fix: pdfminer drops extractable text (#4310)

<!-- CURSOR_SUMMARY -->
> [!NOTE]
> **Medium Risk**
> Changes pdfminer integration to override CID font/CMap handling and
introduces custom stream decoding/parsing, which can affect text
extraction behavior and performance on diverse PDFs (mitigated by
size/mapping caps).
> 
> **Overview**
> Fixes PDFs where **body text was silently dropped** because CIDFonts
used an *embedded Encoding CMap stream* that `pdfminer.six` doesn’t
resolve.
> 
> Adds a bounded embedded-CMap decoder/parser and wires it in via
`CustomPDFCIDFont` + `CustomPDFResourceManager` so `init_pdfminer()`
constructs CID fonts with a parsed CMap (including `WMode`), with
DoS-oriented caps on decompression and total mappings.
> 
> Updates tests with a new fixture-driven regression for both `FAST` and
`HI_RES` strategies plus targeted unit tests for CMap parsing/stream
decoding, and bumps version to `0.22.12` with a changelog entry.
> 
> <sup>Written by [Cursor
Bugbot](https://cursor.com/dashboard?tab=bugbot) for commit
4326b15. This will update automatically
on new commits. Configure
[here](https://cursor.com/dashboard?tab=bugbot).</sup>
<!-- /CURSOR_SUMMARY -->

---------

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

Author: qued

CHANGELOG.md

@@ -1,3 +1,8 @@
+## 0.22.12
+
+### Fixes
+- **Fix fast strategy silently skipping text in some PDFs**: Certain PDF generators (e.g. Prince XML) embed font encoding data in a non-standard way that pdfminer.six does not handle, causing body text to be silently dropped while headings still extract correctly. Added a workaround that reads the embedded encoding data directly.
+
 ## 0.22.11
 
 ### Enhancements

example-docs/pdf/embedded-cmap-cidfont.pdf

@@ -477,6 +477,44 @@ def test_partition_pdf_with_fast_strategy_deduplicates_fake_bold(monkeypatch):
     )
 
 
+def test_partition_pdf_with_fast_strategy_extracts_embedded_cmap_text():
+    """Test that fast strategy extracts text from CIDFonts with embedded CMap streams.
+
+    Some PDF generators (e.g. Prince XML) embed custom Encoding CMaps as PDF streams
+    rather than using predefined CMap names. Without handling this, pdfminer.six silently
+    falls back to an empty CMap and all text using those fonts is lost.
+
+    The test fixture has two fonts: a simple Type1 font (Helvetica) that pdfminer handles
+    fine, and a Type0/CIDFont with an embedded CMap named "Test-Identity-H" that triggers
+    the bug.
+    """
+    filename = example_doc_path("pdf/embedded-cmap-cidfont.pdf")
+    elements = pdf.partition_pdf(filename=filename, url=None, strategy=PartitionStrategy.FAST)
+
+    all_text = " ".join(e.text for e in elements)
+
+    # The Helvetica heading should always be extracted
+    assert "Heading in Helvetica" in all_text
+
+    # These strings are rendered with the CIDFont using the embedded CMap.
+    # Without the fix, they would be silently dropped.
+    assert "This text uses an embedded CMap" in all_text
+    assert "and should be extractable" in all_text
+
+    assert len(elements) == 3
+
+
+def test_partition_pdf_with_hi_res_strategy_extracts_embedded_cmap_text():
+    """Same as the fast strategy test but through hi_res, since both strategies use pdfminer."""
+    filename = example_doc_path("pdf/embedded-cmap-cidfont.pdf")
+    elements = pdf.partition_pdf(filename=filename, url=None, strategy=PartitionStrategy.HI_RES)
+
+    all_text = " ".join(e.text for e in elements)
+
+    assert "This text uses an embedded CMap" in all_text
+    assert "and should be extractable" in all_text
+
+
 def test_partition_pdf_raises_with_bad_strategy():
     filename = example_doc_path("pdf/layout-parser-paper-fast.pdf")
     with pytest.raises(ValueError):