MULTI-OPERAND SHAPES -- also fail, and always at the same typed-lookup wall.

Live ids: A=0040C000  B=0040D000  C=0040E000 (all class 2, all queryable).

A) The third operand of the bind-family calls is an ID, not a flag.
   slot 33 [A,B,t] for t=0..9:
     t=0       -> BadValue (2)      (id 0 is invalid)
     t=1..9    -> BadIDChoice (14)  with bad = t
   i.e. the handler looks t up as a resource id and reports that id as bad.

B) With three VALID ids the lookup still fails -- because it is typed:
   slot33 [A,B,A] -> BadIDChoice bad=0040C000
   slot33 [A,B,B] -> BadIDChoice bad=0040D000
   slot33 [A,B,C] -> BadIDChoice bad=0040E000
   slot33 [A,A,A] -> BadIDChoice bad=0040C000
   (query B still REPLY: the resources themselves are untouched)

   slot40 [A,B,A] -> BadValue bad=0040C000
   slot40 [A,B,B] -> BadValue bad=0040C000
   slot44 [A,B,A,0] / [A,B,B,0] -> BadValue bad=0040C000

   slot18 [A,B,0] -> BadMatch (8);   slot43 [A,B,0...] -> BadValue

So every multi-operand request fails at a *typed lookup of one of its own
operands*.  The operand exists and is queryable, but it does not have the class
the handler asks for, so the lookup rejects it.  That is the same wall as
PUT_TILE, reached from four more directions.

Combined with the single-operand sweep, the whole client-reachable surface --
every implemented slot, with both one- and multi-operand shapes constructed from
each handler's own length check and operand reads -- fails for the same reason:
the server's resource model needs a resource class that its create path will not
issue with a stored id.

(src_base stayed 00000000 and no window changed in any of these runs; the
export render params are unchanged from evidence/export_renderparams_map.txt.)
