ELEMENT-TYPE STORES AND THE RENDER GATE -- measured, and both are dead ends so far.

A) The element-type stores at file 0x1B2B8 (type 1) and 0x1B362 (type 2) are in
   the EXPORT-side render driver (0x1B152), not in arm.  Instrumented through a
   full create -> arm1 -> put -> arm2 -> export cycle:

     TYPE1 / TYPE2 stores fired: NONE

   So no element type is ever written for our requests; the types array at
   rp+0x58 stays zero.

B) The persistent element record at band_array+0xd8 does not change either:

     elem before export : 01050001 00000055 00100000 00000100 00004000 00000000 00003F00
     elem after  export : 01050001 00000055 00100000 00000100 00004000 00000000 00003F00

   Note word 1 = 0x00000055, the heap fill pattern -- the record is PARTIALLY
   UNINITIALISED, so it is not a populated element descriptor and cannot be used
   to infer a field layout.

C) The render-enable flag @0xA6220 is 1 at every step and never changes:

     session init 1 | SET_OP_DEFAULTS 1 | create 1 | arm1 1 | put 1 | arm2 1 |
     window+gc 1 | EXPORT 1

   So the render path is not gated off by that flag.

D) rp (the render-params block, @0xA6000) reads 0 after export -- but that is
   expected: export ALLOCATES it (0x1B18A) and FREES it in its tail
   (CALLA 0x7F63B00).  Reading it from outside is therefore always 0; it must be
   sampled at the in-export breakpoint 0x1B296 (which earlier showed
   rp+0x5a8=src_res_id, rp+0x610=0).

CONCLUSION FROM THIS TURN: the element record is uninitialised, no element type
is ever written, and the render flag is always on -- so nothing in the observed
flow constructs a pipeline at all.  The missing piece remains the 2.3 encoding of
element records (what XIEP_SCALE/CROP/ROTATE/MIRROR/EXPORT carry), which these
probes cannot reveal: a malformed element and a missing element both produce the
same documented BadValue.

NOT PROVEN IMPOSSIBLE -- but no request shape tried so far causes an element type
to be written, and the struct fields do not self-describe.
