strange combination of things causes a Segmentation fault (core dumped)

I have a very complex model which runs in a server and sends me emails when it's done using mailR library in R. It was working fine, but then for a dataset such as this (which I reduced to the minimum here) it caused R to crash with segmentation fault exit. Now, if I use half of the data, the code works, if I remove the generated quantities, it works, and if I don't load mailR it works!

So I tried to reduce the code to the bare minimum that causes the problem (the model doesn't make much sense like this, but it should work; the data is available from the link above) .
Also with this code, it crashes with Segmentation fault, unless I either remove mailR, reduce the data, or very wierdly I remove one out of the three variables that are initialized generated quantities.

So I can live without sending emails from the server, but I think this behavior is really strange, and I guess someone should take a look to see what's happening...

set.seed(123)

library(rstan)
library(mailR)



niter <- 2#1000
chains <- 1#8

load("d.Rda")


model <- "
data {
  int<lower=0> N_obs; 
  int<lower=1> subj[N_obs];    //subject id
  int<lower=1> N_subj;                 //number of subjects
  vector[N_obs] rt;
}


parameters {
  cholesky_factor_corr[2] L_u; 
  vector[2] z_u[N_subj];
  vector<lower=0> [2]  tau_u;     
  real<lower=0> alpha_mu; 
  real<lower=0> alpha_sigma; 
  real<lower=0 > alpha_psi; 
}


transformed parameters {
  matrix[2,2] Lambda_u;
  vector[2]  u[N_subj];         
  real  u_mu[N_subj];         // random intercept and slopes subj
  real  u_sigma[N_subj];         // random intercept and slopes subj

  Lambda_u <- diag_pre_multiply(tau_u,L_u);

  for (i in 1:N_subj){
     u[i] <- Lambda_u * z_u[i];
    u_mu[i] <- u[i][1];
    u_sigma[i] <- u[i][2];

 }




} 

model {
  real psi[N_obs];
  real mu[N_obs];
  real sigma[N_obs];

  alpha_mu ~ normal(0, 10);
  tau_u ~ cauchy(0,100);
  alpha_sigma ~ normal(0, 1);
  alpha_psi ~ normal(0, 250);
  L_u ~ lkj_corr_cholesky(4.0);

  for (i in 1:N_subj){
    z_u[i] ~ normal(0,1);   
  }

 for (i in 1:N_obs){
     mu[i] <- alpha_mu + u_mu[subj[i]] ; 
    sigma[i] <- alpha_sigma + u_sigma[subj[i]]; 
    psi[i] <- alpha_psi; 

  (rt[i]- psi[i])~ lognormal( mu[i], sigma[i]);
  }
}

 generated quantities {
   real A[N_obs];
  real B[N_obs];
  real C[N_obs];
  }
"
ls <- list(
        subj=as.numeric(as.character(d$subj)),
        N_obs=nrow(d),
        N_subj=length(unique(as.numeric(as.character(d$subj)))),
rt= round(d$rt,0)
)



sample <- stan(model_code=model,    
                data=ls, 
                iter=niter,

                init=0,
                chains=chains
                )

Start with the inline rstan::stan call, the supplied model string, and d.Rda, then run the minimal reproduction while varying mailR, the dataset size, and generated quantities as described. Trace the crash from the reproducible script into the RStan sampling entry point. Done means the minimal case no longer causes a segmentation fault, or the issue is narrowed to a documented component interaction.