What changed

• Yield to one macrotask after applying each non-empty, non-final inbound opsBatch.
• Keep batches sequential per filter, so completion and done markers cannot overtake earlier applies.
• Add a regression proving a timer-scheduled user task runs before the next queued inbound apply.

Why

Without a scheduling gap, several already-queued remote batches can monopolize the worker event loop. A foreground read or write that arrives after one batch finishes is then forced to wait behind the next batch even though per-filter ordering does not require that starvation.

This does not make OPFS writes themselves faster. It makes queued sync work cooperative with foreground work.

Fast paths left

• Empty batches and final/done batches do not add a yield.
• Single-batch syncs pay no extra scheduling delay.
• Verification, persistence, acknowledgements, and per-filter ordering are unchanged.
• The existing queue remains the only serialization mechanism; this adds no scheduler state.

Debloat

The original PR was stacked on #164 and included 137 lines of runtime benchmark changes. Those benchmark files and the benchmark ancestry are removed. The review is now 3 files, +83/−1; 74 added lines are the timing/order regression.

Verification

• @treecrdt/sync-protocol: 24/24 tests pass.
• TypeScript build and git diff --check pass.