FAB — the compression

0x00  3     "FAB"
0x03  BYTE  shift (12 in every shipped stream; 10-13 decode)
0x04  WORD  the first word of the flag bit stream
0x06  ...   tokens, with further flag words interleaved

Tokens #

Flag bits are taken least significant first from 16-bit little-endian words.

1                 literal: one byte follows
0 0 a b           short copy: length 2 + (a<<1 | b), distance 256 - next byte
0 1               long copy: two bytes L, H follow
                    distance = 0x10000 - ((H >> (16-shift) | 0xff << (shift-8)) << 8 | L)
                    length   = H & ((1 << (16-shift)) - 1)
                    length 0 -> one more byte n: 0 ends the stream,
                                1 is a no-op, otherwise length n + 1
                    length k -> k + 2

With shift 12 that is a 4,096-byte window: literals cost 9 bits, copies of 2–5 bytes reachable 256 back cost 12, copies of 3–17 bytes cost 18, and copies of up to 256 bytes cost 26.

When a flag word is fetched #

A new flag word is fetched the moment the previous one's sixteenth bit is asked for. So it sits in the byte stream before the data bytes of the token whose flag bit emptied the old word, and an encoder has to reproduce exactly that placement.

Evidence #

The decoder is ff_decode's type-1 branch (file 0x772fa), the same codec the MADS engine's own games use. Every type-1 entry in the 246 shipped MADSPACK files — 793 streams — decodes to exactly its declared size and ends exactly at its declared packed size, with no byte left over.

The shipped packer #

MicroProse's packer seems to work in 40,960-byte blocks: the no-op token turns up 44 times, every time within 223 bytes after the output passes a multiple of 0xa000, and nowhere else.

A packer that does not imitate it still produces valid streams that decode to the same bytes; dos-tools' packer comes out byte-identical for 479 of the 793. That is why a rebuild never re-packs an entry whose contents did not change.