THE FILL-DEPENDENT DATA REACHES THE CPU READBACK -- AND STOPS THERE.

Engine output port 0x5600200, read 384 times per export (pc file 0x12552A),
values captured per fill:

  fill=ff : 384 reads -> FFFF x256, 0000 x65, 0001 x3, 0091 x3
  fill=00 : (cumulative) -> 0000 x213, FFFF x262, 0032 x29, 0091 x14

So the readback IS fill-dependent: a white tile yields ~256 words of 0xFFFF from
the engine, a black tile yields 0x0000.  The data therefore travels:

  photomap (element+0x28 = FF/00)          fill-dependent   [verified]
  accelerator FIFO (0x5600080)             fill-dependent   [verified]
  engine output (0x5600200 -> CPU)         fill-dependent   [verified here]

  staging buffer (0x004FE790)              NOT fill-dependent  [802 identical writes]
  VRAM 0x04000000..0x0401FFFF              NOT fill-dependent  [0 differing bytes]

So the loss is between the CPU's read of the engine output and video memory.  The
routines that read the staging buffer (file 0x1255BA inside the transfer routine,
0x182D94, 0x1288DA, 0x128B0A) are byte/word copy loops, and none of them was
observed producing fill-dependent VRAM traffic.

STATE
  * NO client-chosen XIE image has ever been displayed;
  * export completes only under the opt-in VT1200_ACCEL=1 model;
  * default configuration unchanged; xshow (core X) is the only viewer that has
    shown an image.
