Bug: custom cloud provider still falls back to exhausted OpenHuman budget #2140

Summary

When a user adds a custom cloud model provider/base URL and key, chat/model routing can still surface the OpenHuman backend resource exhausted / budget-exhausted error instead of using the configured custom provider.

Reproduction

1. Exhaust or otherwise hit the OpenHuman/TinyHumans hosted model budget.
2. Add a custom cloud model provider in Settings -> AI, including a custom endpoint/base URL and API key.
3. Select that custom provider/model for chat.
4. Send a prompt.

Expected

The selected custom provider is used for inference. The exhausted hosted OpenHuman budget should not block custom-provider calls.

Actual

The app still reports a resource exhausted / budget-exhausted response, suggesting the request path still uses the hosted OpenHuman backend/provider for at least some selected custom-provider cases.

Notes

Related but not identical to #2075, which covers Composio URLs being built from an LLM base URL. This issue is specifically about model inference routing after adding a custom provider.

No API keys or provider secrets included.