I will create an issue, but let me first better understand what is going on. I have one observation per subject, so I never thought of using (stan)_glmer with (1|subject_id). When I try this with glmer/lme4, Doug yells at me
In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv, :
Model failed to converge with max|grad| = 0.0167366 (tol = 0.002, component 1)
and I do fully understand his anger, since there simply is no variance material to hammer upon. stan_glmer works, but I am not surprised that the (1|subject_id) terms ends up last in varsel; the results otherwise look ok. So it is unclear why you did not think someone would use _glm for this. Is using glmer with a orphaned (1|subject_id) the better choice?
The example shown in the original posting looks like this with stan_glmer, which is much more reasonable.
size solution_terms elpd elpd.se
2 0 <NA> -104 2.1
3 1 op_group -95 4.4
4 2 reflux_preop -92 4.8
5 3 z_weight_preop -91 5.0
6 4 op_group:z_weight_preop -88 5.3
7 5 z_age_op -88 5.4
8 6 op_group:z_age_op -84 5.6
9 7 op_group:reflux_preop -84 5.6
10 8 (1 | subject_id) -84 6.4