(and the full output of the first block still differs, so dct must be different too)
The `sub(5000,14000)` was noticed in disassembly but still didn't give
mag_shift=2 on its own; so I forced mag_shift to 2. On my standard
input file, mag_shift is always 1 or 2.
