duncanmcqueen stops Gemini becoming the invisible default

This fork makes AI model choice follow what a user or deployment has actually configured, rather than quietly falling back to Gemini.

Today, a missing or outdated model choice can send work to Gemini without the user realising it. That is a poor fit for legal teams trying to control providers, costs and troubleshooting.

Duncanmcqueen's proposed change makes the product more explicit. It would:

retain the existing Gemini fallback only for deployments set up with Google sign-in;
otherwise choose from the signed-in user's usable provider settings and saved choices;
show a clear "Select model" state when no suitable option exists;
reject unknown or unavailable choices with an actionable configuration message, rather than silently switching providers;
carry the same approach into chat, table-based work and the Word add-in.

The practical shift is simple: a failed request should point to missing setup, not conceal an unchosen AI provider.