Regression in 1.6.4: unbound vars not filtered

[mainej]

While drafting #468, I noticed another regression. The two problems are quite likely related, but I'm submitting this one separately so I can explain it in isolation.

Consider the same query as in #468, but with the following edit to the or-join:

(or-join
 [?x ?y]
 (and [?x :parent ?z]
      [?z :parent ?y])
- [?y :parent ?x])
+ [?x :parent ?y])

In 1.6.3 this produces:

(["alice" #{[1 3] [1 2]}] 
 ["bob" #{[2 3]}]
 ["clara" #{}])

In 1.6.4 it produces:

(["alice" #{[1 3] [1 2]}]
 ["bob" #{[2 3]}] 
 ["clara" #{[3 nil]}])

It's very surprising to see a nil in the output of a query!

• ❌ When the query is run for "clara", neither part of the or-join binds ?y. In 1.6.3 this would remove all tuples, but it seems that now ?y is bound to nil.

[tonsky]

Hopefully fixed in 1.6.5