part_integrity="cascade" deletes masters that have no matching part rows #1591

With part_integrity="cascade", deleting part rows can delete masters that have no matching part rows, together with everything downstream of those masters.

@schema
class Master(dj.Manual):
    definition = """
    master_id : int32
    """

    class P(dj.Part):
        definition = """
        -> master
        p_id : int32
        """

Master.insert([(1,), (2,)])
Master.P.insert([(1, 1)])          # master 2 has no parts
(Master.P & [{"master_id": 1}, {"master_id": 2}]).delete(prompt=False, part_integrity="cascade")
len(Master())   # 0 — expected 1: master 2 had no matching part row

Cause. The part-to-master lift copies the part's condition to the master whenever the restricted attributes are in the master's primary key (Diagram._apply_propagation_rule_upward, Rule 1). That selects every master satisfying the condition, not the masters of the part rows that matched.

Workaround until the fix ships: restrict the parts by keys that exist, e.g. Master.P & (Master.P & cond).keys(), or use the default part_integrity="enforce".

Fix: the lift restricts the master by the restricted part, master & part_restricted.proj(...), so a master joins the cascade only if one of its part rows did.