[codex] Fix automatic Gemini model discovery

[222wcnm]

Summary

This PR fixes automatic Gemini model discovery while preserving OpenAI-compatible model listing behavior.

Changes

• Route native Gemini /v1beta/models requests to the Gemini API instead of the OpenAI compatibility handler.
• Keep OpenAI /v1/models model listing converted to OpenAI-style response objects.
• Support comma-separated Gemini query key values by selecting one key before proxying.
• Request a larger model list page for OpenAI-compatible model discovery.
• Add focused node:test coverage for Gemini model listing, OpenAI model listing, and query-key selection.

Impact

Clients that call the native Gemini model listing endpoint can auto-refresh available models through this proxy, while OpenAI-compatible clients continue to receive the expected /v1/models shape.

Validation

• node --test

Note: Node emits an existing MODULE_TYPELESS_PACKAGE_JSON warning because package.json does not declare type: module; tests pass.