fix: preserve raw string tool outputs

[gold-silver-copper]

Summary

Fixes #1605.

This changes tool output serialization in rig-core so that Tool::Output = String is preserved as raw text instead of being JSON-stringified again. Non-string tool outputs are still serialized to JSON text as before.

That fixes the double-escaped tool result behavior seen with OpenAI-compatible providers and also restores the intended behavior for string-returning tools that deliberately emit structured JSON for downstream parsing.

What changed

• Added a dedicated tool output serialization path in rig-core
• Preserved String outputs verbatim
• Kept object/number outputs serialized once as JSON text
• Left provider conversion code and ToolResultContent::from_tool_output unchanged

Why this approach

The bug was caused at the generic tool serialization boundary, not in provider-specific message conversion.

Previously, all tool outputs went through serde_json::to_string(&output). For String outputs, that produced a JSON string literal, which then surfaced in tool result messages as quoted/escaped text.

Fixing this at the serializer boundary is lower-risk than adding parser-side unwrapping logic:

• it resolves the reported OpenAI issue
• it keeps the public API unchanged
• it allows String tools that return structured JSON payloads to continue flowing into existing from_tool_output parsing

Tests

Added regression coverage for:

• plain multiline string tool output staying unescaped
• structured JSON returned from a String tool remaining parseable as tool result content
• object outputs still being serialized as JSON text

OpenAI’s chat tool-message schema uses role: "tool", a required tool_call_id, and content as the tool result payload, which indicates provider-facing tool results are raw text/content rather than a JSON-stringified string literal.[1] vLLM follows that same shape: its tool-calling docs say chat templates must handle tool-role messages,[2] and its official OpenAI-compatible example appends tool results as {"role": "tool", "content": result, "tool_call_id": call.id, "name": call.function.name}.[3]

So the Rig fix is aligned with the strongest available evidence: String tool outputs should stay raw when sent back through OpenAI-compatible providers like vLLM.

Sources:

• [1] OpenAI generated chat completion tool-message type
• [2] vLLM tool-calling docs
• [3] vLLM official OpenAI-compatible tool-calling example