- Parse JSON error response from server to get actual used/limit values
- Previously showed 0/0 due to race condition (config fetch vs error)
- AI SDK puts full response body in error.message for non-OK responses
- Updated all quota toasts (request, token, TPM) to use server values