For DELETE (rollback) mode, and given the way fsync works, FULL should not lose committed transactions (plural) but it might lose at most one committed transaction.

Because DELETE depends on deleting a file (not modifying file contents), it much depends on the specific file system's (not SQLite's) journaling behavior.

I don't see how the journal delete depends any more on a specific file system's behaviour than the main data write: If a specific file system can decide to automatically fsync unlink(), it can equally decide to automatic fsync write().

In either case it is clear that (on Linux), if you want guarantees, you need to fsync (the file and the dir respectively).