Mongos socket exceptions against shards that are down - breaks connectivity to whole shard?

I am running mongos with three config servers against two shards.

When I bring a shard down (just by stopping mongod on all nodes of one
shard), I was expecting that that the mongos process will catch this
and then inserts to the missing shard will fail and queries to the
missing shard will fail. These requests may fail gracefully if I have
the partial atribute set in my cursors.

The MongoDB docs say "operations that can be resolved at other shards
will continue to work properly," but whenever I have stopped one shard
completely and then tried to connect to a mongos process by PHP, perl,
or python I get an exception:  "query error: socket exception". I
notice that PHP connections behind Apache (and because their
connections were fairly persistent?) did seem to work as expected for
the life of their connection, I believe.

The mongo command line is a bit more informative:
error: {
	"$err" : "could not initialize cursor across all shards because :
socket exception @ ec2-xxx-.compute-1.amazonaws.com:XXXXX",
	"code" : 14827
}

For me, this means that all the shards are unreachable from my
application if any one shard is completely down. Is there some setting
I am missing that lets mongos or these drivers be more robust?