WITH THE ENGINE-OUTPUT READ PATH MODELLED, THE EXPORT NOW WRITES THE TILE INTO VRAM.

Found and implemented the last link in the data path:

  file 0x125528:  MOVE *B6,*B1 ; ADDK 0x10,B1 ; DSJS A11,0x125528
    B6 = 0x5600200 (engine output port), B1 = a main-RAM staging pointer
    (measured 0x004FE790, ~0x200 words).

  So the engine's OUTPUT port must return the pixels pushed into the FIFO; the
  firmware then stores them.  My model returned a status word there, which is
  why nothing carried through.  Now accel_r returns the FIFO contents for
  0x5600200.

MEASURED AFTER THE CHANGE
  * staging buffer 0x004FE790: receives the tile -- 250 writes of 0xFFFF in a
    white export (earlier my peek showed garbage because it is overwritten later).
  * VRAM writes during one export cycle, white tile:
        before:   241 words
        after:  18,338 words, containing 0xFFFF at e.g. 0x04017830/40/50
    The dominant writer is pc=3F7421E0 (file 0xA843C).
  * export still completes with no error.

So the chain now runs end to end:
  client -> XIE -> photomap -> accelerator FIFO -> engine output -> staging ->
  VRAM writes (18,338 words carrying the tile).

WHAT STILL DOES NOT HAPPEN
  The framebuffer capture does not show the tile: white vs black differ by
  AE 3130 px (Terminal Manager figures + cursor only).  The written addresses
  (0x040178xx) map, at the renderer's VRAM319 pitch of 2048 bits, to about
  (x=48, y=47) -- an area the Terminal Manager repaints, and nothing of the
  client's 64x64 window appears there or at (150,200).

  Either the writes are to a region that is subsequently repainted, or the
  coordinate/pitch mapping between the accelerator's destination and the
  renderer's active picture is still wrong.

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