fix(buffs): use prompt setter in paraphrase buffs to update conversations

[bodapatisaikrishna]

Description

In garak/buffs/paraphrase.py, PegasusT5.transform() and Fast.transform() previously assigned paraphrased messages directly to the private attribute paraphrased_attempt._prompt:

delattr(paraphrased_attempt, "_prompt")  # hack to allow prompt set
paraphrased_attempt._prompt = garak.attempt.Message(
    text=paraphrase, lang=last_message.lang
)

Directly assigning _prompt = Message(...) bypassed the Attempt.prompt property setter. This led to several issues:

1. Unupdated conversations: The prompt setter is responsible for updating self.conversations. Bypassing it meant paraphrased_attempt.conversations was never updated with the paraphrased text and retained the original prompt. Downstream generator executions or reports reading conversations received the unparaphrased text.
2. Incorrect prompt type: Attempt.prompt returned <class 'garak.attempt.Message'> instead of <class 'garak.attempt.Conversation'>.
3. AttributeError on attempt.lang: Attempt.lang expects self.prompt to be a Conversation and accesses self.prompt.turns[-1].content.lang. Calling .lang on any attempt transformed by PegasusT5 or Fast raised AttributeError: 'Message' object has no attribute 'turns'.

Fix

Update both PegasusT5.transform() and Fast.transform() to use the public property setter:

paraphrased_attempt.prompt = garak.attempt.Message(
    text=paraphrase, lang=last_message.lang
)

This matches the correct pattern used in other buffs (garak/buffs/lowercase.py, garak/buffs/encoding.py, garak/buffs/low_resource_languages.py).

Testing

• Added validation in test_buff_load_and_transform in tests/buffs/test_buffs.py verifying that every transformed attempt's prompt is a Conversation and that .lang is accessible without error.
• Added dedicated unit test test_paraphrase_transform_conversation_and_lang in tests/buffs/test_buffs.py.

Verification

• Run unit tests and ensure they pass:

  pytest tests/buffs/test_buffs.py -k test_paraphrase_transform_conversation_and_lang

  Result: 1 passed.
• Run full buff test suite:

  pytest tests/buffs/test_buffs.py tests/buffs/test_buff_config.py

  Result: 14 passed, 1 skipped.
• Run linter and formatting:

  ruff check garak/buffs/paraphrase.py tests/buffs/test_buffs.py
  ruff format --check garak/buffs/paraphrase.py tests/buffs/test_buffs.py

  Result: All checks passed.
• Verify the thing does what it should: paraphrased_attempt.prompt is a Conversation, .lang returns the language string, and paraphrased_attempt.conversations reflects the new paraphrase text.
• Verify the thing does not do what it should not: No regressions in existing buff transform behavior.

[jmartin-tech]

Nice catch this is a clear error from the Conversation support refactor.