THE SLOT-19 LOOKUP IS FILE 0x6B966, NOT 0x66FA0.  And it does not contain the
failure we see.  All previous explanations built on 0x66FA0 are void.

1) The call, from slot 19's own handler:
     rom:0006a7ae CALLA 0x07eab966      ; file = imm - 0x7E40000 = 0x6B966
     rom:0006a7b4 MOVE A8,*A13(0x140),0x1   ; A8 = the returned code

   Arguments pushed immediately before it (slot19 listing 0x6A790..0x6A7AC):
     *A9(0xc0)  -> request payload word 5 (offset 24)
     MOVK 0x3   -> literal 3
     *A9(0x20)  -> request payload word 0 (the resource id)
     A13+0      -> out pointer
     A13        -> frame pointer
     0
     client (*A13(0xffe0))
   i.e. 7 arguments.  0x6B966's frame reads them as
     A13(0xffe0)=client, 0xffc0=arg1, 0xffa0=arg2, 0xff80=arg3,
     0xff60=arg4 (the id), 0xff40=arg5 (=3), 0xff20=arg6 (= payload word 5)

2) 0x6B966's flow (file offsets):
     6B97C push @0x1ad980 (8-bit)               ; core-X-ish table args
     6B984 push @0x1ad990 (8-bit)
     6B98C push the id
     6B990 CALLA 0x07F174A4                     ; returns an object pointer
     6B996 store it, A11 = *A7
     6B99A JRNE 0x6BA8C                         ; non-null -> dispatch on class
     6B99C ... MOVK 0xe                         ; null -> BadIDChoice
     6BA8C A11 = object ; A8 = object->byte4    ; class byte
     6BA90 DEC ; JREQ 0x6B9C4                   ; class 1
     6BA98 DEC ; JREQ 0x6BA1E                   ; class 2
     6BA9C DEC ; JREQ 0x6BA1E                   ; class 3
     6BAA0 JR 0x6BA66 -> MOVK 0xe               ; any other class -> BadIDChoice
     6BA1E (classes 2/3) arg3==NULL ?
             -> require object->byte40 == 0, else MOVK 0xe   (6BA38..6BA40)
        arg3!=NULL -> copy object->byte40 out to *arg3        (6BA2A)
     6B9C4 (class 1) requires arg3 != NULL (else 0xe) and uses a
           two-step deref: object->byte8->byte8 or object->byte12->byte8

   NOTE: this function raises ONLY 0xE and 0xA (6BA40 / 6BAC6).  There is no
   route in 0x6B966 that produces 0x8 before 0x6BBC0.

3) Measurements that conflict with any simple story, recorded as-is:
   a) Sweeping request payload word 5 (arg6) over
        0x0 0x1 0x2 0x100 0x10000 0x80000000 0xFFFFFFFF 0xFFFFFFFE
      in both armI and armF produced BadMatch(8) EVERY time -- arg6 has no
      observable effect.
   b) A breakpoint on the 0xE-build site (0x6BA86) DID fire during a put whose
      reply to the client was ERR8.
   c) Breakpoints on the 0xA site (0x6BAC6) and on a MOVK-0x8 site inside the
      function (0x6BB4C) did NOT fire during that same put.
   d) A breakpoint on the shared reply/logging helper 0x7EFFA5E did not fire at
      all during the put.

   So the ERR8 the client receives is NOT produced by the error-site we can see,
   and site-based attribution is unreliable.  The handler's return value is what
   the framework turns into the protocol error, and the 0x8 must come from a
   callee return (candidates inside 0x6B966: 0x7F174A4, 0x7F16AD0, 0x7EFFA5E).
   Of those, 0x7F16AD0 (file 0xD6AD0) was checked and is a bit-field extractor
   returning 0xFFFF for unknown selectors -- not the source either.

NEXT (bounded): disassemble the rest of 0x6B966 (0x6BBC0..0x6BBE4) and the
0x6BAEA path, and log A8 at every return, to find the 0x8's true origin.  Until
that is known, "BadMatch" cannot be interpreted.
