Currently we have documented for mongoc_client_set_read_concern:

It is a programming error to call this function on a client from a mongoc_client_pool_t. For pooled clients, set the read concern with the MongoDB URI instead.

I believe the reason it is considered a programming error is that it persists on the mongoc_client_t even after being pushed back onto the pool. For example:

   pool = mongoc_client_pool_new (uri);
   client = mongoc_client_pool_pop (pool);
   rc = mongoc_read_concern_new ();
   mongoc_read_concern_set_level (rc, MONGOC_READ_CONCERN_LEVEL_AVAILABLE);
   mongoc_client_set_read_concern (client, rc);
   mongoc_client_pool_push (pool, client);
   client = mongoc_client_pool_pop (pool);
   level = mongoc_read_concern_get_level (mongoc_client_get_read_concern (client));
   printf ("rc=%s\n", level); /* prints "rc=available" */
   mongoc_read_concern_destroy (rc);

This behavior seems surprising and error prone. I think the same applies to write concern and read preference. I think it would be simple enough to have mongoc_client_pool_push reset its state (clear the read/write concern and read preference) before being pushed back onto a pool.