Refusals are retried as malformed output and fail with an unrelated error #43

A refusal from either provider is treated as malformed JSON. Anthropic returns stop_reason="refusal" with empty content, and Responses returns a completed message whose only part is a refusal, so output_text is empty. anthropic._result only checks max_tokens and openai._responses_result only checks status, so the empty text reaches _decode_or_correct, which spends a corrective retry on it and replays the empty string as the assistant turn. Anthropic stop_reason="model_context_window_exceeded" goes the same way.

What a caller sees with n_retry_malformed_structure=2:

• Anthropic: 2 requests, then TypeSafeBadRequestError: 400 messages.1: all messages must have non-empty content except for the optional final assistant message, because the Messages API rejects the empty assistant turn. Nothing in the error says refusal.
• Responses: 3 requests, all refused, then TypeSafeAPIResponseValidationError: 200 Invalid response data at 'answers'.

The README already says a max_tokens cutoff does not consume malformed-output retries. Refusal and context overflow are the same kind of provider-declared non-answer and should raise the same way.

Repro, no network:

git fetch https://github.com/Dnyaneshvn/system-one-adapter-python fix/provider-refusals
git checkout FETCH_HEAD -- tests/test_provider_requests.py tests/test_openai_transports.py
uv run pytest tests/test_provider_requests.py tests/test_openai_transports.py -k "refusal or stop_reason" -q

8 failed on main. The client-level tests go through the real SDK parsers with the HTTP layer stubbed.

Fix on https://github.com/Dnyaneshvn/system-one-adapter-python/tree/fix/provider-refusals:

• anthropic._result raises TypeSafeError("Anthropic response did not complete: {stop_reason}.") for any stop reason other than end_turn, stop_sequence, or None. The max_tokens message is unchanged.
• openai._responses_result raises TypeSafeError("OpenAI response was a refusal: {text}") when a message item carries a refusal part.
• Both run after record_response, so llm_attempts keeps the full response and finish reason.
• 11 regression tests in the existing test files. Chat Completions is left alone since Chat Completions accepts finish_reason="length" as a successful typed evaluation #38 covers finish_reason there.

evals/refusal_handling/ on the branch runs 12 cases (abusive tickets in English, Hinglish and Marathi, a self-harm transcript, a doxxing post, a contract over the context window, two benign controls) through both providers with the HTTP layer stubbed:

On main the 8 unhandled non-answers cost 12 extra requests and 12 corrective retries between them, and every corrective request replayed an empty assistant turn. On the branch they cost 0 of each. The two already-handled outcomes and the two benign controls are unchanged.

Tests: uv run pytest -q (240 passed). uv run pyrefly check: 0 errors.