COMMAND STREAM DECODED -- THE ACCELERATOR IS DMA-DRIVEN, NOT CPU-FED.

Fresh capture during a completing export (VT1200_ACCEL=1 + compression poke):

  0x5600080 : 154 writes   <- a PARAMETER TABLE
  0x5600050 :   1 write    <- a single flush word (0000)

The 0x5600080 stream is a register/value table, not pixel data.  Decoded as
(low byte = parameter index, high byte = value):

  idx 00=02  idx 01=00  idx 02=01  idx 04=FF  idx 05=BF  idx 06=FF  idx 07=F8
  idx 08=01  idx 09=02  idx 0A=00  idx 0B=01  idx 0C=01  idx 0D=02  idx 0E=00
  idx 0F=01  idx 10=00  idx 11=4B  idx 12=4B  idx 13=FF  idx 14=EF  idx 15=FF
  idx 16=EF  ... (values FF/BF/F8/EF/4B look like coefficient/mask entries)

154 words is far too few for 64x64 pixels (256 words minimum), and there is only
ONE data write.  So the pixel traffic never passes through the CPU-visible FIFO
in this run: the engine is expected to fetch the source and write the
destination by DMA from the descriptor programmed into the register block.

CONSEQUENCE
  Believing the earlier 0x7FE-word loop was the pixel path was wrong: that loop
  writes a constant (A11=0) and is a flush.  Implementing "FIFO -> framebuffer"
  as previously planned would not paint anything.

  To paint, the model must decode the descriptor (source address, destination
  address, width/height, strides, and the table uploaded to 0x5600080) and
  perform the copy itself -- i.e. a real device implementation, not a status
  shim.

STATE
  * export completes with no error (previous milestone, still true);
  * the framebuffer is unchanged; no client-chosen image is displayed;
  * model remains opt-in (VT1200_ACCEL=1); default configuration verified intact;
  * xshow (core X) remains the only viewer that has shown an image.
