VERIFIED PROBE (xie/probe_verify.py): one field helper, one reply per request,
raw bytes printed, run TWICE.  Both runs identical.

  SET_OP_DEFAULTS(bpp=1) -> REPLY len=0, head = 01 41 03 00 then all zeros
  create 00401000        -> SILENT (accepted)
  create 00402000        -> SILENT (accepted)
  registered              -> both ids registered, class=2, byte8=1, byte9=1,
                             p44 set, elem(+0xd8) populated, elem.byte0 = 0x01

  slot33 [A,B,0]     -> Invalid code=2 (BadValue)   new REGIDSTORE: none
  slot44 [A,B,1,1]   -> Invalid code=2 (BadValue)   new REGIDSTORE: none
  element records unchanged by both.

  reproducibility: slot33 MATCH, slot44 MATCH across the two runs.

ROOT CAUSE OF THE EARLIER FLAKINESS (found by this script):
  SET_OP_DEFAULTS REPLIES with a 32-byte packet `01 41 03 00` + zeros.  In all my
  earlier ad-hoc runs that reply was never consumed, so it was sitting in the
  socket and got attributed to the NEXT request.  That is why slot 33 "returned
  a reply" with head `01 41 03 00 00...` in some runs -- that packet is
  SET_OP_DEFAULTS' reply, not slot 33's.  Likewise the slot-40 "reply in one run,
  BadValue in another" was the same stale packet.

  So: the 32-byte `01 41` reply carries NOTHING about slot 33/44.  It is the
  op-defaults acknowledgement and its payload is all zeros.

CORRECTION to earlier statements: the element pointer at band_array+0xd8 IS
populated by create (0x3FD7D0 / 0x3FD590) and its byte0 = 0x01 -- it is not
"uninitialised heap fill" as I said; the 0x55 I saw was in a different field
(word 1 of that record).

VERIFIED MODEL NOW:
  create (op=2) -> class-2 resource, byte8=1 (bits/pixel), byte9=1,
                   band array at +44, element pointer at band_array+0xd8,
                   element record populated with byte0=1.
  slot 33 [A,B,0]    and  slot 44 [A,B,1,1]  are BOTH rejected (BadValue),
                   reproducibly, and change no state.

STATUS: still no client-chosen image.  But this is the first measurement set this
session that is reproducible across runs, and it removes a whole class of earlier
"results" that were actually the stale op-defaults reply.
