"_" as a wildcard for attribute produces false positives on large dataset.

I don't have minimal reproducible case yet (db size I see this with is 60k+ datoms).

"False positives": some property entities are included in result set.
Issue is not reproducible with the following snippet, but schema and query have this shape:

(let [conn (ds/create-conn
             {:bar/properties    {:db/valueType :db.type/ref
                                  :db/cardinality :db.cardinality/many}
              :foo/properties    {:db/valueType :db.type/ref
                                  :db/cardinality :db.cardinality/many}
              :property/attribute {:db/valueType :db.type/ref}
              :attribute/name {:db/unique :db.unique/identity}})]
  (ds/transact! conn [{:db/id -1 :attribute/name "A"}
                      {:db/id -2 :foo/properties {:db/id -21 :property/attribute -1}}
                      {:db/id -3 :bar/properties {:db/id -31 :property/attribute -1}}])

  (->>
    (ds/q '[:find [?e ...]
            :in $ ?a
            :where
            [?p :property/attribute ?a]
            [?e _ ?p]]
      @conn 1)
    (ds/pull-many @conn '[*])))

https://clojurians-log.clojureverse.org/datascript/2017-04-06.html
https://clojurians-log.clojureverse.org/datascript/2017-04-07.html

[org.clojure/clojurescript "1.9.229"]
[org.clojure/clojure "1.8.0"]
[datascript "0.15.5"]

might be related:

1. actual dataset has retracted entities
2. https://github.com/tonsky/datascript/issues/214

Start by reducing the reported schema and query to a reproducible case, using the 60k+ datom dataset shape and checking whether retracted entities matter. Compare the wildcard _ query behavior with the discussion linked in the issue and issue #214. Done means the false positives are reproduced and a regression test demonstrates the corrected result set.