Corruption of Rete when removing complex NotNodes

while working further on our drools integration we came across an odd
exception when removing a particular rule:

java.lang.IllegalArgumentException: Cannot remove a sink, when the list of sinks is null
at org.drools.reteoo.ObjectSource.removeObjectSink(ObjectSource.java:159)
at org.drools.reteoo.RightInputAdapterNode.doRemove(RightInputAdapterNode.java:217)
at org.drools.common.BaseNode.remove(BaseNode.java:95)
at org.drools.reteoo.BetaNode.doRemove(BetaNode.java:275)
at org.drools.common.BaseNode.remove(BaseNode.java:95)
at org.drools.reteoo.BetaNode.doRemove(BetaNode.java:280)
at org.drools.common.BaseNode.remove(BaseNode.java:95)
at org.drools.reteoo.RuleTerminalNode.doRemove(RuleTerminalNode.java:387)
at org.drools.common.BaseNode.remove(BaseNode.java:95)
at org.drools.reteoo.ReteooBuilder.removeRule(ReteooBuilder.java:237)
at org.drools.reteoo.ReteooRuleBase.removeRule(ReteooRuleBase.java:371)
at org.drools.common.AbstractRuleBase.removeRule(AbstractRuleBase.java:746)

While stepping through the code it looked like the network was corrupt (there was indeed no
sinks on the ObjectSource, but the node calling removeObjectSink was still linked to it
and claiming it as source).

The rule itself contains multiple NotNodes, checking a condition that looks like this:
not(not(Foo.v = X) and not(Foo.v = Y))

I could track this down to some sort of "loop" in the rete that triggers this when the outer
not node is removed.

When removing BetaNode#doRemove() first walks along 'rightInput':

this.rightInput.remove( context,
builder,
this,
workingMemories );

and eventually in that call it also hits a node that is the direct 'leftInput' of the original beta node.
The removal marks that node as visited in the removal context, and when the 'rightInput.remove' returns to the
beta node it does not visit the leftInput due to this condition in BetaNode#doRemove():

if ( !context.alreadyVisited( this.leftInput ) )

{ this.leftInput.remove( context, builder, this, workingMemories ); }

In other words: before the remove the BetaNode had another node that was both referenced directly as 'leftInput',
as well as an input to the 'rightInput'.

The first removal of the rule "worked", and no exceptions happened. But: any further attempt to re-add the same rule and remove
it again lead to the exception above.

I was able to fix it with the attached patch, reproduced here:

+ boolean needRemoveFromLeft = !context.alreadyVisited( this.leftInput );
this.rightInput.remove( context,
builder,
this,
workingMemories );

• if ( !context.alreadyVisited( this.leftInput ) ) {
  + if ( needRemoveFromLeft ) { this.leftInput.remove( context, builder, this, workingMemories ); }

With this patch applied I could add/delete/add the particular rule repeatedly without problems.

The attached patch also adds an assert in ObjectSource#removeObjectSink(): when removing a sink from an object source with
only one sink the sink was unconditionally replaced with an empty sink, although the argument ObjectSink could be a different
sink than the one in the ObjectSource. For CompositeObjectSinkAdapters this case is checked, but not for single sinks.
I originally suspected that place to be responsible for the problem I observed but the assertion never fired in my tests.

(taken from rules-dev mail "Bug in BetaNode#doRemove()?"