Add integration test that starts the MCP server and calls tools

[raphasouthall]

The current test suite covers the underlying logic (search, graph, triples etc.) but does not test the MCP server layer (`src/neurostack/server.py`) end-to-end. This means a bug in how a tool marshals its arguments or formats its JSON response could go undetected.

Add an integration test that:

1. Spins up the MCP server against a test fixture vault
2. Calls at least `vault_search`, `vault_stats`, and `vault_prediction_errors` via the MCP tool interface
3. Asserts the response structure is correct (keys present, types correct)

Where to look:

• `tests/` — existing test structure and fixtures
• `src/neurostack/server.py` — the MCP tool definitions
• Look at how existing tests set up the database/vault fixture and replicate that pattern

Notes:

• You don't need to test over the network — the MCP tools are regular Python functions decorated with `@mcp.tool()`. You can call them directly in tests after importing from `server.py`, bypassing the stdio transport.
• Check whether there is already a fixture vault in `tests/` or `vault-template/` you can reuse.

Acceptance criteria:

• At least one integration test file under `tests/` that imports and calls MCP tool functions directly
• Tests cover `vault_search`, `vault_stats`, and `vault_prediction_errors`
• Each test asserts response is valid JSON and contains expected top-level keys
• Tests pass in CI (`pytest tests/`)
• No new external dependencies introduced

[clawtom]

I'd like to work on this. My plan: create tests/test_server_integration.py that imports vault_search, vault_stats, and vault_prediction_errors directly from neurostack.server, sets up a temp vault+DB using a fixture similar to the existing populated_db pattern in conftest.py, and asserts each tool returns valid JSON with the expected top-level keys. Will patch neurostack.server.DB_PATH to point to the test DB so tools use test data without hitting the real vault. Does this approach match what you had in mind?