Make sure invalid LLM output doesn't crash build

[Wituareard]

A recent series of failed build attempts has been caused by the LLM mistakenly translating one single link target it never translated before. I'm not saying that this specific issue can't be sufficiently addressed, but we should not let 1 in a 1000 errors crash our build in general.

[anthonybailey]

Well, it definitely shouldn't _crash_! A failed build with clear indication of the error _might_ be ok - i.e. we might choose to have zero tolerance of l10n issues, and leave the existing site up until a fix for new problems can be made. That should be tenable especially if we work from an LLM-derived but pauseai-l10n cache as intended: this provides a last gasp human override by faking an entry. But other policies could work too including publishing some kind of "failed to localized: {en-US text}" placeholder in error cases.

…

On Wed, 8 Jan 2025, 22:26 Wituareard, ***@***.***> wrote: A recent series of failed build attempts has been caused by the LLM mistakenly translating one single link target it never translated before. I'm not saying that this specific issue can't be sufficiently addressed, but we should not let 1 in a 1000 errors crash our build in general. — Reply to this email directly, view it on GitHub <#255>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAAXXDMKJXG6Q7RHBTWTPAL2JWQWRAVCNFSM6AAAAABU24VHG2VHI2DSMVQWIX3LMV43ASLTON2WKOZSG43TMNBWGMYDAMQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>