My latest project is a full language superfamily (the Thakina languages). I'm now at the fourth generation with Highlands Têyisa, and I'm noticing that most of the words are getting far, far shorter ─ for example, *ukomaj became simply we in HT, and probably would be further eroded in the future. Now, this is an extreme case (a more typical example might be *lokanu becoming lwe), but since I have two more generations to go, I'm beginning to get worried. Is there a way that I could restore some of the length and complexity of the words I started with through sound change, or would I have to use compounding / new words?
As a general rule, regular sound changes wear away at words, reducing their information content.
Countering this, morphosyntactic changes restore the lost information.
For example, let's look at Latin. Classical Latin has lots of case marking on nouns; amīcum means "friend" as a direct object, for example, while amīcō means "friend" as an indirect object. But in later Latin, coda m turned into nasalization on the vowel and then disappeared, and u and ō merged into o. Now these two forms were indistinguishable.
As they become less and less distinct, though, Latin-speakers still needed to convey this information. So during this period of merging, Latin-speakers started adding extra words to disambiguate: putting ad "toward" in front of the indirect object. As a result, by the time the forms had entirely merged, Latin had a new way to distinguish them: amico for the direct object and ad amico for the indirect object.
This is how it goes in general. When there's a difference that conveys important information, and that difference is getting worn away by sound change, people will add extra words to disambiguate. Wait long enough, and these extra words will get merged into the original ones: that's how we got "inkpen" in English. Southern US dialects merge
/ɛ/ before nasals, so "pen" became confusingly similar to "pin", and was disambiguated as "ink pen". Eventually these two words merged into a single one.
Sound shifts are to some amount irreversible. Long before your words are completely gone, the rate of homophones rises and the speakers of the language have to deal with it in some way or another.
The most common way of dealing with it is lexical innovation, one part of the homophone pair is replaced by a new word. Sources for that new word may be internal (formation of a compound, using another semantically related word that is already in the language, using some derivational morphology) or external (borrowing from some other language in your world).
There are other possible exits from the dilemma, e.g., using more and more conventionalised multi-word expressions like "to burn with fire" instead of simply "to burn", because your words for "fire" or for "to burn" have become too ambiguous to be used in isolation.