mcmc_pairs sometimes plots wrong number of divergences

[jgabry]

First reported on the Stan forum by user @gregd:

https://discourse.mc-stan.org/t/bayesplot-mcmc-pairs-divergences-error/41457

Here's the code that was given to demonstrate the problem.

library(rstan)
rstan_options(auto_write = TRUE)
library(brms)
library(cmdstanr)
options(mc.cores=2, brms.normalize=TRUE,brms.backend="cmdstanr")
library(bayesplot)

set.seed(123)
x<-runif(100)
y<-sin(12*x)*exp(-3*x)+rnorm(100,0,0.3)
plot(x,y)
y[1:5]<-runif(5,-1.5,-1)
plot(x,y)
points(x[1:5],y[1:5],pch=16,col=2)

bf1<-bf(y~s(x,k=7))
dat<-data.frame(y,x)

m1<-brm(bf1,dat,init=0,control=list(seed=1591897794))
#init and seed for reproducibility
#29 div

np<-nuts_params(m1)
dx<-which(np$Par=="divergent__"&np$Val==1)
length(dx)
#29
np[dx,]
#25 in C2, 4 in C4

vars<-variables(m1)
vars

pc<-pairs_condition(chains=list(1:2,3:4))
pc

mcmc_pairs(m1,pars=vars[2:3],np=np, np_style=pairs_style_np(div_shape=16,div_size=2),condition=pc)
#shows 8, with 4 in each plot
#should show 25 in top right, 4 in bottom left

np[dx[26:29],]
#the 4 in C4
np[dx[26:29],2]

arr<-as.array(m1)
dim(arr)
arr[np[dx[26:29],2],4,2:3]
#the bottom left plot is correct

arr[np[dx[1],2],2,2:3]
#not shown on top right plot
#others also not shown

#same happens with rstan backend