Describe the bug

When using Google Gemini models, if one specific model (e.g., gemini-3.1-pro-preview-customtools) hits a rate limit (429), the OpenClaw gateway appears to block/backoff the entire google provider. This prevents fallback or manual switching to other available models (e.g., gemini-3.0-pro-preview) under the same provider, which likely have independent quotas or are not yet rate-limited.

To Reproduce

1. Configure google provider with gemini-3.1-pro-preview-customtools.
2. Trigger rate limits on this model (e.g., high usage loop).
3. Attempt to use gemini-3.0-pro-preview immediately after.
4. Observe that the request fails or is blocked by gateway backoff logic.

Expected behavior

The rate limit backoff should be scoped to the specific model (or deployment) that returned the 429, allowing other models under the same provider to continue functioning if their quotas are separate.

Additional context

User observed this behavior when gemini-3.1-pro-preview-customtools was rate-limited, but manual switching to gemini-3-pro-preview also failed until the provider cool-down expired.