[Subscription] Index consensus ACK progress by writer - #18835

Description

ConsensusSubscriptionCommitState.stageCommittedAndAdvance previously found the next committed and outstanding entries by scanning the full backlog maps. With N outstanding entries from one writer and reverse-order ACKs, every ACK repeatedly scanned O(N) entries, producing a real O(N^2) ACK path.

I reproduced the behavior by calling the production ConsensusSubscriptionCommitState with a single writer:

Each value is the best of three warmed runs. The baseline grows super-linearly as the backlog doubles; the indexed path does not. The new per-writer TreeSet indexes preserve the existing slot maps for exact ACK removal and make successor lookup logarithmic.

Remote-change check:

• Limit subscription in-flight event accumulation #18055 is still open and adds retained/in-flight event backpressure; it does not change ACK successor scans.
• Merged [Subscription] Reduce consensus WAL replay and ACK contention #18402/[Subscription] Reduce consensus WAL replay and ACK contention (#18402) #18527 reduce WAL replay and ACK lock contention, and merged [Subscription] Reduce consensus commit warning log noise #18450 reduces warning-log amplification; none indexes the commit-state scans.

Testing

• mvn -Ddevelocity.off=true spotless:apply -pl iotdb-core/datanode
• mvn -Ddevelocity.off=true -pl iotdb-core/datanode checkstyle:check -DskipTests
• ConsensusSubscriptionCommitStateTest: 19/19 passed, including a 2,048-entry reverse-ACK regression case.
• git diff --check

The normal full DataNode compile is currently blocked by existing workspace SNAPSHOT/source inconsistencies (SubscriptionQueueAdmission and related classes) and a Windows compiler native-memory failure; the changed manager and targeted test were compiled directly against the existing DataNode build artifacts.

Changed files

• ConsensusSubscriptionCommitManager
• ConsensusSubscriptionCommitStateTest