[Altafk960]

Systematic hardening of PdfParser, closing the serialization half of the max_nesting work (#10023, #10093, #10125, #10127, #10185). Those changes bounded the document-graph traversals and the parse-side recursion (get_value/read_indirect max_nesting), but the serialize path (pdf_repr, PdfArray.__bytes__, PdfDict.__bytes__) still recursed with no depth bound, and two reachable lookups in read_indirect/rewrite_pages could still raise bare KeyErrors on malformed documents.

Changes (2 commits)

1. Bound serialization recursion in pdf_repr — thread a depth counter through pdf_repr and raise PdfFormatError past a conservative MAX_NESTING = 100, mirroring the parse-side max_nesting guard. PdfArray.__bytes__/PdfDict.__bytes__ route through pdf_repr so the bound applies to bytes() as well. The bound is kept conservative because serialization runs several Python frames per nesting level at arbitrary call depth (for example inside Image.save(..., append=True)), and it keeps the stack far below Python's recursion limit. Output for values within the bound is byte-identical — the existing test_pdf_repr byte-exact assertions are unchanged.
2. Raise PdfFormatError for dangling object references — read_indirect now validates the object id against the xref table via check_format_condition instead of leaking a bare KeyError, and rewrite_pages skips unresolvable /Parent references when rewriting an existing document, matching the check_format_condition idiom used throughout the parser.

Tests

3 regression tests: the serialization depth guard (pdf_repr and bytes(), at and just past MAX_NESTING), a dangling /Kids reference raising PdfFormatError instead of KeyError, and a page with a dangling /Parent surviving im.save(..., append=True) without an exception. The full test suite passes.

[radarhere]

It seems unexpected to have MAX_NESTING be entirely separate from the default behaviour of

Pillow/src/PIL/PdfParser.py

Lines 843 to 849 in f58197e

	def get_value(
	cls,
	data: bytes | bytearray | memoryview | mmap.mmap,
	offset: int,
	expect_indirect: IndirectReference | None = None,
	max_nesting: int = 900,
	) -> tuple[Any, int | None]:

and

Pillow/src/PIL/PdfParser.py

Line 1076 in f58197e

def read_indirect(self, ref: IndirectReference, max_nesting: int = 900) -> Any:

[Altafk960]

The two limits guard different operations, so I've made the separation explicit instead of sharing one number. Reading keeps the max_nesting defaults on get_value and read_indirect exactly as 13.0.0 shipped them. Writing stops at 100 because serialization spends several Python frames per nesting level and runs at arbitrary call depth mid-save — RecursionError lands around 250 levels, so a 900 bound on the write path could never fire. Unifying at 100 parses every file in Tests/ identically, but it would tighten the default 13.0.0 published, so I've renamed the constant to MAX_SERIALIZE_NESTING and left the read side alone. If you'd still rather have one shared limit, it's a two-line change. Release notes are now in both PRs as well.

(The three failing checks also fail on current main — the CodSpeed job (Python 3.15), run 40194988170, and the Docker jobs, run 40194988114 — so they don't look related to either PR.)