check_singularity.lme(): errors on models with more than one level of nesting; misplaced parenthesis #937

Two small, independent things in check_singularity.lme(). performance 0.17.1
(= CRAN, = main @ 50552a0), nlme 3.1-170, R 4.6.1.

1. Errors instead of answering, for ~ 1 | A/B

nlme::getVarCov.lme() refuses more than one grouping level
(if (length(obj$groups) > 1) stop("not implemented for multiple levels of nesting")),
so check_singularity() errors out rather than returning TRUE/FALSE:

data(Pixel, package = "nlme")
fit <- nlme::lme(pixel ~ day, random = ~ 1 | Dog / Side, data = Pixel)
nlme::VarCorr(fit) # works fine
#> Dog = pdLogChol(1)
#> (Intercept) 647.2931 25.44196
#> Side = pdLogChol(1)
#> (Intercept) 218.3420 14.77640
#> Residual 233.3463 15.27568
performance::check_singularity(fit)
#> Error in getVarCov.lme(x) : not implemented for multiple levels of nesting

Single-level lme models are fine, and the lmer equivalent
(pixel ~ day + (1 | Dog/Side)) returns FALSE normally.

fit$modelStruct$reStruct looks like the natural replacement: it returns one
block per nesting level, and the blocks are already relative to sigma^2,
so they are dimensionless:

b <- as.matrix(fit$modelStruct$reStruct)
b
#> $Dog
#> (Intercept)
#> (Intercept) 2.77396
#>
#> $Side
#> (Intercept)
#> (Intercept) 0.9356995
lapply(b, function(v) as.matrix(v) * fit$sigma^2) # reproduces VarCorr exactly
#> $Dog 647.2931
#> $Side 218.342

Since this errors rather than misreports, any caller inherits the error unless
it shields itself: insight:::.compute_variances() does, wrapping the call in
.safe(); a direct caller does not.

2. abs() wraps the comparison rather than the variances

L228 of R/check_singularity.R, identical in CRAN 0.17.1 and at HEAD:

any(abs(stats::na.omit(as.numeric(diag(nlme::getVarCov(x)))) < tolerance))

The closing parenthesis of abs() sits after < tolerance, so abs() receives
the logical vector, not the variances:

v <- c(5.4150876, 0.0512695)
v < 1e-5 #> FALSE FALSE
abs(v < 1e-5) #> 0 0 <- abs() of a logical
any(abs(v < 1e-5)) #> FALSE
any(abs(v) < 1e-5) #> FALSE <- presumably intended

It is essentially harmless — abs() of a logical is a 0/1 numeric and any()
coerces back — and the two forms can only differ for a negative variance, which
getVarCov() does not return. But the intent was clearly abs(diag(...)) < tolerance,
as in the sibling methods on L206, L214 and L220. (bbolker made a related remark
about a redundant abs() in the old .merMod line in easystats/insight#878.)

The whole fix is moving that parenthesis: any(abs(...) < tolerance).