TRUSTWORTHY BASELINE ESTABLISHED, and arm's BadMatch localised to AFTER the
shared lookup.

Protocol model used (established last turn, re-confirmed here):
  * create and arm are NO-REPLY requests -- silence means success;
  * only errors emit a packet (type=0 | code | seq | bad | minor | major=0x81);
  * ids MUST come from c.new_id(...) (the client id base changes per connection).

BASELINE -- 3 independent trials, one connection each, each step asserted:

  create  silent
  arm     silent
  put     err BadMatch
  (id=00404000 in all three trials; trace: A_IDSTORE once per trial)

That is reproducible and is now the reference sequence.

ARM FAILURE LOCALISED (4 breakpoints: ID store, lookup success 0x6BBE2,
DD dispatcher entry 0x40008, DD worker call 0x40086):

  arm w9=0x100 (byte8=1) -> silent (success)  trace: LK_SUCCESS
  arm w9=0x300 (byte8=3) -> err BadMatch      trace: LK_SUCCESS
  put w9=0x100 (armed)   -> err BadMatch      trace: (none)

Readings:
 * For BOTH arm cases the shared lookup REACHES 0x6BBE2 (its success return,
  `SUB A8,A8`).  So the lookup does not reject a byte8=3 resource.
 * Arm with byte8=3 still returns BadMatch, and the DD dispatcher (0x40008) and
  its worker call (0x40086) are NOT entered.  So the error is produced after the
  lookup, but not via the device-dependent dispatcher.
 * Put with byte8=1 returns BadMatch WITHOUT reaching 0x6BBE2 -- consistent with
  the gate at 0x6BAF4 failing (1 != 3) and returning at 0x6BAF8.

Also note a correction to the previous turn's instrumented run: that run included
a stray breakpoint at the ROM entry 0x3F200000 which suppressed the trace.  This
run (4 breakpoints, no junk) is the valid one.

NEXT: the arm error is in slot 20's body between the lookup return (0x6AAB8) and
the DD dispatcher, excluding the four MOVK 0x8 sites (0x6ABD2/0x6AC5C/0x6AC96/
0x6ACD4, which do not fire).  Bisect 0x6AAC2..0x6AD1A with valid ids and the
no-reply model.
