Add test coverage for MCP server tools

[YaMachinu]

Problem

src/agentevals/mcp_server.py exposes 5 MCP tools for Claude Code integration — a key part of the product's developer story — but there is zero test coverage for the MCP server:

• No unit tests for any of the 5 tool handlers
• • No integration tests verifying the MCP server starts and responds correctly
• • • No tests for error cases (invalid session IDs, missing eval sets, etc.)
      The MCP server is explicitly called out in the README as a primary interface. Shipping it untested is a reliability risk.

Tools That Need Test Coverage

The 5 MCP tools exposed (from mcp_server.py):

1. list_sessions — list available sessions
2. 2. get_session — retrieve session detail
3. 3. run_evaluation — trigger evaluation against an eval set
4. 4. list_eval_sets — list configured eval sets
5. 5. get_evaluation_result — retrieve evaluation results

Suggested Test Approach

Create tests/test_mcp_server.py with:

# Unit tests using mock TraceManager
async def test_list_sessions_empty():
    ...

async def test_run_evaluation_success():
    ...

async def test_run_evaluation_invalid_session():
    ...

Use pytest-asyncio (already a dev dependency) and mock the TraceManager and evaluator pipeline.