Prevent auto normalise line endings for pdf documents

Author: BobLd

.gitattributes

@@ -3,6 +3,23 @@
 ###############################################################################
 * text=auto
 
+###############################################################################
+# Binary test assets must never be line-ending normalized.
+#
+# Some test PDFs are fully uncompressed (no NUL bytes), so git's text=auto
+# heuristic misdetects them as text and strips CR bytes when storing the blob.
+# That corrupts the binary xref byte offsets, so the file checks out broken on
+# platforms that use LF (Linux/macOS) while Windows autocrlf masks the issue.
+# Forcing these as binary keeps the bytes identical on every platform.
+###############################################################################
+*.pdf binary
+*.png binary
+*.jpg binary
+*.jpeg binary
+*.gif binary
+*.ico binary
+*.snk binary
+
 ###############################################################################
 # Set default behavior for command prompt diff.
 #

CLAUDE.md

@@ -23,6 +23,8 @@ dotnet test --filter "FullyQualifiedName~ClassName.MethodName" UglyToad.PdfPig.R
 
 Test framework: **xUnit**. Test PDFs live in `UglyToad.PdfPig.Rendering.Skia.Tests/Documents/`.
 
+> **Binary test assets must stay binary in git.** `.gitattributes` has `* text=auto`, but uncompressed PDFs contain no NUL bytes, so git's heuristic misdetects them as *text* and strips CR bytes on storage. A PDF's xref uses absolute byte offsets, so dropping CRs shifts the layout and breaks them: the file checks out fine on Windows (`autocrlf` restores CRLF) but corrupt on Linux/macOS (LF), giving `IndexOutOfRangeException` in `MemoryInputBytes.Seek` / `PdfTokenScanner.TryReadStream`. The bug is byte-based, not OS-based — feeding the LF blob to the renderer fails on any OS. `.gitattributes` therefore force-marks `*.pdf` (and `*.png/jpg/jpeg/gif/ico/snk`) as `binary`. After adding a new test PDF, confirm `git check-attr -a <file>` reports `binary: set`; if it was committed before the rule, run `git add --renormalize .`.
+
 #### Image-regression tests (`PdfPigSkiaTest`)
 
 The bulk of the suite renders a PDF page and compares it pixel-by-pixel against a committed golden PNG. Things to know before touching the renderer: