Disabled dangling branches lead to runtime differences in subscription behavior

[bruno-f-cruz]

Discussed in https://github.com/orgs/bonsai-rx/discussions/2006

^{Originally posted by edmundchong September 8, 2024}
I created a frequently updated BehaviorSubject which I pass through some conditional statements. However, there are strange effects from nodes that are commented out (but still connected). Here is a toy example which isolates the issue:

In this case mySubject is only initialized once but is propagated 3 times.

When the commented out node is deleted or simply connected on the other end like the rest of the Condition nodes, everything is fine:
.

The same behavior happens also if I e.g. replace Merge with Zip or CombineLatest or Amb.

Workflow attached below.

foo.bonsai.zip

[glopesdev]

@bruno-f-cruz This regression was introduced in #1324 following an optimization whereby unmerged disabled branches can lead to cancellation of multicast behavior:

bonsai/Bonsai.Core/Expressions/ExpressionBuilderGraphExtensions.cs

Line 982 in 5279025

if (activeBranches == 1) scope.MulticastBuilder.BranchExpression.Cancel();

Unfortunately, the assumptions of the above code are too simplistic (and incorrect). It is quite possible (as your example shows) that the single remaining unmerged branch is in fact a split-and-merge combinator expression which requires the multicast to be preserved for correct behavior.