[Architecture Review][P0] Configurable DeliveryTopology for cluster delivery

[qqeasonchen]

Problem

Two cluster delivery models coexist. EventMeshApplication.enableCluster() uses the sticky local-delivery model and intentionally does not enable partition ownership or cross-instance forwarding, yet the repository still retains PartitionOwnership, ClusterCoordinator, and HttpForwarder — a second partition-ownership / forwarding model. The partition-ownership code itself acknowledges missing generation fencing and incomplete remote offsets.

Risks

• Documentation, code, and actual deployment behavior can diverge.
• The sticky model can cause duplicate reads and amplification at the underlying MQ across instances.
• The partition-ownership model lacks the fencing needed for safe handoff.

Proposed direction

Abstract the cluster mode into a configurable DeliveryTopology:

LOCAL_STICKY_PULL     current default: session pinned to an instance, local pull + local delivery
PARTITION_OWNED_PULL  high-throughput: single partition owner, requires fencing and handoff
FORWARDED_DISPATCH    compatibility: centralized pull + cross-instance forwarding

Each mode must document:

• duplicate-delivery semantics;
• cost / throughput characteristics;
• failure-recovery strategy;
• a corresponding E2E test.

Acceptance criteria

• One explicit DeliveryTopology config selects the mode; inactive topology components are not half-wired.
• Sticky and partition-ownership modes are documented with duplicate/cost/recovery semantics.
• Partition handoff uses fencing (epoch/lease); a stale owner cannot keep delivering.
• E2E tests exist per topology (Meta outage, network split, owner crash).

Related: #5293 (unify cluster delivery topology and fence stale partition owners).

---

Part of the Architecture Review.

[qqeasonchen]

Superseded by #5293 ([Bug] Unify cluster delivery topology and fence stale partition owners), which now defines the full three-mode DeliveryTopology (STICKY_LOCAL_PULL / PARTITION_OWNED_PULL / FORWARDED_DISPATCH). Closing as a duplicate to avoid parallel tracking.