Kyokalith

Changed

• Dirty position write-back moved off the ticking thread and batched into one transaction. A spark profile of a live 12-player server showed DirtyPositionStore.persist accounting for roughly 1% of server-thread wall time: every flush cycle opened a fresh JDBC connection per pending chunk and committed each one separately, so a cycle with a few hundred dirty chunks paid a few hundred file opens and a few hundred fsyncs — on the thread that ticks the world. The cycle now runs on the async scheduler (Schedulers.asyncTimer: runTaskTimerAsynchronously on Spigot/Paper, AsyncScheduler.runAtFixedRate on Folia) and writes the whole batch through a single connection and a single transaction.
  • Nothing about persistence guarantees was relaxed. A dirty flag that fails to land re-opens the cover-and-dig exploit, so the same data is still written on the same cadence — only the thread and the batching changed. A failed batch (e.g. SQLITE_BUSY) falls back to the original per-chunk path, preserving the "one bad chunk must not skip the rest of the cycle" guarantee that has held since 1.x; anything still failing is requeued for the next cycle.
  • Folia/threading safety. The flush task touches no Bukkit API. DirtyPositionStore's shared state was already built on concurrent collections; the write paths (flush, flushAll, clearEpoch) are now additionally serialized by an internal lock. clearEpoch matters specifically here: it and the flush used to be implicitly mutually exclusive because both ran on the main thread, and without the lock an in-flight batch could write back an epoch that NatureRevive had just cleared. The final flushAll in onDisable waits for an in-flight cycle, keeping PlugMan hot-swaps safe.