UPDATE: This ticket has been cloned so that the clone will cover Scenario 1 (as described by Kevin's comment from 03/09/2020) and this original ticket will cover Scenario 2 (as described in the aforementioned comment).
A mongoc_server_stream_t should not be reused after calling mongoc_cluster_run_command_monitored on with the stream.
In two situations, the stream will disconnect and be invalidated:
When that happens, the server_stream for that operation is no longer valid. But we do not always account that correctly.
_mongoc_write_opmsg (incorrectly) always ends unordered bulk operations on error because of
CDRIVER-3305. Once that is resolved, it will need to handle the case of an invalidated server stream.
_mongoc_write_opquery only sets must_stop to true for what it deems is a network error (of which, that check might not be quite right if error labels were applied). But it does not handle the case of a "not master" or "node is recovering" error causing a disconnect.