add developer comment

src/optimization/duality.jl

@@ -86,6 +86,12 @@ function activation_value(σ::ReLU, μᵢ, λᵢ, l̂ᵢ, ûᵢ)
     gᵢûᵢ = @. μᵢ*ûᵢ - λᵢ*σ(ûᵢ)
 
     max = symbolic_max
+
+    # NOTE: ifelse evaluates all of its arguments in advance.
+    # This may seem like a bug since symbolic_max mutates the model
+    # and introduces new variables, but it is fine. Only those
+    # variables that are part of the objective will affect the problem,
+    # and those will only the the ones summed *after* taking the ifelse
     return sum(@. ifelse(l̂ᵢ < 0 < ûᵢ,
                          max(gᵢl̂ᵢ, gᵢûᵢ, 0),
                          max(gᵢl̂ᵢ, gᵢûᵢ)))