feat: pnpm migration, English logs, local temp chat mode

[shabahmd]

Changes

Package Manager: npm → pnpm

• Switched from npm to pnpm (v11.22.0) with packageManager field
• Added pnpm-workspace.yaml (root + public workspace)
• Replaced package-lock.json with pnpm-lock.yaml
• Removed PM2 dependency (ecosystem.config.js, PM2_INSTANCES, PM2_MAX_MEMORY)
• Updated README-en.md deployment instructions

Local/Temporary Chat Mode — single source of truth (551dff6)

• New resolveChatMode() helper (src/utils/chat-helpers.js): the only chat_mode decision point — explicit client 'normal' wins over the flag, then 'local' request or ENABLE_TEMP_CHATS=true, else 'normal'
• Both entry points resolve through it: OpenAI middleware and Anthropic buildInternalRequest (previously hardcoded 'normal')
• generateChatID() now creates upstream chats with the resolved mode (chats/new payload), and both upstream requests use referer /c/local in temp mode — matching real browser traffic
• Previously enabling ENABLE_TEMP_CHATS had no effect: a persistent chat was always created first with hardcoded 'normal'

Fix Anthropic /v1/messages upstream WAF/captcha rejections (551dff6)

• The Anthropic path sent a non-browser-shaped body (missing version/timestamp, skeletal messages, extra session_id/id) → Alibaba risk control fingerprinted it as a bot → HTTP 200 with FAIL_SYS_USER_VALIDATE / RGV587
• New shared buildFeChatMessage() factory emits the full FE 0.2.81 message shape; used by both middleware and Anthropic builder (no duplicated message literals)
• Anthropic body now aligned: single FE-shaped message (history already folded by parserMessages), version: '2.1', top-level timestamp; removed write-only session_id/id/top-level sub_chat_type

Dynamic default model (551dff6)

• Model selection is fully dynamic: /v1/models derives from the live upstream list (+ capability suffixes); request models are alias-matched against it
• Removed stale hardcode: omitted model now resolves to the first upstream t2t-capable model instead of literal qwen3-coder-plus (absent from today's list)

Chinese → English Log Translation

• Translated all logger/console/throw/res strings across 29 source files
• Chinese JSDoc comments preserved
• 0 remaining Chinese strings in active code paths

Files Changed

• 29+7 source files updated across 3 commits
• pnpm-lock.yaml, pnpm-workspace.yaml, public/pnpm-lock.yaml added
• README-en.md updated with pnpm instructions, PM2 removed, temp chat mode documented
• .env.example: documents ENABLE_TEMP_CHATS

Testing

• Full suite: 68/68 pass (node --test --test-force-exit tests/*.test.js)
• New tests: resolveChatMode precedence, middleware stamping, FE message-shape guard
• Verified end-to-end on live server:
  • /v1/chat/completions: no field → local, explicit local ✓, explicit normal opt-out ✓
  • /v1/messages: default, local, and normal all return completions (previously deterministic WAF failure)
  • Models: qwen3.8-max, qwen3.7-max, qwen3.6-plus each resolve to distinct upstream models; omitted model resolves dynamically

[shabahmd]

Well, I prefer my rice boiled, not cooked.

[shabahmd]

Also, I did not know much about PM2, I can re-add them into a new pull request if server hosting is required.