In particular, what has been observed to be the range (in %) of impact that i18n has on effort and time-to-delivery, when it is added well (about mid-way) into the cycle?
Sorry, I don't have detailed research or a percentage. All I can say from - somewhat anecdotal - experience is that it'll be at least a moderate PITA.<p>In general, i18n tools for Java are quite sophisticated so with Java it won't be that much of a problem. Other environments probably require some more manual work.<p>Some IDEs (Intellij IDEA and Eclipse for instance) allow you to extract strings from your source code and externalize those into a property file. However, those tools aren't perfect and will likely still miss some strings. So, at least some manual work will still be necessary.
You'd probably want a reasonably stable version before i18n to avoid much of the back and forth with edits, additions, context queries etc. The more complete the picture your translators have, the better quality your efforts will yield (thus better conversion/engagement/retention) in theory.
Thanks for your replies!<p>We have convinced the customer that it is better to get the English version into production first, get market feedback, get some traction, and then proceed with i18n+l10n.