fix(llm): Claude provider strips tool calls and results from the conversation #3389

src/gaia/llm/providers/claude.py:231, _split_system:

if content is None or content == "":
 logger.debug("Dropping empty %s message for Claude request", role)
 continue
cleaned.append({"role": role, "content": content})

Two independent losses:

1. Whole-turn drop. _build_assistant_message sets content=None when the model
   returned tool calls and no text. That turn is skipped, tool calls included.
2. Silent key strip. The cleaned.append keeps only role and content. A turn with
   both narration and tool calls survives the first check and still loses its
   tool_calls. This is the worse of the two because it looks fine.

There is no outbound translation to Anthropic's shape anywhere in the provider. The three
tool_use references (claude.py:59, :358, :411) are all inbound — parsing Claude's
response back into OpenAI form. The reverse direction does not exist.

Anthropic expects:

• assistant turn: content: [{"type":"tool_use","id":...,"name":...,"input":{...}}]
• result turn: role="user" with content: [{"type":"tool_result","tool_use_id":...,"content":...}]

tests/unit/test_claude_provider.py covers the outbound direction only
(test_tool_use_reencoded_as_lemonade_sentinel); no case feeds the provider an
assistant-message-with-tool_calls or a role=tool message, which is why neither loss is
caught.

Found while investigating #3387 (GAIA never emits a parallel tool call). That benchmark
bypassed this provider, so the two are independent — but this one would discard fan-out
even after #3387 is fixed.