EXPORT PAINTS NOTHING -- the "XIE output pattern" was the window's default
background.  Measured on one MAME instance, one X connection, 64x64 window.

1) Baseline: window created + mapped + GC created, NO export at all:
     BASELINE no-export  nz=512 first=88888888888888882222222222222222

2) Full sequence create -> arm1 -> PUT_TILE -> arm2(v16=0) -> EXPORT,
   then read back the same window:
     AFTER export        nz=512 first=88888888888888882222222222222222
     identical to baseline: True

So every earlier "export OK ... 8888...2222..." result measured the window
background (a stipple), not XIE output.  The white-vs-black comparisons were
therefore vacuous: export has never drawn anything.

3) Why: the export chain is present but has no source image.
   Breakpoints in the export driver (file 0x1B152 = bit 0x3F2D8A90):

     EXPORTENTRY
     EXPORTFLAG f=01                      ; global @0xA6220 non-zero, render taken
     EX_BAND a5=003DC4B0 elem@d8=003FDA10 ; photoflo->band->element IS present
     EX_ELEM a5=003FDA10 f28=00000000     ; element+0x28 == 0

   The driver copies element+0x28 into the render params
   (0x1B238  MOVE *A5(0x28),*A11(0x610)), i.e. it is the imported source
   image.  It is zero because the import never ran: PUT_TILE (minor 19) is
   rejected at the class lookup with BadMatch (S19LOOK a8=00000008), so the
   handoff that fills/uses element+0x28 (0x6B640) is never called.

4) The per-element render loop (file 0x1BC38..0x1BF30) is never reached:
   its breakpoints (ELTYPE / PIPE_T1..T4 / PIPE_LOOPEND / BADFUNC) do not
   fire during the export.

Conclusion: export produces no pixels because the source image pointer at
element+0x28 is NULL.  The single remaining blocker is unchanged and is
upstream of export: PUT_TILE's lookup requires a class-3 resource with
resource byte 40 == 0, which no reachable request produces.
