[hutiefang76]

Parallel producers writing different rows of a DuckDBWritableVector can still update the same packed 64-bit validity word. The non-atomic read/modify/write drops other producers' bits, so valid scalar-function results are returned as SQL NULL (and concurrent null writes can be lost).

Synchronize only the validity word update on the existing buffer, preserving parallel data writes and Java 8 API compatibility. Extend the existing parallel stream regression and add a barrier-based test covering concurrent valid/null writes.

Validation on JDK 17:

• The existing parallel test and the new regression each failed 3/3 against the original implementation.
• Both tests passed 20 repeated runs each after the fix.
• The complete test runner passed all 541 tests. After setting DuckDB threads=1 in the new test to isolate its Java worker barriers, both focused tests passed again; the production change was unchanged apart from comments.
• The two changed files pass clang-format.

AI assistance: DeepSeek assisted implementation and first review; the final diff and actual test output were checked before submission.

[staticlibs]

Thanks for the PR, this is a good catch! The change looks good to me, just a one request to change the lock handling.