I wrote my MA thesis back in 2010 on a possible origin of ah > o sandhi, from a language typology perspective. I am not certain that I still believe it---but the premise in short was that if we took Prātiśākhya authors seriously, then there were ways of pronouncing sandhi h as a velar fricative before a k(h) and a bilabial fricative before a p(h). So aśvax kasya for instance and aśvaf pitur. If so, then sandhi h would be assimilating to more than just the dental, palatal, and retroflex places of articulation. What if something like that occurred prior to a voiced stop? We suspect this might be the case cause of sandhi h > r, which must have gone through some parallel process to retroflex s, agniṣ, but voiced as a pre-Vedic *agniẓ. When voiced frication became impermissible in Indic, this sound became the nearest place-appropriate approximate: r. What if this happened with voiced fricatives of the other places of articulation? In my theory *manas+bhih > manav+bhih > manobhih and similarly Vedic hapax sure duhitar would come from suras duhitar > suraz duhitar > suray duhitar > sure duhitar. Similarly 3rd pl perfect sedur can be explained as sazduh with as > az > ay > e before d. 

The sandhi ah > o then is simply a leveling of the sandhi outcome before labials, while the ah > e was eliminated analogically because so many inflectional endings in Sanskrit are already marked by -e (middles, locatives, datives, etc.) while final o not from sandhi is incredibly rare (u-stem voc). The -e ending would have survived in eastern dialects like Magadhi where it remains as the nom sg. of a-stems.

So why do I no longer fully trust my thesis? Well it is a big theory from a paucity of evidence and although it was a nice puzzle solving exercise, there are likely more parsimonious solutions. For instance, what if sandhi outcomes were conditioned by word-boundaries? My theory treats internal and external sandhi as essentially the same, but this is an unmotivated assumption on my part.

Best,

Caley