An inventory of word endings will not in itself supply an argument for OM syllabification, but it is useful preparation for one. Superficially, there seems to be little restriction on the inventory of consonants that can appear at the end of a word: I find h., m., k, g, n", c, j, n~, t., d., n., t, d, n, p, b, m, y, r, l, v, s", s., s. The only sounds obviously missing are the aspirated stops and h. As for clusters, r is found preceding final k, g, t, d; and the disarticulated nasal (perhaps to be considered a nasalisation of the preceding vowel) precedes sibilants, h. and r. But Sanskrit is famous for its system of sandhi, where word-final sounds undergo neutralizations and assimilations, which are represented in the text. So for example any of the final c, j, t, d, n, l may actually come from any of final t, th, d, dh, depending on the initial segment of the next word.
As for the neutralizations, obstruents lose all inherent laryngeal features of voice and aspiration at the end of a word (e.g., t, th, d, dh all become t), palatals become either velar or retroflex (i.e., a [distributed] feature is lost: c, ch, j -> k or t.), and fricatives other than s lose their [cont] feature, becoming stops (s., s" -> k or t.). It is important to note that this neutralization cannot be purely phonetic in nature: j becomes k in some words, t. in others. This has led Cho (1990:78-79) to argue that there are really two different j, a conclusion for which there is an historical basis but no other synchronic evidence: all j are otherwise indistinguishable. Furthermore s" and h also share this property, and following Cho's approach would require setting up three artificial phonemes to account for all the alternations phonologically. I would prefer to merely recognize that whereas the phonology constrains the set of segments that can appear word-finally, the lexeme itself has a role in determining exactly which segment is chosen. Thus neutralization cannot be left to some utterance-level postlexical process, but must take place at some level where the lexical information in stems is accessible.
It would appear that after the neutralizations, the last consonant of a word must be drawn from the set k, t., t, p, m, y, r, v, s. As for clusters, ns is very frequent, n"s tolerably frequent, and rt and rk rather rare (ns itself is in part a reduction of nts, and n"s comes from underlying nks.) It should be remembered that this inventory is somewhat abstract, because the actual surface forms will be conditioned to a large extent by assimilation to the next segment, and in pausa further neutralizations occur. These are the subject of the next section.