fix(detect): classify tiled scans with garbage OCR as Scanned

Scanned PDFs with JBIG2/tiled image strips were misclassified as
TextBased because no individual image tile exceeded the template
threshold. Now checks aggregate image area per page (≥2M pixels).

Also adds is_garbage_text() check: if a Mixed/template PDF's extracted
text is predominantly non-alphanumeric (<50%), upgrade to Scanned so
callers use proper OCR instead of the garbage text layer.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

src/detector.rs

@@ -197,10 +197,11 @@ pub(crate) fn detect_from_document(
             let analysis = analyze_page_content(doc, page_id);
             pages_actually_sampled += 1;
             log::debug!(
-                "page {}: text_ops={} images={} image_count={} template={} unique_chars={} path_ops={} vector_text={} image_area={}",
+                "page {}: text_ops={} images={} image_count={} template={} unique_chars={} alphanum={} path_ops={} vector_text={} image_area={}",
                 page_num, analysis.text_operator_count, analysis.has_images,
                 analysis.image_count, analysis.has_template_image,
-                analysis.unique_text_chars, analysis.path_op_count, analysis.has_vector_text,
+                analysis.unique_text_chars, analysis.unique_alphanum_chars,
+                analysis.path_op_count, analysis.has_vector_text,
                 analysis.total_image_area
             );
             let is_image_dominated = analysis.image_count > 10
@@ -265,9 +266,8 @@ pub(crate) fn detect_from_document(
 
     // Classification logic
     let (pdf_type, confidence) = if has_template_images && pages_with_text > 0 {
-        // Template-based PDF: has text but images provide essential context
-        // Classify as Mixed with lower confidence
         ocr_recommended = true;
+        // Template-based PDF: has text but images provide essential context
         (PdfType::Mixed, 0.5 + (0.3 * (1.0 - template_ratio)))
     } else if text_ratio >= config.text_page_ratio_threshold {
         ocr_recommended = false;
@@ -383,6 +383,8 @@ struct PageAnalysis {
     image_count: u32,
     /// Number of unique non-whitespace text characters found in string operands
     unique_text_chars: u32,
+    /// Number of unique ASCII alphanumeric bytes (letters + digits) in string operands
+    unique_alphanum_chars: u32,
     /// Number of path construction/painting ops (m, l, c, h, f, re, etc.)
     #[allow(dead_code)]
     path_op_count: u32,
@@ -454,6 +456,11 @@ fn analyze_page_content(doc: &Document, page_id: ObjectId) -> PageAnalysis {
     // outlined text produces thousands of path ops.
     let has_vector_text = path_ops >= 1000 && path_ops > text_ops.saturating_mul(200);
 
+    let unique_alphanum_chars = all_unique_chars
+        .iter()
+        .filter(|b| b.is_ascii_alphanumeric())
+        .count() as u32;
+
     PageAnalysis {
         text_operator_count: text_ops,
         has_images,
@@ -461,6 +468,7 @@ fn analyze_page_content(doc: &Document, page_id: ObjectId) -> PageAnalysis {
         total_image_area,
         image_count,
         unique_text_chars: all_unique_chars.len() as u32,
+        unique_alphanum_chars,
         path_op_count: path_ops,
         has_vector_text,
     }
@@ -801,6 +809,13 @@ fn analyze_page_images(doc: &Document, page_id: ObjectId) -> (bool, u64, bool) {
         }
     }
 
+    // Tiled scans: many small image tiles (e.g., JBIG2 strips) that together
+    // cover the full page. No individual tile triggers the template threshold,
+    // but the aggregate area clearly indicates a scanned/image-backed page.
+    if !has_template_image && total_area >= TEMPLATE_IMAGE_THRESHOLD * 4 {
+        has_template_image = true;
+    }
+
     (has_images, total_area, has_template_image)
 }
 

src/lib.rs

@@ -391,6 +391,16 @@ fn process_document(
         None => (None, LayoutComplexity::default(), false),
     };
 
+    // 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)
+        };
+
     Ok(PdfProcessResult {
         pdf_type,
         markdown,
@@ -444,6 +454,32 @@ fn detect_encoding_issues(markdown: &str) -> bool {
     false
 }
 
+/// Check if extracted text is predominantly garbage (non-alphanumeric).
+///
+/// Broken font encodings produce text like "----1-.-.-.___  --.-. .._ I_---."
+/// where most characters are punctuation/symbols. Real text in any language
+/// has >50% alphanumeric characters.
+fn is_garbage_text(markdown: &str) -> bool {
+    let mut alphanum = 0usize;
+    let mut non_alphanum = 0usize;
+    for ch in markdown.chars() {
+        if ch.is_whitespace() {
+            continue;
+        }
+        // Skip markdown syntax chars that we add (not from the PDF)
+        if matches!(ch, '#' | '*' | '|' | '-' | '\n') {
+            continue;
+        }
+        if ch.is_alphanumeric() {
+            alphanum += 1;
+        } else {
+            non_alphanum += 1;
+        }
+    }
+    let total = alphanum + non_alphanum;
+    total >= 50 && alphanum * 2 < total
+}
+
 /// Analyse extracted items and rects for layout complexity.
 fn compute_layout_complexity(
     items: &[types::TextItem],