the result of bivariate moderation model not the same

Two things. First, you're not using your custom compute plan correctly. It needs to go into your container MxModel, e.g.,

plan <- omxDefaultComputePlan(intervals=TRUE)#<--Note `intervals=TRUE`
plan$steps <- list(
    NM=mxComputeNelderMead(centerIniSimplex=TRUE),
    GD=plan$steps$GD,CI=plan$steps$CI,ND=plan$steps$ND,SE=plan$steps$SE,RD=plan$steps$RD,RE=plan$steps$RD)#<--Note CI step
ACEmodModel  <- mxModel( "ACEmod", modelMZ, modelDZ, funML, multi, Confmod, Conface, plan) #<--N.B.
 
set.seed(100)
ACEmodFit  <-   mxTryHardOrdinal(ACEmodModel,extraTries = 30,intervals=TRUE) #<--No plan

. Notice also what I had to change to get a confidence-interval step in your custom compute plan.

Second, there's a simple explanation for why you're getting different results from your two different runs. As you know, mxTryHard*() randomly perturbs the start values between fit attempts. set.seed() does NOT ensure that every call to mxTryHard*() will produce the same results. Rather, it merely initializes the sequence of pseudorandom numbers in a way that's reproducible. What your script is currently doing is setting the RNG seed, generating a sequence of pseudorandom numbers during the first call to mxTryHardOrdinal(), and then continuing the sequence of pseudorandom numbers in the second call to mxTryHardOrdinal(). If you want your second call to give the same results as your first call, you need to do set.seed(100) again before your second call.

One other thing: since you're using both Nelder-Mead and NPSOL to fit a model that contains equality MxConstraints, you might want to lower the feasibility tolerance from its on-load default, with mxOption(NULL,"Feasibility tolerance",0.001).