fix(ai): prefer configured model after provider switch

[trxuan]

Summary

• keep the explicitly configured chat model at the front of a fetched model list
• restore a saved model only when the current provider also advertises it
• replace a stale cross-provider selection with the configured model instead of showing (list failed) and selecting the provider's first model

Root cause

The chat overlay stores last_model globally. After switching providers, a saved model from the previous provider may not exist in the new provider's successful /models response. The overlay currently treats that mismatch as a failed model fetch and resets model_index to the first server model, even when the configured chat_model is present.

Tests

Added regression coverage for:

• a stale model from another provider falling back to the configured model
• a valid saved model remaining selected
• a manually configured model remaining available when omitted from /models

[vercel]

@trxuan is attempting to deploy a commit to the Faberon Team on Vercel.

A member of the Team first needs to authorize it.

[youxi798]

@trxuan Thanks for the provider-switch fix. I tightened the model state so cached lists stay scoped to the current endpoint, Follow Codex resolves to a discovered model, and background refreshes no longer overwrite an explicit choice.

This is merged and available in the latest Nightly: https://github.com/tw93/Kaku/releases/tag/nightly