[The following clarifcation was received from H.I. in response to the issues raised by Jiblet (aka Malcolm) in the comments to BC8.21. I am posting these paragraphs here with H.I.'s kind permission.]
Indeed, bahuvrīhi compounds are adjectives (except when they are adverbs; bahuvrīhi adverbs are much rarer than bahuvrīhi adjectives, but still not very rare in classical Sanskrit; as adverbs they are indeclinable, with a 'frozen' ending, which is normally that of the neuter acc. sg., morphologically indistinguishable from the neuter nom. sg.). Like other adjectives they can sometimes be 'substantivized'; quite a few names are substantivized bahuvrīhi adjectives. (They have the gender of the substantive that they 'originally' qualified, and which they have come to stand for.) Those bahuvrīhis which are simply adjectives must, like other adjectives, indeed agree with the substantive they qualify.
Malcolm is right that Coulson does not mean to say that all bahuvrīhi compounds are masculine. Coulson says, to quote exactly,
"All bahuvrīhis are essentially adjectival. The compound on which a bahuvrīhi is based is reduced to a stem form, and then inflected to agree with a substantive expressed or understood. The stem form must in the first place be a masculine stem form. Thus a feminine substantive in ā at the end of a bahuvrīhi has its final vowel reduced to short a: e.g. from svalpecchā 'small desire', svalpeccha 'having small desire'."Here we are supposed to understand that "in the first place" implies that this change to a masculine stem form is followed by a subsequent (from the analytical point of view) operation in the formation of the form as it will actually occur in the sentence. This subsequent operation is that referred to in the second sentence I quoted with the words "and then inflected to agree with a substantive expressed or understood." The form in the sentence will be feminine if the substantive qualified is feminine. So svalpeccha is a bahuvrīhi
adjective: now if I wish to say that a monk has small/little desire(s) I may say for instance svalpeccho bhikṣuḥ, if I wish to qualify a nun in the same way I may say svalpecchā bhikṣuṇī.
[Aside: I think the main objection that may be raised to Coulson's explanation is to the sentence 'The stem form must in the first place be a masculine stem-form.' 'In the first place' is not a problem; but it is not accurate to say that we must have this change to a masculine stem-form. The technical name for this change is puṃvadbhāva (literally something like 'becoming as if masculine'). For feminine nouns which are being used as the last member of a bahuvrīhi, puṃvadbhāva is common, it is true, but by no means does it always take place. So Coulson's 'must' is incorrect. The conditions under which puṃvadbhāva does and does not occur for feminine nouns used as last member of a bahuvrīhi are taught by Pāṇini in Aṣṭādhyāyī 6.3.34.]
As far as the word in this verse is concerned, grammar does give us a little help, perhaps. The form vilambakeśī is good, though it is not an exception to what Coulson means when he says that "a polysyllabic feminine -ī is hardly to be found at the end of a bahuvrīhi". Coulson means that it is rare that the last member of a bahuvrīhi is when it stands by itself, not at the end of a bahuvrīhi, a feminine polysyllabic noun in -ī . See his further clarification in chapter 10, to which he refers immediately after the words Malcolm quotes (from the top of p. 119). There he writes (p. 146) "It has already been mentioned that nouns like nadī are rare at the end of an exocentric compound." [I.e. a bahuvrīhi compound.] "Sometimes the difficulty is surmounted by the addition of the adjectival suffix ka: so sapatnīka 'with one's wife', pravṛttabībhatsakiṃvadantīka '(citizens) among whom foul rumours are current'."
But although 'modern' dictionaries (such as Monier-Williams and Apte) do include keśī as a feminine noun, that noun is poorly attested in texts, and the last member of this compound should rather be understood to be the masculine noun keśa. By Pāṇini's sūtra 4.1.54, when the last member of the bahuvrīhi is part of the body of a person described by the bahuvrīhi, the feminine form may optionally be formed with the suffix ṅīṣ (that is, effectively, by substituting -ī, instead of -ā, for the ending -a). (One of the examples given by commentators
on Pāṇini is atikeśī.) The same rule accounts incidentally also for such forms as bhrāntamukhī in SN 4.39d (mukha is of course a neuter -a stem). Other classical Sanskrit authors also use such forms; it seems to me that although Pāṇini says that this is optional there is a clear preference on the part of such authors to use the feminine forms in -ī instead of in -ā (e.g. sukeśī instead of sukeśā, candramukhī instead of candramukhā). A few examples (many could be given), taken from poetry, of keśa becoming -keśī at the end of a feminine bahuvrīhi: sukeśī is used by Kālidāsa in Kumārasambhava 1.60; the feminine plural asitāyatakeśyaḥ is used by Māgha in Śiśupālavadha 10.80; the feminine plural vimuktakeśyaḥ is used by Abhinanda in Rāmacarita 28.125.
On the other hand, if the last member of the compound were veśa/veṣa I am not sure what the justification of the feminine form -veśī/-veṣī rather than -veśā/-veṣā at the end of a bahuvrīhi is. (This may only be my ignorance, which in the field of vyākaraṇa is pretty vast; the 'modern' dictionaries seem to claim that at the end of (bahuvrīhi) compounds for this word too both the forms in -ā and in -ī are allowed.) And, more importantly than what the grammatical rules say, what we see in the actual usage of classical Sanskrit writers is, I think, that feminine bahuvrīhis with veśa/veṣa as final member are used as -ā stems rather than -ī stems. (There may of course be exceptions to this, as there are so many works which I have not read and do not have access to searchable versions of; but as far as my reading goes, and as far as a relatively quick and superficial search of electronically available texts shows, forms in -veśī/-veṣī are, if we set aside the possible occurence here in BC 8.21a, unattested, unlike the forms in -veśā/-veṣā.)
Final note on the grammatical side of the matter, by a real grammarian, Whitney (from his Sanskrit grammar §332a): "Whether a masc.-neut stem in a shall form its feminine in ā or in ī is a question to be determined in great part only by actual usage, and not by grammatical rule."
There is one further consideration that occurs to me: it seems to be normal in Sanskrit to use vilamba or related forms (e.g. present participle vilambamāna; also forms with the same root but a different verbal prefix, notably pralamba/pralambamāna, and forms from the same root but without verbal prefix, i.e. lamba/lambamāna) to describe/qualify a person's hair, but not normal to use the same words to describe a person's garments. At present I can adduce no instance of the latter but quite a number of the former.
Hence I believe we can say that from the point of view of grammar (with the slight caveat that a form -veśī/-veṣī might be grammatically justifiable, though I don't at present know how) and, more importantly, from the point of view of the usage of classical Sanskrit authors, the form vilambakeśī is more plausible than the form vilambaveśī/-veṣī.