fix(storage): close usage stores after lease revocation

[liugddx]

Summary

• release telemetry, model-call, and pricing SQLite resources without re-entering root authority during close
• keep read/write operations lease-bound while making close idempotent after revocation
• close usage stores in root-identity and marker failure tests, and cover reader cleanup after revocation

Part of #2142.

Validation

• npm --workspace @maka/core run build
• npm --workspace @maka/storage run build
• node --test packages/storage/dist/__tests__/usage-stores.test.js (8/8 pass)
• npx biome check packages/storage/src/usage-stores.ts packages/storage/src/__tests__/usage-stores.test.ts
• full storage suite: 734 pass, 12 skip, 1 pre-existing environment-sensitive failure in package-import.test because the current Node process emits the experimental node:sqlite warning; the failure reproduces in isolation and is unrelated to these files

[liugddx]

@Astro-Han

[Astro-Han]

Thanks — the fix is right and matches every sibling store (usage-stores.ts was the only store wrapping close in the lease; the probe-verified idempotency and the new tests all hold, 8/8 green).

One P2 on the concurrency form of the very scenario this PR fixes: an admitted in-flight write at revocation time still makes close reject — the in-flight op fails with invalid_lease, gets collected into failures, and close's throwDeduplicatedFailures (usage-stores.ts:381) upgrades that benign revocation into a shutdown failure (invalid_lease is itself classified as lifecycle/needsDrain:false at :165-167). Reproduced with a probe: close → StorageRootAuthorityError:invalid_lease, which in the host-kernel path becomes a 'did not cleanly close every resource' exit. Pre-existing, but it is the concurrent form of what this PR claims to fix, and it's a small change (fold lifecycle codes into the expected-failure predicate, or filter them before the aggregation). Happy to defer if you'd rather land this first. P3: writerByLease isn't cleaned on close (siblings delete it), and the new test could assert idempotent double-close.

Approving.

[liugddx]

Added aa1fe4ac4 after reproducing the remaining Windows usage-store failure locally.

The failure was not another leaked SQLite lease: Windows rejects renaming a directory that contains the open runtime.sqlite before the assertion begins, and the fixture had placed that setup outside its try/finally. The follow-up now:

• keeps store/owner shutdown inside the failure-safe lifecycle;
• retains the rename/replacement contract on platforms that permit it;
• narrowly skips it on Windows with the platform limitation stated explicitly.

Windows target result: 7 pass / 0 fail / 1 skip for usage-stores.test.js. The new checks are now running.

[liugddx]

The failed required run was unrelated to this PR: the only storage failure was a process crash leaves owned staging that can be reclaimed without touching decoys on Linux (false !== true). Main already fixed that timing assumption in #2385.

Rebased this branch onto current main at 1a7286129, which includes #2385, and retriggered CI. The two PR commits remain isolated on top as 7ca7d4af9 and bcb62c779.