Deliver scheduled cluster messages at their deadline instead of the next poll

[sbking]

Fork-only review draft. The base branch pins Effect-TS/effect main at 212493b so the diff is exactly the two commits under review. Before an upstream PR this rebases over Effect-TS#6441 and the 0003 migration renumbers past its 0003_mssql_schema_fixes.

Summary

Messages saved with a future deliverAt are only discovered by the periodic storage poll, so a scheduled message can be delivered up to a full poll interval after its deadline.

• add a required MessageStorage.nextDeliverAt operation that returns the earliest future scheduled delivery time for a set of shards, with memory and SQL driver implementations
• after every storage read, Sharding re-arms a single replaceable timer for that deadline; when it fires, it runs the existing storage read path
• the wake affects promptness only: if nextDeliverAt fails, the error is logged at debug level and delivery falls back to the existing poll
• use explicit null checks in the memory driver so due-message selection and nextDeliverAt agree when deliverAt is zero

PostgreSQL and SQLite answer the query with MIN(deliver_at) over a partial index on (shard_id, deliver_at) filtered to unprocessed scheduled rows. MySQL and SQL Server have no partial indexes, so they seek per shard (correlated scalar subquery and CROSS APPLY ... TOP 1) over a plain (shard_id, processed, deliver_at, last_read) index, which stays flat as processed rows accumulate. Migration 0003_deliver_at_index creates the index idempotently on all four dialects.

Local container benchmarks used 1 million message rows, 300 assigned shards, and roughly 200 to 180,000 pending scheduled messages. With the selected per-dialect queries and indexes, the deadline lookup stayed below 8ms on PostgreSQL, around 4ms on MySQL, below 0.1ms on SQLite, and around 0.7ms on SQL Server. These are warm-cache local measurements; production calls still pay network latency. Synthetic bulk inserts measured roughly 0.3-2μs of additional index work per inserted row.

Testing

• new tests: a message scheduled before a runner starts is delivered at its deadline instead of the next poll; a failing nextDeliverAt degrades to poll-cadence delivery with the read loop still usable; deliverAt at epoch zero across the due boundary
• pnpm vitest run packages/effect/test/cluster/MessageStorage.test.ts packages/effect/test/cluster/Sharding.test.ts packages/effect/test/cluster/Entity.test.ts (42/42)
• the shared SQL storage suite's nextDeliverAt cases pass against PostgreSQL, MySQL, and SQLite (pnpm vitest run packages/platform-node/test/cluster/SqlMessageStorage.test.ts -t nextDeliverAt); the SQL Server leg runs after the rebase over Support SQL Server in cluster SqlMessageStorage Effect-TS/effect#6441 lands its container fixture
• pnpm check
• pnpm lint