> As a general principle, differences in undoubtedly reconstructed PIE
> proto-forms should reflect Pre-PIE (say, 'Nostratic' or the like)
> diffractions secondarily converging into PIE allotropic variants
That statement is way too strong to make into a "general principle".
Variation must have existed within PIE, but to reduce all variation among IE 'lects into variation that already existed at the proto-language level is equivalent to denying the existence of loanwords (both family-external and family-internal), analogical changes (which are particularly frequent in numerals), affective variation, and any number of other mechanisms that can introduce phonetic distortions into a cognate set. Proto-variation is not somehow a categorically preferrable explanation. In fact it's not an explanation at *all*, it just pushes the question backwards in time.
If you can reconstruct the pre-PIE "diffractions" (and make them independently motivated and not simply circularly assumed) that would lie behind some particular set of variations, then yeah that'll be a good case to consider the variety inherited though.
Here analogical explanations are indeed thinkable. OTTOMH, e.g. Sanskrit _s.as._ "6" might be due to the influence of the same rhyme appearing in _as.t.a:_ "8". Or at an older date, perhaps *w- is original in "6", and all or most examples of *sw- or *s- represent the influence of "7"?
_j.