Tristan's discovery is explained here at flatcap.github.io/linux-ntfs:

If a new record was simply allocated at the end of the $MFT then we encounter a problem. The $DATA Attribute describing the location of the new record is in the new record.

The new records are therefore allocated from inode 0x0F, onwards. The $MFT is always a minimum of 16 FILE Records long, therefore always exists. After inodes 0x0F to 0x17 are used up, higher, unreserved, inodes are used.