feat: initial LexAI Chrome extension scaffold

- WXT + React 18 + TypeScript setup
- Background service worker (LLM API proxy)
- Content script (floating toolbar + text selection)
- Options page (provider/model/API key config)
- Popup UI
- Gitea Actions workflows (CI, preview, release)
- Vitest unit tests + Playwright E2E setup
- Supports: OpenAI, Anthropic, Groq, OpenRouter

src/utils/REFERENCE_NOTES.md

@@ -0,0 +1,21 @@
+# Reference Project Notes
+
+## Scramble (zlwaterfield/scramble) ⭐ 1,647
+- Uses `browserAPI` abstraction for Chrome/Firefox compat
+- Content script listens for messages, calls background to do actual LLM calls
+- **Key pattern:** `replaceSelectedText()` handles both:
+  1. `TEXTAREA` / `INPUT` elements (uses selectionStart/End)
+  2. Generic DOM selection (uses Range API)
+- Context menu integration for triggering actions
+- Good reference for: text replacement, context menu setup
+
+## Proofly (onderceylan/proofly)
+- Privacy-first: uses Chrome's built-in on-device AI (no API key needed)
+- Good reference for: Manifest V3 structure, shadow DOM injection for UI
+
+## Key Learnings for LexAI
+1. **Always handle both textarea AND contenteditable** — different replacement logic needed
+2. **Use shadow DOM for toolbar** — prevents host page CSS from breaking our UI
+3. **Background as CORS proxy** — content scripts can't call external APIs directly
+4. **Context menu = power user feature** — right-click → LexAI actions
+5. **Debounce selection events** — users drag text slowly, don't fire on every pixel