This is not true for roots with plain voiced stops, for example Old Avestan yuxta 'yoked' from Proto-Indo-European *yug-to-, where Bartholomae's law does not apply.
While the idea of voicing affecting the whole cluster with the release feature conventionally called aspiration penetrating all the way to the end of the sequence is not entirely unthinkable, the alternative – the spread of a phonational state (but murmur rather than voice) through the whole sequence – involves one less step and therefore via Occam's razor counts as the better interpretation.
Bartholomae's law intersects with another Indic development, namely what looks like the deaspiration of aspirated stops in clusters with s: descriptively, Proto-Indo-European *leyǵʰ-si 'you lick' becomes *leyksi, whence Sanskrit lekṣi.
Even the example 'swallowed' given above contradicts the usual interpretation of devoicing and deaspiration: by such a sequence, *ghs-to gave, first, *ksto (if the process was already Indo-European) or *ksta (if Indo-Iranian in date), whence Sanskrit *kta, not gdha.
It is not completely clear what the result of a sequence *-gʰdʰ- would have been in Latin, but other evidence suggests lengthening of the vowel plus simplification, so **lēgus[citation needed] (cf.