back-incompatibility ¿caused by factor handling improvements?

[bbolker]

Results differ between 1.1.27.1 and HEAD.

https://stackoverflow.com/questions/71340249/mismatching-results-for-singular-fit-with-different-r-lme4-versions

1.1.27.1:

fit2 <- lmer(total.fruits~(1|reg)+(1|reg:popu)+(1|reg:popu:amd)+
            (1|reg:popu:amd:status),data=Arabidopsis,control=lmerControl(optimizer="bobyqa"))
dput(getME(fit2, "theta"))
t1 <- c(`reg:popu:amd:status.(Intercept)` = 0.150979711638631, `reg:popu:amd.(Intercept)` = 0,
`reg:popu.(Intercept)` = 0.189968995915902, `reg.(Intercept)` = 0.260818869156072
)

same code, HEAD/devel:

t2 <- c(`reg:popu:amd:status.(Intercept)` = 0.15097974334854, `reg:popu:amd.(Intercept)` = 0,
`reg:popu.(Intercept)` = 0.189969036985684, `reg.(Intercept)` = 0.260818797487214
)

see e259823 : this (presumably) modifies the ordering of factors/b-vector/Z-matrix columns, leading to imperfect backward compatibility