The brain readiness probe runs a canned cogitate prompt with a deliberate 2-turn ceiling, so the vendored OpenHands SDK logs `Agent reached maximum iterations limit (N).` at ERROR while our layer treats the run as a success and brain reports ready. Add a diagnostic-only context manager that installs a `logging.Filter` on `openhands.sdk.conversation.impl.local_conversation`, mutates the matching record to INFO, and returns True. The filter is removed in a `finally` on every exit path. Non-diagnostic runs are untouched, and no logger level is changed. A companion solstone INFO explains the condition. It is verbose-only at the default CLI log level while the demoted SDK line still renders, since the demoted record's level check happened before the mutation. Downstream consumers remain unchanged: health classifier, its JS mirror, and support diagnostics still behave as before, while the demoted line no longer carries an ERROR token. Two golden fixture rows pin both forms. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>