Welcome to Princeland depends on that.
Added tests show that NtQueryDirectoryFile mask handling is different from FindFirstFile (even though it doesn't quite match Windows too, the details are different, e. g., on Windows mask "*." will select file "ea" with FindFirstFile but not with NtQueryDirectoryFile). So the correct way is probably do the wildcards matching in kernelbase directly without relying on NtQueryDirectoryFile.
Having files not apprearing in the output of FindFirstFile is not obvious on Windows as well as the mask matches generated short names and that might provide an unexpected output. However, the part with not returning the files with extension when mask requires so looks rather straightforward and depending on that is not probably supposed to semi-randomly break on Windows. While checking the tests before after the changes I was checking that besides not introducing new TODOs no test files are appearing in the missing part in todo tests (as starting to incorrectly missing some files in the output has an obvious regression potential).