fix: use output for python faithfulness statements

[Janhavi Sathe (invocan-jsathe)]

Summary

• Fix a bug in Python Faithfulness where statements were extracted from expected instead of output, causing the scorer to evaluate the wrong text.
• Update both sync and async Faithfulness paths in py/autoevals/ragas.py to route statement extraction through output.
• Add a regression test in py/autoevals/test_ragas.py that uses mismatched output/expected and input-driven mocks so the final score depends on correct routing.

Bug

Faithfulness should score whether claims in the generated answer are supported by context.
In Python, it incorrectly passed expected to statement extraction, so claims were taken from ground truth rather than model output.

Fix

• In Faithfulness._run_eval_async(...): change answer=expected -> answer=output.
• In Faithfulness._run_eval_sync(...): change answer=expected -> answer=output.
• Align sync required-field validation with async by requiring output as well.

Test

Added test_faithfulness_extracts_statements_from_output:

• Uses different output and expected.
• Mocks extract_statements to derive statements from passed answer.
• Mocks extract_faithfulness to derive verdicts from context containment.
• Ensures score behavior reflects correct routing (would fail under old bug).

Validation

• uv run --extra dev --extra scipy pytest py/autoevals/test_ragas.py -k faithfulness_extracts_statements_from_output passed.
• Pre-commit hooks pass on commit.

[Barrett Pyke (barrettpyke)]

Abhijeet Prasad (@AbhiPrasad) - Took a look a this one and it lgtm but wanted to run it past you

[github-actions]

Braintrust eval report

Autoevals (main-1781207099)

Score	Average	Improvements	Regressions
NumericDiff	78% (+0pp)	7 🟢	11 🔴
Time_to_first_token	10.39tok (-0.38tok)	155 🟢	61 🔴
Llm_calls	1.55 (+0)	-	-
Tool_calls	0 (+0)	-	-
Errors	0 (+0)	1 🟢	1 🔴
Llm_errors	0 (+0)	-	-
Tool_errors	0 (+0)	-	-
Prompt_tokens	526.63tok (-0.58tok)	1 🟢	1 🔴
Prompt_cached_tokens	0tok (+0tok)	-	-
Prompt_cache_creation_tokens	0tok (+0tok)	-	-
Prompt_cache_creation_5m_tokens	0tok (+0tok)	-	-
Prompt_cache_creation_1h_tokens	0tok (+0tok)	-	-
Completion_tokens	467.56tok (+2.66tok)	108 🟢	100 🔴
Completion_reasoning_tokens	356.36tok (+2.04tok)	92 🟢	88 🔴
Completion_accepted_prediction_tokens	0tok (+0tok)	-	-
Completion_rejected_prediction_tokens	0tok (+0tok)	-	-
Completion_audio_tokens	0tok (+0tok)	-	-
Total_tokens	994.19tok (+2.08tok)	108 🟢	100 🔴
Estimated_cost	0$ (+0$)	91 🟢	88 🔴
Duration	10.43s (-0.35s)	156 🟢	62 🔴
Llm_duration	11.11s (-0.64s)	160 🟢	57 🔴