BISECT OF file 0x6BA1E..0x6BAA2 -- arm and put take DIFFERENT branches.

10 breakpoints (harness/mame/xie_min4.lua), fired-path per request:
  B=0x6BA1E class2/3 entry   C=0x6BA24 copy byte40 out   D=0x6BA32 byte40==0
  E=0x6BA3E err 0E(byte40)   F=0x6BA66 err 0E(other)     G=0x6BAA2 merge
  H=0x6BAAC store 5a8        I=0x6BAB8 outcell test      J=0x6BAC4 err 0A

  arm, w9=0x100 (byte8=1) :  B C G H            -> succeeded
  arm, w9=0x300 (byte8=3) :  B C G H            -> ERR8 (BadMatch)
  arm, w9=0x300, mode=3   :  B C G H            -> ERR8
  put, w9=0x100 (+ arm)   :  B D G H I          -> ERR8

At B the printed arg1 (`*A13(0xffc0)`, the 2nd-from-last pushed argument):
  arm pushes A13   -> non-zero -> takes C (0x6BA24: *arg1 = obj->byte40)
  put pushes 0     -> zero     -> takes D (0x6BA32: require obj->byte40 == 0)

So the branch is chosen by an argument the CLIENT controls indirectly through
which request is sent, not by the resource state.

Both then merge at G (0x6BAA2) and execute H (0x6BAAC).  From there:
  * put continues to I (0x6BAB8) and then on into the gate at 0x6BAF4.
  * arm STOPS -- I never fires -- and the request returns BadMatch.

So arm's BadMatch is produced between 0x6BAAC and 0x6BAB8, i.e. in
   0x6BAB2  A7 = *A13(0xff20)      ; arm pushes 0xffff here -> 0x0000FFFF
   0x6BAB6  JRLT 0x6BABC
   0x6BAB8  A7 = *A11 ; JRNE 0x6BAEA
For arm that value is 0xFFFF (positive), so JRLT should fall through to 0x6BAB8
-- yet 0x6BAB8 does not fire.  That contradiction is the next thing to resolve;
it may mean the arm frame's *A13(0xff20) is NOT the pushed 0xffff (argument order
differs from the mapping assumed here), or that the failing instruction is the
store at H dereferencing a zero.

NOTE on a related measurement: for arm the copy at C writes obj->byte40 (which is
0 for our resources) to the caller's local, and H then uses that zero as a
pointer:  0x6BAA6 A7 = *A11   (A7 := the copied 0)
          0x6BAAC *(A11+0x120) = *(A7+0x120)   -> reads at 0x120
This is consistent with "arm needs byte40 non-zero", but a direct test (poking
byte40 = the object's own address) did NOT change arm's result, so that theory
is unconfirmed.

Nothing here yet makes arm return OK with byte8=3, so the sequence still cannot
reach the import handoff and src_base remains 0.
