fix: honor user language for AI responses

[sergeybychkovvvpgroup-beep]

Separate PR split out from #291 per maintainer review.

This keeps the IMAP MIME body-part fix focused and moves the unrelated AI language handling change into its own review.

Change:

• honor the user language setting for AI responses

[maathimself]

Thanks, but this branch was cut from #291's first commit rather than from main, so it carries the whole original IMAP/MIME change along with the AI one, its diff touches imapManager.js, emailSanitizer.js, and mail.js, and its commits include 353383c "fix: fetch IMAP body parts individually".

That's the pre-review version, so merging this would re-introduce exactly what we removed in #291: repairResidualQuotedPrintableHtml (corrupts ~56% of real bodies and writes the corruption back to the DB cache), the un-guarded blanket per-part re-fetch, and unwrapEmbeddedMimeText with no recursion bound.

Could you re-cut this off main so it contains only the ai.js change? Since that change is entirely in one file, the simplest path is:

git checkout -B fix/ai-language origin/main
git checkout 9a6d0a7 -- backend/src/routes/ai.js
git commit -m "fix: honor user language for AI responses"
git push --force-with-lease

The AI-language change itself looks good, it just needs to stand alone. Thanks!

[maathimself]

Following up here, since this still carries the #291 IMAP/MIME changes (imapManager.js, emailSanitizer.js, mail.js) on top of the ai.js change. The re-cut from main is still the blocker; the commands in my previous comment should do it.

Two small things to fold in once it stands alone, so the AI change is complete:

• Add a test for the language mapping (aiLanguageInstruction and its allowlist plus fallback). It is a small pure function and easy to cover.
• Decide whether the GTD "waiting" gist summaries (gtdGist.js) should also honor the user's language. Right now /ai/chat gets the language instruction but the gist path does not, so those summaries stay in the email's language. Either behavior is defensible, but it is worth a deliberate call rather than an accidental inconsistency.

The ai.js change itself looks good and is safe: the language value goes through the allowlist and is never interpolated raw into the prompt.