        assert_eq!(cells[1].text, "Glenwood");
        assert_eq!(cells[2].text, "Metro Crossing");
    }

    #[test]
    fn stage2_rejects_cross_line_stacking_into_same_cell() {
        // Two orphans on different rows of the PDF, both equidistant from
        // the same empty cell. Without the same-line guard they'd both stack
        // into that cell ("Shawnee Blue Valley Parkway" run-on); the guard
        // keeps the first orphan and routes the second to the next-nearest
        // empty cell on its own line.
        use crate::tables::StructuredCell;
        // page_h=200. Two empty cells:
        //   cell X (row 0): top-left y=[100, 110], native [90, 100]
        //   cell Y (row 1): top-left y=[112, 122], native [78, 88]
        let mut cells = vec![
            StructuredCell {
                row: 0,
                col: 0,
                rowspan: 1,
                colspan: 1,
                is_header: false,
                text: String::new(),
                page_pt_bbox: [10.0, 100.0, 100.0, 110.0],
            },
            StructuredCell {
                row: 1,
                col: 0,
                rowspan: 1,
                colspan: 1,
                is_header: false,
                text: String::new(),
                page_pt_bbox: [10.0, 112.0, 100.0, 122.0],
            },
        ];
        // Two orphans, different rows of the PDF (y differs by 14pt = a
        // full row), both 2pt outside their target cell — both within
        // cap_y, both equidistant-ish to cell X. Without the same-line
        // guard they'd both land in X.
        //   "Shawnee" should belong to cell X (row 0) — center y=98 is
        //   2pt below X's native min=100.
        //   "BlueValley" should belong to cell Y (row 1) — center y=84
        //   is 4pt above Y's native max=88.
        let items = vec![
            // Shawnee orphan — closer to X (dy=2) than Y (dy=6 from native min=78).
            test_item("Shawnee", 30.0, 94.0, 50.0, 8.0),
            // BlueValley orphan — closer to Y (dy=4) than X (dy=8 from native max=100).
            test_item("BlueValley", 30.0, 80.0, 60.0, 8.0),
        ];
        let claimed: std::collections::HashSet<usize> = std::collections::HashSet::new();

        tsr_assign_orphan_items(
            &items,
            &mut cells,
            &claimed,
            200.0,
            RegionCoordSpace::Standard,
        );
        assert_eq!(cells[0].text, "Shawnee");
        assert_eq!(cells[1].text, "BlueValley");
        assert!(!cells[0].text.contains("BlueValley"));
        assert!(!cells[1].text.contains("Shawnee"));
    }

    #[test]
    fn stage2_allows_same_line_orphans_to_stack_into_one_cell() {
        // Multi-token branch names like "Blue Valley Parkway" are 3 PDF
        // text items at the SAME y-coordinate. They should all stack into
        // the cell their row's branch-name belongs to, not get split
        // across rows by the cross-line guard.
        use crate::tables::StructuredCell;
        let mut cells = vec![StructuredCell {
            row: 0,
            col: 0,
            rowspan: 1,
            colspan: 1,
            is_header: false,
            text: String::new(),
            page_pt_bbox: [10.0, 100.0, 200.0, 110.0],
        }];
        // Three same-line items, all 2pt below the cell's native bottom.
        let items = vec![
            test_item("Blue", 30.0, 94.0, 25.0, 8.0),
            test_item("Valley", 60.0, 94.0, 35.0, 8.0),
            test_item("Parkway", 100.0, 94.0, 45.0, 8.0),
        ];
        let claimed: std::collections::HashSet<usize> = std::collections::HashSet::new();

        tsr_assign_orphan_items(
            &items,
            &mut cells,
            &claimed,
            200.0,
            RegionCoordSpace::Standard,
        );
        // All three same-line orphans stacked into the single empty cell.
        assert_eq!(cells[0].text, "Blue Valley Parkway");
    }

    #[test]
    fn stage2_does_not_overwrite_filled_cells_or_admit_far_orphans() {
        // Stage 2 must only fill EMPTY cells (preserves stage 1's strict
        // behavior on bleed cases) and must reject orphans that fall far
        // outside any cell (prevents pulling a figure title into a table).
        use crate::tables::StructuredCell;
        let items = vec![
            test_item("Real", 50.0, 100.0, 30.0, 8.0),
            // Far orphan — at native y=20 (page bottom edge) on a page where
            // the table sits around native y=92..104 (top-left y=96..108).
            // y-distance to nearest cell is ~70pt, far exceeding the ~12pt
            // cap from median row height.
            test_item("FigureTitle", 50.0, 20.0, 60.0, 8.0),
        ];
        let mut cells = vec![
            StructuredCell {
                row: 0,
                col: 0,
                rowspan: 1,
                colspan: 1,
                is_header: false,
                text: String::new(),
                page_pt_bbox: [40.0, 96.0, 100.0, 108.0],
            },
            StructuredCell {
                row: 1,
                col: 0,
                rowspan: 1,
                colspan: 1,
                is_header: false,
                text: "Pre-filled".to_string(),
                page_pt_bbox: [40.0, 116.0, 100.0, 128.0],
            },
        ];
        let mut claimed: std::collections::HashSet<usize> = std::collections::HashSet::new();
        // Pretend "Real" got claimed by a different cell (won't be re-assigned).
        // Don't claim "FigureTitle" — it's the far orphan.
        claimed.insert(0);

        tsr_assign_orphan_items(
            &items,
            &mut cells,
            &claimed,
            200.0,
            RegionCoordSpace::Standard,
        );
        // Empty cell stayed empty (orphan was too far).
        assert_eq!(cells[0].text, "");
        // Pre-filled cell was not touched.
        assert_eq!(cells[1].text, "Pre-filled");
    }

    // -- recover_startxref_pointer / find_last_valid_xref_table_start ------
    //
    // Direct unit tests on the byte-level scan, addressing review feedback
    // on #230: a coincidental standalone "xref" token that isn't actually
    // followed by a subsection header (start-id + count) must not be
    // treated as a real table — accepting it would let lopdf "succeed"
    // against a bogus offset and silently return garbled/empty content
    // instead of a clean error.

    #[test]
    fn find_xref_rejects_standalone_token_without_subsection_header() {
        // "xref" appears as a real standalone word, but nothing that looks
        // like "<start-id> <count>" follows it.
        let buf = b"Please refer to the xref appendix for details.";
        assert_eq!(find_last_valid_xref_table_start(buf), None);
    }

    #[test]
    fn find_xref_accepts_real_classic_table_header() {
        let buf = b"garbage\nxref\n0 6\n0000000000 65535 f \n%%EOF";
        let pos = find_last_valid_xref_table_start(buf).expect("should find the real table");
        assert_eq!(&buf[pos..pos + 4], b"xref");
        assert_eq!(&buf[pos..], b"xref\n0 6\n0000000000 65535 f \n%%EOF");
    }

    #[test]
    fn find_xref_skips_coincidental_match_and_finds_real_table_before_it() {
        // A coincidental "xref" (no subsection header) appears *after* the
        // real table in the buffer — the scan must not stop at the first
        // (rightmost) standalone token it finds; it must keep looking
        // backward until one actually validates.
        let buf = b"xref\n0 3\n0000000000 65535 f \ntrailer\nsee the xref\n";
        let pos = find_last_valid_xref_table_start(buf).expect("should find the real table");
        assert_eq!(pos, 0);
    }

    #[test]
    fn find_xref_rejects_substring_of_startxref() {
        // "xref" is a substring of "startxref" but isn't a standalone
        // token there (not preceded by whitespace) — must not match, even
        // though a number immediately follows it.
        let buf = b"startxref\n1234\n%%EOF";
        assert_eq!(find_last_valid_xref_table_start(buf), None);
    }

    #[test]
    fn find_xref_rejects_count_run_with_trailing_garbage() {
        // "xref\n0 6garbage" has the right shape (digits, whitespace,
        // digits) but the count run doesn't end at whitespace/EOF — it
        // runs straight into non-digit garbage, so this must not be
        // accepted as a real subsection header.
        let buf = b"xref\n0 6garbage\n%%EOF";
        assert_eq!(find_last_valid_xref_table_start(buf), None);
    }

    #[test]
    fn recover_startxref_pointer_returns_none_without_a_valid_table() {
        let buf = b"Please refer to the xref appendix for details.";
        assert!(recover_startxref_pointer(buf).is_none());
    }
}

#[cfg(test)]
mod rotated_run_region_tests {
    use super::*;
    use crate::types::ItemType;

    fn item(text: &str, x: f32, y: f32, width: f32, height: f32, rotation: f32) -> TextItem {
        TextItem {
            baseline_shift: 0.0,
            text: text.to_string(),
            x,
            y,
            width,
            height,
            rotation,
            advance_known: true,
            font: "Helvetica".to_string(),
            font_tag: "F1".to_string(),
            legacy_symbol_rewrite: false,
            font_size: if rotation == 0.0 { height } else { width },
            page: 1,
            is_bold: false,
            is_italic: false,
            font_weight: None,
            bold_source: None,
            fixed_pitch: None,
            fill_color: None,
            stroke_color: None,
            render_mode: None,
            is_underline: false,
            is_strikeout: false,
            item_type: ItemType::Text,
            mcid: None,
        }
    }

    const STAMP: &str = "arXiv:2301.00001v1 [cs.CL] 1 Jan 2023";

    #[test]
    fn vertical_margin_run_overlaps_only_the_margin_region() {
        // Letter page; region boxes are top-left page coordinates as the
        // layout model reports them: a left-margin strip and the body area.
        let page_h = 792.0;
        let margin = region_bounds(0.0, 0.0, 50.0, 792.0, page_h, RegionCoordSpace::Standard);
        let body = region_bounds(60.0, 0.0, 612.0, 792.0, page_h, RegionCoordSpace::Standard);

        let stamp = item(STAMP, 12.0, 200.0, 20.0, 400.0, 90.0);
        assert!(region_overlaps_item(&stamp, margin));
        assert!(!region_overlaps_item(&stamp, body));
        assert!(region_item_overlap_area(&stamp, margin) > 0.0);
        assert_eq!(region_item_overlap_area(&stamp, body), 0.0);
    }

    #[test]
    fn collect_text_in_region_finds_the_vertical_run_by_its_box() {
        let items = vec![
            item(STAMP, 12.0, 200.0, 20.0, 400.0, 90.0),
            item("Body text", 72.0, 500.0, 50.0, 11.0, 0.0),
        ];
        assert_eq!(
            collect_text_in_region(&items, 0.0, 0.0, 50.0, 792.0, 792.0),
            STAMP
        );
        assert_eq!(
            collect_text_in_region(&items, 60.0, 0.0, 612.0, 792.0, 792.0),
            "Body text"
        );
    }

    #[test]
    fn clockwise_frame_region_bounds_follow_the_turned_page() {
        // A 270° header along the right margin of a Letter page: page box
        // x ∈ [580, 590], y ∈ [664, 700]. In the clockwise-turned frame it
        // sits at x = -(top), y = baseline x.
        let page_h = 792.0;
        let header = item("HEADER", -700.0, 580.0, 36.0, 10.0, 0.0);
        // Top-left region boxes: a strip around the header and the body.
        let strip = region_bounds(
            570.0,
            80.0,
            600.0,
            140.0,
            page_h,
            RegionCoordSpace::Rotated90Cw,
        );
        let body = region_bounds(
            60.0,
            80.0,
            500.0,
            700.0,
            page_h,
            RegionCoordSpace::Rotated90Cw,
        );
        assert!(region_overlaps_item(&header, strip));
        assert!(!region_overlaps_item(&header, body));

        // The legacy heuristic recognises only the counter-clockwise frame
        // (most items at negative y); a clockwise frame is never inferred and
        // must be passed explicitly.
        assert_eq!(
            infer_region_coord_space(&[header.clone()]),
            RegionCoordSpace::Standard
        );
        assert_eq!(
            infer_region_coord_space(&[item("x", 100.0, -200.0, 36.0, 12.0, 0.0)]),
            RegionCoordSpace::Rotated90Ccw
        );
        assert_eq!(
            infer_region_coord_space(&[item("x", 100.0, 200.0, 36.0, 12.0, 0.0)]),
            RegionCoordSpace::Standard
        );
        let strip = collect_text_in_region_in_frame(
            &[header.clone()],
            570.0,
            80.0,
            600.0,
            140.0,
            page_h,
            PageRotation::Cw,
        );
        assert_eq!(strip, "HEADER");
        let misread = collect_text_in_region_in_frame(
            &[header.clone()],
            570.0,
            80.0,
            600.0,
            140.0,
            page_h,
            PageRotation::Upright,
        );
        assert_eq!(misread, "");

        // Bounds round-trip back to the top-left page box they came from.
        for coords in [
            RegionCoordSpace::Standard,
            RegionCoordSpace::Rotated90Ccw,
            RegionCoordSpace::Rotated90Cw,
        ] {
            let b = region_bounds(570.0, 80.0, 600.0, 140.0, page_h, coords);
            let back = extracted_bbox_to_page_top_left(
                [b.x_min, b.y_min, b.x_max, b.y_max],
                page_h,
                coords,
            );
            assert_eq!(back, [570.0, 80.0, 600.0, 140.0], "{coords:?}");
        }
    }

    #[test]
    fn tsr_line_clustering_uses_the_estimated_height_of_unknown_advances() {
        let estimated = STAMP.chars().count() as f32 * 0.5 * 20.0;
        let mut stamp = item(STAMP, 12.0, 200.0, 20.0, estimated, 90.0);
        stamp.advance_known = false;
        let lines = cluster_tsr_cell_text_lines(vec![stamp]);
        assert_eq!(lines.len(), 1);
        // Half the estimated extent, not the 2.5pt floor.
        assert!(
            (lines[0].half_height - estimated * 0.5).abs() < 1e-3,
            "{}",
            lines[0].half_height
        );
    }

    #[test]
    fn vertical_run_without_font_widths_still_overlaps_its_region() {
        // No width information: the advance is unknown, so the run's box is
        // one em wide and zero tall. The estimated height must keep it
        // matchable instead of letting it vanish from region extraction.
        let page_h = 792.0;
        let margin = region_bounds(0.0, 0.0, 50.0, 792.0, page_h, RegionCoordSpace::Standard);
        // Extraction lays a half-em-per-glyph estimate along the run for a
        // width-less font and flags it (the content-stream tests cover the
        // parser side); region matching sees that box like any other.
        let estimated = STAMP.chars().count() as f32 * 0.5 * 20.0;
        let mut stamp = item(STAMP, 12.0, 200.0, 20.0, estimated, 90.0);
        stamp.advance_known = false;
        assert!(region_overlaps_item(&stamp, margin));
        assert!(region_item_overlap_area(&stamp, margin) > 0.0);
        assert!(tsr_region_contains_item(&stamp, margin));
    }

    #[test]
    fn token_splitting_leaves_rotated_runs_whole() {
        let stamp = item("two words", 12.0, 200.0, 20.0, 100.0, 90.0);
        let tokens = split_item_into_token_subitems(&stamp);
        assert_eq!(tokens.len(), 1);
        assert_eq!(tokens[0].text, "two words");
        assert_eq!((tokens[0].x, tokens[0].width), (12.0, 20.0));

        // Upside-down runs advance towards -x: interpolating from the left
        // edge would swap the tokens' cells.
        let flipped = item("two words", 72.0, 500.0, 90.0, 10.0, 180.0);
        assert_eq!(split_item_into_token_subitems(&flipped).len(), 1);

        let body = item("two words", 72.0, 500.0, 90.0, 10.0, 0.0);
        assert_eq!(split_item_into_token_subitems(&body).len(), 2);
    }

    #[test]
    fn token_splitting_spreads_only_an_unknown_zero_width_run() {
        // A known zero advance (an ActualText replacement that painted
        // nothing) keeps its tokens where the run was drawn; a run whose
        // advance is unknown is spread over half an em per character.
        let zero = item("two words", 72.0, 500.0, 0.0, 10.0, 0.0);
        let tokens = split_item_into_token_subitems(&zero);
        assert_eq!(tokens.len(), 2);
        assert_eq!((tokens[0].x, tokens[1].x), (72.0, 72.0));

        let mut unknown = zero;
        unknown.advance_known = false;
        let tokens = split_item_into_token_subitems(&unknown);
        assert_eq!(tokens.len(), 2);
        assert!(tokens[1].x > tokens[0].x + 15.0, "{tokens:?}");
    }
}