THE ENGINE'S OUTPUT GOES TO MAIN RAM, NOT VRAM -- AND OUR PIXELS DO NOT LAND.

Instrumented the read-back loop at file 0x125528 during a completing export:

  0x125528  MOVE *B6,*B1      ; engine output port 0x5600200 -> *B1
  0x12552A  ADDK 0x10,B1      ; advance one word
  0x12552C  DSJS A11,0x125528

Measured destination pointer:
  DST b1=004FE790 b2=3FB2ACF0 a6=00000000
  DST b1=004FE7A0 ...
  distinct b1: first=004FE790 last=004FE980
  inside VRAM (0x04000000..0x1fffffff): FALSE

So B1 is a MAIN RAM buffer (0x004FE790..0x004FE980, ~0x200 words), i.e. the
engine's result is staged in RAM and something else is expected to move it to the
frame buffer.  The 384 reads of 0x5600200 come from exactly this loop (384 == the
a1 argument seen at the routine entry).

CHECKED: after a white export, the staging buffer does NOT contain 0xFFFF --
0x004FE790 holds pointer-like garbage instead.  So either the read-back returns
the wrong thing, or this path is not the one that carries the tile.

STATE
  * export completes without error (opt-in VT1200_ACCEL model);
  * source pixels do reach the accelerator FIFO (0xFFFF / 0x0000 verified);
  * the engine's output is staged to main RAM at 0x004FE790 (destination located);
  * the staged data is not our tile, and nothing writes the tile to VRAM;
  * NO client-chosen XIE image has ever been displayed; white vs black captures
    differ only by Terminal Manager figures and cursor;
  * model opt-in; default configuration unchanged;
  * xshow (core X) is still the only viewer that has shown an image.
