fix(db): give DbClient a close() so test teardown releases the handle

[lovepixel-git]

Refs #284. Addresses the DbClient/SQLite half of that report, which accounts for the bulk of the Windows failures. The path-as-URL bugs and the rest are separate changes, as the reporter proposed.

Root cause

DbClient exposed no way to release its backing resource, so a file-backed SQLite database stayed open for the life of the process. Test teardown then removes the temp directory holding that file. POSIX tolerates unlinking an open file; Windows returns EBUSY. src/__tests__/helpers/createTestDb.ts is imported by 26 test files, which is where most of the reported failures come from.

The helper's own comments already described the defect:

bun:sqlite doesn't expose a close() method on our DbClient interface; on macOS/Linux the file can still be deleted while the handle is open

and the Postgres branch carried a matching TODO: extend DbClient with a close() method to properly terminate the Postgres connection pool. Both are resolved here rather than restated.

CI is ubuntu-latest on every job, so nothing upstream exercises this.

Change

close(): Promise<void> on the interface, implemented by both adapters. SQLite closes the database, which also releases the WAL and SHM siblings created by PRAGMA journal_mode = WAL. Postgres closes the pool.

Teardown closes before unlinking at all five sites that open a file-backed database: the shared helper, plus pluginScheduler, postTypeBuiltInFields, importEndpointGuidance and createDbClient. Every other test database in the repo is :memory:, so that is the complete set.

Two decisions worth your call

Required, not optional. I made close() required rather than close?(). CONTRIBUTING says to update the source of truth and all callers in the same change rather than add a compat path, and the Postgres TODO explicitly asked for pool termination, so implementing both adapters seemed closer to intent than a SQLite-only optional method. Happy to switch it to optional if you would rather keep the interface minimal.

Transaction-scoped clients. wrapSql backs both the owning client and the tx handed to a .transaction() callback. Closing the pool from inside a transaction would tear the connection out from under the enclosing sql.begin(), so I gated it: only the client returned by createPostgresClient closes the pool, and close() on a tx client is a documented no-op. Nothing calls it there today, but the shared wrapper made it reachable, so I would rather it be deliberate than latent.

createFakeDb also gains a two-line no-op close. The as DbClient fakes typecheck through assertion widening, so without it they would satisfy the type while throwing at runtime.

Tests

New: src/__tests__/db/dbClientClose.test.ts, four cases. The EBUSY symptom is not portable, so these assert the invariant instead: after close() the handle is released and later queries reject, close() is idempotent, the directory is removable once closed, and createTestDb's cleanup() closes before removing. That reproduces the regression cross-platform with no Windows machine.

• bunx tsc -b exit 0
• bun run lint clean
• bun test 6534 pass, 1 fail

The one failure is pre-existing

Bundle size budgets > ContentPage-*.js stays under 87.9 kB, at 90043 B against a 90000 B budget. Not from this change: I stashed the whole diff, rebuilt clean main, and got the byte-identical 90043. It only appears once dist/ exists, so a run on a fresh clone will not show it. Flagging rather than folding it in, since it is a separate problem.

Overlap with #194

That PR contains a version of this change inside an 89-file feature branch, currently CONFLICTING and untouched since 2026-07-12. This is the focused version, mirroring how #192/#193/#202 landed separately. Glad to close this if you would rather take it there.