feat(performance): priority scheduler + read coalescing on the SQLite worker (0228 Phase 1)

[crs48]

Implements Phase 1 of exploration 0228.

Context

The natural follow-up to 0227 was "can we have a pool of reader workers + a dispatcher?" The exploration found that the naive version is impossible: the web adapter's opfs-sahpool VFS holds exclusive file handles (a second worker on the same DB falls back to in-memory), and the replica-based version that would give true parallelism is gated on a benchmark proving CPU saturation and on enabling production cross-origin isolation (which has embedding risk). So the doc's recommended "do now" is a scheduler on the single worker — most of the felt latency win at a fraction of the cost.

What landed

A priority scheduler + read coalescing inside SQLiteWorkerHandler (the one chokepoint both the main thread and the data worker share, via connectPort):

• Interactive reads drain ahead of queued writes — a write/sync-apply burst can no longer starve a read. Generalizes the 0227 head-of-line fix to any queued op.
• Identical concurrent reads coalesce into a single execution (request de-dup).
• getSchedulerSnapshot() exposes queue depth for diagnostics.
• No parallelism added (opfs-sahpool is single-connection) — this is ordering only, and serial execution matches SQLite's single-connection model (safer than today's interleave-at-awaits).

New worker-scheduler.ts + 7 deterministic tests proving the ordering/coalescing guarantees (these double as the exploration's "benchmark harness": an interactive read is provably served ahead of queued writes). @xnetjs/sqlite typecheck + full unit suite green.

Deliberately deferred (per the doc's own gating)

• Transaction chunking / cooperative yield — splitting one atomic applyNodeBatch into multiple transactions changes atomicity semantics; correctness-sensitive, wants its own change.
• Live p50/p95 boot probe + explicit bulk-tagging — refinements; the scheduler default (reads→interactive, writes→write) already delivers the benefit.
• Devtools panel — a getSchedulerSnapshot() hook exists; wiring the panel is a follow-up.
• All of Phase 2 (replica pool + dispatcher) — explicitly conditional on a saturation benchmark and on production COOP/COEP; not built speculatively.

The real worker path is exercised by the editor-ux/capture e2e checks.

🤖 Generated with Claude Code

[github-actions]

Preview removed for PR #279.