Per-item fan-out is now held to the documented rates, which makes some maps slower than in 0.5.x (where the SDK's own retries absorbed the throttling until they failed).
Not instrumented yet: the GCP organization-policy and access-context-manager clients outside _iterate, the urllib fallback that lists Azure subscriptions, and the single region-discovery call.
A blocking boto3 or Azure call made on an event-loop thread goes ahead without a bulkhead slot when none is free, so max_concurrency is not a hard cap for that path.
The Azure policy charges a retry-budget token for a retryable response even on the last attempt, which azure-core does not retry.
Learned rates and breaker state live in memory. A new process, including each cloudg map invocation, starts at the configured rates.