[codex] Add MCP access, Chinese-first output mode, and dedicated ingest model

[SipengXie2024]

Summary

This PR bundles four local commits that improve local agent workflows and day-to-day authoring in LLM Wiki.

• add LLM Wiki MCP access and local agent settings
• refactor chat retrieval into a shared helper and fix the sources panel layout
• add a Chinese-first output mode that preserves necessary English terms
• add a dedicated document-processing model config for ingest, merge, and Save to Wiki flows
• ignore local Windows dev helper files

What changed

Local agent / MCP access

• add the local MCP server entrypoint, bridge plumbing, tests, and settings UI
• expose local agent access controls in Settings
• keep the bridge gated so local agent access can be explicitly enabled

Chat and output behavior

• move chat retrieval prompt construction into a shared helper
• add a Chinese-first output mode for Chinese academic workflows that preserves English technical terms, titles, commands, code, and paths
• update output-mode copy and project-creation defaults accordingly

Dedicated document-processing model

• add a separate document-processing model config for ingest-related LLM work
• keep normal chat, deep-research synthesis, lint, dedup, and review on the main model
• route ingest queue, source import, clip ingest, Save to Wiki, page merge, and post-research auto-ingest through the document-processing model path
• add readiness checks and tests for the dedicated ingest model flow

Local repo hygiene

• ignore local Windows shortcut / launcher / tools artifacts used for local development

Validation

• npm run typecheck
• npx vitest run src/lib/document-llm.test.ts src/lib/has-usable-llm.test.ts src/lib/output-language.test.ts src/lib/language-metadata.test.ts

Notes

• existing users keep current behavior by default because document processing still reuses the main LLM until explicitly changed in Settings
• this PR intentionally does not split research/lint/dedup into separate model configs