Bibliography on this phenomenon: Poucet 1966: 140, supplemented by Burman 2018; if you have free time, you can also watch "(False) Etymology and ‘Sabine -l-‘" by Nicholas Zair - he presented it at the 14th Fachtagung of the Indogermanische Gesellschaft in Copenhagen in 2012 (Etymology and the European Lexicon; you can skip the Q&A session at the end, it's not particularly informative).
What do we have?
Indeed, a small number of Latin words have been identified (Bruno 1961 has 39) where one would expect d - based on their IE comparanda - but still encounters l. This is most irregular, based on what we know about Latin so far.
NB: A superficially similar change * -dl- > -ll is very different (it's Inlaut only and then we have a geminate l here, a very common sound change, assimilation). Sihler 1995 (section 223) writes:
*dl, *ld, *nl, *ln, *rl, *ls become L ll.
e.g. Weiss 2020 mentions sella, grallae, and pelluuiae, Sihler 1995 also has sallo. Most importantly, we have IE comparanda to justify *dl in these cases.
Ulixes is typically thought to reflect the Greek dialectal variation (Biville 1995, Weiss 2020: 504, footnote 65; cf. Sihler 1995 section 151).
How to interpret our data?
Hypothetically, one could try to lump all these words together into a Proto-Italic (or 'proto-Latin') phonetic change/law of sorts or one could deal with the words individually or in smaller groups (ad-hoc).
In my opinion, it's all about the plausibility ("the a priori probability") of one's proposal. One needs to examine IE comparanda and how much sense one's proposal makes from what we know from linguistic typology, phonetics, and historical linguistics in general and of the IE family, the Italic group, and Latin itself.
For example, which reconstruction do you find more convincing, Kümmel's PIE *ɖ > PIE *d > Italic *d or Kortland's PIE *t' > PIE *d' (except in Anatolian and Tocharian) > Italic *d?
So, we could continue interpreting this small subset of Latin words as some kind of a dialectal/historical irregularity (a current standard view) or substantially revise our models of Latin, Italic, and IE (and PIE) phonetics (I am aware of three such proposals only, Schrijnen 1914 (cf. Pulju 1995), Lehmann 1986 and Prósper 2019; thanks to Asteroides for bringing to my attention the work of Tim Pulju and Blanca María Prósper).
Lehmann 1986 argued that the d/l variation is a result of PIE d < t' (he subscribed to the glottalic theory). Lehmann himself called this change "an aberrant development." You can read about the glottalic theory e.g. in Wikipedia, if you don't know what it is, and decide for yourself.
Pulju 1995 believes that "Irregular correspondences involving /d/ and /l/ provide evidence for the reconstruction of a rare PIE cluster *dl" [emphasis mine - Alex B.] that "was phonotactically disfavored and was therefore subject to sporadic modification throughout the IE language family." The *dl reconstruction in PIE is not new at all and goes back to Schrijnen 1914. I will leave it up to you to decide whether you find this argument convincing.
A closer look at d/l Anlaut variation
I think it's extremely important to carefully examine relevant data with d/l Anlaut variation (all the etymological data is taken from de Vaan, unless stated otherwise):
- PIE *d corresponds to Latin l (no need or data to reconstruct *dl, either for Proto-Italic or PIE):
lēvir (laevir) 'husband's brother' < PIt. daiwēr < PIE deh2i-uer- 'husband's brother' (Anlaut * d > l)
Italic comparanda: not listed
IE comparanda: Skt. devár-, Pashto lēvar, Gr. δᾱήρ, Arm. taygr, Lith. díeveris / dieverìs, Latv. diẽveris, Ru. déver', SCr. djȅvēr < BSl. * daʔiu̯er-; OHG zeihhur, OE tācor 'husband's brother' < PGm. * taikura-(?).
NB: *dl reconstructions are very unusual, cf. Schrijnen 1914 *dlai̯u̯er (Pulju 1995 *dlāiwēr) His evidence? It is entirely based on one Lithuanian word, laigonas. (I didn't look it up yet myself). Obviously he has to 'explain' how the "objectionable" and "rare" *dl cluster was lost in all the IE languages.
cf. Weiss 2020: 504; also oleo and solium (NB: these two words are examples of intervocalic * d > l) - Matisoff 2013 calls this "internal ‑d‑ � ‑l‑ variation within Latin" or Zair 2012 ""synchronic -d-/-l- variation" (p. 129).
de Vaan writes s.v. odor: "Some forms show the change of intervocalic * d > l, the exact conditions of which are unclear. In this word, it seems that d was preserved in front of ō, but changed to l in front of ē. In view of fulgēre / fulgere, fervēre / fervere, it seems likely that olĕre is older than olēre (cf. Leumann)."
- PIE *dl corresponds to Latin l (via a putative reconstructed *dl stage in Proto-Italic):
longus < PIt. *(d)longo- < PIE *dlongh-o-, the only (?) Anlaut example, cf. Weiss 2020: 178 (Anlaut * dl > l)
Italic comparanda: not listed - how robust is this Proto-Italic reconstruction that is based on one Latin word only? (that being said, take a look at its PIE comparanda below, esp. Skt., OAv. vs. OCS, Greek vs. Hittite)
IE comparanda: Skt. dīrghá-, OAv. darəga-, OCS dlъgъ, SCr. dȕg, Lith. ìlgas, Alb. gjatë 'long' (+ *-tā-); Gr. δολιχός; Hit. * taluki- / talugai- [adj.] 'long'; Hit. zaluknu-zi 'to lengthen' etc.
Phonetic change: total contact regressive assimilation
- PIE *d corresponds to Lat. l (d?)
two Latin words with solid IE cognates and an unclear development in Latin/Italic, lingua and lacruma
lacruma (dacruma - three attestations only, all by the grammarians only - Mar. Vict., Keil 6.26.2) (Anlaut l/d variation)
lacruma < PIt. *d(r)(k)akrunā-? < PIE *drḱ-h2(e)ḱru- 'eye-bitter' (Kortlandt 1985b)
Italic comparanda: not listed
IE comparanda: Skt. áśru-, YAv. asrū [pl.], Gr. δάκρυ, Arm. artasuk' [pl.], Lith. ãšara, OHG zahar, ToB akrūna [obl.pl.] 'tear'; OIr. dér, W. deigr 'tear', Hit. išhahru- [n.] 'tear(s), weeping'
Beekes: "A Hell. form *δάκρῠμα is often assumed to be the source of Lat. dacrŭma, lacrĭma, but see now the discussion in De Vaan 2008 s.v. dacruma ."
Attestations of δάκρῡμα -ατος, τό: Beekes: δάκρῡμα 'mourned for' (Orac. apud Hdt. 7, 169), 'tear' (A.): cf. Montanari (BDAG): tear, lament Aeschl. Pers. 134; Eur. Andr. 92 ( pl. ); cause of weeping: Opp. ( Hdt. 7.169.2) .
cf. pl. δάκρη (δάκρῠ)
Re: δάκρῡμα < *δάκρῠμα in Greek, see Biville 1995: 28-29, 128
- a Greek loan borrowed via another Italic language(s), traditionally explained via "Sabine", starting with Conway 1893, who used the elimination principle; cf. lepesta(e) in Biville "Le passage du grec δ- au l- latin serait un sabinisme", II: 100 or I: 74, ft. 21)
What is "Sabine" exactly? Options galore. 1. A Sabine dialect of Latin; 2. the Sabine language; 3. some other Sabellian (or Sabellic) language, e.g. Oscan (e.g. Sihler 1995)
BUT Weiss 2020: 504 is against the attribution of this etymological irregularity in Latin to Sabine (he calls it a misnomer, see footnote 63 re "the Sabine l theory" or Poucet 1966 calls the Sabine l theory unacceptable ('inacceptable'), a linguistic myth.
- Poucet 1966 mentions Bottiglioni 1943, who reported a similar change/variation in Greek, Etruscan, Sanskrit, Iranian etc., and de Rop 1958 even in the Bantu languages.
Kümmel 2012 also writes that "For non-obstruent stops, it is well known that they sometimes alternate with other non-obstruents, i.e., nasals or liquids (cf. Haider 1983: 86; Stewart 1989: 239f.; Clements & Rialland 2005: 18)" (p. 304) and "Therefore, sporadic alternation of PIE “mediae” with such sounds might reﬂect this state of affairs e. g., *d ~ *l in Hitt. dā- vs. Luwian lā- ‘to take’."
lingua (dingua: one attestation only, Mar. Vict., Keil 6.26.2) (Anlaut l/d variation)
lingua < PIt. *dn̥χ(u)wā- < PIE *dnǵh-uh2 'tongue'
Italic comparanda: O. fangvam [acc.sg.], fancua [nom.pl.] 'tongue' < * fənχu̯ā- < * dh -
IE comparanda: OIr. tengae, MW tafawt 'tongue' < * tnǵh -, Skt. jihvā- 'id.', Av. hizuuā-, Arm. lezow, OPr. insuwis, Lith. liežuvìs, OCS językъ, Go. tuggo, OHG zunga, OIc. tunga < * dnǵh -, ToA käntu, ToB kantwo 'tongue' << * tänkwo.
Comments: How reasonable would it be to posit a *dl reconstruction for Latin lingua taking into account its IE comparanda?
The most promising approach imho is best summarized by Burman 2018:
"Even if the theory of Sabine l is no longer considered credible, modern scholarship still struggles to find a definite explanation of a handful of examples of Latin /l/ where /d/ is expected. A possible reason for our inability to find a unified answer for d/l variation in Latin is that there may not be one. Lingua has been given a plausible individual explanation through analogy to lingere. The solutions to the problem of lacrima, oleo, levir, solium and malus may be easier to find if we approach the words individually instead of as examples of the same change" (p. 58).