[rahul05ranjan]

Fixes #785.

What this changes

Three things, exactly as the issue proposes:

1. The remaining unclosed files close their pool. Every const <name> = createDatabase( variable now gets await <name>.$client.close():
   • 3 files already had a top-level afterAll cleanup — the close is its last statement, the idiom identity-store.integration.test.ts uses;
   • 43 files had no top-level afterAll (only describe-scoped ones) — they gain a module-level afterAll at the end of the file. Registered last, it runs last: Bun runs afterAll hooks first-registered-first, so the close follows every other hook's cleanup. A describe-scoped afterAll would have been wrong here — closing at a describe's end takes the pool from later suites in the same file.
   • 14 files already close via $client.end(...) — left untouched; the pairing accepts either.
2. A guard test keeps the pairing (database-pool-closing.test.ts): every createDatabase variable in server/tests/*.test.ts must have a $client.close() or $client.end( somewhere in the same file. server/tests is outside tsc, so this is the thing that notices file 118 reintroducing the leak.
3. The two stale sentences in scripts/test-server-sharded.ts are replaced with what the issue measured: construction opens nothing, a pool costs its max from its first query and is never reaped, exhaustion is a function of files run rather than imported, and an unclosed pool is a defect in the source rather than a harness ceiling. The sharded runner itself is unchanged and still worth having for wall-clock.

Where it runs

Test infrastructure only; no server code is touched.

• New state that outlives a request? None.
• Second replica? No server change.
• Anything serialised? No.
• Anything fanned out? No.
• New listener, port, schedule? No.

Boundary and audit

• No acting call changed: tests and scripts only.
• No new refusals.
• Nothing new trusted from the client.

Changelog

• No deployment behaves differently: the change is entirely inside server/tests and scripts/.

Proof

No Bun or PostgreSQL on this machine (arm64 Linux CI box), so the local verification is tooling around the text, and the suite itself runs in CI:

• The guard's exact rule, run locally as a Python mirror over all 127 pool-opening files in server/tests: after the edit, zero offenders. Before it, the rule reported the 46 files this PR fixes (the issue counted 55 against an earlier tree; the same rule also finds the 14 $client.end files already satisfied, and correctly ignores scoped pools that are closed in-scope).
• Every edited or new file parses: node --experimental-strip-types --check on all 47 touched .ts files under server/tests (the sharded runner script is not node-parseable even on main — its edit is comment-only).
• No double-close: files using $client.end( got nothing added on top of it; a scan for any variable carrying both a close() and an end( finds none.
• Placement audit: the diff contains exactly 46 inserted close statements — 3 inside an existing top-level afterAll (checked by hand: after the last delete, before the block's close), 43 as the body of a new trailing module-level afterAll. No close was placed inside a describe-scoped hook; the first attempt did that to one file and was reverted and redone with a top-level-only rule.
• What I could not run: bun test itself (needs PostgreSQL); CI is the authoritative run, and the failure modes here are loud — a pool closed too early fails the deletes around it, a syntax error fails the file's import.