Philosophy papers (like David Lewis' Languages and Language) often seem to want to explain the fact that some X is actually the case (e.g. we all use the word "fire" in roughly the same ways) by showing why it would be rational for people to make X the case/preserve that state of affairs X. But this seems potentially problematic:
a) Historians wouldn't generally accept the idea that showing why declaring war was rational for a certain leader explains why he actually did declare war. If the presedent has the policy of always taking the first proposal suggested when he's tired and wants to go home, and P was the first proposal suggested, the fact that it would be rational to do P rather than Q is not the correct explanation for why the president actually did P rather than Q. Similarly if the president never even considered Q, the fact that P serves his interests better than Q seems like an incorrect explanation for his choosing Q.
b) More generally, rationality explanations seem to have exactly the issues that philosophers of biology make a huge deal about when considering evolutionary fitness explanations: the mere fact that some trait would be eliminated by natural selection isn't always the correct explanation for why we don't find it. For example, the fact that humans don't levitate isn't explained by natural selection, but rather by the fact that the total space of mutations available from the original organisms doesn't include that. (That is: plausibly, even if all creatures had had all the offspring they could, and lived as long as they could so there was no culling to generate natural selection, we would *still* find 0 creatures that levitate. )
Applied to the David Lewis case of conventions, this works out in the following way. It *might* be that we get linguistic conventions because once some people are using language a given way, each person works out that it would be rational for them to do the same (this is the nub of lewis' account). Or it might be that most possibilities for doing things differently don't even occur to them - people just brutely follow custom and habit and imitate those around them. (Apparently apes' tool use is like this: chimps in a given area all crack nuts using the same techniques, even though different techniques would work just as well and are used in different areas. Note here that there's no rational benefit to cracking nuts differently from your neighbor). Or it might be some combination of habit plus considerations of rationality. And surely its an empirical matter to find out which.
Given this, I think we can read Lewis either as making a bold empirical conjecture, or as explaining actual behavior by comparing it to the behavior of a simpler, but in some respects similar, model system (as often happens in biology).
bold empirical conjecture: In fact, adherence to linguistic conventions is always produced, not by custom and habit whereby doesn't occur to people to behave otherwise, but by speakers recognizing that it would be rational for them to continue with the convention.
simplified model system explanation: In ideal system S (where there are no limitations on computational power etc and people always behave in the way that best advances their aims), linguistic conventions arise and persist. The actual world of people talking is `relevantly similar enough' to S, for these facts about S to explain how actual linguistic conventions arise and persist in the actual world. [Slot in whatever notion of relevantly similar explains how facts about waves in infinitely deep oceans can explain facts about waves in actual finitely deep oceans].