ROB: Recover a corrupt trailing startxref pointer (closes #3238)

[binggao1230]

Summary

Some PDFs end with a broken trailing cross-reference pointer left behind by a faulty incremental update — the startxref keyword for the final %%EOF has been corrupted (e.g. it reads tartxref), while an earlier, intact startxref from a previous revision is still present further up the file:

trailer
<<...>>
startxref
160113
%%EOF
tartxref      <- corrupt keyword (missing leading 's')
160135
%%EOF

Today PdfReader aborts on such a file with PdfReadError: startxref not found, even though the document is otherwise fully readable from the earlier pointer.

This was noted in #3238, where the suggestion was to "circumvent this by further looking for another startxref". This PR does exactly that: when the startxref keyword expected just above the offset is missing or corrupt, it scans further back for the most recent intact pointer rather than giving up.

Details

• _find_startxref_pos now delegates to a new _find_previous_startxref_pos fallback when the keyword line is not where it should be.
• The fallback walks backwards, and because the offset always sits on the line directly below the keyword, it returns the value read immediately before the recovered startxref.
• The change is purely additive: it only runs when the normal path would have raised, so well-formed files are unaffected. If no earlier pointer exists, the original startxref not found error is still raised.

On the file from #3238 the reader now opens all 11 pages and extracts text correctly, using the valid startxref at 160113.

Test

test_startxref_corrupt_trailing_pointer builds a minimal PDF with a valid startxref/%%EOF followed by a corrupt tartxref trailer and asserts the page is recovered. It runs fully offline (no sample files).

AI use disclosure

Per the project's AI policy: I used AI coding assistance while preparing this PR. Tool: Claude Code. Model: Claude Opus 4.8 (claude-opus-4-8). I've reviewed, run, and verified the change myself and take responsibility for it; the PR text and my replies in this thread are written in my own words.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 97.76%. Comparing base (52545c5) to head (7dee5c7).
⚠️ Report is 174 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3826      +/-   ##
==========================================
+ Coverage   97.73%   97.76%   +0.02%     
==========================================
  Files          55       55              
  Lines       10417    10457      +40     
  Branches     1931     1943      +12     
==========================================
+ Hits        10181    10223      +42     
+ Misses        130      129       -1     
+ Partials      106      105       -1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[stefan6419846]

Thanks for the PR.

Please have a look at the review comments and coverage.

[binggao1230]

Pushed 3d324ba to cover the patch-coverage gap: added direct tests for _find_previous_startxref_pos exercising the same-line keyword form, the offset-below-keyword form, and the two unrecoverable cases (a startxref keyword with no offset above it, and no keyword at all) that raise PdfReadError. All the review threads from earlier should be addressed now — let me know if anything else is needed.