I'm not sure is avoiding trust should be enforced because some cusers could prefer to configure the client to participate in round only with some very specific coordinators and avoid all others. The coordinator identity could be the onion address and if well a coordinator could have multiple addresses, one onion address identifies one and only one coordinator.

I started reading about fidelity bonds because I knew about them but never went deep in my understanding. I think I will have a better idea today, however the problem of calculating the cost of the bond is clear and there seems to be not easily solvable if it is.

I will search in the joinmarket repo for the "directory nodes" idea but yes, your concerns are valid. In the case of DoS attacks we already have the problem because we the same utxo you can already disrupt rounds in all the coordinators at the same time because they don't share their banned uxto list and in the case of discoverability we already have a public announcement mechanism that uses nostr (we could use something different too because trusting nostr relay operators is not wise)