So far I have been giving a fairly traditional account of sandhi, which describes word-finalAllen 1962:25-26). One might expect a neutralization to become more likely at the end of higher prosodic categories, but for an assimilation one would expect the opposite: if it occurs across word boundaries, the unmarked case would be to expect it to occur within word boundaries, as well. Kaisse (1985) also found such rampant word-juncture effects problematic for her theory of sandhi, under which such effects should be lexically or syntactically conditioned. Secondly, many of the neutralizations discussed above, as well as the assimilatory changes called for in the external sandhi chart, do occur word-internally. This is particularly striking when the same process appears to have a disjunctive condition. Obstruents, for example, assimilate in voice to a following obstruent regardless of position (mahat- `big' plus -bhis instr. pl. -> mahádbhih., RV 1.72.9). But before a sonorant, they do so only at the end of a word (as opposed to word-internal mahatâ' instr. sg., RV 1.32.5). Similarly m assimilates fully to stops everywhere, but to glides and l only at word boundaries.
Whitney already pointed out (1889:38) that the major difference between rules of internal and external sandhi, if one abstracts away from the fact that in word juncture one must account for consonant combinations that accidentally do not occur in the morphology, is that sonorants voice preceding obstruents at word boundaries, but not word-internally. This is a strong hint that syllable structure is involved, since one expects obstruents to be able to form onset clusters with sonorants. Rice (1990:307-311) drew the correct conclusion that voicing assimilation is a matter of syllable structure. Obstruents assimilate in voice whenever they find themselves in a coda, and all consonants at the end of a word are necessarily in a coda. Within a word, however, consonants will syllabify with a following sonorant, ending up in an onset, which does not condition the assimilation. Kessler (1993) later demonstrated that for Classical Sanskrit(5) this analysis is appropriate for virtually all sandhi processes that seem to be conditioned by word boundaries.
So one clear advantage of the OM syllabification is that it allows rules of neutralization and assimilation to be stated more naturally. Under this theory, words syllabify according to a principle of maximum onset (Pulgram 1970:47-57, Kahn 1976:42-46, Clements 1990:299), except where morphology adds material to the left (prefixes or compounding stems). Whenever a consonant cannot go into an onset, either because the onset templates forbid a particular cluster or because syllabification does not occur across the pertinent domain (after left-appended material or at the end of words), then it is subject to phonological neutralizations that are in part lexically determined, then to feature-sharing with the next segment. This is virtually a textbook example of universal patterns of consonant assimilation (Cho 1990).
It might not be amiss to give one or two examples of this working word-internally. This table gives the singular and plural of an adjective that ends in a consonant; nouns are essentially identical. It is difficult to find a word that takes all possible inflections with the text of the R.gveda; the example used here is fleshed out with starred forms supplied by analogy (Macdonell 1916:53, Whitney 1889:146-147).
Here one sees that underlying j is preserved as such when it can be in an onset (before a vowel), but is neutralized to a velar and is subject to voicing assimilation when it is stranded in a coda. Verbs follow the same pattern. Grassmann (1875:1115-1118) lists (among others) the following forms for the root yu(n(a))j- `yoke': yunájmi, yun~janti, yunájat, yujyâva, ayujmahi, yuyujre, versus yunaks.i, yunákta, yun"kté, yun"ks.vá, yun"gdhvám, ayukthâs, ayugdhvam, áyuks.âtâm. These alternations are easily accounted for by saying that j can syllabify with a following vowel, m, r, or y; but before s, t, th and dh it must stay in the coda and undergo neutralization to k and subsequent assimilations. This agrees with the fact that stops are found word-initially before sonorants but not before obstruents.
A good deal of information about syllabification can be gleaned from sandhi behaviour itself, once the syllabic theory of sandhi is accepted. The most convincing evidence comes from those situations where adjacent segments give different results word-internally and between words, since this directly indicates that those segments can form an onset sequence, protecting the first element from coda-coditioned changes. I have already mentioned how word-final obstruents lose their own laryngeal features and link to the [voice] feature of the next segment, but do not assimilate to sonorants in the same word, e.g., mahatâ' `big' (RV 7.104.16), satyáh. `true' (RV 1.145.5), s"akráh. `mighty' (RV 7.104.20), asmi `am', RV 1.164.37. Similarly, stops fail to nasalize before word-internal nasals: rátnam `gift' (RV 1.41.6). So it is clear that obstruents syllabify with following sonorants, exactly as seen in word-initial sequences.
Labial sonorants appear to have a special inclination to head onset clusters. Since m can occur in Sanskrit words before n, y, r, l, and v, before which it disarticulates and/or assimilates word-finally, it must form onset clusters with those sounds in e.g. mraks.akr.'tvâ `rubbing to pieces' (RV 8.61.10). Similarly, v must be syllabified with n. (grâ'vn.âm `pressing stones', RV 10.85.4), because âv regularly become au before a consonant at the end of a word (not shown in the sandhi chart). The segment v must also be syllabified with y (suhavyám `fit for invocation', RV 1.74.5), since av becomes o word-finally before y. One might generalize that labial sonorants syllabify with (non-labial?) sonorants.