IMPORT PATH RE-VALIDATED WITH THE TRUSTWORTHY HARNESS.  A long-standing model is
OVERTURNED.

Two trials, one connection each, every reply read and asserted, raw bytes shown.

  trial 1 (white tile)                      trial 2 (black tile)
  SET_OP_DEFAULTS ack REPLY 01 41 03 00 ...  ack REPLY 01 00 03 00 ...   (detail differs)
  create 00401000   SILENT(ok)              SILENT(ok)
  obj               003FDA10                003FD110
  byte8=1 byte9=1 class=2                    byte8=1 byte9=1 class=2
  p44 / elem        set, populated          set, populated
  arm1              SILENT(ok)              SILENT(ok)
  PUT_TILE          ERROR 8                 ERROR 8
  arm2              ERROR 8                 ERROR 8
  EXPORT            SILENT(ok)              SILENT(ok)
  GATE (0x6BAF4) hits during put: NONE      NONE
  RP line           a11=00923CB0 src_base610=00000000   a11=009240F0 src_base610=00000000
  byte8 after all   1                       1
  QUERY_EXPORT      REPLY 01 00 0c 00 04 00 00 00      REPLY 01 00 0c 00 04 00 00 00

THE OVERTURNED MODEL:
  I have been reporting since the middle of this session that
     arm   requires object->byte8 == 1   (0x6ACBE)
     import requires object->byte8 == 3  (0x6BAF4)
  and that PUT_TILE fails AT that gate.

  With a harness that actually asserts replies, the gate at 0x6BAF4 is NEVER
  REACHED during PUT_TILE -- zero hits in both trials.  So PUT_TILE's BadMatch is
  produced somewhere else entirely, and the entire byte8/gate story (including
  the w9=0x300 experiments) was built on data from the stale-reply era.

Also newly verified in this run:
  * arm2 (slot 20, v16=0) returns BadMatch(8) -- it is NOT silent/OK as reported
    earlier;
  * EXPORT is silent(ok) even when PUT_TILE and arm2 both failed;
  * QUERY_EXPORT replies `01 00 0c 00 04 00 00 00` identically in both trials
    (this is its real reply; the 4-word payload follows in the head);
  * src_base (rp+0x610) is 0 in both trials, read at the in-export breakpoint.

CONSEQUENCE: the import failure must be re-located from scratch.  The next step is
NOT the byte8 gate.  Candidates, now that replies are trustworthy: the lookup at
0x7EAB966 returning 8, or a check in slot 19 before/after it.  S19LOOK-style
breakpoints must be re-run under this harness.
