Decide the long-term misbehavior policy for served peers

[gustavovalverde]

Follow-up from #42.

Context

#42 removed the seeder's own misbehavior > 0 eligibility gate and now delegates peer banning entirely to zebra-network, which bans (and removes from the address book) at MAX_PEER_MISBEHAVIOR_SCORE = 100. So peers with a misbehavior score of 1-99 are now servable, matching what zebra itself is still willing to keep and connect to.

Note: zebra's own gossip path (AddressBook::sanitized) drops any peer with misbehavior_score != 0 before sharing it. The seeder is now more permissive than that gossip policy, on the reasoning that served peers are already version- and liveness-verified.

Question

Is full delegation to the ban threshold the right long-term policy, or should the seeder reintroduce a softer misbehavior signal (for example, exclude peers above some fraction of the ban threshold)?

Acceptance

• Confirm from production metrics that serving sub-ban peers does not measurably degrade served-peer quality (client churn / reconnects).
• Decide: keep full delegation, or add a tunable misbehavior ceiling. Document the decision in ADR 004.

Refs: zebra-network MAX_PEER_MISBEHAVIOR_SCORE, AddressBook::sanitized.