The current implentation of write commands is a bit fragile for what concern error handling. In fact it is not safe to throw StaleConfig errors at whichever point of a write command execution on the shard.

The reason is that, as described in SERVER-107482 the mongos up to version 8.2 does not handle correctly top level StaleConfig error properly. It only handles them if they are propagated from the shard to the router as nested write errors. 

So it will be safe for the shard to throw StaleConfig error only once SERVER-107482 is fixed and available both in lastContinuos/lastLTS version. At that point, we know that even in multiversion scenario the routers will always be ready to handle top level StaleConfig.