Optimize watcher shifts and CraftEngine light batches

[Arbousier1]

What changed

• update event-driven tile watchers by shifting only entering/leaving chunk stripes
• cache watcher center, world and configured range so unchanged audits allocate nothing while reload range changes still rebuild safely
• remove stale offline watcher state without allocating online-player/key snapshots
• drain CraftEngine light mutations through reusable double-buffered sets instead of Set.copyOf per flush
• add lifecycle/order coverage for diagonal watcher shifts and exact candidate count

Why

The existing follow-up paths still rebuilt complete watcher windows on every adjacent chunk crossing and allocated an immutable light snapshot for every CraftEngine batch. These are small, independent allocation optimizations that preserve existing behavior and configuration semantics.

Validation

• gradlew check --no-daemon --console=plain
• Paper 26.1.2 compile and tests
• Paper 26.2 compile
• custom-content isolation and Paper-only architecture checks
• deterministic 3x3 diagonal shift equivalence: 5 entering candidates versus 9 full-window candidates

Remaining performance gate

A real CraftEngine installation is required for the light-path runtime A/B; this PR does not claim that unavailable measurement. The planned CraftEngine-native renderer and dropped-item index remain separate large factors.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.