RAW SOCKET DUMP -- the reply parser is correct, but a CREATE CAN FAIL *AFTER*
it has already allocated and registered the object.  That invalidates a large
class of earlier measurements.

Method: one connection, then send raw requests and hexdump whatever comes back
(no parsing assumptions).

  create (minor 8, 12 words):
    00 0e 03 00 | 02 10 47 00 | 08 00 81 00 | 20 00 00 00 | e0 56 41 00 ...
  arm (minor 20, 35 words):
    00 0e 04 00 | 02 10 47 00 | 14 00 81 00 | 00 00 00 00 | 20 00 00 00 ...
  put (minor 19, 23 words):
    00 0e 05 00 | 02 10 47 00 | 13 00 81 00 | 40 57 41 00 | 30 22 29 00 ...

Decoding as an X11 Error packet (type, error_code, sequence, bad_value,
minor_code, major_code):
  create: type=0  code=0x0e (14 BadIDChoice)  seq=3  bad=0x00471002  minor=0x0008  major=0x0081
  arm   : type=0  code=0x0e (14 BadIDChoice)  seq=4  bad=0x00471002  minor=0x0014  major=0x0081
  put   : type=0  code=0x0e (14 BadIDChoice)  seq=5  bad=0x00471002  minor=0x0013  major=0x0081

So the parser's error vocabulary is CORRECT (byte 1 is the error code; bytes 8-9
are the minor; byte 10 is 0x81 = the XIE extension opcode).  That part of the
tooling is sound.

BUT: in this very same run the REGIDSTORE breakpoint (file 0x67B04, the ID-store
inside create) fired -- i.e. the server DID allocate an object and register it
under id 0x00471002 -- and the create request still returned BadIDChoice(14),
with bad_value = that same id.

That means create can fail LATE: the resource is allocated/registered first and a
subsequent validation rejects the request.  Consequences for earlier work:

 * Any sequence where "create = no-reply" was recorded (the reply was very often
   not read back) must be treated as UNKNOWN, and may in fact have been this
   late BadIDChoice.
 * Measurements that attributed an error to arm or put may instead be reporting a
   resource whose create had failed -- the arm/put then correctly answer
   BadIDChoice because the id is not usable, even though an object exists.
 * This is the most likely root of the ERR8 / ERR10 / ERR14 inconsistency across
   runs that caused several retractions.

NEXT REQUIRED FIX before any further conclusions: read EVERY reply and assert it
is a success before proceeding (fail the step otherwise).  The earlier
"create=no-reply" lines must be re-run with reply checking; only sequences whose
create, arm and put all report success are usable evidence.
