test(storage): close agent run stores before cleanup

[liugddx]

Summary

• make the two root-admission test fixtures own every SQLite AgentRun store they open
• close original and reopened stores in reverse order before removing each temporary root
• keep the change limited to test lifecycle cleanup with no filesystem retries or production behavior changes

Part of #2142 Phase 1.

Why

The latest Windows baseline reports six cleanup failures across these two files after their assertions pass. Each test opens an AgentRun store, and two also open a second store, but the fixture removes the root without releasing the shared runtime.sqlite leases. Windows correctly rejects that unlink with EBUSY; POSIX hides the lifecycle omission.

Validation

• npm --workspace @maka/storage run build
• node --test packages/storage/dist/__tests__/agent-graph-supervisor-root-admission.test.js packages/storage/dist/__tests__/claimed-agent-graph-root-admission.test.js (6 pass)
• npx biome check packages/storage/src/__tests__/agent-graph-supervisor-root-admission.test.ts packages/storage/src/__tests__/claimed-agent-graph-root-admission.test.ts
• git diff --check

[liugddx]

The first required workspace lane failed in two unrelated timing-sensitive tests: maka-cu-service did not observe its cancel line within the assertion window, and opencode-free-anonymous observed provider_unavailable before its 42ms shared timeout. The six changed storage tests passed. The Windows artifact independently confirms the intended effect: storage moved from 648 pass / 38 fail / 38 skip to 654 pass / 32 fail / 38 skip, exactly eliminating the six targeted AgentRun cleanup failures, with no remaining maka-graph-supervisor-admission or maka-claimed-graph-admission cleanup errors. I cannot use Actions rerun on the upstream repo from the fork account, so 03d747b5 is an empty commit that retriggers the flaky checks without changing code.

[Astro-Han]

Thanks for another clean fixture-lifecycle fix in the #2339 family — all 7 createSqliteAgentRunStore calls in both files route through openStore(), closing is idempotent behind the refcount/released chain, and 6/6 tests pass. The withTempRoot-scoped tracking is arguably cleaner than #2339's map wrapper.

Three non-blocking notes:

• The 'ci: retry flaky workspace checks' commit (3dcefe9) is an empty commit — zero diff, just a CI retry. Squash will absorb it; worth avoiding message-only pushes in the future.
• store.close?.() leans on close? being optional in the DurableAgentRunStore interface (agent-run-store.ts:142), which is type-consistent but weaker than #2339's mandatory close() — the very invariant this test exists to protect (must close before rm) could silently degrade on POSIX. Consider assert.ok(store.close) or making close mandatory.
• Sibling files still close on the happy path only (regenerate-root-admission.test.ts:27, sqlite-long-term-memory-store.test.ts, agent-graph-supervisor-wakes.test.ts) — a follow-up could apply the same finally-close pattern.

Merging.