Gemini adapter: support wrapFetch / a custom fetch (it is ignored today)

[TheOrdinaryWow]

wrapFetch (#1680) works on the OpenAI, OpenAI-compatible, Anthropic, Mistral, Ollama and OpenRouter adapters, but @tanstack/ai-gemini ignores it. The cause is upstream: @google/genai has no fetch option (googleapis/js-genai#999, with an implementation in googleapis/js-genai#1215 that has had no review since September).

Why it matters to us

We run a server-side support agent behind a provider switch (openai, openai-compatible, anthropic, gemini). Our fetch wrapper does four things, and Gemini is the only provider that loses all of them:

1. A timeout on each HTTP request, separate from the abort for the whole run.
2. A "first request started" hook. We take a usage quota only once a billable call is actually sent.
3. Cleaning provider error bodies, so the API key and the prompt text never reach our diagnostics.
4. A fake fetch injected in tests, and the HTTP latency and status read for a provider health probe.

Asks

• Follow the upstream PR, and pass wrapFetch into the Gemini client once @google/genai supports a custom fetch.
• Until then, write a debug-level warning when wrapFetch is set on an adapter that ignores it. Today it does nothing without a sign, which is easy to miss when you have more than one provider.
• If you would accept it: an adapter option that takes a GoogleGenAI client you built yourself. Users could then pass their own patched client until upstream catches up.

Versions: @tanstack/ai 1.0.1, @tanstack/ai-gemini 1.0.1 (@google/genai ^2.10.0).