## Why

Optional balancing can move recently reassigned partitions again before their new load reports have settled. Initial zero-valued rate reports can make a recovering destination appear idle, while a precomputed balancing plan can become stale after broker lifecycle or controller leadership changes.

## What

Apply a configurable partition reassignment cooldown, require representative post-handoff load reports for optional optimization, and discard stale balancing plans.

## How

Track accepted reassignment batches and observed handoffs in controller-local state. Add `autobalancer.controller.partition.reassignment.cooldown.ms`, defaulting to five minutes; zero disables the cooldown. Bound the whole-cluster healthy-source balancing hold to 30 seconds after broker recovery events.

Require distinct, complete post-handoff load reports spanning the rate meter's five-second tick before optional optimization treats a recovered replica as ready. Keep raw load freshness and mandatory recovery placement unchanged. Excluding a partition from optimization retains its broker's modeled load.

Initial metadata replay and duplicate lifecycle records do not fabricate or extend recovery. Detected slow-source evacuation remains eligible, and broker/controller transitions invalidate queued optional plans.

## Testing

Coverage includes configuration validation and dynamic disabling, accepted versus rejected reassignment batches, expiry, duplicate lifecycle events, controller handoffs, slow-source evacuation, stale plans, and incomplete/duplicate/early load reports.

Local verification passed: the AutoBalancer tagged suites, Checkstyle and Java formatting. The repository's `rat` task was invoked and skipped by upstream configuration.

```bash
./gradlew --build-cache core:S3UnitTest --tests 'kafka.autobalancer.*'
./gradlew --build-cache rat checkstyleMain checkstyleTest spotlessJavaCheck
```

## Reviewer notes

This change is independent of the router/WAL fixes and optional cross-rack placement. Read `PartitionReassignmentCooldown` and `BrokerRecoveryCooldown`, then their integration in `AnomalyDetectorImpl`, followed by the model warm-up rules and tests.

The cooldown bounds optional scheduling; it does not certify storage or application recovery. The issue's measurements do not isolate this PR's individual latency benefit.

Related to AutoMQ#3669.

AI assistance: OpenAI Codex (GPT-6) was used to prepare the contribution split, adapt the changes to main, and update the tests and descriptions.

## Attribution

The initial partition reassignment cooldown was implemented by [@deniszh](https://github.com/deniszh).