ACCELERATOR REGISTER TRACE, WHITE vs BLACK, DIFFED (properly formed watch this time).

Watch: 0x05600000 + 0x200, read/write, logging addr, value and PC.
Both traces: 2436 lines.  diff = 2145 lines, and they are ENTIRELY the data FIFO:

  < ACC 05600080 FFFF 3FB2A0C0     (white)
  > ACC 05600080 0000 3FB2A0C0     (black)
  ... repeated for the ~256 payload words

Every other register line is identical between the two fills:

  0x5600020  0248 / 0258            (constants)
  0x5600030  0032 0033 0432 0233 ... (control sequencing, identical)
  0x5600100  0000 0001 0002 0008 0200 0201 0202 (identical)
  0x5600010  polled identically
  0x5600200  read identically (values differ only in what the model returns)

So: **the ONLY fill-dependent engine port is the data FIFO (0x5600080).**  No
placement, geometry or destination register differs between a white and a black
export -- the engine is never told where to put the pixels per export.

That is why implementing "the missing store" in accel_r/accel_w is not possible
from the traffic: there is no address to key it on.  The engine must place output
using state established outside this per-export register window (its own internal
registers, set once), which is exactly the device model that does not exist.

STATE
  * NO client-chosen XIE image has ever been displayed;
  * VRAM byte-identical for white and black (0 differing bytes / 128 KiB);
  * model opt-in (VT1200_ACCEL=1); default unchanged;
  * xshow (core X) is the only viewer that has shown an image.
