DEFINITIVE TEST: THE FRAMEBUFFER IS BYTE-IDENTICAL FOR A WHITE AND A BLACK EXPORT.

Method (unambiguous, no inference from constant-looking traffic):
  * run the genuine DEC 2.3 client with fill 0xFF, wait for the export to
    complete, then dump VRAM 0x04000000..0x0401FFFF (128 KiB) in 1 KiB chunks
    through rdebug;
  * repeat with fill 0x00 in the same session;
  * diff the two dumps byte for byte.

RESULT
  dumped 131,072 bytes each
  differing byte offsets: 0
  => NO DIFFERENCE.  The client-chosen tile never reaches video memory at all.

This supersedes the earlier, wrong readings ("18,338 VRAM writes carrying the
tile", "the tile is in VRAM at 0x04017830", "250 staging writes of 0xFFFF"): the
VRAM and staging traffic is fill-independent Terminal-Manager/engine constant
activity, and only the DIRECT white-vs-black comparison could show that.

WHERE THE DATA IS PROVEN TO EXIST
  element+0x28 in the server photomap: 0xFF white / 0x00 black  (verified)
  accelerator FIFO writes 0x5600080  : 0xFFFF / 0x0000          (verified)
  everything after that              : fill-independent          (proven here)

So the tile is lost inside the modelled engine path, and no downstream stage
copies it to the frame buffer.

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