Three choices are worth knowing because they change what you operate. The schema is applied, not migrated: every
boot runs one idempotent DEFINE file and records a version record, so a process notices it is looking at a database
some other build prepared instead of discovering it one confusing query at a time. Backfills are explicit one-shot jobs
rather than migration steps. The local engine is memory-backed: its data resets with the pod, deliberately — boot
re-applies the schema and re-runs the canonical seed. Tests use an engine inside the test process: no container, no
port, no shared server, so two tests cannot collide and there is nothing to reclaim afterwards.
Per-worktree isolation comes from the cluster boundary, not from anything extra: dev worktree-env up already keys a
KIND cluster to the worktree path, so its Surreal pod is private by construction and worktree-env down removes it with
the cluster.