FILTER changes a query from sequence to join

[Holger Knublauch]

Hi SPARQL gurus,

In ARQ 2.9.2 I have the following two queries
SELECT *
WHERE {
    BIND (<http://topbraid.org/examples/kennedys> AS ?graph) .
    BIND (42 AS ?dummy)
    GRAPH ?graph {
        ?s rdfs:subClassOf ?o .
    }
}

SELECT *
WHERE {
    BIND (<http://topbraid.org/examples/kennedys> AS ?graph) .
    BIND (42 AS ?dummy)
    GRAPH ?graph {
        ?s rdfs:subClassOf ?o .
        FILTER (?dummy > 0)           # This is different
    }
}

The seemingly trivial addition of the FILTER causes the lower one not to 
work as expected. It will iterate over all graphs instead of using the 
BIND. It's for some reason executing the BIND *after* the GRAPH. A look 
at the debugger reveals the difference in algebra. The upper one is

(sequence
  (extend ((?graph <http://topbraid.org/examples/kennedys>) (?dummy 42))
    (table unit))
  (graph ?graph
    (bgp (triple ?s rdfs:subClassOf ?o))))

and the lower one becomes

(join
  (extend ((?graph <http://topbraid.org/examples/kennedys>) (?dummy 42))
    (table unit))
  (graph ?graph
    (filter (> ?dummy 0)
      (bgp (triple ?s rdfs:subClassOf ?o)))))

What is causing this difference? It feels very inconsistent. How can I 
safely execute the BIND before the GRAPH?

Thanks
Holger