fix(windows): bigint probe identity in runtime preflight; release CC-Switch F3 WAL handles - #7986

Summary

Fixes two Windows-only failures from manual Cross-platform CI run 41021347487 (lane=all, dev 3ea89653f).

• windows 2/11, runtime preflight (product fix). NTFS file IDs can exceed 2^53, the largest integer a JS number holds exactly; the failing run reported ino 10696049116378316. Above 2^53, adjacent IDs compare equal as numbers. src/lib/bun-runtime-preflight.ts identified its own probe directory by number dev/ino, both in-process and in the child script. A replacement directory could therefore pass the "this is my probe" check and be removed. The check now uses lstatSync(path, { bigint: true }). The existing replacement-root test used the same number comparison and failed on CI, so it now compares bigints too. A new deterministic case injects an owned probe and a replacement whose IDs differ only above 2^53. Cleanup must leave the replacement in place. This case fails on the old source and passes on the new one.
• windows 10/11, CC-Switch F3 (test fix). The F3 WAL test wrote rows with writer.prepare(...).run(...) and never finalized those statements. close() therefore did not release live-wal.db, -wal and -shm, and Windows refused to remove the temp tree (EBUSY, 15 s of retries). An isolated probe that did not call the product reproduced the EBUSY. After the statements were finalized, deletion succeeded. The importer closes its own descriptor on the refusal path, so the product needs no change. The test now uses writer.run(sql, ...params), which finalizes each statement. Its assertions are unchanged.

windows 4/11 of the same run (30 min timeout) was a slow runner. On that VM, setup steps took about 4× longer before any test ran (Setup Bun 26 s vs 5–8 s on the other ten shards, Install dependencies 64 s vs 12–18 s), and test durations rose broadly (median 1.36×, p75 2.6×). Neither repository change caused it, so this PR has no code change for it. The batched lane=all dispatch after merge will confirm.

Verification

• bun test tests/lib/bun-runtime-preflight.test.ts before the src change: 15 pass / 1 fail (the new above-2^53 case). After: 59 pass / 0 fail across tests/lib/bun-runtime-preflight.test.ts, tests/service/service-runtime-preflight.test.ts, tests/service/service-desktop-runtime-preflight.test.ts and tests/codex-integration/codex-shim-runtime-preflight.test.ts (Windows 11, Bun 1.4.4).
• bun test tests/providers/cc-switch-import.test.ts on Windows, before: 8 pass / 1 fail (F3 EBUSY at cleanup, 14.9 s). After: 9 pass / 0 fail, F3 49 ms.
• bun run typecheck: 2,243 files, no errors.
• bun run test:changed locally: first run hit its lane deadline (exit 124) on a machine shared with other test lanes; the rerun was stopped deliberately to free the shared host. Import-graph coverage comes from the hosted test changed job on the exact head (pass).
• Fresh independent review at head 9b07b5c: PASS, including a scratch revert of the src that turned the new case red (15 pass / 1 fail) and a five-file focused run (68 pass / 0 fail).
• NOT RUN: full bun run test locally, because other lanes are running Bun suites on the same machine. Hosted Windows coverage comes from the batched lane=all dispatch on dev after merge, since PR CI runs Ubuntu only.

Checklist

• Scope stays focused and avoids unrelated cleanup.
• Docs or release notes were updated when needed. (No user-facing behavior or doc contract changes.)
• Security-sensitive changes were reviewed for secrets, auth, and unsafe defaults. (The preflight change touches no credentials. It narrows what the probe cleanup may delete.)