THE armF STATE, TRACED.  Three distinct lookup codes, one per arming choice.

One connection: I = create op=2 (image), F = create op=1 (photoflo).

  create I 0x40C000 -> OK     REGIDSTORE id=0040C000 obj=003FDC50 ; REGISTER
  create F 0x40D000 -> OK     REGIDSTORE id=0040D000 obj=00388090 ; REGISTER

  arm F (v16=1) then PUT_TILE I  ->  S19LOOK a8=0000000A  (BadAccess, 10)
  arm I (v16=1) then PUT_TILE I  ->  S19LOOK a8=00000008  (BadMatch,  8)
  PUT_TILE F (never armed)       ->  S19LOOK a8=0000000E  (BadIDChoice, 14)

So arming the photoflo genuinely changes WHICH failure the typed lookup reports:
it gets past the stage that produces BadMatch and stops at a BadAccess check.
That is a real state difference, not noise, and it is the only one found.

Caveat on the LOOK breakpoint (0x6BA38): it prints cls=01 in both the armF and
armI cases, and earlier it printed an incrementing 01,02,03 across successive
calls.  That means A7 at 0x6BA38 is NOT the resource struct -- it is the export
element (the same buffer whose +0x28 is src_base).  So LOOK's "cls" and "b40"
columns are not resource fields and must not be read as such; only S19LOOK's
return code (a8 at 0x6A7B4) is trustworthy in that trace.

Registration sites during armF/put: 0 hits (they fire only inside create).

CONCLUSION: arming the photoflo advances the import lookup past BadMatch to
BadAccess.  That is the furthest any configuration has reached.  No configuration
reached the handoff (0x6B640) or wrote src_base.
