0009 — Anti-entropy re-issues intents; rejected intents are evicted

Date: 2026-08-08. Status: accepted. Amends the anti-entropy design in ADR 0003 / PROTOCOL.md.

The original design re-multicast cached intents verbatim (original origin, seq, signature). The simulation suite found a livelock: replay defense tracks the highest seq per origin, so once a consumer has seen any later frame from an intent's author — for example an intent that author sent to a different consumer — a verbatim copy of the earlier intent is stale-dropped forever. The offline consumer can never converge, and bystanders re-offer the intent endlessly; the mesh never quiesces (tests/test_simnet_scenarios.py pins the reproduction).

Accepted fix: a relaying agent re-issues the cached intent body as a fresh int frame under its own origin and seq. The version field inside the body preserves authorship and the total order, so convergence semantics are unchanged; the relayed frame is signed with the same site PSK, which is the same trust domain as the original (ADR 0005 has a single key). Suppression now observes peers' re-issues through the ordinary intent path instead of a special byte-identical duplicate case, which also removes that special case from the node.

Same root class, second instance: bystanders cached intents that their consumer had rejected (self-pair, kind-mismatch, malformed producer id), re-offering a doomed highest-version intent forever. Agents now observe rejection acks on the multicast group and evict/pin the exact rejected version per consumer.