fix: embedding-regression test flaky on Windows (EBUSY: resource busy or locked on graph.db)

[carlos-alm]

Problem

tests/search/embedding-regression.test.ts intermittently fails on Windows with:

Error: EBUSY: resource busy or locked, unlink 'C:\Users\RUNNER~1\AppData\Local\Temp\codegraph-embed-regression-XXXXX\.codegraph\graph.db'

The test already has a 3-attempt retry loop in afterAll (added previously) and calls flushDeferredClose() before attempting deletion, but the file lock is not always released in time even with the retry logic.

Evidence

• Seen in PR fix(native): compute complexity/cfg for prototype method definitions (#1340) #1347 CI run 27055056145 (Windows-latest)
• The base branch fix/native-prototype-extraction-1327 also has a CI run (27055128276) with a different failure, confirming this is pre-existing
• All 2878 test assertions pass; only the cleanup fails

Root Cause

SQLite WAL mode on Windows holds file locks longer than the 3×100ms retry budget allows. The better-sqlite3 db.close() call triggers a WAL checkpoint that can take several hundred milliseconds on Windows.

Fix Direction

Options:

1. Increase the retry count or delay (e.g., 5 attempts × 200ms)
2. Use process.on('exit', ...) to defer cleanup to process exit instead of afterAll
3. Use a retry-based rmSync helper that can wait longer on Windows (up to 2–3s)