I've used NATS before, and I might not have explored all the options, but what I used it for was a pub/sub model of messaging.

I cannot speak to performance of NATS as we used it in low-traffic / low-demand setting. In terms of convenience, when used from Go, iirc, it offered some kind of built-in serialization (unlike eg. ZMQ, where you'd have to add your own).

Debugging lost messages was painful, but I've never used a tool where it wasn't...

NATS is equivalent to MQTT, just a different protocol. It has a persitence layer called Jetstream which is equivalent to Kafka ie. you can rewind through the message stream. The nice thing about NATS really is that it supports websockets unlike Kafka, it supports wildcard subtopics like MQTT, and it bridges very easily to both Kafka and MQTT (bidirectional).

Other fancy stuff is yes, one binary to run it, and it can also through gossiping, support a distributed KV / object store like etcd / s3 except you can register watchers on it (https://docs.nats.io/nats-concepts/jetstream/key-value-store...).