feat: inline language inference in outline generation

[cosarah]

Summary

Embed language inference directly into outline generation (replacing the standalone LLM call approach), move agent profile generation after outlines, and remove the manual language selector along with all hardcoded zh-CN/en-US conditionals.

Related Issues

Supersedes feat/language-inference branch (#381)

Changes

• Pipeline reorder: outline generation now runs before agent profile generation (both server-side and client-side flows)
• Inline language inference: outline LLM output changes from SceneOutline[] to { languageDirective, outlines } wrapper object
• Full-pipeline language directive: languageDirective propagates through outline → agent profiles → scene content → scene actions → chat
• Remove manual language selection: UserRequirements.language field, toolbar language toggle, normalizeLanguage() function
• Remove all zh-CN/en-US hardcodes: prompt construction code uses English only; LLM infers teaching language from user requirement text
• Advanced learner recognition: system prompt identifies advanced foreign language learners (TEM-8, DALF C1, etc.) who should be taught in the target language
• Eval tests: 22 production-sourced test cases (including 3 cross-language PDF cases), gemini-3-flash-preview inference + gpt-4o-mini judge, 21/21 pass

Type of Change

• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)
• Refactoring (no functional changes)

Verification

Steps to reproduce / test

1. Create a course with a Chinese requirement (e.g. "从零学 Python"), verify content is generated in Chinese
2. Create a course with an English requirement (e.g. "Explain photosynthesis"), verify content is in English
3. Verify generation flow order: PDF analysis → Web search → Outlines → Agent generation → Content
4. Verify the language toggle button is no longer in the toolbar
5. Run eval tests: EVAL_INFERENCE_MODEL=google:gemini-3-flash-preview EVAL_JUDGE_MODEL=openai:gpt-4o-mini pnpm vitest run tests/generation/outline-language.eval.test.ts

What you personally verified

• Tested Chinese and English requirements in the frontend — language inference correct
• Verified advanced English learner case (TEM-8 oral fluency improvement)
• Eval tests: 22 cases run concurrently, 21/21 that entered judge all PASS

Evidence

• Eval result: 21/21 (100%) with gemini-3-flash-preview + gpt-4o-mini judge

• TypeScript: npx tsc --noEmit zero errors

• CI passes (pnpm check && pnpm lint && npx tsc --noEmit)

• Manually tested locally

• Screenshots / recordings attached (if UI changes)

Checklist

• My code follows the project's coding style
• I have performed a self-review of my code
• I have added/updated documentation as needed
• My changes do not introduce new warnings

[cosarah]

Eval test cases expanded: 22 → 50

New test dimensions added (28 cases):

Language learning (5) — Japanese/Korean learners, non-zh/en axis (ja→zh), multi-target language learning

Immersive learning (2) — Advanced learners requesting full target-language immersion (Japanese→English, Chinese→French)

Explicit language instruction (2) — User explicitly overrides default language ("请用英文教我", "explain in Chinese please")

Code-switching & bilingual (2) — Mixed zh/en input, explicit bilingual teaching request

Minimal / ambiguous input (2) — Single-word requirement ("微积分"), pinyin romanized input

User profiles (8) — Teacher designing foreign language lesson, parent proxy for IB student, bilingual heritage speaker, professional business English, immigrant integration, tutor with bilingual student, teacher of Chinese-as-foreign-language

Cross-language PDF (7) — English req + Chinese PDF, Chinese req + Japanese/French PDF, Japanese req + English PDF, bilingual PDF, teacher using foreign-language material

Also simplified the language inference prompt (system.md ~80→30 lines, removed redundant pdfLanguageSample). 50/50 with gemini-3-flash-preview + gpt-4o judge.