response-language setting for Ask AI and agent review jobs

[JaeSeoKim]

Summary

Adds a Response Language setting so users can get AI output in their preferred language (e.g. Korean) instead of English, across both Ask AI chat and launched agent jobs (Guided Review / Code Tour / agent code reviews).

Motivation

Ask AI and guide/tour outputs always come back in English (the model default), which makes the AI features awkward for non-English-speaking users during code review. This adds an opt-in language preference that injects a prompt instruction; default is Auto (no instruction, today's behavior).

What's included

Settings UI (AISettingsTab)

• "Response Language" section: Auto (default) + 31 preset languages (native + English labels, SaaS-typical coverage) + Custom… free-text input.
• Stored as the English language name in the cookie plannotator-ai-response-language (consistent with other cookie-backed prefs; works across random ports).

Ask AI

• Optional responseLanguage field on AIContext (all three variants) — rides inside the existing context serialization, so no provider or endpoint changes needed; claude / codex / opencode / pi all pick it up via buildSystemPrompt and buildForkPreamble (forked sessions skip the system prompt, so the fork preamble carries it too).
• The session bakes the language into its system prompt at creation. If settings drift mid-session, a small per-query (Respond in …) tail is appended only until a new session picks it up — a change applies from the very next question, with no duplicate instruction in the steady state.

Agent jobs (guide / tour / review)

• launchJob injects the setting into the job body; responseLanguage added to the jobs-endpoint field allowlist and config threading (Bun server + Pi server mirror).
• agentJobLanguageInstruction() (shared in agent-review-message.ts) is prepended to the guide/tour user message and the review-job user message. Instruction scopes to natural-language output only — file paths, code identifiers, and JSON/schema keys stay as-is, so guide/tour structured output still validates.
• Guide repair jobs are deliberately excluded: they must fix JSON syntax without rewriting content.

Testing

• Unit tests: buildSystemPrompt/buildForkPreamble include/omit the instruction; guide buildCommand prepends it from config.responseLanguage.
• End-to-end: with Korean selected, Ask AI (code review mode) answers in Korean; a real Guided Review job produced fully Korean titles/overviews with valid schema output. Auto reproduces today's prompts byte-identically.
• bun run typecheck clean; full test suite passes.