Don't abort PDF conversion on a single malformed page

[assinscreedFC]

Summary

Closes #2071.

PdfConverter.convert ran the entire per-page pdfplumber loop inside a single try, and on any page error it abandoned the whole loop and fell back to an unguarded whole-document pdfminer.high_level.extract_text call. A single malformed page — for example a form XObject whose /BBox is not a four-number rectangle, which makes the PDF backend raise ValueError: not enough values to unpack (expected 4, got 2) — therefore aborted the entire conversion with FileConversionException, even when every other page was fine.

Change

• Each page's extraction is wrapped in try/except so one malformed page is skipped (logged at debug level via a module logger) while the text recovered from the other pages is kept. page.close() moved to a finally so cached page data is always freed.
• The form_page_count == 0 whole-document pdfminer call is now guarded and falls back to the per-page pdfplumber text already collected.
• The two remaining extract_text fallbacks are guarded so they degrade to "" instead of raising.

The deeper root cause is a robustness gap in pdfminer.six (it unpacks /BBox without validating its length); that is being fixed upstream separately. This PR makes markitdown degrade gracefully regardless of the backend version.

Tests

Added packages/markitdown/tests/test_pdf_invalid_bbox.py (self-contained, building the PDF in-memory): a two-page document whose second page has a malformed form /BBox converts without raising and still recovers the first page's text.

pytest packages/markitdown/tests/test_pdf_invalid_bbox.py packages/markitdown/tests/test_pdf_masterformat.py → passes; black --check clean.

---

AI assistance (Claude) was used; I have reviewed every line and run the tests.

[assinscreedFC]

@microsoft-github-policy-service agree

[noezhiya-dot]

This is exactly the kind of robustness fix that a document converter needs. PDFs in the wild are messy, and one malformed page shouldn't tank the whole conversion.

The changes are well-structured:

• Per-page try/except with debug logging is the right granularity — you get visibility into failures without losing good data.
• Moving page.close() to finally is a good cleanup, prevents cached page data from leaking.
• Guarding the pdfminer fallback calls is defensive but necessary — if pdfplumber couldn't handle it, pdfminer might not either.
• The fallback chain (pdfplumber per-page -> pdfminer whole doc -> already collected chunks -> empty string) degrades gracefully at every level.

One suggestion: consider logging the page-level skip at warning level instead of debug. In production use, a user might want to know that pages were skipped without having to enable debug logging. But that's a minor preference.

The test is thorough — building a malicious PDF in-memory with an invalid /BBox is clever and self-contained. LGTM.