fix(detector): cap decompressed content-stream size before scanning

Stop holding the full inflated page or Form stream in the detector so a
highly compressible Flate stream cannot balloon resident memory.

Co-authored-by: Cursor <cursoragent@cursor.com>

Cargo.toml

@@ -43,6 +43,10 @@ unicode-normalization = "0.1"
 # TrueType font parsing (for Identity-H CID font cmap extraction)
 ttf-parser = "0.25"
 
+# Incremental Flate inflate so detector scans can stop before a highly
+# compressible stream materializes a multi-gigabyte buffer.
+flate2 = "1.1"
+
 # Native builds keep lopdf's parallel parser and CLI logging. Browser WASM is
 # deliberately single-threaded so it works without cross-origin isolation.
 [target.'cfg(not(target_arch = "wasm32"))'.dependencies]

src/detector.rs

@@ -4,6 +4,7 @@
 //! by sampling content streams for text operators (Tj/TJ) without loading
 //! all objects.
 
+use crate::stream_decode::stream_content_for_scan;
 use crate::PdfError;
 use lopdf::{Document, Object, ObjectId};
 use std::collections::{HashMap, HashSet};
@@ -765,10 +766,7 @@ fn analyze_page_content(doc: &Document, page_id: ObjectId) -> PageAnalysis {
 
     for content_id in content_streams {
         if let Ok(Object::Stream(stream)) = doc.get_object(content_id) {
-            let content = match stream.decompressed_content() {
-                Ok(data) => data,
-                Err(_) => stream.content.clone(),
-            };
+            let content = stream_content_for_scan(stream);
 
             // Scan for text operators, collecting raw font names
             let mut page_font_names: HashSet<Vec<u8>> = HashSet::new();
@@ -1303,9 +1301,7 @@ fn scan_xobjects_in_resources(
                 .and_then(|o| o.as_name().ok());
             match subtype {
                 Some(b"Form") => {
-                    let content = stream
-                        .decompressed_content()
-                        .unwrap_or_else(|_| stream.content.clone());
+                    let content = stream_content_for_scan(stream);
                     // Collect raw font names from this XObject's content stream
                     let mut xobj_font_names: HashSet<Vec<u8>> = HashSet::new();
                     let (ops, imgs, paths, fonts) = scan_content_for_text_operators(
@@ -3918,4 +3914,119 @@ mod tests {
             "P3: inherited decodable font should be detected as used"
         );
     }
+
+    fn flate_content(plain: &[u8]) -> lopdf::Stream {
+        use flate2::write::ZlibEncoder;
+        use flate2::Compression;
+        use lopdf::dictionary;
+        use std::io::Write;
+        let mut encoder = ZlibEncoder::new(Vec::new(), Compression::best());
+        encoder.write_all(plain).unwrap();
+        lopdf::Stream::new(
+            dictionary! { "Filter" => "FlateDecode" },
+            encoder.finish().unwrap(),
+        )
+    }
+
+    #[test]
+    fn flate_page_content_still_finds_text_operators() {
+        use lopdf::dictionary;
+        let mut doc = Document::with_version("1.4");
+        let pages_id = doc.new_object_id();
+        let page_id = doc.new_object_id();
+        let font_id = doc.add_object(dictionary! {
+            "Type" => "Font",
+            "Subtype" => Object::Name(b"Type1".to_vec()),
+            "BaseFont" => Object::Name(b"Helvetica".to_vec()),
+        });
+        let content_id = doc.add_object(Object::Stream(flate_content(
+            b"BT /F1 12 Tf (Hello world) Tj ET",
+        )));
+        doc.objects.insert(
+            page_id,
+            Object::Dictionary(dictionary! {
+                "Type" => "Page",
+                "Parent" => Object::Reference(pages_id),
+                "Resources" => dictionary! {
+                    "Font" => dictionary! {
+                        "F1" => Object::Reference(font_id),
+                    },
+                },
+                "Contents" => Object::Reference(content_id),
+            }),
+        );
+        doc.objects.insert(
+            pages_id,
+            Object::Dictionary(dictionary! {
+                "Type" => "Pages",
+                "Kids" => vec![Object::Reference(page_id)],
+                "Count" => Object::Integer(1),
+            }),
+        );
+
+        let analysis = analyze_page_content(&doc, page_id);
+        assert!(
+            analysis.text_operator_count > 0,
+            "bounded Flate decode must still see ordinary page text operators"
+        );
+    }
+
+    #[test]
+    fn flate_form_xobject_still_finds_text_operators() {
+        use lopdf::dictionary;
+        let mut doc = Document::with_version("1.4");
+        let pages_id = doc.new_object_id();
+        let page_id = doc.new_object_id();
+        let font_id = doc.add_object(dictionary! {
+            "Type" => "Font",
+            "Subtype" => Object::Name(b"Type1".to_vec()),
+            "BaseFont" => Object::Name(b"Helvetica".to_vec()),
+        });
+        let form_id = doc.add_object(Object::Stream(flate_content(
+            b"BT /F1 12 Tf (Form text) Tj ET",
+        )));
+        if let Object::Stream(form) = doc.objects.get_mut(&form_id).unwrap() {
+            form.dict.set("Type", Object::Name(b"XObject".to_vec()));
+            form.dict.set("Subtype", Object::Name(b"Form".to_vec()));
+            form.dict.set(
+                "Resources",
+                dictionary! {
+                    "Font" => dictionary! {
+                        "F1" => Object::Reference(font_id),
+                    },
+                },
+            );
+        }
+        let page_content_id = doc.add_object(Object::Stream(lopdf::Stream::new(
+            dictionary! {},
+            b"/Fm0 Do".to_vec(),
+        )));
+        doc.objects.insert(
+            page_id,
+            Object::Dictionary(dictionary! {
+                "Type" => "Page",
+                "Parent" => Object::Reference(pages_id),
+                "Resources" => dictionary! {
+                    "XObject" => dictionary! {
+                        "Fm0" => Object::Reference(form_id),
+                    },
+                },
+                "Contents" => Object::Reference(page_content_id),
+            }),
+        );
+        doc.objects.insert(
+            pages_id,
+            Object::Dictionary(dictionary! {
+                "Type" => "Pages",
+                "Kids" => vec![Object::Reference(page_id)],
+                "Count" => Object::Integer(1),
+            }),
+        );
+
+        let analysis = analyze_page_content(&doc, page_id);
+        assert!(
+            analysis.text_operator_count > 0,
+            "bounded Flate decode must still see Form XObject text operators"
+        );
+    }
 }

src/lib.rs

@@ -37,6 +37,7 @@ pub mod extractor;
 pub mod glyph_names;
 pub mod markdown;
 pub mod process_mode;
+mod stream_decode;
 pub mod structure_tree;
 pub mod tables;
 mod text_quality;

src/stream_decode.rs

@@ -0,0 +1,145 @@
+//! Bounded stream decompression for detector scans.
+//!
+//! `lopdf::Stream::decompressed_content` materializes the full decoded buffer
+//! before any caller can apply a limit. A few megabytes of Flate-compressed
+//! zeros can therefore expand to gigabytes. These helpers stop inflate once
+//! the decoded budget is reached.
+
+use flate2::read::{DeflateDecoder, ZlibDecoder};
+use lopdf::Stream;
+use std::io::Read;
+
+/// Maximum decoded bytes held for a single content stream during detection.
+pub(crate) const MAX_DECOMPRESSED_STREAM_BYTES: usize = 32 * 1024 * 1024;
+
+/// Decode `stream` for scanning, or return an empty buffer when the decoded
+/// size would exceed `max_bytes`.
+pub(crate) fn stream_content_for_scan(stream: &Stream) -> Vec<u8> {
+    match decompressed_content_bounded(stream, MAX_DECOMPRESSED_STREAM_BYTES) {
+        Some(data) => data,
+        None => Vec::new(),
+    }
+}
+
+/// Incremental decode with a hard output cap. `None` means the stream is
+/// larger than `max_bytes` (or not safely decodable within that budget).
+pub(crate) fn decompressed_content_bounded(stream: &Stream, max_bytes: usize) -> Option<Vec<u8>> {
+    let filters = match stream.filters() {
+        Ok(filters) => filters,
+        Err(_) => {
+            return take_if_within_budget(&stream.content, max_bytes);
+        }
+    };
+
+    if filters.is_empty() {
+        return take_if_within_budget(&stream.content, max_bytes);
+    }
+
+    // Plain Flate is the highly compressible case. Detector scans only need
+    // the inflated operator bytes; skip PNG predictors here so inflate can
+    // stop at the budget instead of materializing the full buffer first.
+    if filters.len() == 1 && filters[0] == b"FlateDecode" {
+        return inflate_flate_bounded(&stream.content, max_bytes);
+    }
+
+    if stream.content.len() > max_bytes {
+        return None;
+    }
+    match stream.decompressed_content() {
+        Ok(data) if data.len() <= max_bytes => Some(data),
+        Ok(_) => None,
+        Err(_) => take_if_within_budget(&stream.content, max_bytes),
+    }
+}
+
+fn take_if_within_budget(bytes: &[u8], max_bytes: usize) -> Option<Vec<u8>> {
+    if bytes.len() > max_bytes {
+        None
+    } else {
+        Some(bytes.to_vec())
+    }
+}
+
+fn inflate_flate_bounded(input: &[u8], max_bytes: usize) -> Option<Vec<u8>> {
+    if input.is_empty() {
+        return Some(Vec::new());
+    }
+    match read_bounded(ZlibDecoder::new(input), max_bytes) {
+        Some(data) => Some(data),
+        None if input.len() > 2 => read_bounded(DeflateDecoder::new(&input[2..]), max_bytes),
+        None => None,
+    }
+}
+
+fn read_bounded<R: Read>(mut decoder: R, max_bytes: usize) -> Option<Vec<u8>> {
+    let mut output = Vec::new();
+    let mut buf = [0u8; 16 * 1024];
+    loop {
+        match decoder.read(&mut buf) {
+            Ok(0) => return Some(output),
+            Ok(n) => {
+                if output.len().saturating_add(n) > max_bytes {
+                    return None;
+                }
+                output.extend_from_slice(&buf[..n]);
+            }
+            Err(_) => return None,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use flate2::write::ZlibEncoder;
+    use flate2::Compression;
+    use lopdf::dictionary;
+    use std::io::Write;
+
+    fn flate_stream(plain: &[u8]) -> Stream {
+        let mut encoder = ZlibEncoder::new(Vec::new(), Compression::best());
+        encoder.write_all(plain).unwrap();
+        let compressed = encoder.finish().unwrap();
+        Stream::new(dictionary! { "Filter" => "FlateDecode" }, compressed)
+    }
+
+    #[test]
+    fn small_flate_stream_round_trips() {
+        let plain = b"BT /F1 12 Tf (Hello world) Tj ET";
+        let stream = flate_stream(plain);
+        assert_eq!(
+            decompressed_content_bounded(&stream, MAX_DECOMPRESSED_STREAM_BYTES).as_deref(),
+            Some(plain.as_slice())
+        );
+    }
+
+    #[test]
+    fn highly_compressible_flate_stops_at_budget() {
+        let plain = vec![0u8; 256 * 1024];
+        let stream = flate_stream(&plain);
+        assert!(
+            stream.content.len() < 8 * 1024,
+            "fixture must stay compact on disk, got {} compressed bytes",
+            stream.content.len()
+        );
+        assert!(decompressed_content_bounded(&stream, 16 * 1024).is_none());
+        assert_eq!(
+            decompressed_content_bounded(&stream, 256 * 1024).as_deref(),
+            Some(plain.as_slice())
+        );
+    }
+
+    #[test]
+    fn uncompressed_over_budget_is_skipped() {
+        let stream = Stream::new(dictionary! {}, vec![b'x'; 64]);
+        assert!(decompressed_content_bounded(&stream, 32).is_none());
+        assert_eq!(decompressed_content_bounded(&stream, 64).unwrap().len(), 64);
+    }
+
+    #[test]
+    fn scan_helper_returns_empty_when_capped() {
+        let stream = flate_stream(&vec![0u8; 64 * 1024]);
+        // Production cap is far above 64 KiB, so this still decodes.
+        assert_eq!(stream_content_for_scan(&stream).len(), 64 * 1024);
+    }
+}