I’ve seen somewhere that gettext uses a default threshold of 0.85 when calculating the Jaro distance for measuring similarity. I understand there’s no perfect solution, it was just a suggestion because it’s been hard to follow these changes, I’m not a translator and I help whenever I notice any missing text in my instance, and yet I’m the “top” translator for my language on Crowdin for lack of contribution.
Unfortunately most languages are missing active contributors and I think a week is unlikely to be enough time for a new translation. I try to take a look whenever I get an email from Crowdin, but my language is already missing ~11k words and it’s been hard to keep up.
Pre-translate files with Translation Memory ™ to speed up the translation of the same or similar strings…
It may really improve the workflow if the mentioned fix lands sometime, it also looks like it addresses the fuzzy string scenario.
What happens if the user is customizing the strings locally? Isn’t this already an issue in this case?
Another possible alternative would be to not invalidate by default and have a different workflow when the whole meaning has changed, maybe changing the translation key or having a @discourse-translator-bot invalidate_translations on the commit message, for example, assuming – which may be totally wrong – that changing the meaning for a translation key is an exception.
That all said, Crowdin is really a breeze to work with, none of this would be an issue if more people were contributing to it. I believe many are using the text customization feature as a translation resource.