The same LiteLLM release starts authorizing every Responses API id, not only the ones it minted.

PR #39548 runs authorization on every Responses API id instead of trusting ids the proxy issued, closing a path where a caller could reference someone else's response object. The same release partitions user-key objects into their own in-memory cache (#40713), keeps an open Redis circuit breaker open and quiet on sync read and spend-counter paths (#40624), stops logging the request payload as guardrail_response on pre_call hooks (#39699), and restores credential coverage that a 4.5 entropy limit had silently dropped from secret hiding (#40190). That last one is the kind of bug that hides in plain sight for months.