fix(ai): write analysis, letters and chat replies in the language of …

[Dwarf1er]

Summary

Every system prompt is written in English, so models answer in English even when the resume and job
posting are in another language (seen with French documents). This adds a language-agnostic rule to the
generating prompts so the output follows the language of the documents, and the chat reply follows the
language of the user's message. Nothing is specific to one language.

Changes

• New src/lib/ai/prompts/language.ts with the rules and short reminders.
• Resume review, job match, automation match and cover letter system prompts end with a language rule.
• Each user prompt also gets a one-line reminder beside the documents. A rule in the system prompt
  alone was not enough for a small local model (qwen3.5:9b still answered in English); the reminder next
  to the document fixed it.
• Chat agent prompt gets a language section: reply in the user's most recent message's language, and
  keep tool arguments in the language the user supplied.
• The SCORES: line and the ## section headings stay in English, because jobMatch/parse.ts,
  resumeReview/parse.ts and coverLetter/matchGuidance.ts parse them.
• Cover letter salutation and sign-off are translated into the posting's language.
• New eval suite evals/multilingual (French resume and posting) with language-neutral assertions,
  plus an eval:multilingual script.

Related Issues

Testing

• npm run lint and tsc --noEmit are clean.
• Unit tests: no new failures. 193 tests fail identically on a clean dev in my environment
  (jsdom/localStorage setup), unrelated to this change.
• npm run eval:multilingual against local qwen3.5:9b: review, match and both chat cases pass. A
  French cover letter generated with the app's exact prompt came back in French with a localized
  salutation and sign-off. I could not grade the cover-letter llm-rubric case because I had no DeepSeek
  key.
• Evals so far cover French only. Before the reminder was added, review and match answered in English
  on the same model.
• Also checked manually in the Docker build against Ollama.

Notes

This assumes the resume and posting are in the same language. Mixed-language input uses the language
most of the text is written in.