fix(cloud): wait for the download writer to close before cleanup

[kapelame]

What

An interrupted cloud download can report ENOTEMPTY instead of its network error and leave a staging directory behind. On Node 25.9.0, an async-generator source can reject pipeline() while the destination WriteStream is still opening. The finally block then removes the staging directory concurrently with that open creating the staged file.

Why

The staged output must be closed before its directory can be removed. Preserving the old destination is not sufficient if cleanup hides the actual failure or leaves partial artifacts.

How

Register a close promise before starting the pipeline and await it before removing the staging directory. This preserves the atomic replacement behavior introduced by #3840, including the old destination and caller-provided cancellation reason. No dependency or public API changes.

Test plan

Validation on this isolated download-only branch:

• bunx vitest run src/cloud/download.test.ts from packages/cli: 12 tests passed. The regression uses a manually delayed file-open callback and checks completion ordering, the original error, preserved output, and absence of staging residue.
• bunx oxlint packages/cli/src/cloud/download.ts packages/cli/src/cloud/download.test.ts: 0 warnings and 0 errors.
• bunx oxfmt --check packages/cli/src/cloud/download.ts packages/cli/src/cloud/download.test.ts: both files passed.
• git diff --check.
• The existing commit hooks passed: tracked-artifact and large-file checks, staged-file lint/format, fallow audit --base origin/main --fail-on-issues (no issues in the 2 changed files), type checks for core, Studio, and scripts, and commitlint. The hook type checks used physical copies of build artifacts from unchanged base packages; no CLI or producer artifacts were copied.

Earlier reproduction on macOS arm64 / Node 25.9.0 established that the delayed-open regression fails on the original implementation and passes with this change. A separate 100-call interrupted-stream run produced 19 ENOTEMPTY errors before the fix, with cleanup starting before close in all 100 calls. After the fix, all 100 calls preserved the network error and none cleaned up before close. The previous output survived all calls in both runs. These are the earlier reproduction results, not a new stress run on this isolated branch.

Earlier validation with this fix combined with the separate init-audio and GIF-loop patches passed the full CLI suite (3,213 tests; 3 skipped), CLI build/typecheck, and repository lint/format checks. Those broader checks were not rerun on this isolated download-only branch.

Node 22 and other operating systems were not exercised in the reproduction; it does not establish an identical failure rate on those runtimes. The earlier full workspace test run had separate baseline/environment failures: an existing producer AAC loudness assertion and Player/Studio tests conflicting with Node 25's native Web Storage. The AAC assertion also failed with the base commit's original implementation; disabling native Web Storage passed all 370 Player tests and the 120 tests in the affected Studio files.

Additional verification of the final combined state: after confirming that the original audit checkout contained exactly the three reviewed patches on f6041d7597c8c53d381d4c27feb86b305b3317a4, HYPERFRAMES_NO_TELEMETRY=1 bun run build and HYPERFRAMES_NO_TELEMETRY=1 bun run --filter '*' typecheck were rerun and both passed. Before/after source and diff hashes are identical. These new whole-workspace results cover the final three-fix combination; this standalone branch's own focused tests and hook checks remain documented separately above.

Current-upstream integration check: the three reviewed patches were also applied to a separate checkout of a0a6244be7a40956c5343f5846609e584ee5c6b5. With independent dependencies and a freshly rebuilt current Core runtime, 388 focused tests passed: audio route/transport (97), capture classification (19), render orchestration (222), GIF preparation/compiler (10), and init/download (40). A real local Chrome/FFmpeg smoke check reused the init-generated audio element in a minimal two-second scene with a finite-loop GIF and data-loop="false"; output was 160×90 H.264/AAC, 10 frames, 2.000 seconds, with one GIF playback and a 1.6-second tail hold. The screenshot retry branches were checked by unit tests, not an injected browser failure. Source/diff hashes remained unchanged. This verifies the combined patches against that upstream SHA; the independent commit and its original base remain unchanged.