ACCELERATOR MODEL, FIRST CUT (opt-in, VT1200_ACCEL=1)

1) REGISTER PROTOCOL CAPTURED
   A read/write watch over 0x5600000..0x56001FF during one genuine decclient
   export recorded 8365 accesses:
     0x5600010  read 8192 times   <- status, polled
     0x5600080  written 154 times <- command/data FIFO
     0x5600100  read+written 10   <- control
     0x5600030  written 6
     0x5600020  written 2  (0x248)
     0x5600000  written 1  (0)
   The poll at file 0x122E54 tests bit 1 of 0x5600010, 0xFFF times, and on
   failure prints the ROM string
      "ERROR: timeout occurred on initialization of scalution"
   then returns XIE BadValue(2).

2) DRIVER CHANGE (harness/mame/vt1200.cpp)
   Added an accelerator register block at bit 0x05600000 with:
     accel_r: offset 1 (== 0x5600010) returns 0x0002 ("initialised");
              other offsets return the last written value
     accel_w: latch
   Made it exclusive by splitting the VRAM window around the block, and made the
   whole thing OPT-IN via VT1200_ACCEL so the verified default is untouched.

   Two things learned the hard way while doing it:
   * the address map hands handlers (address - base) >> 4, so 0x5600010 arrives
     as offset 1, not 0x10;
   * an overlapping map entry does NOT take precedence over the VRAM window --
     the range had to be made exclusive by splitting VRAM.

3) RESULT WITH THE MODEL ENABLED
   * the CPU now reads a9=0002 from 0x5600010 (bit 1 set) and the code reaches
     the post-poll path instead of the timeout branch;
   * the machine then FREEZES at frame 1: the export hangs and the guest spins,
     i.e. after passing the "initialisation" wait the firmware blocks in a
     further accelerator wait whose register the model does not satisfy.
   * no image is painted.

4) DEFAULT CONFIGURATION VERIFIED UNCHANGED
   With no environment, the machine boots normally (X ready in 3 s) and the
   framebuffer screenshot is produced.  The model is inert unless VT1200_ACCEL=1.

STATUS.  NO client-chosen XIE image has been displayed.

NEXT: with VT1200_ACCEL=1, sample the guest PC during the freeze to find the
second wait loop, then extend the model with the register it polls (likely the
0x5600080 FIFO's status or a done bit in 0x5600100) -- the same iterative
approach that got past the first gate.
