fix: LSP server should only exit on the exit notification

[prql-bot]

Problem

The LSP server's main_loop returned Ok(()) on any notification:

Message::Notification(not) => {
    eprintln!("got notification: {not:?}");
    return Ok(());
}

Connection::initialize() consumes the initialized notification during the handshake, so the first notification reaching main_loop in a real session is something like textDocument/didOpen or $/setTrace — and the server would shut down the moment a client sent one. Worse, returning early drops the connection's sender, so a follow-up message panics with sending on a disconnected channel and the process exits with code 1.

The existing lsp test only sent an exit notification, so it never exercised a non-exit notification and masked the bug.

Solution

Only the exit notification stops the loop; other notifications are logged and ignored.

Message::Notification(not) => {
    eprintln!("got notification: {not:?}");
    if not.method == "exit" {
        return Ok(());
    }
}

Testing

Added lsp_ignores_non_exit_notification, which sends a $/setTrace notification before exit. Before the fix it fails with success: false, exit_code: 1, and sending on a disconnected channel; after the fix the server logs the notification, continues, and shuts down cleanly on exit. The existing lsp test is unchanged.

Found during the nightly code-quality survey.

[prql-bot]

The fix is correct — exit is the right notification to terminate on, and ignoring others matches LSP semantics. I confirmed both lsp tests pass once the binary is built with --features lsp.

One observation, not a blocker: this regression test won't run in CI. No test job enables the lsp feature — the test-rust matrix uses features: default,test-dbs-external and default (tests.yaml), lsp isn't in default (default = ["cli"]), and no workflow uses --all-features. So the #[cfg(feature = "lsp")] tests — both the new one and the pre-existing lsp test — are never compiled or run in CI; they only guard the fix for developers who run cargo test --features lsp locally.

This gap predates the PR (the existing lsp test has it too), so it's reasonable to leave as-is given lsp is a stub feature. But since the PR's value is the added regression coverage, a maintainer may want to add an lsp-enabled leg to the CI matrix in a follow-up so the test actually protects the fix.