Vedic Sanskrit possesses a set of lexically-restricted devatā dvandva compounds, formally distinct from those of the later Classical language, which raise significant issues for current theories of prosodic constituency by concurrently exhibiting traits of both syntactic phrasehood and lexical wordhood. These compounds' apparently contradictory morphosyntactic and phonological behaviour thus present a prime case study for exploring the syntax-prosody interface and mapping therein. The main goal of this paper is to expand and integrate Match Theory (Selkirk, 2011) with Distributed Morphology (Halle & Marantz, 1994). Building on recent work on categorising heads (Kramer, 2016; Harðarson, 2017), this paper proposes that revising Match Word to Match Stem and thus distinguishing between acategorial roots and categorised stems allows for a unified analysis of devatā dvandva compounds without recourse to stratal constraint reordering. In addition, this paper further underlines the importance of recursive prosodic structure and maximal/minimal projections as distinct targets for prosodic mapping (Ito & Mester, 2012). Finally, this paper presents a tentative pathway for the diachronic development of these compounds into later Classical Sanskrit as involving the loss of these intermediate recursive levels.