WRITE WATCH ON THE ACTUAL src_base SLOT (rp+0x610), plus slots 15 and 22.

Setup: valid c.new_id ids, no-reply model, generated bit addresses.  The rp
pointer was captured at the export breakpoint (0x1B296, where A11 = rp):
  rp = 00913330   ->   rp+0x610 = 00913940
A 32-bit write watchpoint was installed at 00913940.

Sequence and results:

  create/arm1/put/export#1      put = BadMatch, export#1 = BadMatch
  slot15 [id,0]                 silent (success)
  slot22 [id]                   silent (success)
  slot15 [id,id]                silent (success)
  arm2, export#2                silent (success)

  rp+0x610 after export#1 = 00000000
  rp+0x610 after export#2 = 00000000

WRITES OBSERVED AT rp+0x610 (18 events), by PC:
  pc=3FE16CA0  val=00000000   x8   (file 0x3C2D94 - allocator/fill)
  pc=3F2D91F0  val=00000000   x2   (file 0x1B23E - the export copy block)
  pc=3F2D92E0  val=00000055
  pc=3F2D9300  val=00000055 / 00000000
  pc=3F2DBC70  val=0000B055 / 000090E6
  pc=3F2E2C00  val=0000B000
  pc=3F2E2C90  val=00000000

Readings:
 * The only writer inside the XIE export path is 0x1B23E (pc 3F2D91F0), i.e. the
   known `MOVE *A5(0x28),*A11(0x610)` copy, and it writes 0 because
   element+0x28 is 0.
 * The 0x55 / 0xB055 / 0x90E6 values are the heap fill pattern and unrelated
   allocator metadata, not an image pointer.
 * Slots 15 and 22 both return SILENT (success) for [id,0], [id,id] and [id] -
   they are permissive requests - but neither causes any non-zero write to
   src_base.

So: slots 15 and 22 do not populate the source image buffer, and the only
in-code writer of src_base is export's own copy from element+0x28, which stays 0.
Also noted: export fails (BadMatch) if arm2 has not been issued; with arm2, export
is silent/success - consistent with the earlier arm2 gate finding.

CONCLUSION for this line: the src_base slot has been watched directly, across the
full cycle plus slots 15/22, and nothing sets it.  The image source buffer is
never produced.
