That's true only in theory, but not in practice. In practice every inference provider handles errors (guardrails, rate limits) somewhat differently and with dif

That's true only in theory, but not in practice. In practice every inference provider handles errors (guardrails, rate limits) somewhat differently and with different quirks, some of which only surface in production usage, and Google is one of the worst offenders in that regard.