Michael, I was dismal yesterday but I’ve never been a rancorous person. I am from my part alacritous to return to the question at hand.
As far as I can see, this problem revolves around three words and how and on what grounds they operate: hiatus, elision, and sandhi. Also the register makes a difference. Poetry must be distinguished from prose, but rhetoric bridges this gap to some extent so that it is all one continuum. The phonetic or morpho-phonological “tools” that are used must originate in the vernacular language.
However, this must not be understood too strictly. In some of its characteristics, poetry has over a long period of time developed its distinct elevated style (there are some reconstructions to e.g. Proto-Indo-European), and in case of Latin it adopted much via extensive literary borrowing from Greek. Poetry preserved and continually took material from the spoken language different from literary Latin prose which is most often learnt even today. The ancient grammarians and rhetoricians (e.g. Cicero, Quintilian) mention here and there something, but what they say is quite impressionistic and often conflicting. This isn’t surprising as differences of opinion must have existed as well as morpho-phonological development.
Now to return to those three words, there’s the famous and oft-repeated rule uocalis ante uocalem corripitur. This is generally considered word-medially, but the question arises how big or possibly how small the difference was with sandhi, especially in fluent, continuous speech and poetry, of course. In principle I see no opposition that this rule could not have worked in sandhi, too, to some extent. Then again, there’s surely no reason to expect quite as wide-ranging a sandhi in Latin as for instance (quite emblematically) in Sanskrit (which in Sanskrit is always written).
Allen discusses the problem of “vowel junction” in the 4th chapter of Vox Latina, and now that I read it again, it seems quite well-balanced even if it doesn’t claim to solve everything. In the end Allen is inclined to assume synizesis and glide-formation of -ī and -ū (Allen discusses dactylic hexameter). I’d say that if true, the correption might indeed work also on hiatus in word-boundaries. In general, reduction at the end of the word isn’t very uncommon in languages. As an example of synizesis Allen mentions ōdī et amō [oːdi̯etamoː].
This isn’t quite obviously enough, as it works only where the preceding syllable is heavy and only with -ī and -ū. But Allen is reasonably solid in mentioning that words ending in -ō, -ē, and -ā are often common adverbs and conjunctions (e.g. ergō, certē, contrā) and total elision might be expected here. Even this, however, doesn’t solve duplicī aptantur. I don’t know if all possible statistical analyses have already been made on Latin hexameter. Conjecturally at least I’d say that this is somewhat rare occurrence, an “extension” to the principle put forth by Allen. This poses furthermore the question, why such a consummate master as Vergil yielded to such a line (there must be a reason for it). But Michael noted the conjectural rarity of this and already posed the question why Vergil created the line straightaway.
I am given to understand that Luciano Canepari has something to say about Latin sandhi, but I haven’t seen his work and don’t know how much he has to add (La pronuncia neutra, internazionale del latino classico / Manuale di pronuncia italiana 2004).
Finally: why does prodelision occur (potentially) only with es and est? I have no decisive answer and can only offer general assumptions. This would seem to be quite an old phenomenon, and many Indo-European languages have reduction in the copula (in French and Persian the 3rd singular it’s often only [ε] or [e]). I’m not saying, though, that it’s inherited from Proto-Indo-European. The Oscan example víu teremnatust (via terminatast) would suggest that it’s as old as Italic or at least spread areally amongst the closely related languages. I don’t know if it originated when es(t) followed a word ending in -s or in more general contexts (it could be unnecessary to try to trace prodelision in cases like uocitatust 166 BCE [Allen 123]). Couldn’t so unstressed yet so common a word have exceptionally lost its vowel from the beginning of the word?