world-postgres: create run row, slot marker, and run_created event atomically

[tbharthur]

Package: @workflow/world-postgres@5.0.0-beta.34 (@workflow/world@5.0.0-beta.27), Node 22, PostgreSQL 17 (Cloud SQL).

@workflow/world-postgres@5.0.0-beta.34 can create a current slot-numbered run with a legacy ULID event because run creation is split across three autocommits.

allocateEventId() treats absence of a workflow_event_slots row as the signal that a run predates slot numbering and returns wevt_<ulid>. Both run creation paths currently make the run visible before inserting that marker:

INSERT workflow_runs                  -- commits
INSERT workflow_event_slots           -- commits later
INSERT workflow_events (run_created)  -- commits later

The affected paths are the resilient run_started creation path and the ordinary run_created path in packages/world-postgres/src/storage.ts.

A deterministic reproduction is:

1. delay INSERT on workflow.workflow_event_slots with a test trigger;
2. call world.events.create(runId, { eventType: "run_created", ... });
3. as soon as the workflow_runs row is visible, concurrently call world.events.create(runId, { eventType: "run_started" });
4. inspect workflow.workflow_events.

On beta.34 the persisted log contains slot 1 run_created plus a wevt_<ulid> run_started. The runtime later rejects that log because the event ID is not slot-numbered.

Please wrap the run row, slot marker, and first run_created event in one drizzle.transaction() at both creation sites. Preserve onConflictDoNothing() and the resilient path's re-read when a concurrent ordinary run_created wins. The ordinary path must also insert its run_created event inside the transaction and skip the shared tail insertion, otherwise the marker can commit before slot 1 and a concurrent writer can take the first slot.

The invariant should be: for a slot-era run, no transaction can observe the run row without also observing its slot marker and slot-1 run_created event.

We are carrying this as a local pnpm patch (both creation sites wrapped in one drizzle.transaction()) with a real-Postgres concurrency regression; happy to open a PR if that shape is acceptable.