Description

from_dict should be side-effect free, but both GeneratedAnswer.from_dict and ExtractedAnswer.from_dict wrote parsed Document / ChatMessage / Span objects back into the caller's dictionary. Deserializing the same dict a second time (common in replay/logging flows that keep the serialized form around) then crashed:

from haystack.dataclasses import GeneratedAnswer, ChatMessage, Document
ga = GeneratedAnswer(
    data="answer", query="q",
    documents=[Document(content="d")],
    meta={"all_messages": [ChatMessage.from_user("hi")]},
)
s = ga.to_dict()
GeneratedAnswer.from_dict(s)   # OK, but mutates s in place
GeneratedAnswer.from_dict(s)   # AttributeError: 'Document' object has no attribute 'copy'

Document.from_dict already guards against this with data = data.copy(); this extends the same idea using deepcopy, since the mutated values are nested inside init_parameters / meta.

Test plan

• Added test_from_dict_does_not_mutate_input to both TestExtractedAnswer and TestGeneratedAnswer, each asserting the input dict is unchanged after from_dict and that a second deserialization of the same dict still succeeds and is equal.
• The tests fail on main (input mutated / second call raises) and pass with this change.
• Added a release note.

AI disclosure

This change was prepared with the help of an AI coding assistant; the bug was reproduced and the behavior verified before submitting.