fix(openai): raise on empty chat completion instead of clean end_turn

[gadievron]

Root cause

The OpenAI Chat Completions translation (_response_to_unified in utilities/llm/providers/openai.py) guarded an empty choices array but not an empty message. A choice whose message.content is None/empty with no tool_calls fell through to a clean end_turn carrying zero content blocks. For a security tool, an empty end_turn is read as a clean, passing verdict — so a blank completion became a silent false-negative.

This is the same silent-empty-completion shape #207 closed across the other paths, but the chat path never had the guard (it predates #207, which never touched openai.py). The three sibling paths already raise on empty content:

• Responses path — openai.py _responses_to_unified "no usable content" guard
• Anthropic adapter — R4-1 empty-completion guard
• Gemini adapter — empty-candidates guard

Fix

Add the missing no-usable-content guard to the chat path, mirroring the siblings: after the content_filter refusal check (the more specific signal, raised first), if not content_blocks: raise LLMResponseError(...). A tool-use-only response stays valid because content_blocks is non-empty.

Reproduction

A ChatCompletion with one choice, message.content=None, tool_calls=None, finish_reason="stop".

Regression test

tests/test_llm_openai_adapter.py::test_empty_content_raises_llm_response_error — sibling of the existing test_empty_choices_raises_llm_response_error.

# RED (before fix):
FAILED tests/test_llm_openai_adapter.py::test_empty_content_raises_llm_response_error
# GREEN (after fix):
72 passed
# Full suite unchanged otherwise:
2413 passed, 149 skipped, 0 failed   (baseline 2412; +1 = this test)

Compatibility

No API/signature change. Behavior change is narrow: a chat response with neither text nor tool calls now raises LLMResponseError instead of returning an empty end_turn. Tool-use-only and normal text responses are unaffected; refusal/content_filter still raises first.

Surfaced by a post-merge collision review of #206/#207 (this gap is pre-existing, not introduced by either).