[Bug]: MCPToolObservation silently discards structuredContent and result _meta

Is there an existing issue for the same bug?

• I have searched existing issues and this is not a duplicate.

Bug Description

When an MCP tool result carries structuredContent or result _meta, both are silently discarded at conversion: MCPToolObservation.from_call_tool_result keeps only text/image blocks, is_error, and tool_name. Once dropped there, no downstream consumer (event persistence, websocket delivery, protocol-aware clients such as Canvas) can recover them. MCP servers use structuredContent for machine-readable result data and _meta for protocol metadata.

A draft PR with a candidate fix: #4775.

Expected Behavior

The observation preserves structuredContent and result _meta so they serialize with the event, and registered secrets in those fields are masked before persistence or client delivery.

Actual Behavior

Converting a real mcp.types.CallToolResult that carries both fields yields an observation with neither. Reproduced on main at 9d143aac3 (v1.44.1):

$ python repro.py
COMMIT 9d143aac3
PYTHON 3.13.12
STRUCTURED_PRESERVED False
META_PRESERVED False

import mcp.types
from openhands.sdk.mcp.definition import MCPToolObservation

result = mcp.types.CallToolResult(
    _meta={"ui": {"hint": "Bearer token"}},
    structuredContent={"rows": [{"token": "abc"}], "count": 1},
    content=[mcp.types.TextContent(type="text", text="ok")],
    isError=False,
)
obs = MCPToolObservation.from_call_tool_result(tool_name="probe", result=result)
print("STRUCTURED_PRESERVED", getattr(obs, "structured_content", None) is not None)
print("META_PRESERVED", getattr(obs, "result_meta", None) is not None)

Steps to Reproduce

1. Check out main at 9d143aac3 and install with uv sync
2. Run the script above with the repo interpreter
3. Both lines print False: the converted observation carries neither structuredContent nor _meta, and model_dump loses them

Acceptance Criteria

• structuredContent and result _meta survive conversion and event serialization
• Observations serialized before the fields existed still load
• Registered secrets in the new fields are masked, including nested keys, nested values, and JSON scalars
• The fields stay out of the content sent to the model

Installation Method

pip install -e . from a git checkout of main

SDK Version

main branch, commit 9d143aac3 (v1.44.1)

Python Version

3.13.12

Operating System

Linux

Start at MCPToolObservation.from_call_tool_result and run the repro.py example against the stated commit to confirm the fields are dropped. Trace the observation's event serialization and secret-masking paths; done means structuredContent and result _meta survive conversion and loading, nested secrets are masked, and neither field reaches model content.