fix(deps): move PDF parsing to lopdf 0.42, clearing RUSTSEC-2026-0187

.github/dependabot.yml

@@ -10,27 +10,12 @@ updates:
         update-types: ["minor", "patch"]
     # Direct bumps that can never resolve: a parent crate pins the old major and
     # its API type flows into our code. The real fix arrives when the PARENT
-    # ships a new release, so watch for a `scraper` / `pdf-inspector` bump PR and
-    # remove the matching ignore entry then.
+    # ships a new release, so watch for a `scraper` bump PR and remove the
+    # matching ignore entry then.
     ignore:
       # scraper 0.25 (latest) pins ego-tree 0.10; NodeId comes from scraper's
       # ElementRef, so a direct ego-tree 0.11 bump is a type mismatch.
       - dependency-name: "ego-tree"
-      # pdf-inspector 0.1.7 (latest, github.com/firecrawl/pdf-inspector) still
-      # pins lopdf 0.41; a direct bump leaves 0.41 in the tree anyway. lopdf
-      # 0.42 carries the RUSTSEC-2026-0187 fix (stack-overflow SIGABRT on
-      # deeply nested objects, uncatchable by catch_unwind). Contained for us:
-      # `document.sandbox` runs every parse in a child process and is ON in
-      # config.docker.toml, so an abort kills the worker, not the server.
-      # Drop this ignore once pdf-inspector publishes a crate release built on
-      # lopdf >= 0.42 (fixed on their main in 1c32e4bd, unreleased).
-      - dependency-name: "lopdf"
-      # 0.1.7 specifically: it extracts 13,732 fewer characters than 0.1.6 over
-      # our 30-PDF measurement corpus, one document losing six whole pages. The
-      # workspace pin holds 0.1.6; this keeps dependabot from re-proposing the
-      # bad release while still offering 0.1.8 when it lands.
-      - dependency-name: "pdf-inspector"
-        versions: ["0.1.7"]
 
   - package-ecosystem: "github-actions"
     directory: "/"

Cargo.toml

@@ -63,12 +63,13 @@ regex = "1"
 # libraries — preserves the single-static-binary story. The open-core
 # boundary gate asserts lopdf/ttf-parser are ABSENT when `pdf` is disabled.
 #
-# Held at 0.1.6 on purpose. 0.1.7 guts six consecutive pages of some documents:
-# measured over 30 real PDFs it extracts 13,732 fewer characters in total than
-# 0.1.6, all of it one document (arXiv 1412.6980) losing pages 6-11 down to
-# 4-13% of their text. 0.1.6 is the best release we have measured: +23,148
-# characters over 0.1.4 across the same corpus. Re-evaluate on 0.1.8.
-pdf-inspector = "=0.1.6"
+# Was held at 0.1.6 because 0.1.7 gutted pages 6-11 of arXiv 1412.6980. Measured
+# again over 29 real PDFs (papers, IRS forms, RFCs, 10-K, the 756-page PDF spec)
+# scoring word recall against poppler as ground truth: 1.17.0 reaches 98.15% vs
+# 0.1.6's 97.72%, better on 23 of 29 documents, and every page of 1412.6980 now
+# comes back at 97-100%. The raw character count drops slightly only because
+# 1.17.0 rejoins hyphenated line breaks and stops duplicating table rows.
+pdf-inspector = "=1.17.0"
 
 # Diffing (change-tracking / monitor). Myers diff over lines; the parse-diff
 # AST and the unified text surface are both derived from its op stream.
@@ -112,7 +113,7 @@ uuid = { version = "1", features = ["v4", "serde"] }
 url = { version = "2", features = ["serde"] }
 sha2 = "0.11"
 hex = "0.4"
-base64 = "0.22"
+base64 = "0.23"
 
 # Unix process-group kill (browser teardown). Unix-only; already present
 # transitively. Used by crw-renderer's BROWSER_PGIDS group-kill registry.

crates/crw-extract/Cargo.toml

@@ -23,7 +23,7 @@ pdf-inspector = { workspace = true, optional = true }
 # Used ONLY for the pre-parse decompression-bomb guard: structure-parse the PDF
 # and bounded-inflate each FlateDecode stream so a malicious file is rejected
 # before pdf-inspector allocates the full (huge) decompressed payload.
-lopdf = { version = "0.41", optional = true }
+lopdf = { version = "0.42", optional = true }
 flate2 = { version = "1", optional = true }
 lol_html = { workspace = true }
 scraper = { workspace = true }

crates/crw-extract/tests/pdf_tests.rs

@@ -74,3 +74,56 @@ fn error_codes_are_stable() {
     assert_eq!(PdfError::Corrupt("x".into()).code(), "pdf_parse_failed");
     assert_eq!(PdfError::Disabled.code(), "pdf_disabled");
 }
+
+/// RUSTSEC-2026-0187: lopdf <= 0.41 parsed nested arrays with unbounded
+/// recursion, so a ~21 KB PDF whose Catalog holds a 10,000-deep array blew the
+/// stack and aborted the process. A `SIGABRT` is not unwinding, so neither the
+/// `catch_unwind` in `pdf::convert` nor this harness can trap it: on a
+/// regression the whole test binary dies rather than reporting a failure, which
+/// is the loudest signal available. Fixed in lopdf 0.42, which we reach through
+/// pdf-inspector 1.17.
+#[test]
+fn deeply_nested_objects_do_not_abort() {
+    let depth = 10_380;
+    let nested: Vec<u8> = b"["
+        .repeat(depth)
+        .into_iter()
+        .chain(b"]".repeat(depth))
+        .collect();
+
+    let objects: Vec<Vec<u8>> = vec![
+        [
+            b"1 0 obj\n<< /Type /Catalog /Pages 2 0 R /X ".as_slice(),
+            &nested,
+            b" >>\nendobj\n",
+        ]
+        .concat(),
+        b"2 0 obj\n<< /Type /Pages /Kids [3 0 R] /Count 1 >>\nendobj\n".to_vec(),
+        b"3 0 obj\n<< /Type /Page /Parent 2 0 R /MediaBox [0 0 612 792] >>\nendobj\n".to_vec(),
+    ];
+
+    let mut pdf = b"%PDF-1.4\n".to_vec();
+    let mut offsets = Vec::new();
+    for object in &objects {
+        offsets.push(pdf.len());
+        pdf.extend_from_slice(object);
+    }
+    let startxref = pdf.len();
+    pdf.extend_from_slice(
+        format!("xref\n0 {}\n0000000000 65535 f \n", objects.len() + 1).as_bytes(),
+    );
+    for offset in &offsets {
+        pdf.extend_from_slice(format!("{offset:010} 00000 n \n").as_bytes());
+    }
+    pdf.extend_from_slice(
+        format!(
+            "trailer\n<< /Size {} /Root 1 0 R >>\nstartxref\n{startxref}\n%%EOF\n",
+            objects.len() + 1
+        )
+        .as_bytes(),
+    );
+
+    // Reaching the assertion at all is the point: extracting nothing from a
+    // page-less document is fine, aborting the process is not.
+    let _ = pdf::convert(&pdf, false, None, 5 * 1024 * 1024);
+}