CORRECTION + NEW CONTROL FOUND.

A) CORRECTION to evidence/badmatch_gate_located.txt: the "w9=0x300 moves the
   failure from BadMatch to BadAccess" reading was WRONG.  BadAccess is simply
   the UNARMED path:

     w9=0x100, NO arm  ->  put = ERR10 (BadAccess),  lookup gate not reached
     w9=0x100, armed   ->  put = ERR8  (BadMatch),   lookup gate reached

   With w9=0x300 the arm itself fails, so the put was effectively unarmed and
   therefore reported BadAccess.  Nothing was passed.

B) NEW: the arm request SETS object->byte8, and the value is the arm request's
   16-bit field at payload bytes 8..9 (packed here as "v12").  Measured across
   v12 = 0,1,2,3 (v16 = 0..3):

     v12  v16   arm result   object->byte8 after arm
      0    0     OK           0
      0    1     OK           0
      0    2    ERR8          -
      0    3    ERR8          -
      1    0     OK           1
      1    1     OK           1
      1    2    ERR8          -
      1    3    ERR8          -
      2    0     OK           2
      2    1     OK           2
      2    2    ERR8          -
      2    3    ERR8          -
      3    0     OK           3
      3    1     OK           3
      3    2    ERR8          -
      3    3    ERR8          -

   So b8after == v12 exactly.  byte8 is arm-controlled state -- which is
   precisely the byte the lookup gate at 0x6BAF4 compares against 3.

C) BUT with v12=3 (arm OK, byte8=3) the lookup gate is NOT REACHED, and
   neither 0x8 site (0x6BB4A selector check, 0x6BB86 mask check) fires, yet
   put still returns ERR8:

     v12=3 v16=0  arm=OK  put=ERR8  GATE=not reached  sel12=0  mask=0
     v12=3 v16=1  arm=OK  put=ERR8  GATE=not reached  sel12=0  mask=0
     v12=2 v16=0  arm=OK  put=ERR8  GATE=not reached  sel12=0  mask=0

   (the v12=2 case also does not reach the gate here, unlike earlier runs, so
   the gate is only reached under some additional condition)

   Therefore ERR8 has at least TWO distinct origins in slot 19, and the one
   that fires now is EARLIER than the lookup.  Prime candidates are slot 19's
   own geometry checks at 0x6A7D0..0x6A846, which compare the request's
   x/y/w/h against the resource's dimensions.

NEXT (bounded): with v12=3 armed (byte8=3), sweep the tile header x/y/w/h
(payload words 1..4) and watch which early check in 0x6A7D0..0x6A846 rejects,
since the lookup gate is no longer the blocker in this configuration.
