The rise of gemination in Celtic

This study investigates systematically the emergence and establishment of geminate consonants as a phonological class in the Celtic branch of Indo-European. The approach of this study is comparative historical linguistics, drawing on diachronic structuralism combined with aspects of language contact studies and functional approaches to language usage. This study traces the development of geminates from Proto-Indo-European (fourth millennium B.C.), which did not allow geminate consonants, to the Common Celtic period (first millennium B.C.), when almost every consonant could occur as a singleton or as a geminate, and on to the earliest attested stages of the Insular Celtic languages (first millennium A.D.). Although they were prominent in the phonology of Proto- and Ancient Celtic (Gaulish, Celtiberian), ultimately geminates were gotten rid of as a phonological class in the individual Insular Celtic languages. This is probably due to the fact that the contrast between lenited and unlenited sounds took on a central role in Insular Celtic phonology, making gemination a phonetically redundant category. Most instances of geminate consonants in Celtic can be explained by regular sound change operating on inherited clusters of consonants. Each sound change will be discussed in a separate section in a rough chronological order. Effectively, gemination is largely a strategy to reduce the number of allowed consonant combinations. To a limited degree, gemination also had a morphological function, especially in the formation of personal names and in the creation of adjectival neologisms. However, there is a residue of words, especially nouns, in the Insular Celtic languages that defy any attempt at etymologising. They are prime suspects of having been borrowed from prehistoric, substratal languages.


Preliminaries
This study investigates systematically the emergence and establishment of geminate consonants as a phonological class in the Celtic branch of Indo-European. The aim is to come closer to a general theory of this phenomenon in Celtic, in order to permit inferences about the etymology of words that contain geminate consonants in the attested Celtic languages. Geminate sounds have been commonly used as an argument for identifying borrowings from unknown substrate languages. However, before their precise diachronic and synchronic status within the phonological system of Proto-Celtic and the individual languages hasn't been determined, conclusions about layers of loanwords based on geminates are circular.
The approach of this study is diachronic structuralism combined with aspects of language contact studies and of functional approaches to language usage. The working hypothesis is that the inexorable, albeit gradual, rise of geminate consonants as a phonological class across all modes of articulation in the older stages of Celtic was a multicausal process that was fed by a combination of language-internal developments in Celtic and of language-external factors. The focus of this article is on the period up to the beginning of the 1 st millennium a.d. Evidently younger sound changes that occurred in the documented histories of Irish or British and that created further instances of geminates or, at that stage, fortes sounds (for example, assimilations that postdate syncope such as nl > ll in OIr. (Old Irish) tenlach > tellach 'hearth', ld > ll in acaldam > acallam 'conversation, dialogue', or MacNeill's Law in Irish, or provection after syncope in British, or other late sound laws) are not treated here. Reference to them will only be made when they are relevant to clarify points in the prehistoric developments.
The terminology in this article follows the traditional practice in Celtic historical phonology. The term 'geminate consonant' will be used synonymously with 'long consonant'. Geminate consonants are written double in reconstructions, e.g., *ballo-, which is meant as equivalent to a phonetic analysis as [balːo-]. The main contrast between the two classes of Celtic stops is considered to be between 'voiced' (= D) and 'unvoiced/ voiceless' (= T) consonants. Phonetically, the contrast between the D-series and the T-series may rather have been that between 'lenis' voiceless consonants and 'fortis' aspirated voiceless consonants (cf. Stifter, 2017Stifter, : 1191similar Van Sluis, 2019: 3-36;sceptical Martinet, 1952: 201) or, in Eska's (2018) framework of Laryngeal Realism, a contrast in the feature [spread glottis]. Since the sound changes in this article are understood as arithmetic abstractions, notations of reconstructions can usually be transferred easily between alternative frames of references. The phonetic details of these variant descriptions do not make a practical difference here. In addition, it is assumed that intervocalic voiced stops developed non-contrastive, lenited, i.e., fricative allophones already at a very early stage (cf. Schrijver, 2016: 497-499;Stifter, 2017Stifter, : 1189Stifter, -1190. These fricative allophones are not indicated in the reconstructions. The label 'PIE' (Proto-Indo-European) will be reserved for reconstructions that can be securely set up for the protolanguage, while 'qPIE' (quasi-Proto-Indo-European) is an umbrella label for any voreinzelsprachlich, pre-Celtic reconstruction that may have arisen in the long period between the break-up of Proto-Indo-European and the emergence of Proto-Celtic (PC), for instance during an extended Western Indo-European period. qPIE reconstructions will be given in standard PIE phonology, even where this may be anachronistic. Mutatis mutandis, similar considerations apply to the label 'qPC' (quasi-Proto-Celtic).
Examples are cited from the standard lexicographic collections. As a rule, no special reference to these sources is made and etymologies are only discussed in cases where it is necessary. The reference points for Proto-Indo-European are Lexikon der Indogermanischen Verben (LIV = Kümmel & Rix, 2001), Nomina im Indogermanischen Lexikon (NIL = Wodtko, Irslinger & Schneider 2008) and Indogermanisches Etymologisches Wörterbuch (IEW = Pokorny, 1959). For Proto-Celtic, the standard handbooks are Matasović's Etymological Dictionary of Proto-Celtic (2009) and Schumacher's (2004) lexicon of primary verbs, Die keltischen Primärverben. Old Irish words are cited from the electronic Dictionary of the Irish Language (eDIL = Toner et al., 2019; albeit with occasional spelling normalisation to an idealised Old Irish standard; for criticism of the inconsistent spelling of headwords in eDIL see Griffith et al., 2018: 7), Welsh and Cornish from Geiriadur Prifysgol Cymru (GPC Online), and Breton from Favereau (2016). Gaulish data is taken from Delamarre (2003) and from Delamarre (2007) for personal names; Celtiberian data from Jordán Cólera (2019). Lepontic and Cisalpine Celtic examples are taken from Lexicon Leponticum . Ancient Celtic personal names are cited either as abstract stems or in an idealised nominative singular, even if that form is unattested. Since this article is concerned with the predesinential parts of words, this procedure has no consequence for the main argument. Due to the nature of the sources, the semantics of ancient Celtic cognates is occasionally uncertain. The position adopted here is that the choice of ancient cognates is guided by plausibility, i.e., formal phonological and morphological correspondence with words in other languages. If new evidence should emerge that casts doubt on these equations, the items would have to be removed.
This study consists of two big conceptual parts. Chapters 2-8 are devoted to etymological gemination, i.e., types of gemination that arose by regular sound change operating on inherited words. The sections within this first part are arranged in an approximate, but not strict, chronological order. The principle of thematic coherence occasionally overrides chronology. For instance, morphological processes such as gemination arising from inflection, derivation and compounding are treated in single sections, even if they may relate to extended periods of time. The second part of the study is concerned with non-etymological gemination. 'Non-etymological' means that such types or instances of gemination cannot be described by regular sound laws, but they arose non-predictably from pragmatic contexts such as addressing, through sound-symbolic neologisms, or are due to borrowing.

Previous research
The first dedicated study of geminates in a Celtic language was in response to an observation in Germanic linguistics, namely that obstruents before *n seem to have become geminates in Proto-Germanic. This is known as 'Kluge's Law' (proposed by Kluge, 1884; for a modern take, see Kroonen, 2013: xxxiv-xxxv). Stokes (1891; revised in 1891-3) took his inspiration from this recently postulated Germanic rule and tried to apply it to Old Irish. It is evident from Stokes' discussion that he had no clear understanding of the phonetic reality of Old Irish double consonant spellings such as cc, and many of his explanations no longer stand up to a modern understanding of Celtic historical phonology. From an early date, the notion of the operation in Celtic of a sound change comparable to Kluge's Law was met with scepticism (e.g., GOI (Grammar of Old Irish = Thurneysen, 1946) 92-93;Sjoestedt, 1926: 19-20;Martinet, 1952: 197-198).
Zupitza (1900) meant progress over Stokes' attempts insofar as he also recognised other sources for Irish geminates. Building on a very thorough study of spellings in the major Old and Middle Irish sources, he rightly identified many instances of geminates as the products of assimilation across morphemic boundaries, but he also saw a major source of geminate stops in clusters of obstruents + *n. Like Stokes, many of his etymologies have been rejected in the meantime. In a similar vein, Pedersen (VKG (Vergleichende keltische Grammatik) i 158-161, 476-477) distinguishes between double consonants arising from assmilation in word formation, while also being inclined to operate with a Celtic version of Kluge's Law. In addition, he recognises doubling of consonants in terms of endearment. Sommerfelt (1954) considered substratal influence responsible for establishing the contrast between short and long consonants in Irish, as well as that between non-palatalised and palatalised consonants. Kuryłowicz (1957) adopted a very different strategy in order to explain those instances of gemination in Celtic and Germanic that are not caused by assimilation. Critical of the notion of Kluge's Law, he saw a different, single principle behind the phenomenon. Starting from a core where pairs of simple vs. geminate consonants were the product of regular sound change (l : ll, n : nn), Kuryłowicz thinks that this opposition was extended as a morphological marker to other sounds as well.
New life was breathed into the hypothesis of the geminating force of *n by Lühr, 1985, who adduced further instances of geminate occlusives as evidence for that rule. She had to explain the numerous counterexamples through analogy (Lühr, 1985: 345). In addition, Lühr's hypothesis allows for geminates that are due to expressivity. The idea of a Celtic version of Kluge's Law was also advocated by Bammesberger, 1998. Like a century before, most scholars remained sceptical to the usefulness of such a law in Celtic.
De Bernardo Stempel (1999: 508-522) provides a useful list of Old Irish words with gemination. She distinguishes between inherited geminates, i.e., those geminates that arose through regular sound change, and other cases, which she largely ascribes to expressiveness. Her discussion also includes a detailed refutation of Lühr (1985). In a 2010 article, De Bernardo Stempel turns her attention to geminates in the ancient Continental Celtic languages, in particular to the possibility of gemination resulting from accentuation. Like in the case of Old Irish, she distinguishes between etymological and non-etymological gemination. This useful dichotomy between etymological and non-etymological gemination will inform the present article, too. McCone (2005: 406-407) touches briefly on the subject of geminates in the context of substratal influence on Insular Celtic. His remarks are specifically directed at word-initial geminates and their role in the emergence of initial mutations. He maintains that, except perhaps for rr, ll, nn, the status of geminated initial consonants in the historically attested Insular Celtic languages is dubious, and he downplays their significance as an indicator of substrate influence.
The status of gemination in the phonology and phonetics of the Insular Celtic languages and the part it played, or did not play, in establishing the system of initial mutations, especially in British Celtic, was the topic of a heated debate that extended for more than five decades, from the middle of the 20 th until the early 21 st century (see the more detailed summary in Van Sluis, 2019: 15-24). The debate grew out of an exchange between Jackson (1953: 473-480, 545-548, 565-573, 634-638;1960;1967: 307-308, 317-323) and 1966); cf. also Kortlandt (1982), Feuth (1983), Koch (1989). A satisfactory solution to the problem was proposed by Harvey (1984) and Russell (1985) that removed the notion of gemination as a genuine mutation in the history of Irish and British. Afterwards, the debate shifted away from gemination towards the related question of the relationship between the spirant mutation and nasalisation in the British Celtic languages, with notable contributions by Thomas (1990), Sims-Williams (1990, McCone (1996;92-96), Schrijver (1999), and Isaac (2004;.
Although it can be seen from this brief history of research that many scholars have contributed to the study of Celtic gemination or to aspects of it in the past, a comprehensive modern account of the phenomenon in a diachronic perspective is lacking. The following study aims at filling this gap, by conducting a systematic analysis of all the phonological sources and internal and external contexts in which geminate sounds arose in the early history of the Celtic languages.

Proto-Indo-European and Pre-Celtic
Celtic geminate stops cannot have been inherited from Proto-Indo-European (PIE), since the Indo-European protolanguage famously did not have geminate consonants as a phonological class. This prohibition went so far that even accidental geminates across the word boundary were prone to simplification. It is believed that the so-called 'mobile s' of Indo-European arose in that way (Mayrhofer, 1986: 120-121). However, this is not the whole story, because geminate sounds were not entirely absent from some registers. A handful of strongly affective words with geminates can be projected back to the protolanguage. Celtic languages inherited some of these words, again for use in strongly affective contexts.
After some far-reaching simplifications of the phoneme inventory, especially the merger of the Indo-European velar and palatal series of stops and the merger of the mediae and mediae aspiratae (Schrijver, 2016: 497-499;Stifter, 2017Stifter, : 1189Stifter, -1190 and the loss of laryngeals, the Pre-Celtic system of consonants was reduced to a relatively simple 15-phoneme inventory that probably did not yet distinguish geminates as a phonological category. All sound changes discussed in the following chapters operate on the basis of the post-PIE, pre-Proto-Celtic sound inventory in Table 1.
Marginal phonemes, i.e., phonemes that occur only in very restricted contexts and which are allophones of core sounds of the system, are put in brackets here and in later sections. In keeping with the traditional practice in comparative Indo-European and Celtic studies, the phonemes /j/ and /w/ will be represented by the symbols i̯ and u̯ in the main part of this article.
From their virtual lack in the ancestral stages of the language, it follows that geminates must have been acquired as a phonological class after the break-up of the Proto-Indo-European parent language. It will become clear from this study that the emergence of geminates occurred gradually. There are around twenty regular, phonological sources for geminate consonants of Proto-and Common Celtic origin, plus one that is exclusive to Irish. In addition, a number of non-phonological sources (sound symbolism, loans) will be identified.

Etymological gemination: Proto-Celtic and Common Celtic
The phonetic class among which geminates started to arise earliest in Celtic are resonants. The starting point are several inherited consonantal clusters of common occurrence. They provided the input for assimilatory processes that created long, i.e., geminate resonants either already in Proto-Celtic or so early after the split into individual branches that the changes spread across the entire speech community, leading to identical outcomes everywhere.

*ln > *ll
The assimilation of *ln > *ll is carried through in all attested Celtic languages and can therefore be securely assigned to Proto-Celtic, bolstered by a large number of convincing etymologies. The input for this change is predominantly morphological, i.e., nominal or verbal derivatives where suffixes with n-attached to roots ending in -l.
Other possible examples are uncertain.
Finally, two important morphological processes in ancient Celtic languages must be discussed separately. One of them may have arisen, and the other certainly arose as a consequence of the assimilation *ln > ll.
In Celtiberian, this sound change resulted in a productive inflectional and derivational pattern (Jordán Cólera, 2019: 611-614). The Celtic suffix -on-, particularly common with an individualising function among personal names, inflected originally with full paradigmatic ablaut, i.e., lengthened grade *-ū < *-ō(n) in the nominative and full grade *-on-and zero grade *-n-in other cases. In the majority of cases, ablaut was eradicated in Celtiberian by generalising the lengthened grade across the oblique cases as well. However, the original state of affairs is still preserved in words with roots or presuffixal stems ending in l. The Celtiberian script does not write geminate consonants, but Motta (1981) convincingly argued that written nom. abulu, gen. abulos or nom. statulu, gen. statulos hide phonological /abulū abullos/ and /statulū statullos/. The nominative continues PC °lū < PIE *°lō(n), while the genitive is the regular outcome of the zero-grade in the oblique stem, namely °lnos. Other Celtiberian words may also hide original *ln behind a written single l. This has, for instance, been proposed for kelaunikui, which Jordán Cólera (2019: 690) explains as *kelnH-mno-.
On a more speculative note, the same process can be invoked to explain the emergence of the common hypocoristic suffix *-llo-for names in the ancient Celtic languages. I illustrate my ideas with Gaulish, but the basic steps are valid generally. It is well-known that the morphology of shortened names often does not observe meaningful morpheme boundaries (Schmitt, 1995: 424), but that dithematic compound names can be truncated in the middle of the second element, irrespective of the meaning and the transparency of the formation. Stüber et al. (2009: 256) cites Adnema as a shortened form of a compound with *nemes-'sky' or *nemeto-'sanctuary' as second element, and Verca as a shortening of a compound of *u̯ er-'on, upon' and a second element starting with *k-. At the same time, shortened names often appear as on-stems (see Stüber, 2004 for that suffix in general). For example, the compound name Boudilatis (perhaps 'having the fury/heat of booty/victory') consists of the two lexemes *bou̯ di-'victory, booty' and *lāti-'warrior fury'. This could be truncated to the short name *bou̯ dilo-(not attested), which could in turn be 'individualised' as *bou̯ dil-on-. The suffix, which originally was fully ablauting, would have led to a paradigm nom. sg. *bou̯ dilū, acc. sg. *bou̯ dilonam (perhaps attested as Bodilo) with full and lengthened grades in the strong cases, and an oblique stem *bou̯ dill-< *bou̯ dilØn-with zero-grade suffix and regular assimilation. The oblique stem then provided the springboard for newly thematised *bou̯ dillo-(attested as Boudillus), whence *-llo-could be reanalysed as a suffix in its own right added to the i-stem *bou̯ di-. The notional connection with *bou̯ dilāti-, or a compound with any other second element starting in *l-, had been lost at that stage. The derivation was felt to operate directly on the first element alone, which could be extended to nouns other than i-stems. An alternative, slightly variant explanation of the suffix *-llo-starts from monothematic, originally adjectival names in *-lo-, which were likewise 'individualised' by the addition of *-on-, e.g., *kamulo-'servant' or 'champion' → 'individualised' *kamulon-, *kamuln-→ *kamullo-, attested as Camullus.
In a further step, the presence of gemination in a suffix with -l-provided the starting point for the analogical introduction of consonantal gemination into other suffixes, in particular the very common suffixes in *-ko-, to which doublets in *-kko-were then created. This explains the observation (section 9.2.4.) that some geminate sounds show a notable propensity to occur in suffixes.

*sm > *mm
The Proto-Celtic language showed a tendency towards weakening of word-internal s (Stifter, 2012: 541-542;Stifter, 2017Stifter, : 1192, so much so that in various word-internal clusters with resonants, it was prone to disappear through assimilation to the resonant or, in other words, with compensatory lengthening of the resonant. However, only one such development can lay a secure claim on Proto-Celtic age by virtue of being attested for Celtiberian, namely *sm > *mm. (1) The PIE pronominal dative singulars *Hi̯ osmōi̯ 'to whom' and *tosmōi̯ 'to him' show up, via PC *i̯ ommūi̯ and *sommūi̯ (with generalisation of the stem allomorph *so-), as iomui and somui in Celtiberian.
It is interesting to note that while Celtic *s is weak and feeble in sonorant contexts, it is strong and 'aggressive' in contexts with stops. That is to say, especially when following an obstruent, that sound is 'weakened' and eventually completely assimilated to the *s. However, as soon as m is involved as well, it is the one that finally prevails. In less flowery language, *Tsm (where T is any obstruent) becomes *mm already in Proto-Celtic. Neuter verbal abstracts in *-s-man-are one class of words where this can be observed. One example illustrates the pan-Celtic treatment of such complex clusters:  (Hamp, 1974).
Another example is supported by three languages: (4) PC *garsman 'shout, cry' > *garmman > OIr. gairm, W, Corn., Bret. garm, and perhaps Gaul. garma [n]. For this explanation, I assume that *s in the cluster *rsm assimilated to the *m, and that the resulting *rmm contrasted with more common *rm to such an extent that it did not undergo the spirantisation to *rṽ in British. If, however, the *s assimilated to the preceding *r rather than to the following sound and the resulting *rrm merged with *rm before British spirantisation, unspirantised Brit. garm may be a borrowing from Irish (personal communication Paulus van Sluis).
Numerous other examples of neuter verbal abstracts of this type are conveniently collected in Stüber (1998: 45-83;2015: 114-115) and need not be repeated here.
Clusters with two nasals are particularly common in the present stem formation of verbs.
(6) OIr. ro·finnadar 'to find out' ultimately continues the PIE root *u̯ ei̯ d-'to see'. First, the n of the inherited nasal infix formation *u̯ i-n(e)-d-became fossilised as *u̯ ind-, and then another nasal suffix was added to this neo-root, i.e., PC *u̯ ind-nu-, which directly underlies the Old Irish present stem.
(7) Structurally similar, although involving a guttural instead of a dental, is OIr. srennaid 'to snore', probably a denominal verb from the noun srenn 'snoring' < *srenk-no-or *srengʰ-no-(the precise character of the guttural is difficult to determine).
In addition to these relatively common sources for the geminate resonants ll, mm and nn mentioned in the foregoing, a number of rarer consonant clusters may have fed into the creation of marginal geminate obstruents through regular phonological change.

*dk > *kk
A handful of Insular Celtic examples illustrate the assimilation of *dk > *kk at the end of root syllables. Since this assimilation is also found across the composition boundary in Gaulish (see section 6.2.), it is likely that the change is already Proto-Celtic.
A fortiori one might expect that *kk would also continue *gk and *tk, but examples are difficult to identify.

Etymological gemination: developments across branches with identical or similar outcome
Further assimilations, predominantly involving clusters with resonants, took place in a staggered fashion after the end of the Proto-Celtic linguistic unity. From what the fragmentary documentation of the ancient Celtic languages allows us to see, none of these changes were carried through in all Celtic languages. Gaulish and, more importantly, Celtiberian sometimes reflect a more conservative stage. In the Insular Celtic languages, however, the changes discussed in this chapter are completely carried through. This is not a proof for a genetically close relationship between Goidelic and British, but it may be a function of their later attestation, i.e., tendencies towards assimilation, which had their kernel much earlier, had enough time to unfold completely in these languages.
What unites these assimilations and those of the previous section, is that they occur word-internally, but not word-initially, and that they happened after the vocalisation of syllabic resonants. There is no single overarching rule that accounts for all contexts.

*sn > *nn
Two forms in Celtiberian appear to indicate that, unlike *sm, *sn had not assimilated to *nn already in Proto-Celtic, even though there is ample evidence from the three other branches that the change was carried through before the historical period. The two items that show the conservative behaviour of Celtiberian are: (1) PIE *pr̥ h₃-sneh₂-> *φrasnā > *φrannā > OIr. rann, W rhan, MCorn. ran, Bret. rann 'part, share'. If Cib. arznas in Botorrita I derives from the same preform (Jordán Cólera, 2019: 107, 121), it would demonstrate that the assimilation to *nn was not of Proto-Celtic age, but the divergent syllabification of the initial resonant is to be noted.
(2) The comparison of OIr. trén, ogam TRENA-'strong', which continues PC *treχsno-, with Gaul. Trenos indicates that clusters of the type *-Csn-were subject to an early reduction to simple *-n-at least at the Core Celtic stage. Cib. masnai, also in Botorrita I, has been suggested to continue either *mak-snā-'enclosure' oder *mad-snā-'breaking' (see Wodtko, 2000: 244). In either case, the treatment of the triple consonant cluster deviates from that in *treχsno and the cluster -sn-is evidently maintained.
(13) Matasović (2009: 416;2020: 337-338) sets up *u̯ esnālā-, derived from PIE *u̯ esr/n-'spring', as the preform for OIr. fannall, W gwennol, Bret. gwennel 'swallow'. I have instead proposed a substratal loan from a preform *u̯ annell- (Stifter, 2010). Apart from cases that are recent, productively formed compounds, the precise nature of the clusters in the examples above is uncertain. Are they inherited instances of -rs-or did they also arise through secondary processes or through re-analysis? This question is all the more legitimate since Gaulish does provide examples that show the operation of the change *rs > *rr. In order to reconcile the contradictory evidence, one can speculate that the change was just under way in Gaulish at the beginning of the historical period.
(7) PIE *u̯ erso-'being on a high point' > OIr. ferr 'better' This is the same etymon as the preceding item, but with a different ablaut grade. See also W gwell in 3.7. (1).

*ls > *ll
Like in the case of the preceding change, Celtiberian provides an example that shows the unassimilated cluster, namely VELSAM. Due to the lack of a clear analysis of the word, it is uncertain whether its internal cluster is the retention of an inherited sequence or is the result of a secondary development (see Jordán Cólera, 2019: 908 for proposals, none of which is compelling). Other names in Celtiberian sources with the sequence -ls-such as belsa, belsu, kelse are probably borrowings from Iberian. There is no secure example in Celtiberian of the change *ls > *ll having taken place. As for Gaulish, Delamarre (2007: 194)  In the Insular Celtic languages, at any rate, the change has been carried through: (1) PIE *melso-, *ml̥ so-'deceit, fault' > *melso-, *malso-> *mello-, *mallo-> OIr. mell 'destruction, confusion, error', W mall 'destruction, evil'; perhaps also the first element mallo , mello-in Gaulish personal names.

*sr
The case of *sr is different. The alleged change *sr > *rr in OIr. errach 'spring' < *u̯ esr-(VKG i 82) is highly questionable. Not only is there no good parallel for this change (see the arguments below), but the loss of the initial *u̯ would also be unprecedented. Errach is better analysed as an adjectival derivate of err 'hinder part, extremity, tail' (see 3.2. (4)), i.e., 'that which is at the end (of the winter)'. The spelling dírruidiguth 'derivation' beside dírṡuidigud (both in the St Gall glosses, namely Sg. 53a11 and 188a8) for the verbal noun of díṡruthaigidir 'to derive' does not prove the Proto-Celtic assimilation of *sr > *rr since the verb is an Old Irish calque on Latin dērīuāre, which shows the synchronic 'strengthening' or devoicing of r caused by lenited *ṡ.
The best examples of inherited *sr show a single r as the outcome in Irish and, to a lesser degree, in British, but the details of the intermediate stages are not entirely clear. It is possible that *s became *ð before *r, perhaps passing through a stage *z. This is best illustrated by the feminine numeral *tisres > Gaul. tidres, OIr. teóir, MW teir '3' (Kim, 2008: 160−161 (1) PIE *kēs-reh₂-'tool for combing' > *kīsrā-> *kīrā-> OIr. cír 'comb'.
(4) Delamarre (2003: 128-129) explains the Gaulish compound element craro-as 'hornet' < PIE *kȓ̥ h₂sro-(cf. Lat. crābrō), but this is pure speculation since its meaning is unknown. If the etymology is correct, it would show the loss of *s. Of these, the Gaulish and British forms are irrelevant here since they continue the full-grade allomorph of the stem. Only the allomorph seth(V)r-in Old Irish is potentially relevant for the present question. It has been suggested that *-θr-in the oblique cases continue directly the zero grade *-sr-of the suffix. However, this cannot be used as proof for the regular treatment of *sr in Celtic since these forms may have been remodelled after the other kinship terms, which all have oblique stems in -thr-from *-tr (McCone, 1994: 277-278, 283).

*sl > *ll
Only ambiguous evidence is found for the single relevant example in Ancient Celtic. In the Insular Celtic languages, at any rate, the change has been carried through.
Geminate -ll-is also the outcome in Old Irish verbal compounds when root-initial *sl-comes to stand between two vowels. For instance, the compound verb *to-sli-i̯ e/o-'to earn' comes out as deuterotonic do·slí, but prototonic ·tuilli. In British, in contrast, root-initial *sl-in such contexts is treated like lenited l, e.g., the root *slad-'to slay' occurs in the compounds W ymladd 'to kill, fight', Corn. omladh 'to fight', or OBret. anlaedam 'I attack'. In the case of W dyrllid, Bret. dellit 'to earn' < *to-ro-sli-i̯ e/o-, ll is not a trace of *sl, but is due to the deleniting effect of r upon a following l.

*nl > *ll
Only examples from Irish come to mind for this assimilation. Two are compounds with the preverb *en-+ a second element starting in *l-. It is not excluded that their behaviour is analogical after other preverbs and that the change occurred as late as the immediate prehistory of Irish.
3.7. *rl > *ll? I know of two proposals for the assimilation of *rl > *ll, neither of which is entirely compelling.
3.8. *ld > *ll? None of the examples for the notion that *ld became *ll in Proto-Celtic are conclusive, and there is one strong argument against it. I will start with the alleged examples in favour of that change.
(2) OIr., W coll, OCorn. colled, Bret. koll 'destruction, damage' has been compared with PGerm. *halta-'lame' < PIE *kol(h₂)do-, from the same root *kelh₂-'to strike' as the preceding item. Because of the semantic distance between the languages, it is not necessary to operate with identical formations. Like before, the Celtic word could continue a formation with a nasal suffix, i.e., *kolh₂no-with loss of the laryngeal by Saussure's Law. Thurneysen (GOI 95) and Hamp (1974e: 196) compare Lat. culpa, but the notion of PC *ll < *lp was rejected in section 2.5. Hamp also sets up the alternative reconstruction *kol(d)no-without further discussion.
(4) In any case, PIE *meldo-'mild, soft' > OIr. meld, later mell 'pleasant' is decisive evidence that the cluster was retained unchanged up to Old Irish.
The changes in section 3.1.-section 3.8. were major additional sources of geminate resonants, especially in the Insular Celtic languages. Although they include no obstruent sounds, through their number they reinforced geminates as a phonological class in Celtic. Kuryłowicz (1957: 141-144) makes a similar point about the pivotal role that geminate resonants arising from soundchange played in establishing gemination as a marked phonological category, but his arguments for gemination as a morphological process as a whole are not convincing and rely only on a small number of examples.

'tau Gallicum'
In this section, a diversity of phonologically related clusters are treated together, although they belong to several different chronological layers. Clusters of dentals involving voiceless dentals, of dentals followed by s, and of s followed by s across a morphological boundary, are grouped together under a single heading, glossing over various issues that are not relevant to the present question. Already at the Proto-Indo-European stage, clusters of dentals developed an excrescent medial sibilant on an allophonic level (section 3.9.1-section 3.9.2.). These clusters merged with clusters of dentals with original *s over the course of the ancient Celtic period (section 3.9.3-section 3.9.6. In the ancient languages, their behaviour and orthographic representation is still differentiated, which indicates that they had not all fallen together indiscriminately yet. Depending on their origin, they may appear as biphonemic assibilated or as geminate sibilant sounds, traditionally referred to as tau Gallicum (see Eska, 1998 for more details). All these clusters end up as s in the Insular Celtic languages, via an intermediate stage of geminate or 'strong' *ss.
The general outlines of these developments are well-known (e.g., McCone, 1996: 48, 99;Stifter, 2017Stifter, : 1192; the intricate minutiae do not need to detain us in this article, which focusses on the bigger, systemic picture of Celtic sound developments. Seven different types of input lead to the same outcome in the Insular Celtic languages. Completeness is not envisaged here, only a few notable examples will illustrate each type.

*Ds > *tˢ > ss
Proto-Celtic must have possessed a large number of examples of this change as part of the morphologically regular formation of the s-subjunctive of strong verbs, and of the related formation of the future (continued in Old Irish) and the desiderative (attested in Gaulish). Here I will limit myself to the subjunctive, which was formed by adding the suffix *-se/o-directly to the final consonant of the root (Schumacher, 2004: 49-57). Due to the phonotactic constraints of Celtic, this then underwent various types of changes. In the case of dentals, the result in Proto-Celtic was tau Gallicum. Because of the fragmentary or relatively late transmission of most Celtic languages, evidence for this formation is only plentifully attested in Old Irish, but sporadic evidence is found in the other languages as well. Two examples will suffice as an illustration for Irish: (4) In Celtiberian, robiseti 'may cleave (?)' may continue *bitˢe/o-< *bʰidʰ-se/o-from the root *bʰei̯ dʰ-'to split, cleave' (Schumacher, 2004: 224-225).
It is conceivable that this category was still productive in the ancient Celtic languages and that new finds of texts will furnish more examples in the future.
(5) In British Celtic, the category had been largely abandoned by the period of medieval attestation. The only lexical fossiles are MW gwares 'he may succour' and ryres 'he may run', two compounds of *ret-se/o-.
The British languages sometimes have st where the other Celtic languages have ss. Schrijver (1995: 410-430;) explains such cases as continuing *-sst-, i.e., clusters that arose when a suffix with -stV-was added to stems or roots ending in s or, occasionally, a dental. The evidence for this morphological structure is circular, since it is only recoverable from the fact that the British languages have st. Possible instances are: (5) qPIE *h₂u̯ es-steh₂-'spending the night' > *u̯ esstā-> OIr.

*s-s > *ss
Although inherited clusters of two consecutive s were simplified to a single s according to a well-known Indo-European sound rule, where such a formation was synchronically transparent, double ss could be reintroduced in the individual Celtic languages.
The subjunctive and future stems of most roots where this rule could apply (PC *gus-'to choose' < *gȇu̯ s-, *k u̯ is-'to see' < *k u̯ ei̯ s-, *tau̯ s-'to be silent' < *th₂eu̯ s-, *u̯ os-'to spend the night'< *h₂u̯ es-, all after Schumacher, 2004) have been rebuilt in such a way that the phenomenon can no longer be observed in Irish.

'Original' *ss
The etymological origin of tau Gallicum is not always clear. Occasionally its origin may have been onomatopoetic. A case in point is Gaul. bussu-if it means 'kiss' or 'lip', which seems to have a parallel in Irish (Bérla na Filed) bus, pus 'lip' and in Southern German Busserl 'kiss'.

Stops plus *n?
Sequences of stops + *n have been claimed by some scholars to be a source for geminate stops ('Kluge's Law'; e.g., Stokes, 1891; Stokes, 1891-3, Zupitza 100; VKG i 158-161; Lühr, 1985;Bammesberger, 1998; see the section on previous research in the introduction). It is not possible in the context of this study to discuss all the proposed items (for example, Lühr, 1985 discusses 72 relevant forms). Suffice it to say that the proposals are rarely convincing (cf. GOI 92-93), that they can be accounted for in alternative ways, or that they face counterexamples (see the discussions by Kuryłowicz (1957: 131-132) and Strachan (1891-4)). For almost every geminate occlusive, a combination of that same simple occlusive + n can be found that must be reconstructed as such for Proto-Celtic and which therefore runs counter to the predictions of Kluge's Law in Celtic. 2 I will illustrate these general objections with only a few examples.
(2) Gaulish attests to countless examples of the patronymic suffix -ikno-'son/daughter of' < * i-kn-o-, whose semantic core is the PIE root *ken-'to originate from'. The common ancient Celtic onomastic suffix *-kko-has nothing to do with this, but is rather due to 'onomastic' gemination of the common adjectival suffix *-ko-, in analogy to other onomastic suffixes with geminates, such as -illo-< *-ilno-or -ullo-< *-ulno-(see chapter 9.). 2 A noteworthy side aspect of Lühr's study (1985: 300-302) is that she sets up geminate *pp for Proto-Celtic, which, unlike its simplex counterpart, she assumes to have been preserved. This differential treatment has an exact parallel in Japanese. A potential candidate is Gaul. *lapparo-, which can be postulated as the source of Fr. lapereau 'bunny', against Fr. lievre 'hare' < Lat. leporem.
(3) On a related note, the complex suffix *-o-gn̥ (h₁)-o-'born from' regularly develops to *-agno-in Goidelic. This is attested in ogam inscriptions as -AGNI, and as the common suffix -án in Old Irish, not **-aggo-> OIr. **-ac, as would be predicted by Kluge's Law. I do not enter the vexed question whether *-agno-is regularly reflected as -an in British Celtic, or whether that suffix has a different origin.

Etymological gemination: developments across branches with divergent outcome
A small number of simplifications or assimilations of clusters led to very different outcomes in the sub-branches of Celtic, or they occur so sporadically that they are better classified as isolated developments rather than proper sound changes.

*zd
Proto-Celtic *zd continues the rare PIE cluster *sd. No evidence is known how this cluster was treated in Celtiberian.
The ultimate outcome of *zd in Old Irish is ⟨t⟩ = /d/, apparently via an intermediate stage *dd. How early this assimilation occurred is unclear. In the case of *zg, the spelling TASEGAGNI, probably for OIr. Tadgán, in the ogam inscription I-KIK-002 (= CIIC 28) seems to indicate the retention of the sibilant until very recently before Old Irish. The ultimate outcome of *zd in the British Celtic languages is *θ (W ⟨th⟩, Corn. ⟨th⟩, Bret. ⟨zh⟩), which is also the regular outcome of *tt. A stage *dd like in Irish is precluded since that yields d in the British Celtic languages. One way of explaining the British development is to assume that *zd passed through the intermediate stage *tt, even though the required change *zd > *tt is untrivial. It finds potential support in Gaulish where the stage *tt appears to be attested in Gall-Lat. petia, pettia 'piece', which can hardly be separated from PC *k u̯ ezdi-'part, share', and in Gall-Lat. bottia 'blister, bump, boss' and *botto-'knob, button', which underlies various words in the Romance languages and which could be cognate with OIr. bot. Isaac (2004: 74-76) proposes a different pathway from PC *zd to Brit. *θ and the Gaulish reflex. Schrijver (1995: 376) is agnostic about the intermediate stages in the development.
However, the precise relationship of these words remains to be clarified.

*χt
A handful examples appear to attest to the irregular assimilation of PC *χt (of diverse origins) > *tt, instead of retaining them as two separate segments as is the norm. Since examples for this assimilation are isolated across the family and within the languages, it is best to view them as sporadic simplifications of the clusters, the motivation for which remains obscure.
(1) The name of the Ibero-Celtic people Vettones in Central Spain can be understood as *u̯ eχt-on-, a form with the individualising suffix -on-built on the stem seen in Gaul. O 'Brien (1954; has drawn attention to three Irish words that appear to exhibit a similar behaviour: (2) OIr. utlach 'lapful' stands beside uchtlach, synchronically derived from ucht 'lap' < PIE *pektu-'breast'. It is conceivable that the internal -ch-was lost through dissimilation against the final -ch.

Etymological gemination: morphological gemination in derivation and inflection
Many instances of gemination treated above (e.g., *ln > *ll, *sm > *mm, *sn > *nn) occur across morpheme boundaries and are in fact indirectly the result of processes of word formation. Evidence for a few more such changes arising through derivation or inflection is usually restricted to a single branch within Celtic. Therefore, it will be apposite to discuss them separately.

*n-n > *nn
A group of words with nn in the Celtic languages are best analysed as derivates with a nasal suffix from roots ending synchronically in n.
Although the phenomenon appears to be pan-Celtic, the evidence for it is often confined to a single language or two languages at best. It is noteworthy that the resulting sequence -nn-was not simplified, as may have been the case with early instances of *-mm-> *-m-(see 6.3. (2)). Perhaps these words were formed at a time when geminates were already established as a phonological class in the language.
(1) The double nn in the Gaulish personal names Adgennos, Congennolitanos, Adgonnetios has in the past either been ignored or explained as a sporadic gemination of *-geno-< *gȇnh₁o-'born' in personal names. However, onomastic gemination is typically coupled with shortening, but none of these names show shortening. Furthermore, all examples of -nn-are found in compounds with preverbs as first members, whereas compounds with other elements, e.g., Cib. mezukenos, Gaul.
(3) The Gaulish name element rinno-of unknown meaning, the rare Irish adjective renn 'swift, hasty', and W rhyn 'rigid, stiff, brave, rough, cold' may form an equation, although the divergence in meaning renders this less than obvious. If the Irish adjective represents the original meaning, one can think of a formation from the root *h₃rei̯ H-'to whirl', perhaps from a hypothetical nasal-infixed verb *rinati 'to run', comparable with Goth rinnan < *h₃ri-n-H-, with the addition of the adjectival suffix *-no-to the present stem. Alternatively, the formation could derive from the noun *rei̯ no-'great flowing mass of water' < *h₃rei̯ (H)-no-, with ablaut and with the addition of a second nasal suffix.
(4) The Celtiberian name Stennoco (in Latin script) is a derivative of the shape *sten-n-o-from the on-stem name stenū < *stenōn with a single n (Jordán Cólera, 2019: 613-614). The name ⟨stena⟩ in Celtiberian script could likewise stand for *stennā, since the vernacular script does not graphically mark geminates. This development is parallel to that of the genitive abulos < *abul-nos in section 2.1.
(5) The first element of Gaul. sonnocingos has been thought to contain the oblique stem of the Indo-European word for 'sun'. If this doubtful proposal is right, the structure of the word must be something like *su(h₂)n-no-, i.e., a derivative with a nasal suffix (which still would not explain the vowel o).

Other
It is conceivable that (very rare) combinations of nominal stems ending in *-b + the athematic dative and instrumental plural endings *-bis and *-bos may have led to geminate *-bbwithin nominal paradigms, but no such examples are attested from ancient Celtic. Probably there were other, rare contexts of similar character in which geminates arose through regular processes of inflectional and derivational morphology.

Etymological gemination: morphological gemination in composition
Compounding, that is the univerbation of two or more lexical items to form a new lexical item that is more than just the sum of its constituent parts, is an important process of word formation of ancient and medieval Celtic languages. When in the course of compounding two consonants come into contact with each other, new geminates can emerge, either because the consonants were identical or similar already before, or because assimilation takes place.

Assimilations across morpheme boundaries
A special source for geminates are obstruent clusters where the first obstruent assimilates totally to the articulation of the second.
In the majority of instances, however, the consonants across morpheme boundaries are different and some sort of assimilation takes place, either assimilation in voice or in articulation, or in both. This is readily observable in verbal and nominal compounds where the first element is a preverb ending in a stop, in particular *ad-and *eχs-. The latter occurs in the s-less allomorph *ek-in at least a subgroup of such formations (see the discussion in Russell (1988: 118-121), how and why *ek-and other preverbs lost their *-s). Examples come from all branches of Celtic. However, the way how cross-morphemic clusters are treated differs in complex ways, not only between the branches, but sometimes even within a single language. This difficult matter deserves a detailed discussion.
The behaviour of the preverb *ek(s)-is special. The expected development in front of a voiceless consonant is that the first *k, via the intermediate stage *χ, was either lost or assimilated to the *s. This is shown by the agreement of Gaul. escingo-'warrior, infantrist' < *eχs-keng-'striding out' (but cf. also extincon of unknown meaning) and OIr. escarae 'enemy' < *eχs-karant-(however, its lack of syncope points to a recent date of formation); cf. also OIr. sesca 'sixty' < *su̯ eks-kont-. Celtiberian provides the comparable example eskenim, perhaps 'foreigner', but in this case in front of a voiced stop, if it goes back to *eχs-gen-i-. For the position before *t, a parallel is provided by *trek-stu-> *treχstu-> *trestu-> OIr. tress 'contention, fight', W tres 'battle'. However, in young, as it were nachgrundsprachlich, compounds, it seems that the consonantal part of the preverb assimilated totally to the following consonant, perhaps in parallel with the treatment before voiced consonants (see the following paragraph). This may have occurred separately in the already differentiated Celtic branches after the break-up of Proto-Celtic. Old Irish has only three relevant examples: etaim 'chance, opportunity (?)' < *eχs-tud-sman-or *eχs-dī-tud-sman-'act of falling out', ettech 'refusal' < *eχs-teg-o-, verbal noun of as·toing 'to refuse', and etal, etail 'pure, sinless' < *eχs-tol-o/i-'being outside of desire'. The latter has a parallel in W ethol 'chosen' < *ettol < *ek-tol-. This treatment of *k-t > *tt across compositional boundaries contrasts with the regular development of pre-Celtic *Kt > PC *χt within simplex lexemes, and must be due to morphophonological analogy after the model of voiced sounds.
When the second element was a voiced obstruent, the outcome differs decisively between the languages and no comprehensive picture emerges. Old Irish shows full assimilation in such cases. Verbal and nominal compounds provide countless examples; it will suffice to mention a few representatives, e.g., OIr. ·eipir 'says' < *eχs-ber-or ·acair 'sues, accuses' and acrae 'act of suing, bringing an action' < *ad-gar-. In order to explain the difference from how *eχs-is treated before voiceless consonants in old formations seen above, the development before voiced consonants may have been the following: *eχs-b° > *eγzb° > *eγb° with loss of 'sandwiched' *z and with subsequent assimilation to *ebb-. As OIr. naidm 'act of binding, bond' < *nad-man-and maidm 'act of breaking' < *mad-man-demonstrate, *d did not assimilate regularly to a following *m. However, in analogy to the verbal compounds cited above, *dm did become *mm in verbal compounds with ad-as first preverb, e.g., the verb ad·midethar 'to aim at, evaluate' has the prototonic stem ·aimdethar and the verbal noun ammus < *ad-med-. In Stifter (2017b: 221) I tentatively suggested that, in contrast to *dm, *tm may have regularly given *mm in Celtic, in order to explain OIr. amm 'time' < *amman-< PIE *h₂et-men-'the act of going around', but alternative explanations are available for this word, for which see the cited article. The case of the compound *eχs-med-is ambivalent. From its prototonic stem a neo-simplex verb éimdid 'to refuse, reject' was created, which is frequently written with mh, indicating lenition of the m. On the other hand, the further-derived compound fo·émid, for·émid 'to be unable, fail; refuse' never shows such a spelling and is in fact once transmitted with a double mm. One of the two stems must have undergone analogical change.
The evidence is not so resounding in sheer numbers in British, and it seems to differ between the classes of sounds involved and according to chronological layers of word formation. Clear examples of full assimilation are found for the preverb *ad-followed by b-. W, Corn., Bret. aber 'river-mouth' < *abber-< *ad-ber-is reflected the same in all three languages and must be an old formation; W aberth 'sacrifice' < *ad-ber-tā-and W abwyd 'bait, lure' < *ad-+ bwyd 'food' show the same treatment. However, the evidence for -d followed by a g-is ambivalent. Some examples with *ad-in intensifying and preverbal function display the plain simplification of the geminate: W agarw 'rough, stern, bitter' < *ad-garu̯ o-(cf. OIr. acarb 'very rough'), W agwrdd 'strong, mighty' (related to Hisp.-Lat. gurdus 'dolt'?); W agwedd 'manner, fashion' (beside gwedd 'appearance' < *u̯ idā-'look') can only have been formed after initial *u̯ -had developed into an occlusive. On the other hand, a handful of examples appear to illustrate first assimilation of *dg to *gg and then fricativisation to *χ, effectively making it fall together with the outcome of *kk. These are the hapax W achwir 'true, genuine', seemingly a compound of *ad-+ gwir 'true', W achlan 'all, total', if it is from *ad-+ glan 'clean', and W achwre, achre 'part of a roof or fence; covering', perhaps from *ad-+ *u̯ regi-, related with OIr. fraig 'wall'. Again several of these words feature g-that arose word-initially rather late before PC *u̯ -. Compounds of the root PC *gab-'to take' do not provide conclusive evidence. Schumacher (2004: 321) argues that the inherited stem was replaced by *kab-in British, e.g., W dyrchafael, MCorn. drehevel 'rising, ascending' continue *to-ro-ud-kab-aglārather than *-ud-gab-.
The outcomes are also rather diverse for the preverb *eχs-. It is evident that some of the developments must be due to analogy, but it is difficult to determine with certainty what the regular inherited treatment was. Some examples suggest that in such cases the outcome was a fricative, e.g., W differ, MCorn. difres 'to defend, protect' < *dī-eχs-ber-, and W dichlyn 'to choose, pick', Bret. dilenn 'to choose, select' < *dī-eχs-glenn-(cf. OIr. as·gleinn, ·eclainn 'to examine' < *eχs-glenn-; see Hamp, 1974c andHamp, 1974d Even though I have been writing these reconstructions with a medial -s-, chiefly for reasons of etymological transparency, it is not certain that the -s-was retained or, if it was, if it played a role in the sound changes. Schrijver (1995: 376;1999: 2), on the other hand, suspects that the development in such contexts was "*-ksb-> *-xsb-> *-sb-> *-hb-> *-p-> *-f-" (and analogically for *-ksg-), i.e., that -s-played a crucial role. In view of W achlan < *agglan < *ad-glanoetc. mentioned above (if that is the right explanation), I consider the possibility that these British forms involved the asigmatic allomorph *ek-of the preverb (see the following paragraph), with full assimilation to the initial of the following element, i.e., *dī-ek-glenn-> *dīgglenn-> *dīχlennetc. In view of aber < *abber-< *ad-ber-, it must then be assumed that the case of the labial in W differ < *dībber-< *dī-ek-ber-is analogical to the guttural. Another piece of evidence for the development of *eχ(s)-g°, unfortunately without insights into the crucial intermediate stages, is MBret. elas 'gizzard, liver', corresponding to OIr. eclas 'gizzard, oesophagus, stomach' < *eχ(s)-glasso/ā- (Hayden & Stifter, 2022). Because of the regular disappearance of the guttural before l in Breton, it permits no inference about the precise treatment of the cluster.
That there existed an asigmatic allomorph *ek-in British Celtic, or at least in the stage immediately preceding Welsh, is evident from other formations. This allomorph is most conspicuously visible in W eglwg 'conspicuous, visible' < *ek-lukofrom the root *leu̯ k-'to become bright'. Moreover, a series of words starting with e-, all exclusive to Welsh, can conceivably be analysed as compounds with *ek-, giving evidence of a sequence of reanalyses of such formations. They were therefore probably only formed productively within the Welsh language during the historical period. They include eglan 'sea-shore' < *egglannā-< *ek-+ *glannā-(glan 'shore'), egwan 'very weak' < *ek-+ *u̯ anno-(gwan 'weak'), as well as the evidently late neologisms egwyl 'respite' (16 th -17 th century, beside gŵyl 'holiday') and egwal 'cabin' (18 th -century, beside gwâl 'lair, den'); furthermore eban 'feeble' beside ban 'top' and edif 'greedy' beside difiog 'voracious' (Russell, 1988: 120). At first glance, the first of these appear to contradict the claim above that *ek-g° became *egg° and ultimately *eχ°. However, since these formations are late and secondary, various analogical steps are involved. Eglwg 'conspicuous' was synchronically analysable as consisting of eg-+ a lenited allomorph of *llwg 'light, visibility'. Eglan etc. can therefore have been formed as compounds of the same eg-+ -lan, the lenited allomorph of glan, etc. Since, on the other hand, the relationship of the simplex glan to the compound eglan could also be analysed as that of a prefix *e-+ unmutated base glan, the way was then free to form e-ban from ban and e-dif from dif°.
The situation is very different in Gaulish. In the certain instances of the preverb *ad-followed by *b-or *g-, assimilation is typically lacking, e.g., adgarios, adgariontas, adgarie < *ad-gar-(contrast OIr. acrae 'act of suing, prosecution', ·acair 'to sue, prosecute; bewitch' < *aggar-of the same structure), Adbugiounus < *ad-bug-, Adbogius < *ad-bog-(contrast OIr. apach 'corpse, remains' < *abbou̯ g-, possibly of the same structure). Cisalpine Gaulish shows the same treatment if aśkoneti(o) represents the name written Adgonnetius < *ad-gonn-et-in the Latin script. Assimilation is likewise missing between d and m, as the name Admina (aśmina in the Lepontic script) attests. The name Annamat(i)us and the placename Annamatia are undoubtedly to be connected with the name that is more commonly written Adnamat° 'against the enemies', but note that attestations for the assimilated variants do not come from the Gaulish heartland, but from 'marginal' areas such as Noricum or Pannonia where influence from other languages is conceivable. De Bernardo Stempel (2010: 69) cites the divine epithet Agganaicus < *ad-gen-aki̯ oof Jupiter (Pavia, 2 nd c. a.d.) as an example for the assimilation of dg > gg. Perhaps this reflects a late development in Cisalpine Gaulish, but the number of additional changes required for the etymology casts doubt on the relevance of this example. The preponderant lack of assimilation in most of the Gaulish evidence means either that assimilation of voiced clusters across the morpheme boundary was not a pan-Celtic phenomenon, or that these compounds were morphophonologically so transparent that the assimilation could be easily undone by reanalysis. Its ambiguous writing system renders the pertinent Celtiberian evidence meagre and difficult to interpret. The gentilic name abo[..]kum, which has been emended to aboiokum on the basis of Abboiocum in a Latin inscription, has been explained as *ad-bog-i̯ o-ko-by Prósper (2005: 252-254). The verbal form usabituz, perhaps 'let him cut out', may continue *uts-ad-bi-tūd (Schumacher, 2004: 226-231). While the deficiency of the Celtiberian script with regard to writing obstruent clusters leaves it undecidable if *d has been assimilated to *b in these cases or if it is just not graphemically expressed, the spelling Abboiocum in the Latin script indicates a genuine geminate resulting from assimilation, provided the etymology is correct.
The examples above are all taken from compounds where the identity of the involved elements is well established and would also have been easily recoverable for native speakers. This transparency of the elements may have entailed a special, analogical treatment across the morpheme boundary. It is therefore possible that geminates that are not transparently analysable as resulting from assimilation may show different outcomes. This is conceivably the case in the British Celtic languages, as Anders Jørgensen (personal communication) reminds me. As argued in the sections above, sometimes what must have been voiced geminates in prehistory are reflected by unlenited sounds, sometimes the outcome appears to be spirant sounds, at least in the case of *gg, as if an intermediate stage had consisted of a voiceless geminate (cf. Russell, 1988: 115-125). W achlan and eglan show contradictory behaviour within a single language. The conditions for the divergent treatment are not always obvious. When the outcome of the geminate in Welsh is a single voiced stop, but lacks parallels in Breton and Cornish, it may have been formed more recently than words with parallels, where the outcome is a voiceless fricative.
At this point it is apposite to look at the treatment of other possible examples of voiced geminates, especially of *gg, in the British languages, when they do not occur across transparent morpheme boundaries. They allow placing the ambiguity of the treatment of *gg in a bigger picture.
(3) Another item is equally ambiguous: W gwraig, OCorn. grueg, Bret. gwreg 'woman' speak in favour of a preform *u̯ rakī or *u̯ rakū with k. On the other hand, the spelling variants of the rare Irish word frac, fracc, frag 'woman' are most straightforwardly interpreted to stand for /g/ < *u̯ raggā-'(old) woman'. This interpretation finds support in Scottish Gaelic fràg 'a kind woman'. While the spellings frac and fracc could conceivably stand for /frak/ < *u̯ rakkā-, the spelling frag and ScGael. fràg would remain isolated in that case. A uniform explanation that accounts for all Gaelic forms is preferable. In addition, the British Celtic languages possess the bye-form W gwrach, OCorn. gurah, Bret. gwrac'h 'old woman'. This could either reflect *u̯ rakkā with 'expressive' gemination of the *k of *u̯ rakī/ū or, if my analysis above of bychan etc. as *biggo-is accepted, it could reflect *u̯ raggā-and correspond directly to the Irish form. Phonaesthetically, -ch /x/ acquired a negative connotation especially in Welsh (Rodway, 2019;Wmffre, 2007: 59;Zimmer, 2000: 278). This may have been prompted by loans from Irish with their frequent suffix -ach that conferred a particularly alien and in consequence pejorative feeling on such words (Sims-Williams, 2011: 183-184). This has been suggested to explain the choice of the 'phonemaestheme' -ch and the semantics of gwrach. However, while this phonetic attitude is specifically Welsh, the negative connotations of gwrach appear to go back already to Proto-British, since it is equally attested in all three British languages. This detracts from the phonaesthetic explanation of gwrach.  Schrijver, 1995: 308) and OBret. legh, Bret. lec'h 'place'. Several explanations are possible. The forms with the fricative could continue a sporadically geminated *-gg-; or they could continue by-forms with suffixal *-s-of unclear function added to the stem, e.g., *u̯ riχsā and *leχsā. Finally, the -c'h of Breton could be due to a phonetic strengthening in sandhi of final *-γ > *-χ in Late Proto-British *g u̯ reγ and *leγ, comparable to the strengthening of *-h > *-χ seen in Bret. dec'h 'yesterday' < *deh < *γdes(i), in contrast to W doe (Schrijver, 1995: 390). The latter appears to be the simplest explanation. It only entails the extra assumption that the stem form of the singular was then also extended to the plural. In consequence, these words are not relevant for the treatment of geminates.
Several items seem to exhibit the simplification of *dd > d, like across the morpheme boundary seen above.
(6) This is more difficult to argue for a number of other words in which *dd can be set up for etymological or comparative reasons. For W rhwd, OBret. rod 'rust; 'mud, filth', and the Cornish placename Polroad, Polrode, possibly from *ruddo-< *rudzdo-see section 6.1. above.
(8) Another possible instance is the equation OIr. gat 'theft' and Bret. gad 'hare', which leads to the reconstruction of the common preform *gaddo/ā-(in Stifter, 2021, I linked the two semantically distant words through the popular belief that hares steal milk).
While in Irish it is the rule that geminates that resulted from the assimilation of two stops were subsequently simplified to a single stop, the foregoing discussion makes it unavoidable to conclude that geminate clusters were treated differentially in British. *bb and *dd seem to have been reduced to single voiced stops, but *gg may have become the voiceless fricative *χ. Where, on the other hand, single *g results in Welsh, this may rather be due to a late analogy. Clearly, more research needs to be done on the treatment of geminates, and it cannot be guaranteed that all relevant examples were taken into consideration in this study, since no exhaustive search was carried out.

*RR
This paragraph brings together a handful of diverse instances in individual languages where geminate resonants do -or do not -arise as a consequence of identical sounds coming into contact across the composition boundary; phenomena which do not fit precisely into any of the preceding sections.
(1) A special, language-internal case of identical sounds across the morpheme boundary is Gaul. petorritum 'four-wheeled wagon'. The geminate rr must have arisen from metathesis of earlier *petru-rito-, which contains the composition form *petru-< *ku̯ etru-of the numeral '4' (itself metathesised from earlier *k u̯ etur-), and a nominal formation of the root *ret-'to run'. Alternatively *k u̯ etur-could be an archaism preserved in this compound, or petor-is simplified from the younger Gaulish form *petu̯ or-with influence from the cardinal *k u̯ etu̯ ores '4'.
(3) A potential parallel is OIr. neim 'poison', which inflects as a neuter n-stem in the singular. The underlying formation would be expected to be *nem-men-(root *nem-'to apportion' + suffix *-men-), but since the Old Irish word descriptively continues *nem-en-, it is attractive to operate with the same early simplification of *mm here. A different possibility for both *kommen-and *nemmen-is to assume the loss of the medial consonants in cases where the second syllable came to stand in the zero grade, i.e., *kommn-> *komn-, with subsequent generalisation of the simplified stem *komen-(personal communication Michael Weiss). Such a strategy has been employed to explain OIr. gein 'birth' < *genen-← *gȇnh₁men-(see 5.1. (1)).
At the end of the foregoing developments, the phonological system of Celtic had morphed into the one in Table 2.
There was a contrast between s and a strong sibilant, although that contrast does not mirror that of simple vs. geminate consonants elsewhere. It is probable that single voiced stops already had lenited allophones in intervocalic position, but this is not indicated in the system above. On the whole, geminated voiced stops were rare except across the morpheme boundary (an observation already made by Martinet, 1952: 198).

Etymological gemination: Proto-Goidelic *NT > *DD
Notwithstanding the many diverse developments presented so far, one of the most common sources for geminate voiced stops in Irish are Proto-Celtic clusters of nasal + voiceless stop that developed into the corresponding voiced geminates after the separation of Goidelic from the rest of the Celtic languages, e.g., Proto-Goidelic *gg < PC *nk. At some point in the prehistory of Irish, the voiceless stops in such clusters assimilated in voice to the preceding elements, while the nasals assimilated in the mode of articulation to the following stops (GOI 126-127;McCone, 1996: 106-109;Schrijver, 1993: 35-39). The two met, as it were, in the middle and a geminate voiced obstruent resulted. While geminate voiced stops are a rarity in Proto-Celtic and in other ancient European languages (Martinet, 1952: 198), thanks to this change they are very common in Primitive Irish.
In ogam inscriptions of the classical period (5 th -7 th century a.d.), the resulting voiced stops are written with the letters for D and G, e.g., DECCEDDA < PC *dekantos or TOGITTACC < *tonketāko (the double spelling DD or of the other consonants has nothing to do with geminate sounds, but is an orthographic convention in ogam that is independent of the phonological nature of the consonant). In Old Irish orthography, they are usually expressed by ⟨c⟩ and ⟨t⟩ (and by ⟨p⟩ for /b/ arising in other contexts). Preceding Proto-Celtic short *a and *e (the latter of which had been raised allophonically to mid-high *ı before the tautosyllabic nasal) were lengthened in the process and fell together as é. They are retained as such in accented syllables, but are shortened along with other long vowels in unaccented syllables. As a consequence, words with initial éc-and ét-very commonly go back to compounds consisting of the negative prefix *an-'un-' + etyma in c-and t-. (4) and by the Old Irish preverb ceta· < *kanta < *km̥ th₂ 'together with' (cf. OW cant 'with') and by the preposition/preverb etar < *enter 'between'.
Other forms have received less attention in this dispute, even though they are equally relevant for determining the outcome of PC *inC. McCone (1996: 107-108) mentions a couple of "problematic instances with unlengthened stressed vowel", namely ecor 'arrangement' (the verbal noun of in·cuirethar), tecosc 'instruction', do·ecmaing 'to befall', do·ecmalla 'to collect', and conjugated and therefore stressed forms of the preposition etar 'between', such as etruinn < *enter-snī (uel sim.). McCone wonders if "the following o (plus r or m/v) played a role in the loss of the nasal", but he arrives at no other conclusion than that "the precise conditioning remains unclear". He refers to GOI 518-519 in this context, but does not actually quote or discuss Thurneysen's alternative solution, even though it merits a closer look.
Thurneysen proposes that "these examples can best be explained by assuming that in them the preposition had at one time the form in-" (relevant forms are also discussed by Armstrong, 1976: esp. 64-66). The Celtic preposition with the meaning 'in' had been inherited from Indo-European as *eni (Dunkel, 2014: 224-225). It is attested as a plain preposition in the most archaic form in Celtiberian eni, and in a few nominal compounds in other Celtic languages, e.g., Gaul. Enignus, OIr. ingen 'daughter', Ogam INIGENA < *enigenā, OIr. inis, W ynys, Bret. enez 'island' < *eni-sth₂-ih₂-. However, in the Gaulish preposition in it appears with loss of final -i and with unexpected raising. This same form *in underlies also the ordinary Old Irish preposition i N and Welsh yn. That the Irish preposition continues the high vowel *i follows from the fact that *en with mere apocope of final *-i would have resulted in **a N in Old Irish, like the masculine infixed pronoun -a N < *-en did. Dunkel (2014: 223) suggests that the vowel of Gaul. in reflects the raised allophone *ɪ of *e before a nasal in tautosyllabic position, i.e., before consonant. 3 In Irish, *ɪn is usually kept distinct from original *in, but in this case it has to be assumed that mid-high *ɪ was further raised to *i.
In verbal composition, the situation is even more complex. The allomorphs of the preverbs have been conflated to such a degree that a clear distinction is not always possible. This is not aided by the fact that the presentation and morphological analysis of verbal forms in eDIL and other handbooks is often imprecise or incorrect. The archaic allomorph *eni-shines through OIr. do·infet 'to inspire' and its 3sg. present subjunctive ·tinib < *to-eni-su̯ izd-, but otherwise evidence for it is hard to come by and hard to distinguish from the more common allomorphs *ande-and *in-(reconstructed preverbs are written with a final hyphen in order to distinguish them formally from prepositions and adverbs). The 2sg. imperative of in·cosaig/in·coisig 'to signify, indicate' is inchoisc. As is evident from the lenited -ch-, this cannot be from *in/en-kom-sech-but must either go back to *enikom-sk-or *ande-kom-sk-. eDIL quotes no attestations with -d-, e.g., *·indchoisc. The lack of such forms in a verb that is well attested in early sources can be taken as indirect evidence that it involves *eni-.
The most common allomorph in verbal compounds appears to be *ande-. It seems that in the earliest period it occurred as ind· in pretonic position and as ·in(d)-in stressed position, as opposed to *in-that appeared as in· and ·in-respectively. Ultimately both merged in in· in pretonic position and cannot always be kept apart in other positions either. A precise description of all contexts is not intended here. Finally, and to complicate matters even further, it will become clear from the following that beside *eni-, *andeand *in-, there was also a fourth allomorph *en-with a much more restricted domain. Leaving aside *eni-, whose presence in a handful of Irish compounds was demonstrated above, it can be shown by minimal pairs of phonological micro-contexts that a three-way contrast between *ande-, *in-and *en-needs to be made. The distinction between *ande-and *in-is required to account, for instance, for the different behaviour of in(d)·lá 'to enter into, arrange, etc.' and its prototonic stem, represented by the 3sg. subj. pres. ·indell < *ande-la-(not **·ell < **·in-lauel sim.), versus in·loing 'to join, bring together, put upon, etc.' and its prototonic form ·ellaing < *in-long- 3 There are other seemingly sporadic examples of i where e is expected, but none is parallel to *in in distribution. The 3 rd singular of the copula OIr. is, W ys must come from *isti < *esti < *h₁esti, but these seem to be independent developments in Goidelic and British Celtic. In Old Irish, the i is the result of raising of unaccented vowels before palatalised *s (Griffith, 2016: 48-51), in Welsh it is an instance of i-affection of *e (Hamp, 1974b: 33;Schrijver, 1995: 265-268). In Gaulish, PC *esti may in fact be attested as esi (L-98 1a9) with unraised e. On the other hand, the 1 st singular of the copula appears as imi (L-120), immi (G-13) in Gaulish, but since the corresponding Old Irish form am continues *emmi < *h₁esmi with unchanged vocalism, this raising must be specific to Gaulish. 4 Pace Dunkel (2014: 224) who derives it from *en-dʰe. The comparison of OIr. indel 'preparation, machinery' with Welsh annel 'trap, deception' < *ande-lo-, and Gaul. ande-leave no alternative than to reconstruct PC *andi/ande < *n̥ dʰi.
The reason why *in-instead of *en-is set up here for the preforms, at least for deuterotonic verbal forms when the preverb stands before the accent, is that *en-in unstressed position would have become **an·. 5 On the other hand, the presence of the allomorph *en-(or potentially even *an-!) is required in other contexts, for instance by verbs such as con·éitet 'to accompany' < *kom-en-tei̯ g-; or in·túaisi 'to listen to' whose prototonic stem must be from *en-tou̯ stī-, e.g., 3pl. ·éitset; or by the augmented subjunctive stem of in(d)·fét 'to tell, relate', e.g., 1sg. pres. subj. ·écius < *en-kom-u̯ ei̯ d-s-. All of these verbal forms show the regular change of *en-> ébefore a Proto-Celtic voiceless stop described at the beginning of this chapter. In the case of ·ellaing, no decision can be made between *in-and *en-, since the outcome would be the same. Ultimately, in this way a fourfold allomorphy of *eni-, *en-, *in-and *ande-can be demonstrated in verbal morphology, with a number of distributional restrictions on their occurrence, as illustrated in Table 3.
While all the phonological developments invoked so far are trivial in the diachrony of Old Irish, this is not the case for the words quoted by Thurneysen (GOI 518-519) and McCone (1996: 108-109), namely ecor 'arrangement', tecosc 'instruction', do·ecmaing 'to befall', do·ecmalla 'to collect' and etruinn. These cannot go back to *en-koroetc. since it was just seen that this would have regularly resulted in **écor etc. The logical conclusion, after the alternatives have been excluded, is that they must continue *in-koro-, 6 *to-in-kom-sk u̯ o-, *to-in-kom-ink-, *to-in-kom-la-, and *inter-snī respectively, with the development of *in-k-> *igg-in a first step, and then with regular lowering of *igg-before a back vowel > *egg-. The same solution applies to itge, itche 'request, petition' < *in-tech-i̯ o-< *in-tek u̯ -i̯ o-, only with the regular absence of lowering. 7 What about the distribution of *invs. *en-? Perhaps its obscured rationale is that *in-was originally at home in pretonic position, that is, as a plain preposition and before the accented part of the verbal complex, while *en-occupied the stressed part of the verbal complex, as reflected in con·éitet, ·éitset or ·écius, and was used in most verbal nouns. In a subsequent development, the unstressed variant intruded by analogy into the stressed portion of the verb, which explains cases such as do·ecmaing or ecor. The replacement of *enby *in-in this environment is not a regular, but a sporadic, analogical process. Operating with a sporadic replacement is not completely arbitrary or random. That there could be a formal distinction between pretonic and tonic allomorphs of preverbs is commonplace elsewhere in Old Irish, for instance in the alternation do· vs ·to-or a H /as· vs. e(s)-. It is equally well known that the two allomorphs could influence each other.
In the case of a H /as·/e(s)-, the spread of the pretonic allomorph into tonic position is observable in such verbs as as·beir, ·epir, where the imperative is found as apair already in Old Irish, or in pronominal forms such as 1sg. asum or 3sg. masc./neut. as instead of archaic es-. In the case of do·/to-, the tonic allomorph was frequently used in pretonic position, in archaic and in archaising orthography, for example archaising to·beir for do·beir.
The presence of two allomorphs side by side with each other in one paradigm is not isolated in the wider Celtic perspective either. It is securely attested for the prehistory of the Old Irish preposition do 'to, for' and for the preposition/preverb de 'from', the distribution of neither of which follows predictable rules throughout their respective paradigms. In the case of do, the allomorph PC *dū < PIE *doh₁ (Dunkel, 2014: 148-149) underlies the plain preposition and conjugated forms such as dúinn 'to us' < *dū-snī(s) or dúib 'to ye' < *dū-su̯ ī(s), as well as British Celtic *dī and perhaps Gaulish duci, whereas *do < PIE *do underlies conjugated forms such as duit 'to you' << *do-tī < *do-toi̯ or dó 'to him/it' < adverbial *do. In the case of de, its allomorphs PC *dī < pre-Celtic *dē and *de < *de occur without apparent distributional rationale in several compound verbs, e.g., ·díltai 'to deny' < *dī-slondīvs. dermat 'forgetting' < *de-ro-mento-(but see the critical discussion in Dunkel, 2014: 148-156).
A side remark: The preposition/preverb PC *enter is always continued with a short vowel in OIr. etar (cf. Lash, 2017 for other Old Irish allomorphs of it). This is expected for the preposition in pretonic position, but the short vowel is not

*kom-en-tei̯ g--con·éitig *en-kom-u̯ ei̯ d-s--·écius
expected in inflected forms of the preposition and when the preverb occurs in nominal compounds. For instance, the 1pl. is attested as etruinn, not as expected **étruinn, or the word for 'boundary-ditch, fence' is etarbae, not expected **étarbae. It might look attractive to derive those forms from an innovatory allomorph **inter, in which the more basic local preposition *in had analogically caused the replacement of inherited *en by *in. However, such a scenario is excluded. There is no possibility how the hypothetical preform **inter would have led to *ed'er with lowered *e-in the first syllable. It is therefore more economic to assume that the preform is indeed the traditional *enter > *ēd 'er, 8 which via vowel shortening and depalatalisation in pretonic position resulted in *eder. In a further step, this shortened pretonic variant replaced the original tonic variant *éter everywhere. The complete replacement across the board of all tonic allomorphs by the pretonic ones in this scenario nicely illustrates the randomness of analogical change. In the case of several other prepositions, the language tolerated the coexistence of dual stems in unstressed and stressed prepositions: co vs. cuc-, amal vs. saml , dar vs. tor-or, indeed, as demonstrated above, i vs. ind-/and.

Etymological gemination: Insular Celtic mutations
When we broaden the perspective from the level of isolated lexemes to that of accentual units and constituent phrases, we can see that some of the sandhi-effects across words that ultimately became the initial mutations of the Insular Celtic languages, can be analysed as the same or similar types of geminations that arose word-internally. These sandhi-related processes, namely the interaction of word-final sounds with word-initial sounds, thus extended the positions where geminate sounds were phonotactically permissible from word-internal to word-initial position. To all extents and purposes, internal and external sandhi behave equivalently. A comprehensive diachronic description of the emergence of mutation in Insular Celtic is outside the aims of this study. The present section only aims at highlighting the fundamental parallels of certain types of mutations with word-internal gemination. Only sandhi involving two interacting consonants is relevant for the question of gemination. Leniting contexts, where an initial consonant originally followed a word ending in a vowel, are therefore ignored here.

Irish
In the following, the examples will be drawn mainly from contexts that underlie mutations in Irish. Traditional grammars of Irish mention a mutation which they call 'gemination' (e.g., GOI 150-153). This is a misnomer, both synchronically and diachronically. Diachronically, other sandhi contexts also led to geminate sounds in Primitive Irish in the appropriate contexts. Synchronically, the occasional geminate spellings of certain sounds are best regarded as markers of non-lenition . Instead, the main effect of this mutation, which finds no graphic expression in genuinely Old Irish sources, but which is directly observable in Middle and Modern Irish written sources and in Modern Irish pronunciation, is the prefixation of h-to a following word if it starts with a vowel. Therefore this mutation is called 'aspiration' here (cf. Stifter, 2009: 65).
The sandhi contexts that are relevant for the diachronic study of gemination are therefore nasalisation, aspiration, and non-mutation, i.e., the appearance of the unmutated radical sound. It is essential to acknowledge that in addition to the morphosyntactically recognised mutations, the absence of an overt change, i.e., 'non-mutation', also is a mutation in its own right. From the point of view of diachronic phonology, aspiration and non-mutation are for the most part just two sides of the same coin. In the vast majority of cases, both continue contexts where the mutated word was preceded by a word ending in -s. 9 At the time when final syllables had not been apocopated yet, the effects of non-mutation must have been identical to those of aspiration. The split between aspiration and non-mutation depends on the context: either intervocalic or adjacent to a consonant. Aspiration occurs when the final *-s > *-h combined with the initial vowel of the following word in Primitive Irish after the shift of the syllable boundary. When a consonant followed, the -h merged with that consonant. In most cases, this must have led to a phonetically slightly longer pronunciation that, at least in the case of the resonants, but perhaps also of stops, meant phonetic similarity to or identity with word-internal geminates. However, the effect was different if the following word started with *u̯ -. In that case the chain of events was *-s u̯ -> *-h u̯ -> *# hu̯ -> OIr.
-Ø f-, where the outcome f (phonologically equivalent to a geminated *u̯ !) is identical to that of the internal group *-su̯ -, but is not phonetically similar to single *u̯ , which rather gets lost. On the level of surface phonetics this means that when *-h stood immediately before a consonant, it was absorbed by it and non-mutation ensued.
Synchronically, aspiration is only caused by proclitic monoor disyllabic particles that end in a vowel. It is conceivable that aspiration was originally also caused by Old Irish inflectional forms that ended in a vowel, for example the accusative plural of masculine and feminine inflectional classes, but there is no written evidence for this. This effect can only be conjectured. Non-mutation, in any case, occurs after all classes of words that ended in a consonant and in phrase-initial position.
Through the outlined processes, non-lenited initials and non-nasalised initial voiced stops came to be phonetically identifiable with internal geminates. The phonetic salience of geminates word-internally, and the rise of geminates word-initially, must have mutually reinforced each other. Sandhi-generated geminates will have considerably increased the overall token frequency of gemination in speech and strengthened its phonological status in the system. In contrast, gemination was rare in absolute word-final position.
In the case of the nasal mutation, several contexts need to be distinguished. A probably weakly articulated nasal sound in final position was attached and merged with initial consonants. It assimilated fully to initial resonants. That this sound must have been phonologically *-n can be gleaned from the different treatments of inherited initial and internal clusters with *m, which do not turn into geminates, e.g., *mrogi-> OIr. mruig 'land', *mligeti > mligid 'to milk', or *kom-rigo-> OIr. cuimrech 'binding, bond'. Furthermore, this -n is directly visible in the nasal mutation n-on initial vowels. It is therefore evident that, like in Gaulish and British Celtic, final PC *-m in the ending of accusative singulars, genitive plurals, and neuter nominative singulars, had become *-n in the prehistory of Irish. There are not many inherited word-internal clusters of *n + *r or *l that illustrate their identical treatment to that across the word boundary. The best examples are furnished by compounds with the preverb *en-'in', where it hadn't been replaced by other allomorphs. Practically speaking, the only good examples are those with *l, which indeed show the expected assimilation to *-ll-, namely OIr. ellam and ellach (see section 3.6.). For *-nr-, the only potential example that I am aware of is OIr. eirr 'chariot-fighter'. This has been suggested to continue *en-ret-'he who runs into (the battle)' (see NIL 577 n. 3), but the analysis *ers-sed-'he who sits at the back' is preferable because of its parallel to arae 'chariot-driver' < *are-sed-'he who sits at the front' (see 3.2. (4)). Nasalisation of voiceless initial stops has exactly the same outcome as in word-internal position, i.e., a voiced single stop results synchronically. It can be surmised that the sound was a geminate at an intermediate stage (cf. chapter 7.). Nasalisation of voiced initial stops does not lead to geminates, but to prenasalised stops.

British and Gaulish
In contrast to the situation outlined for Irish, there is a broad consensus that gemination of initial sounds played no role in the emergence of British Celtic mutations (Harvey, 1984;Russell, 1985;Schrijver, 1999;Sims-Williams, 1990;Sims-Williams, 2008;Thomas, 1990).
Very occasionally, external sandhi phenomena that look similar in structure to Insular Celtic mutations can be found written in Gaulish inscriptions, e.g., reguccambion < *regū-k (u̯ ) ' kambion 'and I straighten the crooked' in the inscription from Chamalières (L-100). The final c of reguc is probably enclitic *-k u̯ e for 'and' that merged with the initial c-of cambion 'crooked' (differently De Bernardo Stempel, 2010: 69).
9. Non-etymological gemination: onomastic gemination 9.1. Vocative morphophonology While phonological developments represent the largest source for geminates, other factors may also have contributed to the increase of geminate sounds in the language. One such factor is the pragmatics of onomastic morphology. It is a general observation that, especially in the ancient Celtic data, gemination is especially frequent in anthroponomastics, but not among the 'long' dithematic compound names of the Indo-European type, which often have martial and heroic connotations, but rather among more colloquial short or hypocoristic names (Stüber et al., 2009: 37-38). This morphophonological behaviour is not limited to Celtic, but is a feature of many ancient Indo-European naming systems (cf. Schmitt, 1995: 425, 618, 620;and Ellis Evans, 1967: 296-297, 376 with earlier literature). Masson (1986: 220) stresses the central pragmatic importance of the vocative as the context in which gemination of stem-final consonants could arise, in personal names but also in expressions that belong to colloquial registers.
The origin of this process of onomastic gemination may lie in 'vocative reduction' or 'vocative truncation', one method of the formal marking of vocatives (see Daniel & Spencer, 2009: 628-629 for this and other types of vocative marking). Descriptively speaking, vocatives can be shorter than the forms of names that fill the regular argument slots in a sentence (cf. the pertinent examples cited in Daniel & Spencer, 2009: 629;Janson, 2013: 224-231;Schmitt, 1995: 419-425, 618). Indo-European languages bear this out in many ways. The shortening can take the form of the reduction of the number of segments or of the reduction of moras, for example in the Proto-Indo-European vocative of o-stems such as *u̯ iHros → *u̯ iHre 'man'; or the vocative of -eh₂-stems, caused by laryngeal loss in pausa, e.g., qPIE *g u̯ enah₂ → *g u̯ ena 'woman', e.g., in OCS voc. ženo; for consonant stems, cf. Greek nom. Σωκράτης (Sōkrátēs) → voc. Σώκρατες (Sōḱrates), which shows both stress shift and mora reduction.
In other languages, whole syllables are dropped at the end. A contemporary illustration is the neo-vocative of Russian whereby male names in -a lose the final vowel, e.g., nom. Saša → voc. Saš. This habit has recently been borrowed into Georgian (Amiridze, 2022: 1-2). At the same time, this pragmatic shortening of names in situations of address -no doubt connected with the emotional urgency typical of conversations -can be counteracted on other levels of the speech act. The reduction or loss of segments, especially at the very end, is sometimes balanced by increasing the moraic count further to the front of the word. This is illustrated by the Georgian neo-vocative, where the loss of the final vowel is accompanied by the lengthening of the root vowel, e.g., Šota → voc. Šoot, or Gvanca → voc. Gvaanc (Amiridze, 2022: 2), so that the overall moraic count of the name remains the same. Vowel lengthening is a typologically common process of vocative marking (Daniel & Spencer, 2009: 629). I believe that consonant gemination is a comparable phenomenon that is equally linked to vocative truncation, with the difference that, instead of lengthening a vowel as in Georgian, it is accompanied by the lengthening of the final consonant. Gemination can thus be regarded as a process that compensates for phonological or morphological loss, perhaps as a way of making up towards the end of the utterance for the overall reduction in the moraic structure. Due to the absence of suitable textual genres in the surviving documentation of the ancient Celtic languages, the connection between gemination and vocatives cannot be demonstrated with actual examples, but the relatively high number of geminates in personal names is indirect evidence of this tendency.

Gemination in Ancient Celtic names
Once gemination had been established as a morphological feature of personal names in a highly specific context, namely in the address of persons in intimate or informal speech acts, it could then be transferred also to contexts outside of vocatival function (for the generalisation of vocatives to other contexts more broadly, see Stifter, 2013). Examples of shortened names with gemination in core syntactical functions are well attested in ancient Celtic, e.g., Gaul. Eppo, reflecting a compound name with *epo-'horse' as first member, Blattia ← *blātu-'flower', or Sammus ← *samo-'summer'. Commonly the 'root' (or rather onomastic basis) of such names is monosyllabic. Sometimes the etymology of Gaulish names is not entirely certain. For instance, the Gaulish name element poppo-has been interpreted as 'cook' < PIE *k u̯ ok u̯ o- (Delamarre, 2003: 252), and the names Peccia, Peccio have been compared with Ogam Irish QECIA, QECEA < PIE *k u̯ ek⁽ u̯ ⁾i̯ o-'couragious, strong' (Delamarre, 2003: 247), related with W pybyr 'eager, vigorous, brave' < *k u̯ ek u̯ ro-. Onomastic gemination is the best explanation in such cases.
Conceptually related is gemination in kinship terms and in generic nouns for persons that are intimately known, e.g., OIr. macc 'son, boy' < *mak u̯ k u̯ o-vs. ungeminated *mak u̯ o-in Gaul. mapon, W, Corn., Bret. mab; or W geneth 'girl' < *genettāvs. Gaul. geneta, OIr. geined < *genetā-(more on this in section 11.1. on symbolic gemination). Gaulish has the personal name Matta, perhaps created from *mati/u-'good', but this word also appears in Raeto-Romance as a generic term for 'girl'. Maybe it had a similar generic function already in spoken Gaulish, from where it was retained as a substratum loan in Romance.
But also names with a larger phonetic body show gemination. A consonant, usually the final consonant before the ending, can be geminated without concomitant shortening of the name. Suffixes with voiceless velars are common in all Celtic languages. They continue earlier *-ko-or *-kȏ-added to vocalic stems, whence more complex suffixes such as *-iko-, *-uko-, * īko-, or the very productive *-āko-arose in the Celtic languages. In Gallo-Greek inscriptions, -Vkko-is occasionally found instead of expected -Vko-, especially when the vowel is -i-: Δονικκα, Ουαλικκο(νε), Ουηϐρουκκου, and the frag- Under close inspection, their distribution turns out to be heavily skewed. The following results are based on a small sample and are therefore only preliminary.
As Table 4 shows, some geminates have a clear predilection for suffixes, while others occur only in the root or in the semantically meaningful portion of the names. -ll-(whose possible origin as a suffix was discussed in section 2.  to the onomastic stem *att-, apparently from the etymon PIE *atto-'daddy' (see 1. (1)). Other geminates are too rare to draw clear conclusions.
-mm-occurs only once in a root, and -rr-is found two or three times, apparently always in roots.
The two examples of -pp-are too damaged to draw any conclusions. The most striking observation, however, is that voiced obstruents do not occur as geminates at all in this corpus. This links in with the overall impression that emerges from this study, namely that geminate voiced stops arose only very late through processes in the individual languages or that they are chiefly found in lexemes that are suspect of borrowing.
On a more general note, mainly based on anecdotal observations, geminates seem to be more common in Gaulish short names or in names formed with suffixes. This may be a special characteristic of Gaulish and seems to be less common in Irish. The distribution of geminate sounds in ancient and medieval Celtic names deserves research on a much wider and more diverse material basis, but this goes beyond the aims of this article. There are some practical methodological limitations to identifying relevant examples in the written record. The Lepontic script of the Cisalpine Celtic languages and the Celtiberian script do not distinguish graphically between single and geminate consonants. Geminates in these languages can only be identified when vernacular names are also transmitted in the Latin or Greek alphabets, either in epigraphy or in manuscript texts. For instance, the Celtiberian name ⟨lubos⟩ in the Celtiberian script has a counterpart in the genitive Lubbi in the Latin alphabet, but, to complicate matters, Lubus is also found in epigraphy. Cisalpine Celtic aśkoneti(o) in vernacular writing has a parallel in Adgonnetius, kasilus corresponds to Cassillus, and esanekoti possibly contains the adjective *kotto-'old'. In other cases, such as koimila and anteśilu, the lack of a Latin parallel does not permit to decide if the l is single or geminate. Furthermore, the frequency of gemination in the attested written corpus of ancient and medieval Celtic languages may give a distorted picture. Dithematic names without onomastic gemination, which may be typical of the small aristocratic elite, may be overrepresented in our available sources. It is conceivable that the frequency of short or hypocoristic names with onomastic gemination was higher among the non-aristocratic population and accordingly in everyday spoken language.
While the examples in this section are mostly taken from Gaulish, the formal processes of vocative truncation counteracted by compensatory gemination are valid for all Celtic languages. These processes established gemination as a morphophonological feature in the naming system. Through them, names, especially high-frequency variants such as hypocoristics, came to contain a relatively high proportion of geminate stops, in contrast to the rest of the lexicon, where gemination was most prominent among resonants. This relative increase in geminate stops will in turn have led to an overall reinforcement of the status of geminates in the phonological system of Celtic languages.

Non-etymological gemination: accent-related or graphic gemination
De Bernardo Stempel (2010: 71-79) has observed that many words in ancient Celtic, especially in Gaulish, display etymologically unexplained gemination, in particular in posttonic position -assuming that her hypotheses about the placement of the accent on the penultimate syllable in Gaulish are accurate. It is not possible here to repeat all of the sub-types she discusses, but a few select examples shall suffice. 11 The putative position of the Gaulish accent will be indicated by an acute accent: simiuisonna 'the name of a month' < *sēmi-u̯ ēs-ón-ā 'half spring (?)', ogronno-'the name of a month' < *ou̯ gróno-'cold one (?)'; uxello-'high' < *uχsélo-< *upselo-, cf. OIr. úasal, W uchel, Bret. uhel < *ou̯ χselo-, apparently with the same suffix, but with a different ablaut grade in the root.
De Bernardo Stempel thinks of genuine phonological gemination that is dependent on the stress. While this is possible, I do not want to exclude the alternative possibility that, at least in a subset of the material, such spellings could be a merely graphic convention in order to replicate Gaulish accentuation on the penultimate syllable within the framework of Latin suprasegmental phonotactic rules, which only permit stress on the penultimate syllable when it is positionally long.
The reverse of this rule or tendency is the simplification or degemination of geminates in pretonic position (De Bernardo Stempel, 2010: 79). A good candidate for this is the group of Gaulish names around Biracos (stress on ā) beside unsuffixed Birros, the latter being a likely cognate of OIr. berr, W byr, Corn. ber, Bret. berr 'short'.

'Expressiveness'
Another factor -psychological and therefore outside the domain of regular phonetic change -played a role in the increase of geminate sounds. A number of instances of geminates are found in words that have, or are believed to have, an affective or emotive quality. This category of gemination is commonly referred to as 'expressive gemination ' (e.g., De Bernardo Stempel, 2010: 80;De Bernardo Stempel, 1999: 508-521;Lühr, 1985: 275-276;Stüber et al., 2009: 260;Kuryłowicz, 1957: 132, 138, 142-144; these references are not meant to be exhaustive). In some kinship terms an emotional involvement is evident, especially where the creation of a geminate occurs only in a single language, e.g., OIr. macc 'son, boy' < *mak u̯ k u̯ o-, which contrasts with the older, ungeminated *mak u̯ o-that is continued in Gaul. mapon, W, Corn., Bret. mab; or W geneth 'girl' < *genettāvs. Gaul. geneta, OIr. geined < *genetā-. Perhaps *u̯ raggā-or *u̯ rakkā-'old woman' (see 6.2. (3)) and Gaul. ninno-, OIr. nen 'handmaid', OW Nennius, Bret. nen(n) < *ninno/ā-(?) 'servant (?)' can be included in this category as well. In these cases, there is no discernible phonological reason for the divergent development in just a single language. It is undeniable that a rule of arbitrary gemination, whether we call it expressive or hypocoristic, must have been synchronically operative at least in kinship terms. This was already evident in the words *atta 'dad' and *mamma 'mum', which have to be reconstructed with geminates even for Proto-Indo-European, a language that otherwise actively avoided geminate sounds. Gemination in personal names may be of a similar nature.
Interjections are a class of words that combine expressiveness with a performative aspect, which can find formal expression in gemination.
However, expressiveness as an explanatory strategy is often extended beyond these narrowly circumscribed cases to words where emotional involvement of the speaker is not easily discernible. For example, we may ask ourselves why so many Celtic adjectives for basic concepts such as 'small', 'old', 'weak' display geminates. Most of them cannot be captured by an emotion-based concept of expressiveness. Instead I propose to use the more neutral, descriptive term 'symbolic gemination'. It takes its motivation from the concept of sound symbolism that refers to a perceived resemblance between the phonological structure of a word and its meaning. Sound symbolism can have a strong element of iconicity and onomatopoeia, but, like almost anything in language, a lot of the symbolism is arbitrary and can be triggered by culturally and grammatically specific conditions that, in the case of prehistoric languages, must remain obscure to us.
The etymologies discussed in this study so far comprise phonologically or morphophonologically motivated instances of gemination in inherited words (if we consider onomastic gemination as a morphophonological process). 'Symbolic' or 'iconic' gemination, on the other hand, is not triggered by phonetic cues, but is motivated by pragmatic considerations such as the semantics of words, or it has no discernible motivation at all -or at least none that is recoverable to us. It translates a special relationship, which speaker and hearer have to a concept, into linguistic markedness of words.
I am aware of the danger that any non-native classification of words as 'symbolic' may be arbitrary. I have therefore decided to restrict the category of symbolic gemination in Celtic to two types of words: onomatopoetic nouns (in fact a single example) and adjectives. Other scholars may apply different criteria.

Adjectives
One semantic class of words in which phonologically obscure gemination is noticeably frequent, are comparable adjectives, very often those for basic concepts such as physical qualities. My speculative approach to explaining this descriptive fact is that the iconic component of gemination allowed to give expression to uncertainty about the precise magnitude of those qualities, i.e., it was an iconic way of 'hedging one's bets' (in the same way as speakers of German tend to qualify adjectives by irgendwie 'somehow', in order not to commit themselves too strongly to a specific value). In addition to their symbolic or iconic component, most of the adjectives in this category are also notable for being etymologically opaque. Many of them are not only found in Insular Celtic, but have parallels in Gaulish. Even though the contexts in which they appear in Gaulish, namely preponderantly in personal names, do not allow us to determine their semantic content beyond all doubt, the formal correspondence with words known from Insular Celtic renders their interpretation as adjectives fairly certain. For practical purposes, the material will be arranged in groups that reflect their distribution. Some adjectives are found in several languages, and some are exclusive to a single sub-branch. In those cases that are found across several branches, borrowing between the branches can never be excluded.

Inherited formations
Only a subset of adjectives with geminates is of more or less good Indo-European inheritance. They are being repeated in this panorama for the sake of the overall argument, even though they have already been discussed in previous sections.
(1) PC *dallo-'blind' > Gaul. dallo-, OIr., W, Corn., Bret. dall. As discussed in section 2.1. (5), it is attractive to explain this as PIE *du̯ l̥ no-, but the lack of preserved *u̯ in the potential Gaulish examples is worrying. If the word is of Indo-European origin, its geminate can be explained through regular sound change from *ln.
(2) qPC *dorro-> Ir. dorr 'harsh, rough' and the abstract doirr 'anger'. Both words are only attested late. An Indo-European etymology *dor-so-from the root *der-'to tear' has been suggested, but such an explanation is by no means unavoidable.
(13) If the Gaulish names Suallius/a, Sualius are cognate with Ogam Ir. SUVALLOS and OIr. súaill, súail 'small, trifle', they probably have to be analysed as compounds of *su-'good' + an element *u̯ al(l)-i-(in the case of OIr. súail(l) hardly identical with the root *u̯ al-'ruler, to rule'). It is uncertain if the geminate ll is original or caused by individual secondary factors within each language.
(2) A mechanically reconstructed immediate preform of MIr. láitir 'strong, powerful' would be something like PrimIr. *lāddiri-. Perhaps the conspicuous fact that the degrees of comparison, e.g., comparative laiteri, and the derived abstracts láitire and láitirecht show no syncope, is an indication that the preform was more complex, e.g., *lāde/idVri-. No analysis suggests itself for any of those reconstructions.
(4) MIr. prapp 'sudden, swift' could be a language-internal neologism that makes iconic use both of the phonologically highly marked sound p and of gemination, or rather its reflex, i.e., a voiceless unlenited intervocalic stop. Alternatively, it may reflect a symbolically reduplicated Latin loan, namely rap-rapidus (personal communication, Paulus van Sluis).
Where do the adjectives without Indo-European pedigree come from? What is notable about this collection of adjectives, which strongly relies on Irish, is that, apart from being formally united by gemination, they form several semantically coherent subgroups. They relate to basic physical concepts; to physical defects; and they are symbolic of swiftness and agitation. Gemination was phonologically original in the subset that continues Indo-European formations. From there, it can have spread as a semantic marker within the semantic field. External influence need not be invoked. It is evident that words such as *biggo-'small', *buggo-'soft', *laggo-'weak' constitute a 'phono-semantic' class whose structure was self-replicating. That such a process was productive, is evident from words such as prapp 'rapid, quick', which, if it isn't a substratal loan, can have been created sound-symbolically at a late date.
Another reason for ascribing these words to Urschöpfungen within Celtic, and not to loans from unknown languages, has to do with the typological tendency of adjectives to be less amenable to borrowing than nouns. The fact that such a number of conceptually central adjectives display gemination points rather to sound-symbolism as the driving force. Once gemination had been established as a phono-semantic marker for basic adjectives, there must have been enough internal momentum to create new material through this morphophonological process, without taking recourse to borrowings from a substrate.

Unclear gemination
A small number of nouns and verbs of certain or likely Indo-European origin possess geminates that cannot be explained straightforwardly from their putative preforms by phonological developments. Neither do these words manifest an affective character. For some of them, I will make tentative morphological proposals, usually involving some sort of derivational process that can account for the gemination. However, these attempts are merely speculative. In some words, the gemination has to remain unaccounted for for the time being.

nouns
(1) PIE *h₁eh₂no-'ring' > *āno-→ *ānnii̯ o-> OIr. áinne. This item is related to Lat. ānus, Arm. anowr. Maybe the derivation involves the addition of the suffix -no-, i.e., *ān-no-, to explain the geminate, or it is due to the syncope of a preform such as *ān-in-i̯ o-. The lack of Osthoff shortening and the palatalisation speak in favour of the latter option.
(3) The relationship among OIr. ícc 'healing' < *īkkā-and W, OCorn. iach, Bret. yac'h 'healthy' < *i̯ akko-or *i̯ ekko-is notoriously difficult, as is its extra-Celtic relationship, if there is any, with Myc. a₂-ke-te-re, ja-ke-te-re, Gr. ἄκος 'cure, medicine'. Matasović (2009: 171) and Zair (2012: 68) tentatively and fully conscious of the formal problems think of *i̯ h₂ko-as a starting point, but neither addresses the question of the geminate. Schrijver (1995: 103-104) dismisses the link with Greek and instead suggests, also tentatively, a formation *i̯ et-ko/ā-from the root *i̯ et-'to position onself'. Although it cannot account for the vocalism of the Irish word, it explains the geminate. Finally, I want to join in the tentative chorus myself. A possible source for the geminate could be a Proto-Celtic adjectival formation *i̯ ek-ko-from the root *i̯ ek-'to speak'. The semantic connection would be the premodern notion of healing through words of power. But I need to stress that aside from explaining the geminate, this etymology creates other issues with the initial vowel.
(7) qPC *truddo-from the root PIE *treu̯ d-'to push' is set up in IEW 1095 as the preform of OIr. troit 'fight, battle', but neither word-formation, stem-class nor vocalism are satisfactorily explained by this. If an Indo-European connection is abandoned, the word can also be set up as *troddV-or *trontV-.

verbs
Only a select number of Old Irish verbs will be included here.
(2) On the evidence of ModIr. dímhigean 'contempt', OIr. do·meicethar 'to despise, condemn' has a medial /g/. This precludes the reconstruction ?*mik⁽'⁾né/n-h₂-, tentatively proposed in LIV 429. Operating with the formation *mi-n-k(h₂)-e/o-, where the position of the infixed n would have to be secondary, results in PrimIr. *migge/o-. This explains the medial g of OIr., but leaves the stressed vowel e unaccounted for. W edmygaf 'to admire, honour' and Bret. dismeg 'opprobrious' < *-mik-seem to continue a formation without a nasal infix, unless all go back to a preform *migg-.

Possible loanwords
After having scrutinised the possible sources and the types of geminate consonants in the lexicon of the Celtic languages, the question to what extent geminate sounds can be relied upon as indicators for the substratal origins of words can finally be broached. This applies of course only to words for which no etymological explanation within the Indo-European paradigm can been found, and to words for which no case -how ever flimsical -for a sound-symbolic neologism can be made. Some words with geminates are evidently foreign, a case in point being *katto-'cat', which is a trans-European Wanderwort, perhaps ultimately originating in a Bronze-Age Afro-Asiatic language such as Nubian. At the same time, this word is atypical for the substratum loans that we are interested in here since it refers to a foreign concept brought in from outside, not to one that was encountered in situ by incomers.
After having dealt in the preceding sections with words for which some sort of language-internal explanation can be found, we are left with a list of items that have no good etymology and no obvious source. This makes them suspect of being loans. The collection below expands on Matasović's (2009: 441-443) list of suspected non-Indo-European loanwords in Celtic, enlarged by additional examples that have been found in non-systematic searches. Some of the words have manifest parallels in languages outside of Celtic, but formal irregularities prohibit their reconstruction for the Indo-European protolanguage or even for a common Western Indo-European subnode: (1) qPC *ātti-(?) > OIr. áitt 'place'. A variety of Indo-European etymologies have been proposed for this word, all requiring the 'Kluge's Law' treatment *-tn-> *-tt- (Pedersen, VKG i 161: *pōthni-;Klingenschmitt in Lühr, 1985: 303: *ō-i̯ et-nā-;Bammesberger, 1998: *pōtni-). Since the validity of Kluge's Law for Celtic is not accepted in this article, the word is only mechanically projected back to a possible Proto-Celtic pre-form, without further analysis.
(30) qPC *met(t)o-(?) > OIr. meth 'decay', W meth 'failure, error', Corn. meth, Bret. mezh 'shame'. The irregular correspondence between the Irish and the British words suggests that one branch must have borrowed the word from the other.
In the absence of an extra-Celtic cognate, establishing even the Celtic preform is impossible. If it was *meto-, Irish meth is regular and the word was borrowed into British (thus Bauer, 2015: 71-73). If it was *metto-with a geminate, it is the other way round. Because of its isolation, the word may be a substratal loanword, in which case the probability of *metto-increases.
(32) W pwll, Corn. poll, Bret. poull 'hole, pit; pool, pond'. OIr. poll 'hole' is evidently a borrowing; usually its source is believed to be Welsh. A mechanical back-projection results in qPC *k u̯ ullo-, which is phonologically implausible, since labiovelars were delabialised before *u in Proto-Celtic. A borrowing from OE pól 'pool' is phonologically and genetically unlikely: it cannot explain the vocalism of the Welsh and Breton words and a loan from Old English would not be expected to be present in all three British languages. Alternatively, it could be a substratal borrowing from the so-called 'p-language', a pre-Celtic language postulated to have been present in Ireland as late as the first half of the 1 st millennium a.d. by Schrijver (2000;2005; see also Stifter forthc.).
(37) qPC *tanno-'green oak' > Gaul. tanno-, Bret. glastannen. OIr. tinne, which is glossed as 'holly, elder' in some glossaries, should be kept separate. The word is perhaps only an artificial construct to press the ogam letter into an arboreal scheme. Formally and semantically, Germanic Old Saxon danna 'pine', OHG tanna 'fir wood' have to be kept apart.
On the other hand Matasović, for whom the Celtic words for 'swallow' are inherited, reconstructs *u̯ ennālā-< *u̯ esn-, derived from the Indo-European word for 'spring'. However, this reconstruction cannot explain the vocalism of the Irish word, nor of the Gaulish, if it belongs here. The different reconstructions at least agree in setting up a geminate *-nn-for the preform.

Assessment
We can now proceed to a provisional assessment of this collection. Most of these words are isolated and lack parallels outside Celtic. Where parallels exist, they are often in Germanic (but this may be partly a consequence of how the data was compiled). It is therefore conceivable that these words may have been borrowed from unknown languages in the prehistory of Celtic, in Western Europe or on the Western Archipelago. Unlike the adjectives in section 11.3., the borrowing of nouns is typologically a trivial phenomenon. The case for borrowing can be made more rigorously if it can be shown that the candidate words share specific phonological or morphological features apart from gemination. A number of relatively unspecific observations can indeed be made (I will occasionally also include words from previous sections in this discussion): Many of these words are monothematic and monosyllabic (if we disregard the Proto-Celtic ending), where gemination is found at the end of the stem or root syllable. From a Proto-Celtic perspective, the geminate sound occurs in the onset of the second syllable that also contains the ending. Typical examples are *bratto-or *gobbo-. Examples with more than one syllable are *ētt s (V)lūm(b)-, *gobann-, *kappilo-, *karrekā-, *kīmmukko-, *krokkeno-, *mekkVno-, *u̯ annellā-; *kau̯ anno-takes a special place since *-anno-can be isolated as a suffix (see Jørgensen forthc.).
Another feature is the preponderance of short vowels, or rather the absence of long vowels, in the root syllables (cf. Van Sluis forthc.). Including disyllables, the most common vowels are a (11) and o (6 or 7); e (5) and u (4 or 5) lag somewhat behind. Intriguingly, i is absent from the collection, unless the obscure Gaul. ill(i)o-is a substratal borrowing. This is in contrast to the adjectives in section 11.3., where i is not uncommon. Exceptions with long vowels are *ātti-, *ētt s (V)lūm(b)-, *kīmmukko-, *nūsso-, items that are all only attested in the Insular Celtic languages. Maybe they reflect borrowings from a local prehistoric language in these islands. Perhaps items such as *rou̯ kko-and *u̯ ei̯ ttā-, that have been mechanically reconstructed with a Proto-Celtic diphthong, can be referred to the group with long vowels as well. Among the words with geminate stops, the majority have voiceless sounds (the comparative dearth of geminate voiced stops was already highlighted by Martinet (1952: 198)). Exceptions are *gaddo/ā-, *gobbo-, *skaddo-and *sle/iggii̯ o-. *Braddo-, *kladdo-and *tru/oddo-could be further instances, but since they are restricted to Irish, their *dd could in theory continue earlier *zd. *Kloggo-is a special case since it may be onomatopoetic. Among geminated resonants, preponderantly nn and rr are found.
The words have been preponderantly adopted as o-or ā-stems.
From the words included in the group with possibly Indo-European origin (section 12.1.), *krutto/ā-, *makko-and *tru/oddo-would also fit formally into the group of potential loanwords. On the other hand, *φlikkā-and *φrikko-stand apart with their radical i.
One must never lose sight of the fact that the borrowings need not come from a single source. The formal difference between some of the words can be an indication that more than one donor language is involved. The phonetic shape of the extant words owes as much to the target phonology of the Celtic languages to which it had to be adapted, as it does to the donor languages. Depending on when the borrowing occurred, the donor language need not have possessed geminates at all: if, at the time of borrowing, the receiving language already had allophonic lenition of inherited intervocalic stops, any foreign unlenited intervocalic stop could have been perceived as a geminate.

The insular outlook
At the end of the developments described in the foregoing chapters, the phonological system of Celtic had morphed into the one in Table 5.
Automatic, positional allophones are in brackets. This system reflects, in a somewhat idealised fashion, the state of affairs in the ancient Celtic languages and in the prehistoric Insular Celtic languages before the emergence of phonological lenition. For the descriptively p-Celtic languages, *p has to be substituted for *kʷ. The superscript question marks after the geminated labiovelar stops indicate their somewhat uncertain character. They cannot have been very numerous in the first place, but if they existed (for instance as a precursor to OIr. macc 'son'), it is unclear if they were true geminates, i.e., *makʷkʷo-, or rather sequences of velars + labiovelars, i.e., *makkʷo-. The relative frequencies of geminate stops, especially of the voiced stops, may have differed between the individual Celtic languages.

Irish
After it had taken such a long time to build up a phonological system that had an opposition of length for almost every consonant, it is surprising to see how quickly this system disintegrated even before the full attestation of the medieval Celtic languages set in. Chiefly responsible for the quick abandonment of gemination, at least on the phonological level, was the rise of the opposition between unlenited vs. lenited sounds in the Insular Celtic languages. The emergence of lenition first as an allophonic feature of single stops in intervocalic position may well at first have been a side effect of the strong opposition between single vs. geminate consonants, caused by a maximum phonetic polarisation of the manner of articulation (see also Martinet (1952: 198-203, 215-216) for a more phonetically oriented account of the loss of gemination and for the systemic implications of gemination as a prominent phonetic class). However, after the contrast lenition : non-lenition had emerged as the central phonological opposition in Insular Celtic, the gemination of stops lost its phonological significance and turned into a phonetic feature that was concomitant with non-lenition, indicated by the bracketed length marks in Table 6 below. The old opposition geminated : ungeminated was translated into the new phonemic opposition unlenited : lenited only among resonants. The resulting phonemic system of Primitive Irish is shown in Table 6.
For the ancient, medieval and early modern Gaelic languages, information about the phonetic status of the sounds can only be inferred indirectly from the written sources. The fact that single and double resonants were distinguished fairly consistently in manuscripts up to the modern period is an indication that the graphic distinction corresponded to a real-life phonetic distinction. At no point in the history of Irish can a comparable consistency be observed for the writing of unlenited obstruents. This is not the place to discuss the intricacies of Old, Middle and Early Modern Irish spelling conventions.  Suffice it to say that, for example in Old Irish, the representation of word-internal unlenited stops by single or double letters appears to have been chiefly a matter of personal discretion of each scribe. This is exacerbated in later copies of Old Irish texts by the interference of a host of alternative spelling conventions for Modern Irish.
For the modern languages, the earliest recordings of traditional speakers allow some insight into the phonetic realisation of the sounds. For Irish, these are the recordings on wax cylinders by Rudolf Trebitsch in 1907 (kept at the Phonogrammarchiv of the Austrian Academy of Sciences) and the recordings by Wilhelm Doegen on shellacs, carried out between 1928-1931Conroy et al., 2009). Recent phonetic studies of the recordings of Donegal Irish in the Doegen collection shed some light on the phonetic length of unlenited sounds (Wheatley & Iosad, 2021). It appears that in the early 20 th century, a contrast in length was still perceptible for some sounds, best preserved among resonants where the researchers noted perceptible length distinctions. In contrast, there was no significant length distinction in obstruents. These findings agree with the orthographic obervations made above.

British
The details of the prehistoric developments of geminated sounds are rather different in the British Celtic languages and can only be sketched here. A striking difference to Irish is the sound *p as the regular outcome of the labiovelar *kʷ in the phonological system of British Celtic. It seems that the voiced labiovelar *gʷ was reinterpreted as an allophone of *w and that the two occurred in complementary distribution. When the merger of *w and *gʷ happened is unclear.
The transformation of the Proto-British system ( Table 7) into that of the individual neo-Celtic languages Welsh, Cornish and Breton has been the subject of a long controversy (see the section on previous research in the introduction). From a structural point of view, the operation of intervocalic lenition on single stops removed single voiceless stops from intervocalic position for a short period. In response to this, a chain-shift was set in action in which gemination lost its phonological significance and became merely allophonic with non-lenition, or, in other words, geminate voiceless stops refilled the now empty phonotactical slot of simple voiceless stops. Geminated voiced *b: and *d: fell together with the new simple *b and *d, which themselves were the results of the earlier lenition of simple *p and *t. 12 The case of *g: is less clear. It may also have developed into *g, thereby falling together with the product of lenited simple *k, but in some cases it seems to have merged with *k: in yielding *x (see the discussion in section 6.2.). Unlike in Irish, there was probably never a stage where geminates occurred word-initially as allophones of unlenited sounds. See Van Sluis (2019: 30-35) for an account that differs in the details with regard to initial consonants.
The operation of syncope in Late Proto-British brought previously intervocalic obstruents and fricatives into contact with each other and thus created new complex consonantal clusters, which were prone to assimilation. The outcomes of this process called 'provection' in British historical phonology are voiceless stops, e.g., the Welsh river name Calettwr < *kaled † đuβr 'hard-water' < *kaleto-dubro-, Bret. klopenn 'skull' < *klog † benn 'stone-head' < *kloko-k u̯ enno- (Harvey, 1984: 98-99). By virtue of two segments coalescing into one, one could expect that the assimilation product was a long sound at first, at least phonetically and at least for a short period of time. Van Sluis (2019: 69-74) argues that word-external provection (i.e. in sandhi) also produced long consonants up to Early Middle Welsh.
However, the ultimate result word-internally is a phonologically single, voiceless stop (Harvey, 1984: 99). Provection triggered yet another chain-shift by which existing single voiceless stops, i.e., the old voiceless geminates, underwent a new round of phonetic weakening when intervocalic inside a word, after resonants, or within accentual units. This weakening, which has the appearance of a 'secondary lenition' (Greene, 1956: 289), but which is traditionally called 'spirantisation', resulted in the introduction of voiceless fricatives into the system, i.e., old *pp > *p > *f etc. In order to grasp the bewildering aspects of the diachronic treatment of geminates in British Celtic it is quintessential to rigorously differentiate between phonology and phonetics. Different classes of sounds (voiced stops, voiceless stops, resonants) were degeminated first phonologically, then phonetically, at different times (Isaac, 2004: 70). Like in the Gaelic languages, geminate resonants remained longest as a phonological class and their simplification only occurred during the historical period.
The system immediately before the emergence of the individual British languages was therefore the one in Table 8. (ː) indicates that length may have remained for some time as a phonetically concomitant feature, even though it played no role in phonology (cf. also Schrijver, 2011: 30-33).

Conclusions
Geminate consonants were a distinct and prominent phonological class in the history of the Celtic languages, namely in the attested ancient Celtic languages of the Continent, as well as in the reconstructable stages immediately before the attestation of the medieval Celtic languages in the insular world.
Geminates also played an important role on the interface between phonology and morphophonology. The phonology of the historically attested Insular Celtic languages, on the other hand, has evolved into different systems where gemination as a phonetic feature is concomitant to other oppositions at best.
While sound change in inherited lexical items can explain the regular emergence of the core of this Common Celtic phonological class, it cannot account for all the types of gemination and for the entirety of the examples. By and large, a dichotomy is discernible in the Celtic lexicon between, on the one hand, words with geminate resonants and a 'strong' sibilant, and, on the other hand, words with geminate stops. Very crudely put, geminates of the first group tend to have good Indo-European etymologies, whereas those of the second group are less amenable to explanations along traditional Indo-European principles, unless they occur across a morphemic boundary especially in compound formations. In the latter, assimilations across the morpheme boundary contributed to the rise of geminates, but to different degrees in the different Celtic languages. The evidence especially of ancient Celtic personal names indicates that the various sounds showed a skewed distribution between the root and the suffixal part of words. This observation illustrates that gemination had acquired a morphological function in addition to its purely phonological status. Voiceless geminate stops were fairly common in Proto-Celtic, but voiced geminate stops are virtually absent from Proto-Celtic. Words in which the latter did not arise across morpheme boundaries, and in Goidelic as the product of the change *NT > *DD, are mostly either strongly suspect of being sound-symbolic neologisms or borrowings from unknown sources.
Although there is no clear-cut demarcation line between the groups, it is useful for analytic purposes to divide the material into one with phonologically explicable gemination and one with gemination that has no known diachronic source. In many of the steps outlined in this survey, gemination can be seen as a phonetic strategy to reduce the number of phonotactically admissable consonant clusters while retaining the moraic structure of the words. Etymological gemination is thus a compensatory reaction on the prosodic level to the loss of phonological segments. In the case of non-etymological gemination, two possible sources are conceivable: either they are loans from prehistoric substratal languages, or they are Urschöpfungen, i.e., neologisms created within Proto-Celtic or in the individual languages. For typological reasons it has been suggested above that nouns rather belong to the first category, whereas adjectives belong to the second.
Over the long period of at least one and a half thousand years, a variety of factors conspired and various processes reinforced each other to slowly increase the number of types, and the number of tokens, of geminate consonants in the Celtic languages, but there was not just a single pathway to this. Some of the earliest steps in this direction constitute 'uneven' phonological change, when, for instance, *sm and *ln became *mm and *ll already in Proto-Celtic, whereas *sn and, perhaps, *sl did not immediately follow suit. The material assembled in this investigation also demonstrates that the emergence and treatment of geminated voiced stops went rather diverse paths in the Celtic languages.
In addition to sound changes that provide purely phonological factors for the rise of gemination, there are also psychological drivers, namely borrowing, symbolism and pragmatics.
In previous scholarship, expressiveness has been claimed to be an important factor, but this has not proved to be a useful concept on any grand scale. While the emotional attitude of speakers clearly plays a role in kinship terms, that segment of the lexicon accounts only for a tiny portion of instances of gemination. After the systemic establishment of gemination as a phonological category, it may rather have been the concept of sound symbolism that occupied a central role in the further spread of gemination, in particular in the creation of adjectives describing physical reality. It is very rare that we find, as it were, 'spontaneous' gemination in words of Indo-European inheritance, i.e., gemination that can neither be explained by phonological nor by psychological factors.
Because of the imbalance in documentation between Continental and Insular Celtic languages, it is difficult to assess if the frequency of gemination, especially of obstruent sounds, is higher in the latter than in the former, that is to say, if their frequency increased over time. However, it is noteworthy that geminated voiced stops are surprisingly rare, albeit not absent, in Gaulish. They may have been reinforced as a numerically substantial class only in Insular Celtic. Throughout the observable history and the reconstructable prehistory of Celtic, it is always the resonants that have been particularly prone to the creation of new instances of gemination (or rather 'fortification' during the younger stages of the Insular Celtic languages, when one can no longer speak of true geminates). However, when 'full saturation', as it were, of gemination was achieved in the Insular Celtic languages, it was quickly abandoned as a phonologically distinct class. Of the data analysed in this study (excluding personal names), over 50 lack an obvious inherited, i.e., Indo-European explanation. We may allow for the possibility that a subgroup of words with geminates was created language-internally through the operation of analogy and sound symbolism. This view is taken here especially for adjectives. However, the overall number of unetymologised words is too large to ascribe them entirely to those factors. It is suggestive that at least a subgroup of them, in particular nouns, were borrowed from prehistoric local precursor languages in Western Europe and in the Western Archipelago. Geminates are only one sign for substratal loans, and ideally they should go hand in hand with other markers for non-inherited origin, such as other unusual phonological and morphological features. The question to what extent the number of unetymologised words with geminates in Celtic is different (larger, equal, or lower) than in other languages of Europe, especially the Germanic (cf. Kuiper's (1995: 68-72) source 'A2' for loans into that language branch) and Italic languages and Greek, goes beyond the present article and needs to be investigated in a broader, comparative context.
In any case, from the foregoing it emerges that Celtic is an Indo-European branch that seems to contain a substantial body of words with geminate sounds borrowed from unknown substratal languages. The question imposes itself if this observation may be linked with the fact that in the Celtic languages intervocalic single stops underwent weakening through lenition, that is to say that unlenited single intervocalic stops of the donor language were perceived by speakers of Celtic as their only available unlenited intervocalic stops, namely geminated. The evidence for such a scenario is not clear-cut: Single intervocalic voiced stops were probably already allophonically lenited in Proto-Celtic. This would render foreign unlenited voiced stops particularly prone to be replaced by geminates in the borrowing process, but in fact voiced geminate stops constitute only a small minority among the words with unknown provenance in the Celtic languages. It is therefore best to assume that the donor language or languages did indeed have geminates.

Ethics and consent
Ethical approval and consent were not required.

Data availability
All data underlying the results are available as part of the article and no additional source data are required.

Overview
This article presents consistent and comprehensive evidence to support a scenario of the development of the reconstructed consonantal phoneme inventories of Proto-Indo-European and Proto-Celtic that motivates the emergence of phonologically long (aka geminate) consonants in the Celtic branch and the later restructuring of length oppositions in Insular Celtic languages.
The article argues extensively and mostly convincingly for the emergence of long consonants (predominantly resonants and secondarily voiceless stops) in Proto-Celtic mainly as a result of assimilation of consonant clusters, both original clusters and, notably, clusters arising through word-formation processes and therefore morpheme encounters. These developments are dealt with in detail distinguishing common phonetic outcomes within the Celtic branch from divergent outcomes of original clusters (Sections 3 and 4) and different chronologies and kinds of wordformation processes involved (Sections 5 and 6). The paper also highlights the relatively smaller evidence for long voiced plosives in Proto-Celtic and in earlier attested Celtic languages such as Gaulish, a fact which might be related to the divergent developments of voiceless and voiced stops in lenition contexts in the two Insular Celtic branches.
The author suggests that reconstructible long consonants for which etymologies cannot be grounded on attested forms in Indo-European languages can be due to other factors, such as affective connotations and sound symbolism, especially in word classes and semantic fields where these phenomena are purportedly more frequent (address forms of kinship terms and personal names, adjectives denoting physical qualities). What cannot be arranged in any of these files is reasonably, albeit tentatively, ascribed to prehistoric borrowing (lexical copying) from unknown substratal languages. A merit of the article is that it thoroughly engages with the reconstruction and motivation of forms for which long consonants might be based on affective connotations, sound symbolism, and even borrowing (Sections 9, 11, 13), so that, in an attempt to avoid circularity, these residual groups are not just unaccounted for left-overs.
Doubtful issues and problematic cases are not concealed (Sections 10 and 12).
The comments that follow are therefore not to be meant as major criticims, but rather as suggestions that, it is hoped, may help improve the clarity and persuasiveness of the paper. They are grouped under three headings: 1) request for consistent terminology, 2) comments on specific points, 3) typos and minor slips.
Reference is made below to pages in the pdf file and to numbered sections; abbreviations are the same as the paper's, if not given in full.

1)
In general, there are a few terminological inconsistencies, probably due to cumulative drafting, that could be addressed, to the reader's benefit: I suggest that a single term, either "plosive" or "stop" or "occlusive", be used throughout. If these terms are not synonyms in the author's usage, this should be clarified. Similarly for "fricatives" and "spirants" (e.g. p. 20 § 6.2 "spirant sounds"). "Obstruents" is yet another term that in some loci but possibly not throughout the paper is used for stops/ plosives/ occlusives (e.g. cf. p. 22 "a geminate voiced obstruent resulted", p. 36 "intervocalic obstruents and fricatives"), while obstruent is in fact a term that includes both stops and fricatives.
Either "velar" or "guttural" should be chosen (again, there does not seem to be any difference between their usage, so a single term would be preferable).
For the mid-high front vowel in Goedelic, ɪ is used on p. 23, which is the IPA symbol for mid-high front (unrounded), while on p. 22 ( § 7 and § 7 (4)) a different symbol ɩ, possibly rather used for high central unrounded, is used instead.
On a slightly different note, although the widespread terminology "mediae and mediae aspiratae" (p. 6) for the consonant inventory of PIE has a long tradition, it could be improved through greater consistency, allowing conversion with the terms used for the daughter languages and in general.

2) Comments on specific points:
p. 4 (Preliminaries): "*ballo-, which is meant as equivalent to a phonetic analysis as [balːo-]." The phonetic transcription [balːo-] (rather [balːo]-) here seems to imply that a long consonant in a reconstructed form is not considered phonologically relevant (e.g. an allophone between vowels in this example); this is not generally the case, as reconstructed forms are rather closer to broad, phonemic transcriptions (put between slashes //). Anyway, this point should be better clarified. Also, it should be clarified if any stress and syllabic border is meant or implied in reconstructed forms.
p. 4 (Preliminaries): "The main contrast between the two classes of Celtic stops is considered to be between 'voiced' (= D) and 'unvoiced/voiceless' (= T) consonants". It should be made clear from the outset (and not simply implied or taken for granted) whether the main phonological contrast between the two classes of stops (voiced vs. unvoiced) for the purposes of the paper is deemed corresponding, i.e. not a different opposition, to a phonetic contrast between singleton and long (= lenis and fortis respectively). This is not so, but what follows here seems to suggest the contrary.
p. 4 (Preliminaries) and throughout: The label "quasi-" in qPIE and qPC is slightly misleading in this context, as it does not correspond to usual practice both for attested and reconstructed languages: in fact, modifiers distinguishing different phases usually point to chronological order (pre-, post-, early, late, old, middle, modern), while "quasi" refers to an approximation from either direction, or is rather synonymous with pre-. Therefore, quasiPC could be meant as a phase of late Indo-European approximating the separation of the Celtic branch. Clearer labels would be useful.
p. 6, Table 1: p is inserted in the bilabial fricative case, but I suppose this is a slip: a fricative should be represented by a symbol for a fricative (e.g. for a voiceless bilabial fricative [ɸ]); [p] is a plosive that possibly lies behind a fricative consonant at this stage. This development from PIE has not been mentioned here before, however (see the paragraph listing the main changes in the consonant system from PIE to Proto-Celtic -beginning with "After" and ending with " Table 1 in Section 1", p. 6). So, either p should be put in the right column (plosive) or fricativisation of the bilabial plosive should be mentioned and a fricative put in this slot in Table 1 (see § 2.1 (14) for an instance of this development).
p. 7 § 2.1 (9) "*salʲ †nʲī-": The use of the symbol † to signal the position where a vowel has undergone syncope in the prehistory of Irish should be mentioned.
p. 14 § 3.9.1 (3): "constitutes the root of OIr. impersonal verb do·cuäs 'has gone'." It is rather an impersonal verb FORM, traditionally a singular passive (perfect) form, something which is different from an impersonal verb (i.e. a verb that is not construed with a nominative argument). The point is not totally irrelevant here since the Indo-European formation with *-to-lies regularly behind OIr. preterite/perfect passive forms. From the point of view of both their etymology and the Old Irish paradigms, the examples in § 3.9.2 (4) and § 3.9.5 (3) are the same; it is only the English translations that need to resort to "impersonality" with an intransitive verb such as 'to go'. Perhaps a different translation, anyhow, could be envisaged, e.g. 'there has gone', 'someone has gone'.
p. 17 § 5.1 (6) "Old Irish has two distinct verbs that influenced each other formally: in PIE *sn̥ -n(e)-h₂-'to obtain' [...] PIE *su̯ n̥ h₂-'to sound' > *su̯ ana-acquired geminate *nn probably after its model, i.e., *su̯ anne/o-> OIr. seinnid 'to sound, make music'." This kind of analogy at an early stage is frankly hard to believe, if the verbs were unrelated and, at the time when gemination was taken over by the second verb, length was distinctive for *n. Although mutual influence between the two verbs can be conceived, either a different explanation could be envisaged for geminate n in seinnid 'play music', or a later analogy should be invoked.
p. 21 § 6.3 (1) "Alternatively *ku̯ etur-could be an archaism preserved in this compound, or petor-is simplified from the younger Gaulish form *petu̯ or-with influence from the cardinal *ku̯ etu̯ ores '4'." Since these alternatives are simpler, it would be better to mention them first, and then put forward the hypothesis which implies two successive metatheses.
p. 22 (end of § 6.3) "There was a contrast between s and a strong sibilant..." If this is mirrored in the opposition between s and ts, it seems rather an instance of fricative vs. affricate. Otherwise, this point could be better qualified.
p. 22 (end of § 6.3) "On the whole, geminated voiced stops were rare except across the morpheme boundary." If that is so, it could be wiser to put geminated voiced stops on a different level as opposed to geminated voiceless stops: Table 2 puts all geminated consonants except nasals and liquids between brackets, which means that they are all considered allophones of the corresponding singletons, but if geminated voiced stops are so infrequent and limited to morpheme boundaries, one wonders whether they can be put on a par with the other allophones and should not be considered simply consonant clusters (where sub-phonemic alternations were rather between assimilated and non-assimilated clusters, than between short and long consonants). The point is that consonant encounters in word formation do not automatically create geminated allophones, but only (allowed) consonant clusters. The absence of true geminated allophones for voiced stops at this stage could be linked to their later divergent developments ( § 6.2).
p. 23, fn. 3 "3 rd singular of the copula": insert "present indicative" p. 24 Caption of Table 3: "preverbal 'in'" looks strange, as forms between single inverted commas are usually translations (is this what is meant here?) On a different note, the reconstructed forms in the Table are not straightforward, since allomorphy of the preverb is not clearly grounded on the basis of reconstructed contexts but rather on the basis of OIr. outcomes and in particular stress positioning. An independent reason, either phonetic or other, should be put forward for each previous stage allomorphy. Also, at this point it is not entirely clear how relevant the reconstruction of these allomorphs is for the main purpose of the paper. Since the discussion is meant to justify why *in+voiceless stop does not turn into i/e+voiced stop in OIr. in some cases, the starting point and the detour of the discussion should take a different stance.
p. 24 fn. 5 This reconstruction assumes that *et is the particle that occurred in main clauses between preverb and verb-root and after a simple inflected verbal form. This option should be clarified.
p. 24 ( § 7) "*igg-before a back vowel > *egg-." Why a back vowel and not a lower non-front vowel? Does this occur with a following u?
p. 25 § 8.1 "the outcome f (phonologically equivalent to a geminated *u̯ !) is identical to that of the internal group *-su̯ -, but is not phonetically similar to single *u̯ , which rather gets lost." It gets lost if internal, but not if initial.
p. 26 § 8.2 While it is generally acknowledged that gemination did not play a role for the opposition between lenition and non-mutation, as it did in Goedelic, it is commonly held that it played a role for Welsh fricativisation of voiceless consonants (voiceless spirantisation; see the summary in Willis 2009 and § 14.2 in this article). At least some lines here about this development would clarify the picture and the author's stance in its regard and in relation to Section 14.2.
p. 27 § 9.2 "in the root or in the semantically meaningful portion of the names". Suffixes and endings are also meaningful (as all morphemes, by definition), so here it would be better to put the morphological predilection in different terms (e.g. semantically intense part?). Usually content vs. function word is used for lexemes, but that would not easily work for morphemes in name formation.
p. 27 § 9.2 "in Gaulish short names or in names formed with suffixes." This should be clarified: it is not entirely clear to what kind of names these are contrasted (names formed by composition, supposedly?).
p. 27 § 9.2 "came to contain a relatively high proportion of geminate stops, in contrast to the rest of the lexicon, where gemination was most prominent among resonants." This appears to be true for voiceless geminate stops, but not voiced ones (as per Table 4).
p. 28 § 10 "The reverse of this rule or tendency is the simplification or degemination of geminates in pretonic position". This simplification would not be warranted in the second scenario just mentioned, i.e. that gemination after penultimate stressed syllable is due to Latin stress position rules.
p. 29 § 11.1 "restrict the category of symbolic gemination in Celtic to two types of words: onomatopoetic nouns (in fact a single example) and adjectives." Cross-linguistic parallels for these kinds of expressive gemination would be useful for the sake of the argument.
p. 29 § 11.3.1 "Only a subset of adjectives with geminates is of more or less good Indo-European inheritance. They are being repeated in this panorama for the sake of the overall argument, even though they have already been discussed in previous sections." It is not entirely clear why forms where gemination emerges as the outcome of sound change are listed in this section where expressive gemination in adjective bases is deemed to occur and where one expects to find instances of adjectives where "phonologically obscure gemination" is identified (beginning of Section 11.3). Rather than supporting the overall argument, it puzzles the reader.
p. 30 § 11.3.2 (9) "qPC *kokko-" This word (with geminate consonant) is attested in Greek and Latin (and attributable to Galatian, Freeman 2001: 16), and probably refers primarily to the kermes oak berry (Delamarre 2003: 120 'grain rouge, cochenille'), so one wonders whether it can be an instance of expressive gemination in adjectives. The development of the colour meaning is secondary.
p. 30 § 11.3.2 (11) *messamo-. It is not entirely clear to me why this item is put here. Its formation is different from the other cases in this Section, as gemination in all likelyhood occurs as a result of word formation.
p. 31 § 11.3.3 "They relate to basic physical concepts; to physical defects; and they are symbolic of swiftness and agitation." Could taboos have played a role?
p. 31 end of § 11.3.3 "Another reason for ascribing these words to Urschöpfungen within Celtic, and not to loans from unknown languages, has to do with the typological tendency of adjectives to be less amenable to borrowing than nouns." This is a crucial point in the argument and therefore, although it is commonplace, it deserves non-trivial mention and updated references (e.g. Matras 2007, Matras & Adamou in press).
p. 35 § 14 "*makʷkʷo-, or rather sequences of velars + labiovelars, i.e., *makkʷo-." From a phonetic point of view, it is rather unlikely that secondary articulations like labialisation occurred twice, so to say. Generally, geminate consonants are produced with a longer constriction duration than singletons (Ladefoged and Maddieson 1996); phonological gemination of complex articulations such as affricates, for example, usually consist of a longer stop gesture with delayed release (e.g. [tːs]) rather than two successive articulations stop+fricative ([tsts], Ladefoged and Maddieson 1996: 92). Anyhow, typological parallels for gemination including two successive labialisations should be mentioned in support of the first option. (The articulation of Modern Irish palatal vs. neutral tense (long) resonants, in varieties where the length distinction is preserved, could also be compared, although this is not an exact phonetic parallel). [Note that this comment argues for the opposite view as Martin Kümmel's review. *kk w is meant to stand in contrast to single *k w , rather than to *k w k w ] p. 35 § 14.1 "However, after the contrast lenition : non-lenition had emerged as the central phonological opposition in Insular Celtic, the gemination of stops lost its phonological significance and turned into a phonetic feature that was concomitant with non-lenition, indicated by the bracketed length marks in Table 6 below." If the emergence of lenition "may well at first have been a side effect of the strong opposition between single vs. geminate consonants", after the phonologisation of lenition, geminate stops could rather lose their phonetic status of long or tense consonants, but the effects of their earlier phonologisation were maintained and continued in fact in the phonological status of (non-lenited) stops. In other words, if the opposition between single vs. geminate stop turned into fricative vs. stop, this is not a real instance of dephonologisation, as would be the case if it were maintained only in some environments, rather a re-phonologisation. This at least is in line with what is stated at the beginning of Section 8.1.

3) Typos and minor slips:
p. 5 (previous research) Stokes > Stokes' p. 5 (previous research) assmilation > assimilation p. 10 § 3.1 (2) oder > or p. 20 § 6.2 (3) bye-form > by-form p. 22 (end of § 6.3) "in the system above": p. 4: Caution is advisable when quoting from Favereau's dictionary. He modernizes Old Breton (with accompanying phonetic transcription according to dialect!) and makes headwords out of material only possibly attested in e.g. place-names. It is always a good idea to check rarer words in Hemon's GIB and online in DEVRI.
p. 6, § 2.1, example (4) and following: It might be a good idea to briefly mention Hill (2010-12). It relates directly to the discussion of words with syllabic l+n here and in the nasal presents below. If Hill is correct, the proposed reconstructions here will not work. p. 7, § 2.1, example (9): It may be worth mentioning the existence of MBret. sallaff 'to salt', adj. sall 'salted'. A derivation from French saler 'to salt' is possible (with Bret. sall 'salted' being back-formed from the verb). However, the Breton -ll-is surprising and the match to the proposed PCelt. form is intriguing. Could be cognate with OIr. sall instead.
p. 8, § 2.2, example (3): It would be better to include Middle Breton cam 'step' here instead of the Modern Breton derivative (especially since the Breton suffix -enn may cause confusion for the reader).
p. 10-11, § 3.1, example (8): The lack of syncope seems to be secondary if the reconstruction in the preceding entry is to be trusted.
p. 11, § 3.1, example (13): Reference to the discussion of Egurtzegi and Ariztimuño (2013) would be suitable here, since it appears to eliminate some of the evidence in favour of substratal status.
This leaves us with mellt as the only reasonably solid example of preserved *-ld-(as Brit. *-lt-). Given that we have examples of vacillation between -ll and -llt in Middle Welsh (MW gwellt next to OW guel, MW guell, MW gwyllt next to MW guyll, gwyll, MW deall next to dallt), it might rather be the case that the final -t in mellt is unetymological.
p. 20, discussion of the outcome of voiced geminates in Brit.: Allowing an analogical treatment of clusters across morpheme boundaries may allow us to treat the outcome of *zd as having passed through the stage *dd, only to become devoiced to *tt and spirantized and thus being part of a general devoicing of voiced geminate clusters (cf. Jørgensen 2022). This requires that e.g. Bret. gad is a borrowing (presumably from Irish). p. 21, example (4): It would make sense to standardize Breton gwrec'h, grec'hent to either forms in gwr-or gr-.
p. 30, § 11.3.2, example (1): It seems much more likely that the -an in bychan, etc. is the diminutive suffix, cf. the meaning.
p. 30, § 11.3.2, example (4): This contradicts the change of *gg > *kk > *x presented earlier in the text. I have suggested (Jørgensen 2022: 145) that the actual reflex of *buggo-in Breton is bouc'h 'blunt' and that bouk is a borrowing from Irish.
p. 33, § 13.1, example (26): It may be better to leave out W croen and OCorn. croin, since they cannot continue *krokken-directly anyway, whatever the explanation. Also, as far as I can see, a reconstruction *krokkenno-would fit both Irish and Breton better (MCorn. crohen tells us nothing and Gallo-Lat. crocina does not fit anyway and may have suffix substitution). The word consistently rhymes as having /-enn/ in Middle Breton and is mostly spelled in accordance with this. It does not otherwise behave like a singulative in -enn, forming its plural with internal i-affection, so it does not appear to have been influenced by this suffix.
comprehensive overviews, not all cases can be discussed in all details but there are no serious shortcomings.
Remarks on single points are given under the heading numbers of the original paper. again.
3.9.5 For the development of *st > ss an intermediate stage [ts] is unnecesssary, so its asumption depends on (Gaulish) attestations of ðð/θ etc. Are there examples other than tuððos/tuθos for this? In any case, such a development in Gaulish does not necessarily imply that Insular Celtic had the same metathesis.
3.10 Fn. 2 If Proto-Celtic really had (acquired and) preserved *pp in contrast to simple *p, wouldn't it be useful to include this possible (but very rare) phoneme in the phoneme tables?
4.1 Would it be possible to set up a relative chronology by which *zd > *dd > British (and Gaulish) *tt was earlier than the rise of younger *dd (which was not devoiced)? Devoicing of geminates is a rather natural change (and probably part of Kluge's Law in Germanic), and the examples of a voiceless outcome under 6.2 seem to show the same, and might be older, as discussed there.
6.1 What could be the reason why original *dʱsdʱ > *dzdʱ does not develop like original *tst > Celtic *ts? With only one example, some kind of analogical influence is difficult to exclude.
6.2 Here the question arises what kind of "morpheme boundaries" are meant: formations with prefixes are originally compounds (at least they are in Sanskrit where external sandhi is applied), so it is actually boundaries of phonological words, not (word-internal) morphemes. Could this make a difference for the outcome of clusters, and would it be possible to assign conflicting results to full vs. incomplete univerbation?
11.3.1 (13) The vocalism of Slav. *munaga-has been argued to be regular from a zero grade (Matasović 2004) -still not directly comparable to Celtic *men-and Germanic *man-but not really "irregular".
14. There is hardly any real difference between *kʷkʷ and *kkʷ, as it is quite improbable that it was actually pronounced without lip rounding in the first half of the long [k]-gesture and with rounding in the second, and it is even more improbable that such a [kkʷ] could stand in contrast to [kʷkʷ].
14.1 The recordings of Trebitsch have been published in Lechleitner & Remmer (2004), so this edition should be cited here.
14.2 Tables 7 and 8: Why is labial-velar /gʷ/ present in table 7 but gone in table 8? Is there not still a unlenited : lenited correlation with /w/ in the attested languages, more or less like /b/ : /β/? Probably, the glide should be classified as labio-velar rather than "bilabial", and one may also consider adding another labio-velar phoneme /hʷ/ (cf. Primitive Irish /xʷ/), especially since this possibly does not always go back to a cluster *hw (at least according to Jørgensen 2012).
The relative length of voiceless obstruents may be argued to have remained relevant until the split of vowels (reflected) in (Modern) Welsh into a long/tense in originally open syllables (preceding simple voiced consonants and some fricative-stop clusters) and short/lax variant in originally closed syllables (preceding clusters, original geminates, voiceless consonants and m, ng), so it must have played a role in phonology.