fix(#80): keep MCP server alive until client disconnects

[randomm]

Root Cause

rmcp::serve_server().await only handles the MCP initialization handshake and returns a RunningService handle. The actual message serving (tools/list, store_memory, etc.) runs as a background tokio task. The previous code dropped the handle immediately, cancelling the background task before any tool calls could be served.

Fix

Add service.waiting().await to keep block_on alive until the client disconnects:

let service = rmcp::serve_server(handler, (stdin, stdout))
    .await
    .map_err(|e| Error::Config(format!("MCP server error: {}", e)))?;
service.waiting().await
    .map_err(|e| Error::Config(format!("MCP server task error: {}", e)))?;

This matches the pattern used by reference MCP server implementations with rmcp v1.2.

Changes

• src/mcp/server.rs: Add service.waiting().await — core lifecycle fix
• src/mcp/tools.rs: Make StoreWrapper and methods pub(crate) for test access
• src/mcp/tests.rs: Add #[tokio::test] exercising ingest/search/list data path
• Cargo.toml: Add tokio to [dev-dependencies] for #[tokio::test]

Quality Gates

• cargo fmt --check ✅
• cargo clippy -- -D warnings ✅
• cargo test ✅ (162 passed)
• cargo test --no-default-features ✅ (154 passed)
• cargo clippy --no-default-features -- -D warnings ✅

Fixes #80
Part of #72