On pg. 81 fn. 8 I noted an experiment by Joaquin Romero which showed that temporal reduction of an s was enough to induce subjects to perceive the segment as rhotic. The late eminent phonetician J. C. Catford also addressed the phonetics of rhotacism in a 2001 article. He claims that "it is generally assumed that the fact that intervocalic -z- (< -s-) came to be reinterpreted as an r implies that the /r/ of the language at that time must have been a fricative or approximant ɹ, or at least that such an r must have been an acceptable pronunciation of /r/." I'm not sure that I have ever run into this specific claim before, but, if true, it would imply that the Classical Latin trill pronunciation was at least in part an innovation.
The most interesting part of the Catford article, however, deals not with Latin, but with the phonetics of syllabic r in Sanskrit which is described in the native tradition as jihvamūla 'tongue-root', i.e velar. Catford argues that this refers to a "molar r" of the sort quite commonly found in American English. The upshot of the article is the claim that the various sorts of rhotacism found in IE languages are best explained if PIE *r was not, or not exclusively, a trill.
Incidentally Catford suggests the neologism paleophony for "the reconstruction and study of ancient pronunciation".
Catford, J.C. 2001. On Rs rhotacism and paleophony. Journal of the International Phonetic Association 31.2, 171–85.