bug: AnswerBuilder.run() mutates input documents' meta dict

[sachinn854]

Describe the bug

AnswerBuilder.run() permanently modifies the meta dict of the original input Document objects.
After calling run(), every input document gains a source_index key (and a referenced key when
reference_pattern is set) — even though the caller never asked for that.

Error message

No error is thrown — this is a silent correctness bug.

Expected behavior

doc.meta should remain unchanged after run() returns. The source_index key should only appear
on the copy of the document inside GeneratedAnswer.documents, not on the original input document.

To Reproduce

from haystack import Document
from haystack.components.builders import AnswerBuilder

doc = Document(content="Paris is the capital of France.", meta={"source": "wiki"})
builder = AnswerBuilder()
builder.run(query="Capital of France?", replies=["Paris."], documents=[doc])

print(doc.meta)
# {"source": "wiki", "source_index": 1}  ← original was mutated

Additional context

Root cause: answer_builder.py line 207:

doc_meta: dict[str, Any] = doc.meta or {}

doc.meta or {} returns the SAME dict object when meta is non-empty (truthy). The next line
doc_meta["source_index"] = idx + 1 then mutates the original doc.meta directly.

dataclasses.replace(doc, meta=doc_meta) was clearly intended to avoid mutation, but the
shallow alias defeats it.

Fix: doc_meta: dict[str, Any] = dict(doc.meta)

Note: chat_prompt_builder.py even has an explicit comment "use dataclasses.replace to avoid
in-place mutation" — answer_builder.py was missed. Similar bug was fixed in Document.from_dict
via PR #11330.

FAQ Check

• Have you had a look at our new FAQ page?

System:

• OS: Windows 11
• GPU/CPU:
• Haystack version: main branch
• DocumentStore:
• Reader:
• Retriever:

[julian-risch]

Thanks, feel free to open a PR!

[sachinn854]

PR raised: #11375

[sachinn854]

Note: there's a duplicate PR #11378 for the same issue.