Trial-marketing. Twitter users generally tend to form relationship (follow/see/RT)-graphs stick within language boundaries (and to a lesser-extent, national identity boundaries) - by using language or region as a proxy for a "clean" subset of users you can get some experimental results that you won't get by using random user selection, for example. Also, users unaffected by this change will still be able to see other users' replies and could then share screenshots of said replies which would defeat the point of this exercise - by using language boundaries means it's less likely that such a thing would happen (e.g. if this was Sweden and Norway instead of US+Japan, would you really bother to go through Swedish tweets to find replies for other users?)

In short: it sounds like they're uncertain about pushing this change through on everyone, and want to trial it first.

Twitter has gone-back on ideas they've thrown around before - like when they said they would hide other users' tweets' Like +RT counts.