According to CGEL (and all other sources I looked at so far, such as Quirk et al. and Biber et al.), the following are morphological compounds, i.e. compound adjectives, and not syntactic composites where one adjective appears as a pre-head modifier in an adjective phrase:
dark-blue icy-cold pale-green red-hot white-hot
The possibility that some of these could be syntactic constructions is not even discussed. In sections dealing with adjective phrases (AdjPs), no possibility is mentioned of one adjective appearing as a dependent of another, and the lack of that possibility is not commented upon.
(As I will argue below, in the section A possible case of an adjective+adjective syntactic construction, there seem to be good reasons for preferring the syntactic analysis in at least some cases.)
How the analogous issue is treated in the case of nouns
CGEL's silence as far as syntactic combinations of adjectives is in contrast to how the corresponding issue is treated in the case of nouns, i.e. in the case of the distinction between composite nominals and compound nouns. One reason for CGEL's more extensive treatment in that case, presumably, is that nouns appear as modifiers of other nouns very frequently, especially as complements (a flower seller) but sometimes also as modifiers (a cotton shirt). Presumably for that reason, CGEL spends some time to explain how one distinguishes these syntactic constructions from morphological compounds such as ice cream or sunrise. In short, CGEL says that
What distinguishes the syntactic construction from the compound noun is that the component parts can enter separately into relations of coordination and modification. (p. 449)
So London colleges is not a compound noun because the following are all possible:
coordination of the modifier: various [London and Oxford] colleges
coordination of the head: various London [schools and colleges]
right-node raising (RNR): [two London and four Oxford] colleges
submodification: two [south London] colleges
stacked modification: two London [theological colleges]
CGEL says that these five constructions
thus provide diagnostic tests for distinguishing between syntactic constructions and compounds. In principle, satisfaction of any of the tests is sufficient to demonstrate that a sequence of elements forms a syntactic construction.
None of the five constructions are possible with ice cream or sunrise. You can't say ice and custard creams when you mean ice cream and custard cream, and you can't say sunrise and set when you mean sunrise and sunset. The other four types of construction also fail.
There is much further discussion in CGEL, which includes an acknowledgment that there are difficulties with this treatment, such as borderline cases. This is followed by a defense of retaining the distinction between composite nominals and compound nouns despite these difficulties.
Of interest is what CGEL says about non-syntactic criteria:
Various non-syntactic criteria have been proposed as differentiating between composite nominals and compound nouns, as in such pairs as composite black bird ("bird which is black") and blackbird ("species of bird"):
 i stress: the composite nominal has primary stress on the second element (blackˈbird), while the compound has it on the first (ˈblackbird),
ii orthography: the composite nominal is written as two orthographic words,the compound as one.
iii meaning: while the meaning of the composite nominal is straightforwardly predictable from the component parts, that of the compound is not—it is specialised, denoting a particular species,
iv productivity: in the composite nominal the dependent can be replaced by any other adjective that is semantically compatible with the head, whereas there is a quite limited number of compounds with the form Adj + bird.
The correlation between these criteria and the syntactic tests of coordination and modification is, however, very imperfect, and since we are concerned with the delimitation of a syntactic construction we will naturally give precedence to the syntactic tests in the many cases of divergent results.
A possible case of an adjective+adjective syntactic construction
But now consider the fact that we can say dark and light blue+noun (examples here, here). If we were talking about nouns, this would already be sufficient to establish that dark blue is a syntactic composite. But arguably, we can also have submodification (?very light and very dark blue crayons, RNR (?Kim's very light and Pat's very dark blue crayons) and stacked modification (?dark saturated blue crayons).
Given all this, what is the argument for dismissing any possibility of adjectives being dependents of other adjectives, in particular for dismissing the analysis of dark blue in dark blue crayons as an adjective phrase as opposed to a compound adjective (which would then be hyphenated, dark-blue)?
Here is what CGEL says about compound adjectives of the type adjective+adjective (pp. 1656–1658):
4.3 Compound adjectives
Our initial classification here is again based on the category of the central element. We distinguish three categories: adjective-centred (cholesterol-free, red-hot), verb-centred (fun-loving, MIT-trained, germ-resistant), and a residual category centred on nouns (highbrow) or having the form of preposition + noun (upmarket).
4.3.1 Adjective-centred compound adjectives
■ Adjective + adjective
There are far fewer established compounds of this form. We distinguish two cases.
 bitter-sweet deaf-mute shabby-genteel Swedish-Irish syntactic-semantic
The components here are of equal status. The last two illustrate highly productive patterns, both of which are predominantly used in attributive function: Swedish-Irish trade, a syntactic-semantic investigation. In general these can be glossed with coordinative and: "bitter and sweet", "deaf and mute", etc. In some, however, there is an understood "between" relation: "trade between Sweden and Ireland".
 dark-blue icy-cold pale-green red-hot white-hot
As with nouns, it is not always easy to distinguish between coordinative and subordinative types, and some analysts include these with the coordinative type. We regard them as subordinative because we take the first element to be semantically modifying the second: icy, red, and white, for example, have an intensifying role. Bright, dark, light, pale occur productively with colour terms. Compounds combining two colour terms, such as blue-grey or orange-red probably belong here too, with the first identifying a particular shade of the colour denoted by the second.