STAGING BUFFER RE-MEASURED WITH THE CORRECT SPAN -- STILL NOT FILL-DEPENDENT.

Correction first: the previous staging watch used length 0x400 address units
(128 bytes) but B1 walks 0x004FE790..0x004FE980 (496 bytes), so most stores fell
outside the watch.  Re-measured with a 512-byte span:

  fill=ff : 4399 writes  0000 x1960, FFFF x260, 5555 x256, 4100 x114
  fill=00 : 4699 writes  0000 x2123, FFFF x260, 5555 x256, 4100 x127

The 0xFFFF count is IDENTICAL (260) for both fills, so the FFFF words in the
staging buffer are constant engine/manager traffic, not the client's tile.  The
totals differ slightly (allocation bookkeeping) but no value tracks the fill.

Meanwhile the engine output read IS fill-dependent (0x5600200: FFFF x256 for a
white export).  So the divergence is real and located precisely at the point
where the CPU's read of the engine output is stored: the store lands at B1, but
what ends up in that buffer is not what the read returned.

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 configuration unchanged;
  * xshow (core X) is the only viewer that has shown an image.
