I always thought that the documentation is the protocol.

In other words, the source of truth would be what is written on the website labeled as "Specification" under https://microsoft.github.io/language-server-protocol/specifications/specification-current, and not some implementation detail in a vscode-languageserver-node library.

That said, in the client implementation for Sublime Text we already have a workaround for exactly this use case (more precisely one part of the proposed changes), because apparently there is already a language server that violates the current specs for this return type.

Of course I don't know how other clients handle it, or whether they would crash if null is returned and they try to iterate the results directly.