Gemini plugin and Vercel providers build the whole history as an array on every send #3869

Part of #3868 (follow-ups to #854).

After #3852, OpenAI chat, OpenAI Responses and Anthropic request bodies are built one journal row at a time and streamed, so no whole-history array exists during a send. Two providers are a declared exception:

• Gemini plugin (plugins/google-gemini): it sends through @google/genai generateContent/generateContentStream, which take the full request object, so the plugin reads every row into an array first.
• Vercel (packages/providers/src/openai-vercel): the AI SDK's streamText/generateText take a message array.

Effect

Peak memory during a send on these two providers grows with context size. Retained memory between turns is unaffected, because the array is released when the call ends.

Options

• Gemini: the plugin is now single-auth, so it could POST the REST endpoint directly with a streamed body, the same way the Responses route does.
• Vercel: route sends through the OpenAI-compatible chat path and keep the AI SDK only where it adds something, or drop the provider if it overlaps fully with the OpenAI provider.

Until then, the array must be released at call end and must not be duplicated.