[codex] Normalize i18n and prompt baseline

[heikog]

What changed

• Normalizes backend locale handling around Accept-Language parsing and English fallback behavior.
• Converts core LLM prompt templates from Chinese-only wording to an English baseline with explicit response-language instructions.
• Updates frontend locale initialization so unsupported or saved zh values fall back to en.
• Uses max_completion_tokens for GPT-5 model names while keeping max_tokens for older models.

Why

This makes the app safer to run with non-Chinese locales and modern OpenAI models, without mixing in Graphiti or simulation-round UI changes.

Validation

• python3 -m compileall -q backend/app
• git diff --check