fix(vision): preserve adaptive OCR fallbacks

docs/rust-api.md

@@ -380,6 +380,22 @@ provides the same native-only behavior through the OCR result/provenance
 shape; `Force` renders every selected page. Learned layout intentionally
 returns an explicit unsupported error in this lightweight pipeline.
 
+For ambiguous mixed pages, `Auto` privately retains clean native fragments
+instead of discarding them when OCR is selected. After recognition it compares
+script-agnostic text quality, OCR confidence, character overlap, and material
+new coverage. Exact native text wins over a duplicate or weak OCR hypothesis;
+complementary image-backed text is fused; and pages where both candidates are
+weak recommend the hosted document pipeline. Public native-only extraction
+continues to suppress pages marked unreliable, and clean text documents pay no
+renderer or model-initialization cost.
+
+In `Auto`, pages routed only for suspicious font encoding or vectorized text
+first get a bounded positioned-text probe through PDFium. A credible recovered
+text layer with sufficient geometric page coverage skips rasterization and
+model loading for that page; garbled, partial, or insubstantial recovery
+continues through OCR. Recovered tables are reflected in the same document
+metadata as tables found by the primary extractor.
+
 The one-call API keeps the most recently used verified OCR engine in process.
 Long-lived workers therefore verify the pinned artifacts and build the ONNX
 sessions once, then reuse those loaded sessions across documents. The cache is

src/lib.rs

@@ -459,8 +459,15 @@ 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)
-        .map(|(result, _)| result)
+    extract_pages_markdown_mem_impl(
+        buffer,
+        pages,
+        None,
+        &MarkdownOptions::default(),
+        false,
+        false,
+    )
+    .map(|(result, _)| result)
 }
 
 #[cfg(all(feature = "ocr", not(target_arch = "wasm32")))]
@@ -476,6 +483,7 @@ pub(crate) fn extract_pages_markdown_mem_for_ocr(
         password,
         markdown_options,
         markdown_options.strip_headers_footers,
+        true,
     )
 }
 
@@ -485,6 +493,7 @@ fn extract_pages_markdown_mem_impl(
     password: Option<&str>,
     markdown_options: &MarkdownOptions,
     strip_repeated_headers_footers: bool,
+    preserve_ocr_candidates: bool,
 ) -> Result<(PagesExtractionResult, u32), PdfError> {
     validate_pdf_bytes(buffer)?;
     let (doc, page_count) = load_document_from_mem_with_password(buffer, password)?;
@@ -663,7 +672,15 @@ fn extract_pages_markdown_mem_impl(
 
         results.push(PageMarkdown {
             page: page_0idx,
-            markdown: if needs_ocr { String::new() } else { md },
+            // 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,
         });

src/vision/fusion.rs

@@ -6,6 +6,7 @@ use std::time::Instant;
 use thiserror::Error;
 
 use crate::markdown::{to_markdown_from_items_with_rects_and_page_count, MarkdownOptions};
+use crate::text_quality::{detect_encoding_issues, is_cid_garbage, is_garbage_text};
 use crate::types::{ItemType, TextItem};
 use crate::PageMarkdown;
 
@@ -91,6 +92,73 @@ pub struct FusedPages {
     pub ocr_time_ms: u64,
 }
 
+/// Origin of a trustworthy native-text candidate retained for adaptive OCR.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub(crate) enum NativeCandidateOrigin {
+    /// Text produced by pdf-inspector's normal native extractor.
+    Extractor,
+    /// Positioned text independently recovered through PDFium.
+    Pdfium,
+}
+
+impl NativeCandidateOrigin {
+    fn description(self) -> &'static str {
+        match self {
+            Self::Extractor => "native extraction",
+            Self::Pdfium => "PDFium native recovery",
+        }
+    }
+}
+
+#[derive(Debug, Clone, Copy)]
+struct TextCandidateQuality {
+    alphanumeric_chars: usize,
+    score: f32,
+}
+
+/// Clean native text retained while an ambiguous page is compared with OCR.
+#[derive(Debug, Clone)]
+pub(crate) struct NativeFallbackCandidate {
+    markdown: String,
+    quality: TextCandidateQuality,
+    origin: NativeCandidateOrigin,
+}
+
+impl NativeFallbackCandidate {
+    /// True when an independent native recovery is substantial enough to
+    /// cancel OCR for recoverable font/vector routing reasons.
+    pub(crate) fn is_complete_recovery(&self) -> bool {
+        self.quality.alphanumeric_chars >= 40 && self.quality.score >= 0.68
+    }
+
+    pub(crate) fn markdown(&self) -> &str {
+        &self.markdown
+    }
+
+    pub(crate) fn is_stronger_than(&self, other: &Self) -> bool {
+        self.quality.alphanumeric_chars > other.quality.alphanumeric_chars
+            || (self.quality.alphanumeric_chars == other.quality.alphanumeric_chars
+                && self.quality.score > other.quality.score)
+    }
+}
+
+/// Validates and scores native Markdown for possible post-OCR comparison.
+///
+/// This intentionally uses script-agnostic evidence. A native candidate only
+/// needs to be trustworthy, not necessarily complete: a clean native header
+/// can still be fused with an image-backed OCR body.
+pub(crate) fn assess_native_candidate(
+    markdown: String,
+    origin: NativeCandidateOrigin,
+) -> Option<NativeFallbackCandidate> {
+    let quality = assess_text_candidate(&markdown)?;
+    (quality.alphanumeric_chars >= 8).then_some(NativeFallbackCandidate {
+        markdown,
+        quality,
+        origin,
+    })
+}
+
 /// Converts positioned OCR spans to Markdown through pdf-inspector's existing
 /// deterministic geometry, reading-order, table, and Markdown pipeline.
 ///
@@ -116,13 +184,46 @@ pub fn ocr_page_to_markdown(
 /// OCR replaces pages whose native extraction was already rejected. On clean
 /// native pages (for example in `Force` mode), normalized duplicate OCR blocks
 /// are removed and only genuinely additional blocks are appended. Pages that
-/// needed OCR but still have no credible local result recommend the hosted
+/// needed OCR but still have no credible OCR result recommend the hosted
 /// document pipeline instead of silently presenting an empty result as final.
 pub fn fuse_ocr_pages(
     native_pages: &[PageMarkdown],
     ocr_run: &OcrRun,
     document_page_count: u32,
     options: &OcrFusionOptions,
+) -> Result<FusedPages, OcrFusionError> {
+    fuse_ocr_pages_impl(
+        native_pages,
+        ocr_run,
+        document_page_count,
+        options,
+        &BTreeMap::new(),
+    )
+}
+
+/// OCR-pipeline fusion with trustworthy partial native candidates.
+pub(crate) fn fuse_ocr_pages_adaptive(
+    native_pages: &[PageMarkdown],
+    ocr_run: &OcrRun,
+    document_page_count: u32,
+    options: &OcrFusionOptions,
+    native_candidates: &BTreeMap<u32, NativeFallbackCandidate>,
+) -> Result<FusedPages, OcrFusionError> {
+    fuse_ocr_pages_impl(
+        native_pages,
+        ocr_run,
+        document_page_count,
+        options,
+        native_candidates,
+    )
+}
+
+fn fuse_ocr_pages_impl(
+    native_pages: &[PageMarkdown],
+    ocr_run: &OcrRun,
+    document_page_count: u32,
+    options: &OcrFusionOptions,
+    native_candidates: &BTreeMap<u32, NativeFallbackCandidate>,
 ) -> Result<FusedPages, OcrFusionError> {
     validate_options(options)?;
 
@@ -178,16 +279,31 @@ pub fn fuse_ocr_pages(
                     document_page_count,
                 );
                 let ocr_markdown = preserve_ocr_line_breaks(&ocr_markdown, local);
-                let (markdown, source) = if native.markdown.trim().is_empty() || native.needs_ocr {
-                    (ocr_markdown, PageContentSource::Ocr)
+                let (markdown, source, adaptive_recommends_hosted) = if let Some(candidate) =
+                    native_candidates
+                        .get(&page_number)
+                        .filter(|_| native.needs_ocr)
+                {
+                    let choice = choose_adaptive_content(
+                        candidate,
+                        &ocr_markdown,
+                        local.ocr.mean_confidence,
+                        options.hosted_recommendation_confidence,
+                    );
+                    warnings.push(choice.warning);
+                    (choice.markdown, choice.source, choice.recommend_hosted)
+                } else if native.markdown.trim().is_empty() || native.needs_ocr {
+                    (ocr_markdown, PageContentSource::Ocr, false)
                 } else {
-                    merge_native_and_ocr(&native.markdown, &ocr_markdown)
+                    let (markdown, source) = merge_native_and_ocr(&native.markdown, &ocr_markdown);
+                    (markdown, source, false)
                 };
                 let weak_ocr = local
                     .ocr
                     .mean_confidence
                     .is_none_or(|confidence| confidence < options.hosted_recommendation_confidence);
-                let recommend_hosted = native.needs_ocr && (markdown.trim().is_empty() || weak_ocr);
+                let recommend_hosted = native.needs_ocr
+                    && (markdown.trim().is_empty() || weak_ocr || adaptive_recommends_hosted);
                 if native.needs_ocr && markdown.trim().is_empty() {
                     warnings.push("OCR produced no usable text".to_string());
                 }
@@ -244,6 +360,155 @@ pub fn fuse_ocr_pages(
     })
 }
 
+struct AdaptiveContentChoice {
+    markdown: String,
+    source: PageContentSource,
+    warning: String,
+    recommend_hosted: bool,
+}
+
+fn choose_adaptive_content(
+    native: &NativeFallbackCandidate,
+    ocr: &str,
+    ocr_confidence: Option<f32>,
+    weak_ocr_threshold: f32,
+) -> AdaptiveContentChoice {
+    let ocr_quality = assess_text_candidate(ocr);
+    let weak_ocr = ocr_confidence.is_none_or(|confidence| confidence < weak_ocr_threshold);
+    if weak_ocr || ocr_quality.is_none() {
+        return AdaptiveContentChoice {
+            markdown: ensure_trailing_newline(native.markdown()),
+            source: PageContentSource::Native,
+            warning: format!(
+                "kept trustworthy {} because OCR was weak or unusable",
+                native.origin.description()
+            ),
+            recommend_hosted: true,
+        };
+    }
+
+    let ocr_quality = ocr_quality.expect("checked above");
+    let overlap = content_overlap(native.markdown(), ocr);
+    let ocr_novel_chars = ocr_quality
+        .alphanumeric_chars
+        .saturating_sub(overlap.shared_chars);
+    let material_novelty =
+        ocr_novel_chars >= 2 && ocr_novel_chars * 8 >= ocr_quality.alphanumeric_chars.max(1);
+    let native_substantially_covered =
+        overlap.shared_chars * 4 >= native.quality.alphanumeric_chars.max(1) * 3;
+
+    if native_substantially_covered && !material_novelty {
+        return AdaptiveContentChoice {
+            markdown: ensure_trailing_newline(native.markdown()),
+            source: PageContentSource::Native,
+            warning: format!(
+                "kept trustworthy {} because OCR added no material coverage",
+                native.origin.description()
+            ),
+            recommend_hosted: false,
+        };
+    }
+
+    // A shorter, materially lower-quality OCR hypothesis should not displace
+    // exact native text even when its mean confidence happens to be high.
+    if ocr_quality.score + 0.12 < native.quality.score
+        && ocr_quality.alphanumeric_chars * 10
+            <= native.quality.alphanumeric_chars.saturating_mul(11)
+    {
+        return AdaptiveContentChoice {
+            markdown: ensure_trailing_newline(native.markdown()),
+            source: PageContentSource::Native,
+            warning: format!(
+                "kept higher-quality {} after comparing OCR",
+                native.origin.description()
+            ),
+            recommend_hosted: false,
+        };
+    }
+
+    let (markdown, source) = merge_native_and_ocr(native.markdown(), ocr);
+    let warning = match source {
+        PageContentSource::Native => format!(
+            "kept trustworthy {} because OCR duplicated its content",
+            native.origin.description()
+        ),
+        PageContentSource::Fused => format!(
+            "fused trustworthy {} with complementary OCR",
+            native.origin.description()
+        ),
+        PageContentSource::Ocr => unreachable!("merge never returns OCR-only content"),
+    };
+    AdaptiveContentChoice {
+        markdown,
+        source,
+        warning,
+        recommend_hosted: false,
+    }
+}
+
+#[derive(Debug, Clone, Copy)]
+struct ContentOverlap {
+    shared_chars: usize,
+}
+
+fn content_overlap(first: &str, second: &str) -> ContentOverlap {
+    let mut first_counts = BTreeMap::<char, usize>::new();
+    for character in normalized_content_chars(first) {
+        *first_counts.entry(character).or_insert(0) += 1;
+    }
+    let mut second_counts = BTreeMap::<char, usize>::new();
+    for character in normalized_content_chars(second) {
+        *second_counts.entry(character).or_insert(0) += 1;
+    }
+    let shared_chars = first_counts
+        .iter()
+        .map(|(character, count)| (*count).min(second_counts.get(character).copied().unwrap_or(0)))
+        .sum();
+    ContentOverlap { shared_chars }
+}
+
+fn normalized_content_chars(text: &str) -> impl Iterator<Item = char> + '_ {
+    text.chars()
+        .flat_map(char::to_lowercase)
+        .filter(|character| character.is_alphanumeric())
+}
+
+fn assess_text_candidate(markdown: &str) -> Option<TextCandidateQuality> {
+    if markdown.trim().is_empty()
+        || is_garbage_text(markdown)
+        || is_cid_garbage(markdown)
+        || detect_encoding_issues(markdown)
+    {
+        return None;
+    }
+
+    let alphanumeric_chars = markdown
+        .chars()
+        .filter(|character| character.is_alphanumeric())
+        .count();
+    if alphanumeric_chars == 0 {
+        return None;
+    }
+    let visible_chars = markdown
+        .chars()
+        .filter(|character| !character.is_whitespace())
+        .count()
+        .max(1);
+    let density = alphanumeric_chars as f32 / visible_chars as f32;
+    let length_score = (alphanumeric_chars as f32 / 160.0).min(1.0);
+    let nonempty_lines = markdown
+        .lines()
+        .filter(|line| !line.trim().is_empty())
+        .count()
+        .max(1);
+    let line_score = (alphanumeric_chars as f32 / nonempty_lines as f32 / 12.0).min(1.0);
+    let score = (0.45 + length_score * 0.25 + density * 0.20 + line_score * 0.10).min(1.0);
+    Some(TextCandidateQuality {
+        alphanumeric_chars,
+        score,
+    })
+}
+
 /// Converts recognized line polygons to ordinary PDF-space text items.
 fn ocr_text_items(page: &RoutedOcrPage) -> (Vec<TextItem>, usize) {
     let mut discarded = 0usize;
@@ -778,6 +1043,10 @@ mod tests {
         }
     }
 
+    fn native_candidate(markdown: &str) -> NativeFallbackCandidate {
+        assess_native_candidate(markdown.to_string(), NativeCandidateOrigin::Extractor).unwrap()
+    }
+
     #[test]
     fn scanned_page_uses_geometry_ordered_ocr_and_provenance() {
         let native = [native(0, "", true)];
@@ -930,6 +1199,133 @@ mod tests {
         assert_eq!(result.pages[0].provenance.source, PageContentSource::Native);
     }
 
+    #[test]
+    fn adaptive_fallback_keeps_native_text_when_ocr_is_weak() {
+        let native = [native(0, "", true)];
+        let run = run(vec![routed_page(
+            1,
+            vec![span("Inv0ice total uncertain", 10.0, 0.3)],
+            Some(0.3),
+        )]);
+        let candidates = BTreeMap::from([(
+            1,
+            native_candidate("Invoice total: $420.00\nPayment received\n"),
+        )]);
+
+        let result =
+            fuse_ocr_pages_adaptive(&native, &run, 1, &OcrFusionOptions::new(), &candidates)
+                .unwrap();
+
+        assert_eq!(result.pages[0].provenance.source, PageContentSource::Native);
+        assert_eq!(
+            result.pages[0].markdown,
+            "Invoice total: $420.00\nPayment received\n"
+        );
+        assert!(result.pages[0].provenance.hosted_recommended);
+        assert!(result.pages[0].provenance.warnings[0].contains("OCR was weak"));
+    }
+
+    #[test]
+    fn adaptive_fallback_prefers_exact_native_text_over_duplicate_ocr() {
+        let native = [native(0, "", true)];
+        let run = run(vec![routed_page(
+            1,
+            vec![span("Invoice total 420.00 Payment received", 10.0, 0.98)],
+            Some(0.98),
+        )]);
+        let candidates = BTreeMap::from([(
+            1,
+            native_candidate("Invoice total: $420.00\nPayment received\n"),
+        )]);
+
+        let result =
+            fuse_ocr_pages_adaptive(&native, &run, 1, &OcrFusionOptions::new(), &candidates)
+                .unwrap();
+
+        assert_eq!(result.pages[0].provenance.source, PageContentSource::Native);
+        assert!(!result.pages[0].provenance.hosted_recommended);
+        assert_eq!(result.pages[0].markdown.matches("Invoice").count(), 1);
+        assert!(result.pages[0].provenance.warnings[0].contains("no material coverage"));
+    }
+
+    #[test]
+    fn adaptive_fallback_fuses_short_novel_ocr_content() {
+        let native = [native(0, "", true)];
+        let run = run(vec![routed_page(
+            1,
+            vec![span("Status ready 42", 10.0, 0.98)],
+            Some(0.98),
+        )]);
+        let candidates = BTreeMap::from([(1, native_candidate("Status ready\n"))]);
+
+        let result =
+            fuse_ocr_pages_adaptive(&native, &run, 1, &OcrFusionOptions::new(), &candidates)
+                .unwrap();
+
+        assert_eq!(result.pages[0].provenance.source, PageContentSource::Fused);
+        assert!(result.pages[0].markdown.contains("42"));
+        assert!(!result.pages[0].provenance.hosted_recommended);
+    }
+
+    #[test]
+    fn adaptive_fallback_recommends_hosted_for_high_confidence_garbage_ocr() {
+        let native = [native(0, "", true)];
+        let garbage = "a@@b%%c&&d==e~~".repeat(12);
+        let run = run(vec![routed_page(
+            1,
+            vec![span(&garbage, 10.0, 0.99)],
+            Some(0.99),
+        )]);
+        let candidates = BTreeMap::from([(1, native_candidate("Invoice total 420\n"))]);
+
+        let result =
+            fuse_ocr_pages_adaptive(&native, &run, 1, &OcrFusionOptions::new(), &candidates)
+                .unwrap();
+
+        assert_eq!(result.pages[0].provenance.source, PageContentSource::Native);
+        assert!(result.pages[0].provenance.hosted_recommended);
+        assert!(!result.pages[0].markdown.contains("@@"));
+    }
+
+    #[test]
+    fn adaptive_fallback_fuses_native_header_with_scanned_body() {
+        let native = [native(0, "", true)];
+        let run = run(vec![routed_page(
+            1,
+            vec![
+                span("Quarterly account report", 10.0, 0.96),
+                span("March revenue 420 units", 30.0, 0.96),
+                span("April revenue 510 units", 50.0, 0.96),
+            ],
+            Some(0.96),
+        )]);
+        let candidates = BTreeMap::from([(1, native_candidate("# Quarterly account report\n"))]);
+
+        let result =
+            fuse_ocr_pages_adaptive(&native, &run, 1, &OcrFusionOptions::new(), &candidates)
+                .unwrap();
+
+        assert_eq!(result.pages[0].provenance.source, PageContentSource::Fused);
+        assert_eq!(result.pages[0].markdown.matches("Quarterly").count(), 1);
+        assert!(result.pages[0].markdown.contains("March revenue 420 units"));
+        assert!(result.pages[0].markdown.contains("April revenue 510 units"));
+        assert!(result.pages[0].provenance.warnings[0].contains("complementary"));
+    }
+
+    #[test]
+    fn native_candidate_scoring_is_multilingual_and_rejects_garbage() {
+        assert!(assess_native_candidate(
+            "請求書 合計金額 4200円\n支払済み\n".to_string(),
+            NativeCandidateOrigin::Extractor,
+        )
+        .is_some());
+        assert!(assess_native_candidate(
+            "a@@b%%c&&d==e~~".repeat(12),
+            NativeCandidateOrigin::Extractor,
+        )
+        .is_none());
+    }
+
     #[test]
     fn force_mode_appends_only_additional_ocr_blocks() {
         let native = [native(0, "Native title\n", false)];

src/vision/pdfium.rs

@@ -2,9 +2,11 @@
 
 use std::path::Path;
 
-use firecrawl_pdfium::{Pdfium, PixelFormat, PixelPoint, RenderConfig};
+use firecrawl_pdfium::{PageChar, Pdfium, PixelFormat, PixelPoint, RenderConfig};
 use thiserror::Error;
 
+use crate::types::{ItemType, TextItem};
+
 use super::{
     PageRenderer, PageTransform, RenderBufferError, RenderOptions, RenderPixelFormat, RenderedPage,
 };
@@ -65,6 +67,15 @@ pub struct PdfiumRenderer {
     pdfium: Pdfium,
 }
 
+/// Positioned native text recovered from one selected PDF page.
+#[derive(Debug)]
+pub(crate) struct PdfiumTextPage {
+    pub(crate) page: u32,
+    pub(crate) page_width: f32,
+    pub(crate) page_height: f32,
+    pub(crate) items: Vec<TextItem>,
+}
+
 impl PdfiumRenderer {
     /// Loads PDFium using `firecrawl-pdfium`'s documented discovery chain.
     pub fn load() -> Result<Self, RenderError> {
@@ -100,6 +111,56 @@ impl PdfiumRenderer {
         self.render_pages_impl(pdf_bytes, pages, password, options)
     }
 
+    /// Extracts positioned native text from selected 1-indexed pages.
+    ///
+    /// This is deliberately separate from rendering: callers can probe a
+    /// suspicious embedded text layer before paying for rasterization and
+    /// OCR. A page-level text failure is treated as an unavailable recovery
+    /// candidate so the caller can continue to its normal OCR fallback.
+    pub(crate) fn extract_text_pages(
+        &self,
+        pdf_bytes: &[u8],
+        pages: &[u32],
+        password: Option<&str>,
+    ) -> Result<Vec<PdfiumTextPage>, RenderError> {
+        const MAX_TEXT_CHARS_PER_PAGE: usize = 250_000;
+
+        if pages.is_empty() {
+            return Ok(Vec::new());
+        }
+        if pages.contains(&0) {
+            return Err(RenderError::InvalidPageNumber);
+        }
+
+        let document = self.pdfium.load_document(pdf_bytes.to_vec(), password)?;
+        let page_count = document.page_count();
+        if let Some(&page) = pages.iter().find(|&&page| page as usize > page_count) {
+            return Err(RenderError::PageOutOfBounds { page, page_count });
+        }
+
+        let mut recovered = Vec::with_capacity(pages.len());
+        for &page_number in pages {
+            let page = document.page(page_number as usize - 1)?;
+            let page_size = page.size();
+            let text = match page.text_with_limit(MAX_TEXT_CHARS_PER_PAGE) {
+                Ok(text) => text,
+                Err(error) => {
+                    log::debug!(
+                        "page {page_number}: positioned native text recovery unavailable: {error}"
+                    );
+                    continue;
+                }
+            };
+            recovered.push(PdfiumTextPage {
+                page: page_number,
+                page_width: page_size.width,
+                page_height: page_size.height,
+                items: text_chars_to_items(text.chars(), page_number),
+            });
+        }
+        Ok(recovered)
+    }
+
     fn render_pages_impl(
         &self,
         pdf_bytes: &[u8],
@@ -145,6 +206,106 @@ impl PdfiumRenderer {
     }
 }
 
+fn text_chars_to_items(chars: &[PageChar], page: u32) -> Vec<TextItem> {
+    #[derive(Debug, Clone, Copy)]
+    struct Bounds {
+        left: f64,
+        bottom: f64,
+        right: f64,
+        top: f64,
+    }
+
+    fn flush(items: &mut Vec<TextItem>, text: &mut String, bounds: &mut Option<Bounds>, page: u32) {
+        let Some(bounds) = bounds.take() else {
+            text.clear();
+            return;
+        };
+        if text.is_empty() {
+            return;
+        }
+        let width = (bounds.right - bounds.left) as f32;
+        let height = (bounds.top - bounds.bottom) as f32;
+        let x = bounds.left as f32;
+        let y = bounds.bottom as f32;
+        if !x.is_finite()
+            || !y.is_finite()
+            || !width.is_finite()
+            || !height.is_finite()
+            || width <= 0.0
+            || height <= 0.0
+        {
+            text.clear();
+            return;
+        }
+        items.push(TextItem {
+            text: std::mem::take(text),
+            x,
+            y,
+            width,
+            height,
+            font: "PDFium native text".to_string(),
+            font_size: height.max(1.0),
+            page,
+            is_bold: false,
+            is_italic: false,
+            is_underline: false,
+            is_strikeout: false,
+            item_type: ItemType::Text,
+            mcid: None,
+        });
+    }
+
+    let mut items = Vec::new();
+    let mut text = String::new();
+    let mut bounds: Option<Bounds> = None;
+    for character in chars {
+        let Some(value) = character.unicode else {
+            flush(&mut items, &mut text, &mut bounds, page);
+            continue;
+        };
+        if value.is_whitespace() {
+            flush(&mut items, &mut text, &mut bounds, page);
+            continue;
+        }
+
+        let rect = character.loose_bounds.normalized();
+        if !rect.left.is_finite()
+            || !rect.bottom.is_finite()
+            || !rect.right.is_finite()
+            || !rect.top.is_finite()
+            || rect.width() <= 0.0
+            || rect.height() <= 0.0
+        {
+            flush(&mut items, &mut text, &mut bounds, page);
+            continue;
+        }
+        text.push(value);
+        bounds = Some(match bounds {
+            Some(bounds) => Bounds {
+                left: bounds.left.min(rect.left),
+                bottom: bounds.bottom.min(rect.bottom),
+                right: bounds.right.max(rect.right),
+                top: bounds.top.max(rect.top),
+            },
+            None => Bounds {
+                left: rect.left,
+                bottom: rect.bottom,
+                right: rect.right,
+                top: rect.top,
+            },
+        });
+    }
+    flush(&mut items, &mut text, &mut bounds, page);
+    items.sort_by(|first, second| {
+        first
+            .page
+            .cmp(&second.page)
+            .then(second.y.total_cmp(&first.y))
+            .then(first.x.total_cmp(&second.x))
+    });
+    items
+}
+
 impl PageRenderer for PdfiumRenderer {
     type Error = RenderError;
 
@@ -237,6 +398,17 @@ fn bgr_to_rgb_in_place(
 #[cfg(test)]
 mod tests {
     use super::*;
+    use firecrawl_pdfium::{PagePoint, PageRect};
+
+    fn page_char(value: char, bounds: PageRect) -> PageChar {
+        PageChar {
+            unicode: Some(value),
+            code: value as u32,
+            bounds,
+            loose_bounds: bounds,
+            origin: PagePoint::new(bounds.left, bounds.bottom),
+        }
+    }
 
     #[test]
     fn bgr_pixels_are_converted_to_rgb_in_place() {
@@ -263,4 +435,34 @@ mod tests {
             Err(RenderBufferError::InvalidBufferLength { .. })
         ));
     }
+
+    #[test]
+    fn invalid_character_geometry_splits_text_runs() {
+        let chars = [
+            page_char('A', PageRect::new(0.0, 0.0, 8.0, 10.0)),
+            page_char('X', PageRect::new(10.0, 0.0, 10.0, 10.0)),
+            page_char('B', PageRect::new(20.0, 0.0, 28.0, 10.0)),
+        ];
+
+        let items = text_chars_to_items(&chars, 1);
+
+        assert_eq!(
+            items
+                .iter()
+                .map(|item| item.text.as_str())
+                .collect::<Vec<_>>(),
+            ["A", "B"]
+        );
+    }
+
+    #[test]
+    fn coordinates_that_overflow_f32_are_discarded() {
+        let left = f64::from(f32::MAX) * 2.0;
+        let chars = [page_char(
+            'A',
+            PageRect::new(left, 0.0, left + 1.0e30, 10.0),
+        )];
+
+        assert!(text_chars_to_items(&chars, 1).is_empty());
+    }
 }

src/vision/pipeline.rs

@@ -1,20 +1,31 @@
 //! One-call native extraction and OCR pipeline.
 
-use std::collections::BTreeSet;
+use std::collections::{BTreeMap, BTreeSet};
 use std::path::{Path, PathBuf};
 use std::sync::{Arc, Mutex, OnceLock};
 use std::time::Instant;
 
 use thiserror::Error;
 
-use crate::{MarkdownOptions, PageOcrReasons, PdfError};
+use crate::text_quality::{
+    analyze_text_quality, detect_encoding_issues, is_cid_garbage, is_garbage_text,
+};
+use crate::{
+    MarkdownOptions, PageOcrReasons, PdfError, OCR_REASON_SUSPECTED_GARBLED_TEXT,
+    OCR_REASON_VECTOR_TEXT,
+};
 
+use super::fusion::{
+    assess_native_candidate, fuse_ocr_pages_adaptive, NativeCandidateOrigin,
+    NativeFallbackCandidate,
+};
 use super::oar::onnx_runtime_library_path;
+use super::pdfium::PdfiumTextPage;
 use super::{
-    fuse_ocr_pages, route_ocr_pages, run_ocr_pages, FusedPageMarkdown, HttpModelDownloadError,
-    HttpModelDownloader, ModelAcquireError, ModelStore, ModelStoreError, OarOcrEngine, OarOcrError,
-    OcrFusionError, OcrFusionOptions, OcrMode, OcrOptions, OcrRoutingError, OcrRun, OcrRunError,
-    PdfiumRenderer, RenderError, RenderOptions, PP_OCR_V6_SMALL,
+    route_ocr_pages, run_ocr_pages, FusedPageMarkdown, HttpModelDownloadError, HttpModelDownloader,
+    ModelAcquireError, ModelStore, ModelStoreError, OarOcrEngine, OarOcrError, OcrFusionError,
+    OcrFusionOptions, OcrMode, OcrOptions, OcrRoutingError, OcrRun, OcrRunError, PdfiumRenderer,
+    RenderError, RenderOptions, PP_OCR_V6_SMALL,
 };
 
 #[derive(Debug, Clone, PartialEq, Eq)]
@@ -210,7 +221,7 @@ pub fn process_pdf_with_ocr_mem(
 
     let mut page_markdown_options = options.markdown.clone();
     page_markdown_options.include_page_numbers = false;
-    let (native, page_count) = crate::extract_pages_markdown_mem_for_ocr(
+    let (mut native, page_count) = crate::extract_pages_markdown_mem_for_ocr(
         buffer,
         selected_pages_zero_indexed.as_deref(),
         options.password.as_deref(),
@@ -223,13 +234,99 @@ pub fn process_pdf_with_ocr_mem(
         return Err(OcrPipelineError::InvalidSelectedPage { page: invalid });
     }
 
-    let routed = route_ocr_pages(
+    let initially_routed = route_ocr_pages(
         options.ocr.mode,
         page_count,
         &native.pages_needing_ocr,
         selected_pages.as_deref(),
     )?;
 
+    // The OCR extractor retains clean native fragments on pages that still
+    // require OCR. Remove them from the ordinary native result now so Off mode
+    // preserves the existing suppression contract, while keeping a private
+    // candidate for post-OCR comparison and geometry-aware fusion.
+    let mut native_candidates = BTreeMap::<u32, NativeFallbackCandidate>::new();
+    for page in &mut native.pages {
+        if !page.needs_ocr || page.markdown.trim().is_empty() {
+            continue;
+        }
+        let page_number = page.page + 1;
+        let markdown = std::mem::take(&mut page.markdown);
+        if may_preserve_extractor_candidate(page_number, &native.ocr_reasons_by_page) {
+            if let Some(candidate) =
+                assess_native_candidate(markdown, NativeCandidateOrigin::Extractor)
+            {
+                native_candidates.insert(page_number, candidate);
+            }
+        }
+    }
+
+    let mut routed = initially_routed.clone();
+    let mut renderer = None;
+    let mut recovered_natively = BTreeSet::new();
+    if options.ocr.mode == OcrMode::Auto {
+        let complete_recovery_candidates =
+            native_recovery_candidates(&initially_routed, &native.ocr_reasons_by_page);
+        let mut native_probe_pages: BTreeSet<u32> =
+            complete_recovery_candidates.iter().copied().collect();
+        native_probe_pages.extend(
+            native_candidates
+                .keys()
+                .filter(|page| routed.contains(page))
+                .copied(),
+        );
+        if !native_probe_pages.is_empty() {
+            let native_renderer = PdfiumRenderer::load()?;
+            let recovered = native_renderer.extract_text_pages(
+                buffer,
+                &native_probe_pages.iter().copied().collect::<Vec<_>>(),
+                options.password.as_deref(),
+            )?;
+            renderer = Some(native_renderer);
+
+            for page in recovered {
+                let Some(markdown) =
+                    credible_native_recovery(&page.items, page_count, &page_markdown_options)
+                else {
+                    continue;
+                };
+                let Some(candidate) =
+                    assess_native_candidate(markdown, NativeCandidateOrigin::Pdfium)
+                else {
+                    continue;
+                };
+                let may_skip_ocr = complete_recovery_candidates.contains(&page.page)
+                    && candidate.is_complete_recovery()
+                    && native_recovery_covers_page(&page);
+                if may_skip_ocr {
+                    let Some(native_page) = native
+                        .pages
+                        .iter_mut()
+                        .find(|entry| entry.page + 1 == page.page)
+                    else {
+                        continue;
+                    };
+                    native_page.markdown = candidate.markdown().to_string();
+                    native_page.needs_ocr = false;
+                    native_candidates.remove(&page.page);
+                    recovered_natively.insert(page.page);
+                } else {
+                    match native_candidates.entry(page.page) {
+                        std::collections::btree_map::Entry::Vacant(entry) => {
+                            entry.insert(candidate);
+                        }
+                        std::collections::btree_map::Entry::Occupied(mut entry) => {
+                            if candidate.is_stronger_than(entry.get()) {
+                                entry.insert(candidate);
+                            }
+                        }
+                    }
+                }
+            }
+            routed.retain(|page| !recovered_natively.contains(page));
+        }
+    }
+
     let ocr_run = if routed.is_empty() {
         OcrRun {
             pages: Vec::new(),
@@ -239,7 +336,10 @@ pub fn process_pdf_with_ocr_mem(
     } else {
         // Resolve the native renderer before any network request so a missing
         // PDFium installation cannot trigger a model download it cannot use.
-        let renderer = PdfiumRenderer::load()?;
+        let renderer = match renderer {
+            Some(renderer) => renderer,
+            None => PdfiumRenderer::load()?,
+        };
         let engine = cached_ocr_engine(&options.ocr)?;
         run_ocr_pages(
             &renderer,
@@ -256,7 +356,20 @@ pub fn process_pdf_with_ocr_mem(
         .markdown(page_markdown_options)
         .render_dpi(options.render.dpi)
         .hosted_recommendation_confidence(options.hosted_recommendation_confidence);
-    let fused = fuse_ocr_pages(&native.pages, &ocr_run, page_count, &fusion_options)?;
+    let mut fused = fuse_ocr_pages_adaptive(
+        &native.pages,
+        &ocr_run,
+        page_count,
+        &fusion_options,
+        &native_candidates,
+    )?;
+    for page in &mut fused.pages {
+        if recovered_natively.contains(&page.page) {
+            page.provenance
+                .warnings
+                .push("recovered a credible positioned native text layer before OCR".to_string());
+        }
+    }
     let pages_recommending_hosted = fused
         .pages
         .iter()
@@ -265,6 +378,14 @@ pub fn process_pdf_with_ocr_mem(
         .collect();
     let markdown = assemble_document_markdown(&fused.pages, options.markdown.include_page_numbers);
 
+    let mut pages_with_tables = native.pages_with_tables;
+    for page in &fused.pages {
+        if markdown_has_table(&page.markdown) && !pages_with_tables.contains(&page.page) {
+            pages_with_tables.push(page.page);
+        }
+    }
+    pages_with_tables.sort_unstable();
+
     Ok(OcrPdfResult {
         markdown,
         pages: fused.pages,
@@ -273,9 +394,9 @@ pub fn process_pdf_with_ocr_mem(
         pages_routed_to_ocr: routed,
         pages_recommending_hosted,
         ocr_reasons_by_page: native.ocr_reasons_by_page,
-        pages_with_tables: native.pages_with_tables,
+        pages_with_tables: pages_with_tables.clone(),
         pages_with_columns: native.pages_with_columns,
-        is_complex: native.is_complex,
+        is_complex: native.is_complex || !pages_with_tables.is_empty(),
         processing_time_ms: elapsed_ms(started),
         render_time_ms: fused.render_time_ms,
         ocr_time_ms: fused.ocr_time_ms,
@@ -360,6 +481,239 @@ fn normalized_cache_path(path: PathBuf) -> PathBuf {
     absolute
 }
 
+fn native_recovery_candidates(routed: &[u32], reasons: &[PageOcrReasons]) -> Vec<u32> {
+    let reasons_by_page: BTreeMap<u32, &PageOcrReasons> =
+        reasons.iter().map(|entry| (entry.page, entry)).collect();
+    routed
+        .iter()
+        .copied()
+        .filter(|page| {
+            reasons_by_page.get(page).is_some_and(|entry| {
+                !entry.reasons.is_empty()
+                    && entry.reasons.iter().all(|reason| {
+                        reason == OCR_REASON_SUSPECTED_GARBLED_TEXT
+                            || reason == OCR_REASON_VECTOR_TEXT
+                    })
+            })
+        })
+        .collect()
+}
+
+fn may_preserve_extractor_candidate(page: u32, reasons: &[PageOcrReasons]) -> bool {
+    reasons
+        .iter()
+        .find(|entry| entry.page == page)
+        .is_some_and(|entry| {
+            !entry.reasons.is_empty()
+                && !entry
+                    .reasons
+                    .iter()
+                    .any(|reason| reason == OCR_REASON_SUSPECTED_GARBLED_TEXT)
+                && entry.reasons.iter().any(|reason| {
+                    reason == crate::OCR_REASON_SCANNED || reason == OCR_REASON_VECTOR_TEXT
+                })
+        })
+}
+
+fn credible_native_recovery(
+    items: &[crate::TextItem],
+    document_page_count: u32,
+    options: &MarkdownOptions,
+) -> Option<String> {
+    let text = items
+        .iter()
+        .map(|item| item.text.as_str())
+        .collect::<Vec<_>>()
+        .join(" ");
+    if is_garbage_text(&text) || is_cid_garbage(&text) {
+        return None;
+    }
+
+    let markdown = crate::to_markdown_from_items_with_rects_and_page_count(
+        items.to_vec(),
+        options.clone(),
+        &[],
+        document_page_count,
+    );
+    let markdown = remove_duplicate_table_lines(&markdown);
+    let structured_uniform_ascii = is_uniform_case_structured_ascii(&text, &markdown);
+    let quality = analyze_text_quality(items);
+    if !structured_uniform_ascii && (quality.has_encoding_issues || detect_encoding_issues(&text)) {
+        return None;
+    }
+    (!markdown.trim().is_empty()).then_some(markdown)
+}
+
+fn native_recovery_covers_page(page: &PdfiumTextPage) -> bool {
+    const VERTICAL_BANDS: f32 = 6.0;
+
+    let width = page.page_width;
+    let height = page.page_height;
+    if !width.is_finite() || !height.is_finite() || width <= 0.0 || height <= 0.0 {
+        return false;
+    }
+
+    let mut min_left = width;
+    let mut max_right = 0.0_f32;
+    let mut min_bottom = height;
+    let mut max_top = 0.0_f32;
+    let mut positioned_items = 0usize;
+    let mut occupied_bands = BTreeSet::new();
+    for item in &page.items {
+        if !item.text.chars().any(char::is_alphanumeric)
+            || !item.x.is_finite()
+            || !item.y.is_finite()
+            || !item.width.is_finite()
+            || !item.height.is_finite()
+            || item.width <= 0.0
+            || item.height <= 0.0
+        {
+            continue;
+        }
+        let left = item.x.clamp(0.0, width);
+        let right = (item.x + item.width).clamp(0.0, width);
+        let bottom = item.y.clamp(0.0, height);
+        let top = (item.y + item.height).clamp(0.0, height);
+        if right <= left || top <= bottom {
+            continue;
+        }
+
+        positioned_items += 1;
+        min_left = min_left.min(left);
+        max_right = max_right.max(right);
+        min_bottom = min_bottom.min(bottom);
+        max_top = max_top.max(top);
+        let center = (bottom + top) * 0.5;
+        let band = ((center / height) * VERTICAL_BANDS)
+            .floor()
+            .clamp(0.0, VERTICAL_BANDS - 1.0) as u8;
+        occupied_bands.insert(band);
+    }
+
+    positioned_items >= 6
+        && (max_right - min_left) / width >= 0.15
+        && (max_top - min_bottom) / height >= 0.35
+        && occupied_bands.len() >= 3
+}
+
+fn is_uniform_case_structured_ascii(text: &str, markdown: &str) -> bool {
+    if !text.is_ascii() || text.contains('$') || text.chars().any(char::is_control) {
+        return false;
+    }
+
+    let letters: Vec<_> = text
+        .chars()
+        .filter(|character| character.is_ascii_alphabetic())
+        .collect();
+    if letters.len() < 200 {
+        return false;
+    }
+    let uniform_case = letters
+        .iter()
+        .all(|character| character.is_ascii_uppercase())
+        || letters
+            .iter()
+            .all(|character| character.is_ascii_lowercase());
+    if !uniform_case {
+        return false;
+    }
+    if markdown_has_table(markdown) {
+        return true;
+    }
+
+    let nonempty_lines = markdown
+        .lines()
+        .filter(|line| !line.trim().is_empty())
+        .count();
+    let visible_chars = text
+        .chars()
+        .filter(|character| !character.is_whitespace())
+        .count()
+        .max(1);
+    let structural_chars = text
+        .chars()
+        .filter(|character| {
+            character.is_ascii_digit()
+                || matches!(
+                    character,
+                    '{' | '}'
+                        | '['
+                        | ']'
+                        | '('
+                        | ')'
+                        | '<'
+                        | '>'
+                        | '_'
+                        | '='
+                        | '+'
+                        | '*'
+                        | '/'
+                        | '\\'
+                        | '|'
+                        | '&'
+                        | '^'
+                        | '%'
+                        | '#'
+                        | '@'
+                        | '~'
+                )
+        })
+        .count();
+    nonempty_lines >= 4 && structural_chars * 20 >= visible_chars
+}
+
+fn markdown_has_table(markdown: &str) -> bool {
+    markdown.lines().any(|line| {
+        let trimmed = line.trim();
+        trimmed.starts_with('|')
+            && trimmed.ends_with('|')
+            && trimmed
+                .split('|')
+                .filter(|cell| !cell.is_empty())
+                .all(|cell| !cell.is_empty() && cell.chars().all(|ch| ch == '-'))
+    })
+}
+
+fn remove_duplicate_table_lines(markdown: &str) -> String {
+    let mut output = String::new();
+    let mut adjacent_table_row = None;
+    for line in markdown.lines() {
+        let trimmed = line.trim();
+        let is_table_line = trimmed.starts_with('|') && trimmed.ends_with('|');
+        if is_table_line {
+            if !trimmed.contains("|---") && trimmed.matches('|').count() >= 4 {
+                let canonical = canonical_table_text(trimmed);
+                if !canonical.is_empty() {
+                    adjacent_table_row = Some(canonical);
+                }
+            }
+        } else if trimmed.is_empty() {
+            // Keep adjacency across the blank line emitted after a table.
+        } else {
+            let duplicate = adjacent_table_row
+                .as_ref()
+                .is_some_and(|table_row| *table_row == canonical_table_text(trimmed));
+            adjacent_table_row = None;
+            if duplicate {
+                continue;
+            }
+        }
+        output.push_str(line);
+        output.push('\n');
+    }
+    while output.ends_with("\n\n\n") {
+        output.pop();
+    }
+    output
+}
+
+fn canonical_table_text(text: &str) -> String {
+    text.replace('|', " ")
+        .split_whitespace()
+        .collect::<Vec<_>>()
+        .join(" ")
+}
+
 fn assemble_document_markdown(pages: &[FusedPageMarkdown], include_page_numbers: bool) -> String {
     let mut document = String::new();
     for (index, page) in pages.iter().enumerate() {
@@ -427,6 +781,25 @@ pub enum OcrPipelineError {
 mod tests {
     use super::*;
 
+    fn recovery_item(text: &str, x: f32, y: f32, width: f32, height: f32) -> crate::TextItem {
+        crate::TextItem {
+            text: text.to_string(),
+            x,
+            y,
+            width,
+            height,
+            font: "PDFium native text".to_string(),
+            font_size: height,
+            page: 1,
+            is_bold: false,
+            is_italic: false,
+            is_underline: false,
+            is_strikeout: false,
+            item_type: crate::types::ItemType::Text,
+            mcid: None,
+        }
+    }
+
     #[test]
     fn cache_path_is_stable_when_a_relative_directory_is_created() {
         let current = std::fs::canonicalize(std::env::current_dir().unwrap()).unwrap();
@@ -442,6 +815,130 @@ mod tests {
         assert_eq!(before, after);
     }
 
+    #[test]
+    fn native_recovery_is_limited_to_recoverable_routing_reasons() {
+        let routed = [1, 2, 3, 4];
+        let reasons = [
+            PageOcrReasons {
+                page: 1,
+                reasons: vec![OCR_REASON_SUSPECTED_GARBLED_TEXT.to_string()],
+            },
+            PageOcrReasons {
+                page: 2,
+                reasons: vec![crate::OCR_REASON_SCANNED.to_string()],
+            },
+            PageOcrReasons {
+                page: 3,
+                reasons: vec![OCR_REASON_VECTOR_TEXT.to_string()],
+            },
+            PageOcrReasons {
+                page: 4,
+                reasons: vec![
+                    OCR_REASON_SUSPECTED_GARBLED_TEXT.to_string(),
+                    crate::OCR_REASON_SCANNED.to_string(),
+                ],
+            },
+        ];
+
+        assert_eq!(native_recovery_candidates(&routed, &reasons), vec![1, 3]);
+    }
+
+    #[test]
+    fn native_recovery_requires_text_coverage_beyond_a_header() {
+        let header = PdfiumTextPage {
+            page: 1,
+            page_width: 600.0,
+            page_height: 800.0,
+            items: (0..8)
+                .map(|index| recovery_item("HEADER", index as f32 * 60.0, 740.0, 50.0, 12.0))
+                .collect(),
+        };
+        assert!(!native_recovery_covers_page(&header));
+
+        let complete = PdfiumTextPage {
+            page: 1,
+            page_width: 600.0,
+            page_height: 800.0,
+            items: vec![
+                recovery_item("Top one", 40.0, 700.0, 180.0, 12.0),
+                recovery_item("Top two", 260.0, 680.0, 180.0, 12.0),
+                recovery_item("Middle one", 40.0, 400.0, 180.0, 12.0),
+                recovery_item("Middle two", 260.0, 380.0, 180.0, 12.0),
+                recovery_item("Bottom one", 40.0, 100.0, 180.0, 12.0),
+                recovery_item("Bottom two", 260.0, 80.0, 180.0, 12.0),
+            ],
+        };
+        assert!(native_recovery_covers_page(&complete));
+    }
+
+    #[test]
+    fn uniform_case_guard_requires_structured_content() {
+        let table_text = "STATUS CODE 100 READY ".repeat(20);
+        let table_markdown = "|STATUS|CODE|\n|---|---|\n|READY|100|\n|READY|200|\n|READY|300|\n";
+        assert!(is_uniform_case_structured_ascii(
+            &table_text,
+            table_markdown
+        ));
+
+        let prose = "THIS IS ORDINARY UPPERCASE PROSE WITH NATURAL WORDS ".repeat(20);
+        let prose_markdown = prose
+            .split_whitespace()
+            .collect::<Vec<_>>()
+            .chunks(8)
+            .map(|line| line.join(" "))
+            .collect::<Vec<_>>()
+            .join("\n");
+        assert!(!is_uniform_case_structured_ascii(&prose, &prose_markdown));
+    }
+
+    #[test]
+    fn recovered_markdown_drops_plain_duplicates_of_table_rows() {
+        let markdown = "|Date|Value|Status|\n|---|---|---|\n|April 1|42|ok|\n\nApril 1 42 ok\n";
+
+        assert_eq!(
+            remove_duplicate_table_lines(markdown),
+            "|Date|Value|Status|\n|---|---|---|\n|April 1|42|ok|\n\n"
+        );
+    }
+
+    #[test]
+    fn recovered_markdown_keeps_nonadjacent_repeated_table_text() {
+        let markdown = "|Date|Value|Status|\n|---|---|---|\n|April 1|42|ok|\n\nSummary follows.\n\nApril 1 42 ok\n";
+
+        assert_eq!(remove_duplicate_table_lines(markdown), markdown);
+    }
+
+    #[test]
+    fn extractor_candidates_require_clean_partial_content_reasons() {
+        let reasons = [
+            PageOcrReasons {
+                page: 1,
+                reasons: vec![crate::OCR_REASON_SCANNED.to_string()],
+            },
+            PageOcrReasons {
+                page: 2,
+                reasons: vec![OCR_REASON_SUSPECTED_GARBLED_TEXT.to_string()],
+            },
+            PageOcrReasons {
+                page: 3,
+                reasons: vec![OCR_REASON_VECTOR_TEXT.to_string()],
+            },
+            PageOcrReasons {
+                page: 4,
+                reasons: vec![
+                    crate::OCR_REASON_SCANNED.to_string(),
+                    OCR_REASON_SUSPECTED_GARBLED_TEXT.to_string(),
+                ],
+            },
+        ];
+
+        assert!(may_preserve_extractor_candidate(1, &reasons));
+        assert!(!may_preserve_extractor_candidate(2, &reasons));
+        assert!(may_preserve_extractor_candidate(3, &reasons));
+        assert!(!may_preserve_extractor_candidate(4, &reasons));
+        assert!(!may_preserve_extractor_candidate(5, &reasons));
+    }
+
     #[test]
     fn off_mode_extracts_native_text_without_runtime_side_effects() {
         let bytes = std::fs::read("tests/fixtures/thermo-freon12.pdf").unwrap();
@@ -488,9 +985,30 @@ mod tests {
         let result = process_pdf_with_ocr_mem(&bytes, OcrPdfOptions::new()).unwrap();
 
         assert!(result.pages_routed_to_ocr.is_empty());
+        assert!(result.markdown.trim().is_empty());
         assert_eq!(result.pages_recommending_hosted, vec![1]);
     }
 
+    #[test]
+    fn partial_native_scan_text_is_retained_only_inside_ocr_orchestration() {
+        let bytes = std::fs::read("tests/fixtures/scan_with_native_header_text.pdf").unwrap();
+        let public = crate::extract_pages_markdown_mem(&bytes, None).unwrap();
+        assert!(public.pages[0].needs_ocr);
+        assert!(public.pages[0].markdown.trim().is_empty());
+
+        let (ocr, _) = crate::extract_pages_markdown_mem_for_ocr(
+            &bytes,
+            None,
+            None,
+            &MarkdownOptions::default(),
+        )
+        .unwrap();
+        assert!(ocr.pages[0].needs_ocr);
+        assert!(ocr.pages[0]
+            .markdown
+            .contains("Order Detail Report by Account"));
+    }
+
     #[test]
     fn password_is_redacted_and_used_for_native_extraction() {
         let options = OcrPdfOptions::new().password("secret123");

tests/ocr_tests.rs

@@ -2,7 +2,8 @@
 
 #[cfg(feature = "ocr")]
 use pdf_inspector::vision::{
-    process_pdf_with_ocr_mem, ModelDownloadPolicy, OcrPdfOptions, PageContentSource,
+    process_pdf_with_ocr_mem, ModelDownloadPolicy, OcrPdfOptions, OcrPipelineError,
+    PageContentSource,
 };
 use pdf_inspector::vision::{
     ModelStore, OarOcrEngine, OcrEngine, OcrMode, OcrOptions, PageTransform, RenderPixelFormat,
@@ -127,7 +128,16 @@ fn complete_ocr_pipeline_routes_and_assembles_a_scanned_fixture() {
 
     assert_eq!(result.pages_routed_to_ocr, vec![1]);
     assert!(!result.markdown.trim().is_empty());
-    assert_eq!(result.pages[0].provenance.source, PageContentSource::Ocr);
+    assert!(result
+        .markdown
+        .contains("Order Date Item Code Description Status Unit Cost\n\n03/14/2024"));
+    assert!(result.markdown.contains("$482,110.40\n\n05/02/2024"));
+    assert_eq!(result.pages[0].provenance.source, PageContentSource::Fused);
+    assert!(result.pages[0]
+        .provenance
+        .warnings
+        .iter()
+        .any(|warning| warning.contains("complementary OCR")));
     assert_eq!(
         result.pages[0].provenance.ocr_model.as_ref().unwrap().name,
         PP_OCR_V6_SMALL.id
@@ -135,6 +145,50 @@ fn complete_ocr_pipeline_routes_and_assembles_a_scanned_fixture() {
     assert_eq!(repeated.markdown, result.markdown);
 }
 
+#[cfg(all(feature = "ocr", feature = "render-pdfium"))]
+#[test]
+fn auto_recovers_credible_native_text_before_loading_ocr_models() {
+    let Some(_renderer) = load_renderer() else {
+        return;
+    };
+
+    let bytes = std::fs::read("tests/fixtures/shinagawa_identity_h.pdf").unwrap();
+    let ocr = OcrOptions::new()
+        .mode(OcrMode::Auto)
+        .model_directory("/models/must-not-be-read")
+        .model_downloads(ModelDownloadPolicy::Offline);
+    let result = process_pdf_with_ocr_mem(&bytes, OcrPdfOptions::new().ocr(ocr)).unwrap();
+
+    assert_eq!(result.pages_recommended_for_ocr, vec![1]);
+    assert!(result.pages_routed_to_ocr.is_empty());
+    assert!(result.markdown.contains("羽田空港新飛行経路"));
+    assert!(result.markdown.contains("|4月30日|有|81.0|"));
+    assert!(result.markdown.contains("※１ 最大騒音レベル"));
+    assert!(result.pages_with_tables.contains(&1));
+    assert_eq!(result.pages[0].provenance.source, PageContentSource::Native);
+    assert!(result.pages[0].provenance.ocr_model.is_none());
+}
+
+#[cfg(all(feature = "ocr", feature = "render-pdfium"))]
+#[test]
+fn auto_rejects_garbled_native_recovery_and_continues_to_ocr() {
+    let Some(_renderer) = load_renderer() else {
+        return;
+    };
+
+    let bytes = std::fs::read("tests/fixtures/shifted_cipher_tounicode.pdf").unwrap();
+    let ocr = OcrOptions::new()
+        .mode(OcrMode::Auto)
+        .model_directory("/models/must-not-be-read")
+        .model_downloads(ModelDownloadPolicy::Offline);
+    let error = process_pdf_with_ocr_mem(&bytes, OcrPdfOptions::new().ocr(ocr)).unwrap_err();
+
+    assert!(matches!(
+        error,
+        OcrPipelineError::ModelAcquire(_) | OcrPipelineError::ModelStore(_)
+    ));
+}
+
 fn recognize(
     model_directory: &std::ffi::OsStr,
     pages: &[RenderedPage],