(work in progress)
The following are not hard rules so much as recommendations born of years of experience—and suffering.
It is advised to stop as soon as the etymology reaches the first Latin lemma. There is no need to go back further to Proto-Italic or beyond. If one wishes to categorize Romance words by their longer-range relationships, one can always use dercat. In the future, it should be possible to have this automated anyway.
Most Romance nouns correspond best to the accusative form of their Latin ancestors. Latin nouns are, however, lemmatized in their nominative forms. In most cases the difference is fairly trivial and not worth worrying about. The Italian figlio, figlia can be said straightforwardly to derive from Latin fīlius, fīlia (nominative forms).
There are cases, however, when the Latin nominative and accusative differ in some important way- for instance in the presence or absence of certain consonants or even an entire syllable (which may affect the stress position). In such cases it is advised to link to the Latin nominative lemma whilst displaying the accusative form which gave rise to the Romance descendant. For instance, the etymology for Italian città reads 'inherited from Latin cīvitātem', which links to the Latin nominative lemma cīvitās.
Most Romance verbs are lemmatized in the infinitive, while Latin verbs are lemmatized in the 1SG present indicative form. The approach suggested above for verbs works just as well here. For instance, the etymology for Spanish tacer reads 'inherited from Latin tacēre', which links to the Latin lemma taceō.
If a Romance word has more or less the same sense as its Latin antecedent, which is usually the case, then there is no need to gloss the latter.
There is no need to list random cognates in a Romance etymology. One or two may be provided when they are in some way rare, interesting, or genuinely useful to have on-hand. Some examples are:
If the Romance words derive from an alternative form of the Latin lemma, it is recommended to list them as descendants under that lemma with a label such as 'reflexes of the variant X'. See monachus#Descendants or impleo#Descendants for examples.
It is advised to list reconstructed Vulgar Latin forms in the descendants section of Latin entries rather than among the derived terms, where they are liable to be lost in a sea of regular Latin derivatives. Placing them in the descendants section has the additional benefit of allowing the use of desctree or see desc.