In my local testing with the new parsoid languageconverter implementation it seems like we might be double-converting messages: we don't actually have any localizations for en-x-piglatin, so when we ask for those we get the 'en' fallback, but then we parse that message with the *legacy* parser, and the user language variant is still in the RequestContext so we end up running the message through LanguageConverter. If the message was already present in the variant, we would have converted it twice. Converters *try* to be idempotent, so this might not immediately be noticible, but the pig latin converter is *not* idempotent so it should be pretty obvious.

Moving the language variant into the parser options will help some, as it will force the legacy converter not to look into the request context for language conversion. We could/should also specify the 'disable language conversion' flag when messages are parsed.

Change #1236396 had a related patch set uploaded (by C. Scott Ananian; author: C. Scott Ananian):

[mediawiki/core@master] WIP: Disable content conversion when parsing interface messages

https://gerrit.wikimedia.org/r/1236396