pymer4's `model.ranef_var` reports standard deviations (`sd__(Intercept)`,
`sd__Observation`), not variances, so the ICC variance ratios must square them.
The previously released values were SD ratios.
Main-study ICCs (observed k = 9.2 audio, 8.9 text):
ICC(2,1) audio .312 [.229, .378] text .241 [.170, .299] p = .093
ICC(2,k) audio .806 [.732, .848] text .738 [.646, .792] p = .120
superseding .333/.821 (audio) and .295/.788 (text) with p = .127/.113.
Bootstrap CIs (2,000 stimulus resamples) and permutation p-values (2,000
iterations) were recomputed on the corrected scale; Krippendorff's alpha is
unaffected and unchanged.
Note that `analysis/common/_lme4.py`'s `lmer_varcomp` returns VarCorr's `vcov`,
i.e. variances already, and is correctly left unsquared.
