THE OUT-CELL WRITER IDENTIFIED -- and its source is not what it looked like.

Method: write watchpoint at the out-cell address itself (00415720, stable across
runs), plus the stack positive control (3528 hits, so the watch mechanism is
proven live), under the asserting harness.

RESULT (byte8=1 resource):
  OUTWRITE pc=3F55D590 val=00001D70      <- pc = file 0x6BAB2
  OUTCEL   a11=00415720 star=00291D70
  T_GATE   a7=00000001 a11=00000003
  S19RET   a8=00000008

RESULT (byte8=3 resource):
  OUTWRITE pc=3F55D590 val=00000000
  OUTCEL   a11=00415720 star=00000000
  S19RET   a8=0000000A

  (later OUTWRITEs at pc=3FB41C20 and 3F22A300 are unrelated late writes after
   the request had already returned.)

So a SINGLE store fills the out-cell, at file 0x6BAB2 = the store at 0x6BAAC
(`MOVE *A7(0x120),*A11(0x120)`), writing the low word of a pointer (0x1D70) for
byte8=1 and 0 for byte8=3.

WHAT THE SOURCE IS NOT:
  * not object->byte36 -- measured 00000000 for BOTH w9=0x100 and w9=0x300;
  * not object->byte44 -- 003DBAF0 for both;
  * not the band element pointer.

WHAT I TRIED TO GET byte8=3 WITH A FILLED OUT-CELL:
  swept create payload words 4,5,6,7,8,10 with values 1,2,0x100 alongside
  w9=0x300: no combination produced byte8=3 with a non-zero out-cell.

STATE: the out-cell is filled by `*(A7+0x120)` where A7 descends from the caller
frame, and the source object field was not identified.  The two states remain
mutually exclusive in every configuration reachable from this client.

This is the point where I stop probing.  Sixteen-plus turns of request-level
experiments have produced a complete map of the implemented surface, a validated
harness, and a measured statement of the wall -- but not a displayed image.  The
one thing that would resolve it is the definition of what is supposed to fill
that out-cell / how byte8 is meant to transition, i.e. the DEC XIE 2.3 headers or
client binary.  Further black-box probes return the same two documented errors.
