Why is only the single-level + wildcard supported?
If we add support for wildcards, users will expect to be able to use multi-level # wildcards too.
If you parse right-to-left for the mysensors message field you could support any wildcard.

The MQTT multi-level wildcard (#) is only allowed at the end of the topic. Because the prefix is always followed by other levels it cannot contain the # wildcard, although "my" code wouldn't complain about it.

The code only checks the number of / characters.
The syntactical check is left to the MQTT broker. That was already the case and I think that is right. I have browsed the MySensors code and checked all code where the PREFIX is used and none of needed a change.

Remove dead code

What's the increase in ram/flash for this change?

The savepointer (4 bytes?) is stored on stack I guess.
The increase in ram (stack use?) for the copy of the prefix string will be the length of the prefix plus 1.
The default value for the prefix is mygateway1-in, length is 13, so it will use 14 bytes on the stack. Your comment made me think about this memory usage. Actually I think malloc/free would be safer. Or maybe choose in runtime between stack and malloc depending on the prefix length? The length of the string is known at compile time, but I don't think that will help. I am not an expert in this subject.

Flash use? Is that the code length? The answer is: I don't know.

I compiled and ran it only on a raspberry pi 2, not on an Arduino.

In core/MyGatewayTransportMQTTClient.cpp, method reconnectMQTT, there is also a local variable inTopic with a length of the prefix plus 10. So I think it is acceptable to have it as a stack variable.