When sending saga timeouts the current partition information is not automatically added to the timeout message. When such a saga timeout message is processed, the partition to be used for the CosmosDB cannot be extracted from the incoming saga timeout message.
Ensure that the timeout data state that is passed to RequestTimeout contains the partition key information. The Cosmos DB Persistence Usage with non-default container sample demonstrates this approach.
Support for automatically resolving partition information for saga timeouts is being considered. If this workaround affects the application, add a thumbs-up or share the use case on issue #720.