What

Let foreground engine reads run between background inbound-sync batches in dedicated-worker wa-sqlite clients.

#166 creates an event-loop gap between remote apply batches. This PR adds the small priority scheduler that decides what may use that gap.

Ordering rules

• Engine read APIs are foreground.
• Inbound sync appendMany calls are explicitly background.
• Local writes, generic runner queries, and unannotated RPCs remain normal priority.
• A foreground read may bypass queued background work only.
• A normal/local write is an ordering barrier: later reads cannot jump it.
• At most eight foreground jobs bypass an earlier background job, preventing sustained reads from starving sync.
• A running SQLite batch is never interrupted; prioritization happens only between RPC batches.

Direct and shared-worker runtimes keep their existing FIFO paths and pay no scheduler overhead. Public write options expose only background; foreground remains an internal read classification, so callers cannot make writes jump sync work.

Why dedicated-worker only

The earlier revision also changed the shared worker's global queue, but its benchmark used one client/port and therefore did not exercise cross-tab scheduling. That unproven scope and its wire-protocol field are removed. A future shared-worker change should begin with a two-port ordering test and benchmark.

Performance evidence

Focused run 29203415286 completed four counterbalanced normal/background trials per case. Values are background-vs-normal deltas between trial medians; negative latency is better.

Every empty-tree trial improved read p95 (memory: 11.3–14.6%; OPFS: 14.4–17.9%). Populated cases stayed near neutral. The largest median latency cost was 4.0% remote-batch p95. Four trials are directional evidence, not a claim of statistical certainty.

The harness's total-duration field included the final integrity read, so this decision deliberately does not use its derived throughput values. Per-operation latency samples and the final-count integrity check are unaffected.

Debloat

• 8 files, +251/-32, down from the original 19 files, +1,029/-70.
• Production changes are +113/-32; the rest is a changeset and focused tests.
• Shared-worker scheduling and RPC priority serialization are removed.
• Benchmark infrastructure is not part of this production PR.
• A compact composition regression proves an engine read is classified ahead of two background writes.

Verification

• Priority scheduler and engine priority composition: 7/7 deterministic tests.
• treecrdt-sync: 7/7.
• sync-sqlite's changed backend tests: 3/3.
• TypeScript builds for interface, protocol, auth, sync-sqlite, sync, benchmark, and wa-sqlite.
• Formatting and git diff --check pass.
• Full GitHub build, browser, playground E2E, native-node, and Postgres CI is green.

Dependency

Stacked on #166. Review only this incremental diff.