Add end-to-end tests for compaction-only since optimization

Outbox coverage for since optimization is missing a real end-to-end positive case. Current tests prove the negative cases well enough, but the positive compaction-only behavior is only guarded indirectly. That leaves the compaction optimization path under-tested relative to the amount of routing logic involved. Expected coverage: * a subscription is fully EOSE on compaction only * the outbox applies since optimization at the correct point in the lifecycle * dedicated-only routes do not optimize * mixed dedicated and compaction routes do not optimize * the positive compaction case is proven through the public outbox behavior rather than only unit-level classification