Solved:
I just had this exact issue and took TWO steps to address it; I stopped using C/C++ stdlib apis and ::DeleteFile(..), and switched to:
::MoveFileEx(src,dest,MOVEFILE_WRITE_THROUGH);. See: MOVEFILE_WRITE_THROUGH
h = ::CreateFile(DELETE | SYNCHRONIZE,OPEN_EXISTING,FILE_FLAG_DELETE_ON_CLOSE | FILE_FLAG_OPEN_REPARSE_POINT); ::CloseHandle(h);
The above are pseudo calls showing the relevant flags, specifically note that there is NO sharing on the CreateFile call used to achieve delete.
Together they gave me better precision on the rename and delete semantics. They are working in my code now and have improved the precision and control from other threads/processes (watching the file-system for changes) interjecting actions on the file due to latencies [or sharing] in the other rename and/or delete APIs. Without that control, a file set to delete when its last kernel-handle was closed might actually languish open until the system was rebooted, and you might not know.
Hopefully those feedback snippets might prove useful to others.
Addendum: I happen to use hardlinks for a portion of the work I do. It turns out that although you can create hardlinks on a file that is OPEN, you cannot delete ANY of them until all handles to ANY of the underlying data-stream(s) to that NTFS file are closed. That is weird since:
Which would lead you to think that only the last hardlink should be non-deletable while the kernel has one or more open-file handles referring to the hardlinked NTFS File's MFT-Entry/ATTRs. anyway, just another thing to know.