fix(items): guard None refusal in ItemHelpers.extract_last_content

[jawwad-ali]

Summary

ItemHelpers.extract_last_content is typed -> str, and its ResponseOutputText branch already coerces a possibly-None text to "" (added in #3394). The sibling ResponseOutputRefusal branch, however, returned last_content.refusal unguarded — the one remaining path in this helper that can return None despite the -> str contract.

refusal is typed str by the Responses API schema, but the same provider-gateway (e.g. LiteLLM) and model_construct streaming paths that surface a None text can surface a None refusal. This guards it with or "", matching the text branch directly above and the sibling helpers extract_text, extract_refusal, and text_message_output, which all already use or "".

Test plan

• Added test_extract_last_content_tolerates_none_refusal_content in tests/test_items_helpers.py, mirroring the existing test_extract_text_tolerates_none_text_content: it builds a ResponseOutputRefusal.model_construct(refusal=None, ...) and asserts extract_last_content returns "" (and not None).
• uv run pytest tests/test_items_helpers.py — 35 passed.
• uv run ruff format and uv run ruff check are clean; uv run mypy src/agents/items.py passes.

Issue number

N/A — this completes the None-guarding introduced in #3394 (same function, text branch) and #3375 (text_message_output); the refusal branch is the lone remaining unguarded path in a helper typed -> str.

Checks

• I've added new tests (if relevant)
• I've added/updated the relevant documentation
• I've run make lint and make format
• I've made sure tests pass

[seratch]

We've been receiving this patch several times but we don't need the change. #3751 clarifies why we don't accept it.