As the title mentions, the native LSP configuration needs to set refresh=1 when calling ncm2#complete, here:

This is because some LSP servers (e.g. https://texlab.netlify.com/) refresh possible completions on every call and do not return every completion initially, so the caching means that completion results are incomplete and do not get updated as the user continues to type.

I've made this an issue rather than a PR as I don't know how broadly true this is (i.e. other LSP servers seemed to work correctly without this change), so I'm uncertain if this ought to be a global change or something configurable per-server.

To reproduce: Use texlab with the native LSP implementation. Note that you get completions when initially typing \, but continuing to type (e.g. trying to complete \hypersetup by typing \hy, etc.) will lead to empty completions. Hitting space and backspace will refresh the completion list with correct completions.