When performing a rolling downgrade on a version 3.6.7 cluster to version 3.4, once the config servers and shards have all been downgraded, the remaining version 3.6 mongos nodes can no longer connect to the cluster.  This can be seen in the mongos.log:

<timestamp> W NETWORK [UserCacheInvalidator] remote host has incompatible wire version: IncompatibleServerVersion: Server min and max wire version are incompatible (0,5) with client min wire version (6,6).

However, when connecting to a version 3.6.7 mongod and running the command db.isMaster(), the resulting document shows "minWireVersion" : 0, "maxWireVersion" : 6.

This makes it impossible to perform a rolling downgrade from 3.6.7 to 3.4.x without taking an outage, as all mongos nodes become unusable from the time that the config servers are downgraded until the time that at least one mongos is downgraded.